Tax Policy and the Economy, Volume 18 (NBER Tax Policy and the Economy)

18 Tax Policy and the Economy 2004 National Bureau of Economic Research Edited by James M. Poterba The Character and D...

Author: James M. Poterba

22 downloads 1038 Views 1MB Size Report

This content was uploaded by our users and we assume good faith they have the permission to share this book. If you own the copyright to this book and it is wrongfully on our website, we offer a simple DMCA procedure to remove your content from our site. Start by pressing the button below!
Report copyright / DMCA form

DOWNLOAD PDF

18 Tax Policy and the Economy

2004

National Bureau of Economic Research Edited by James M. Poterba The Character and Determinants of Corporate Capital Gains How Fast Do Personal Computers Depreciate? Concepts and New Estimates Tax Policy and Education Policy: Collision or Coordination? A Case Study of the 529 and Coverdell Saving Incentives Reported Incomes and Marginal Tax Rates, 1960–2000: Evidence and Policy Implications

NBER The MIT Press

The (Un)changing Geographical Distribution of Housing Tax Benefits: 1980–2000

TAX POLICY AND THE ECONOMY 18

TAX POLICY AND THE ECONOMY 18 edited by James M. Poterba

National Bureau of Economic Research The MIT Press Cambridge, Massachusetts London, England

NBER/Tax Policy and the Economy, Volume 18, 2004 ISSN: 1531-3468 ISBN: Hardcover 0-262-16226-1 ISBN: Paperback 0-262-66184-5 Published annually by The MIT Press, Cambridge, Massachusetts 02142-1407 © 2004 by the National Bureau of Economic Research and the Massachusetts Institute of Technology. All rights reserved. No part of this book may be reproduced in any form by any electronic or mechanical means (including photocopying, recording, or information storage and retrieval) without permission in writing from the publisher. Standing orders/subscriptions are available. Inquiries, and changes to subscriptions and addresses should be addressed to MIT Press Standing Order Department/BB, Five Cambridge Center, Cambridge, MA 02142-1407, phone 617-258-1581, fax 617-253-1709, email [email protected] In the United Kingdom, continental Europe, and the Middle East and Africa, send single copy and back volume orders to: The MIT Press, Ltd., Fitzroy House, 11 Chenies Street, London WC1E 7ET England, phone 44-020-7306-0603, fax 44-020-7306-0604, email [email protected], website http://mitpress.mit.edu In the United States and for all other countries, send single copy and back volume orders to: The MIT Press c/o Triliteral, 100 Maple Ridge Drive, Cumberland, RI 02864, phone 1-800-405-1619 (U.S. and Canada) or 401-658-4226, fax 1-800-406-9145 (U.S. and Canada) or 401-531-2801, email [email protected], website http://mitpress.mit.edu This book was set in Palatino by Achorn Graphic Services and was printed and bound in the United States of America. 10

9

8

7

6

5

4

3

2

1

NATIONAL BUREAU OF ECONOMIC RESEARCH Officers Michael H. Moskow, Chairman Elizabeth E. Bailey, Vice Chairman Martin Feldstein, President and Chief Executive Officer Susan Colligan, Vice President for Administration and Budget and Corporate Secretary Robert Mednick, Treasurer Kelly Horak, Controller and Assistant Corporate Secretary Gerardine Johnson, Assistant Corporate Secretary

Directors at Large Peter C. Aldrich Elizabeth E. Bailey John H. Biggs Andrew Brimmer John S. Clarkeson Don R. Conlan George C. Eads Jessica P. Einhorn

Martin Feldstein Jacob A. Frenkel Judith M. Gueron Robert S. Hamada George Hatsopoulos Karen N. Horn John Lipsky Laurence H. Meyer

Michael H. Moskow Alicia H. Munnell Rudolph A. Oswald Robert T. Parry Richard N. Rosett Marina v. N. Whitman Martin B. Zimmerman

Directors by University Appointment George Akerlof, California, Berkeley Jagdish Bhagwati, Columbia Ray C. Fair, Yale Michael J. Brennan, California, Los Angeles Glen G. Cain, Wisconsin Franklin Fisher, Massachusetts Institute of Technology Saul H. Hymans, Michigan

Marjorie B. McElroy, Duke Joel Mokyr, Northwestern Andrew Postlewaite, Pennsylvania Uwe E. Reinhardt, Princeton Nathan Rosenberg, Stanford Craig Swan, Minnesota David B. Yoffie, Harvard Arnold Zellner, Chicago

Directors by Appointment of Other Organizations Richard B. Berner, National Association For Business Economics Gail D. Fosler, The Conference Board A. Ronald Gallant, American Statistical Association Richard C. Green, American Finance Association Thea Lee, American Federation of Labor and Congress of Industrial Organizations Robert Mednick, American Institute of Certified Public Accountants

Angelo Melino, Canadian Economics Association Jeffrey M. Perloff, American Agricultural Economics Association John J. Siegfried, American Economic Association Josh S. Weston, Committee for Economic Development Gavin Wright, Economic History Association

Directors Emeriti Carl F. Christ Lawrence R. Klein Franklin A. Lindsay

Paul W. McCracken Peter G. Peterson

Bert Seidman Eli Shapiro

Relation of the Directors to the Work and Publications of the NBER

1. The object of the NBER is to ascertain and present to the economics profession, and to the public more generally, important economic facts and their interpretation in a scientific manner without policy recommendations. The Board of Directors is charged with the responsibility of ensuring that the work of the NBER is carried on in strict conformity with this object. 2. The President shall establish an internal review process to ensure that book manuscripts proposed for publication DO NOT contain policy recommendations. This shall apply both to the proceedings of conferences and to manuscripts by a single author or by one or more co-authors but shall not apply to authors of comments at NBER conferences who are not NBER affiliates. 3. No book manuscript reporting research shall be published by the NBER until the President has sent to each member of the Board a notice that a manuscript is recommended for publication and that in the President’s opinion it is suitable for publication in accordance with the above principles of the NBER. Such notification will include a table of contents and an abstract or summary of the manuscript’s content, a list of contributors if applicable, and a response form for use by Directors who desire a copy of the manuscript for review. Each manuscript shall contain a summary drawing attention to the nature and treatment of the problem studied and the main conclusions reached. 4. No volume shall be published until forty-five days have elapsed from the above notification of intention to publish it. During this period a copy shall be sent to any Director requesting it, and if any Director objects to publication on the grounds that the manuscript contains policy recommendations, the objection will be presented to the author(s) or editor(s). In case of dispute, all members of the Board shall be notified, and the President shall appoint an ad hoc committee of the Board to decide the matter; thirty days additional shall be granted for this purpose. 5. The President shall present annually to the Board a report describing the internal manuscript review process, any objections made by Directors before publication or by anyone after publication, any disputes about such matters, and how they were handled. 6. Publications of the NBER issued for informational purposes concerning the work of the Bureau, or issued to inform the public of the activities at the Bureau, including but not limited to the NBER Digest and Reporter, shall be consistent with the object stated in paragraph 1. They shall contain a specific disclaimer noting that they have not passed through the review procedures required in this resolution. The Executive Committee of the Board is charged with the review of all such publications from time to time. 7. NBER working papers and manuscripts distributed on the Bureau’s web site are not deemed to be publications for the purpose of this resolution, but they shall be consistent

viii

Relation of the Directors to the Work and Publications of the NBER

with the object stated in paragraph 1. Working papers shall contain a specific disclaimer noting that they have not passed through the review procedures required in this resolution. The NBER’s web site shall contain a similar disclaimer. The President shall establish an internal review process to ensure that the working papers and the web site do not contain policy recommendations, and shall report annually to the Board on this process and any concerns raised in connection with it. 8. Unless otherwise determined by the Board or exempted by the terms of paragraphs 6 and 7, a copy of this resolution shall be printed in each NBER publication as described in paragraph 2 above.

CONTENTS Introduction: James M. Poterba xi Acknowledgments

xv

THE CHARACTER AND DETERMINANTS OF CORPORATE CAPITAL GAINS 1 Mihir A. Desai and William M. Gentry HOW FAST DO PERSONAL COMPUTERS DEPRECIATE? CONCEPTS AND NEW ESTIMATES 37 Mark E. Doms, Wendy E. Dunn, Stephen D. Oliner, and Daniel E. Sichel TAX POLICY AND EDUCATION POLICY: COLLISION OR COORDINATION? A CASE STUDY OF THE 529 AND COVERDELL SAVING INCENTIVES 81 Susan Dynarski REPORTED INCOMES AND MARGINAL TAX RATES, 1960–2000: EVIDENCE AND POLICY IMPLICATIONS 117 Emmanuel Saez THE (UN)CHANGING GEOGRAPHICAL DISTRIBUTION OF HOUSING TAX BENEFITS: 1980–2000 175 Todd Sinai and Joseph Gyourko

INTRODUCTION James M. Poterba Massachusetts Institute of Technology and NBER

This volume represents the eighteenth installment of the NBER’s Tax Policy and the Economy conference series. This series communicates current academic research findings in the areas of taxation and government spending to policy analysts in both the government and the private sector. Tax Policy and the Economy papers address issues that have an immediate bearing on current policy debates as well as questions that are of longer-term interest. All of the research described at Tax Policy and the Economy meetings has some connection to policy analysis in the field of public finance. The five papers in this year’s volume touch on topics that are as old as the income tax itself, such as the measurement of depreciation allowances, as well as on currently emerging issues, such as the tax treatment of income on assets that have been saved to pay for expenses associated with higher education. The first paper, by Mihir Desai and William Gentry, investigates “The Character and Determinants of Corporate Capital Gains.” The paper starts with the observation that the tax and other determinants of corporate capital gain realizations have received far less research attention than the analogous determinants of individual capital gain realizations. The authors note that this is surprising, since corporate capital gain realizations average nearly thirty percent of individual realizations. The paper then shows that aggregate corporate capital gain realizations track aggregate individual gains quite closely, and it also finds a substantial elasticity of corporate capital gain realizations with respect to the corporate capital gains tax rate. The authors document this behavioral elasticity using both aggregate time series data and panel data on individual firms. They conclude that high corporate capital gains tax rates may discourage firms from raising cash by selling physical assets or the securities of other companies, thereby inducing a “lock-in effect.” Their analysis of firm-level data on sales of physical assets and securities represents a novel investigation of the asset turnover behavior of U.S. firms.

xii

Introduction

The next paper focuses on a perennial issue of tax policy design: the estimation of asset depreciation rates. Mark Doms, Wendy Dunn, Stephen Oliner, and Dan Sichel are the authors of “How Fast Do Personal Computers Depreciate? Concepts and New Estimates.” This paper uses a large new data set consisting of the second-hand prices of personal computers manufactured by five major firms to study the rate at which these computers lose their resale value. Their estimates suggest that a personal computer loses roughly half of its remaining value with each year of use. This decline in value is the result of both physical depreciation and revaluation. These estimates of asset value decay rates can be used to evaluate the depreciation provisions of the current income tax code. At low inflation rates, current rules generate depreciation benefits that are similar to the loss in value experienced by PCs. At higher inflation rates, however, current depreciation allowances would fall below actual depreciation. This paper offers a framework for evaluating depreciation allowances and for using asset price data to calibrate tax depreciation parameters. The third paper, by Susan Dynarski, is “Tax Policy and Education Policy: Collision or Coordination? A Case Study of the 529 and Coverdell Saving Incentives.” This paper analyzes the benefits of recently-created tax incentives for college saving, when viewed from the perspective of participating families. The key insight of this paper is that one must recognize the interplay between tax incentives to accumulate assets to finance higher education and the incentives embodied in college financial aid rules. In particular, Dynarski shows that the standard treatment of college saving under the income tax and financial aid rules that applied prior to 2003, and that was under debate and possible reform at the time of the conference, could make some households worse off if they saved in a taxfavored saving account. This outcome is the result of poor coordination between financial aid rules and tax rules. This possibility was particularly important for “aid-marginal” households, those for whom a substantial increase in the family’s financial assets could result in a substantial decline in the family’s financial aid package. The next paper is Emmanuel Saez’ “Reported Incomes and Marginal Tax Rates, 1960-2000: Evidence and Policy Implications.” This paper addresses a central question of tax policy debate, with a particular bearing on the problem of revenue estimation: how do changes in marginal income tax rates affect reported taxable income? Saez’ new empirical analysis is based on repeated cross-sections of tax filings. These data sets provide a large sample of taxpayers in each year, thereby allowing greater precision than smaller panel data sets in estimating some aspects of taxpayer behavior. The results suggest two important conclusions. First, there appear to be stark differences in taxpayer responses to different tax

Introduction

xiii

reforms. While the Tax Reform Act of 1986 and the tax increase in 1993 were associated with large taxpayer responses, earlier tax reforms, notably those in the 1960s, appear to have resulted in only minimal changes in taxpayer behavior. Thus it appears as though the elasticity of taxable income with respect to the marginal tax rate is a time-varying or context-dependent parameter. Second, the taxable income elasticity appears to be greater at higher income levels than at lower levels. This implies that for the purpose of revenue estimating, it may be more important to consider potential behavioral responses at high incomes than at lower incomes. Finally, the last paper focuses on the long-standing issue of tax subsidies to housing. Todd Sinai and Joseph Gyourko study “The (Un)changing Geographical Distribution of Housing Tax Benefits: 1980–2000.” Their paper offers a careful analysis of the income tax expenditures associated with owner-occupied housing. They compute the difference in the total income tax liability that households in different states would face under two different assumptions about the tax treatment of housing. The first is the status quo, and the second is an alternative tax regime in which homeowners are taxed on their imputed rental income while claiming deductions for interest expense and property taxes, as under the current tax rules for itemizes, as well as for maintenance costs, which are not currently deductible. The results show that homeowners in coastal California, New England, and several mid-Atlantic states with high house prices, receive the largest per-household subsidies. The broad geographical pattern of subsidies has remained reasonably stable over time, but rising dispersion in house prices has increased the dispersion in perhousehold tax subsidies. The cross-state dispersion in tax subsidies to owner-occupied housing is greater today than twenty years ago. Each of these papers illustrates the type of policy-relevant research that is carried out by the affiliates of the NBER Public Economics Program. Each provides important background information for policy analysis, without making recommendations about the merits or demerits of particular policy options. I hope that each of these papers will provide useful input to various participants in the policy process who are concerned with the design of tax policy.

ACKNOWLEDGMENTS In planning and organizing this year’s Tax Policy and the Economy conference and the associated volume, I have incurred debts to many individuals. Martin Feldstein, President of the NBER, has been an active supporter of this conference throughout its history. Conference Department Director Carl Beck and Rob Shannon did an outstanding job in updating the invitation list and in disseminating information about the conference to potential participants. The members of the NBER Conference Department, notably Lita Kimble, Rob Shannon, and Carl Beck, handled conference logistics with efficiency and good cheer. Helena Fitz-Patrick oversaw the publication process with outstanding attention to detail and with exceptional speed and efficiency. I am also grateful to Professor N. Gregory Mankiw, who is currently on leave from Harvard University’s Department of Economics to chair the President’s Council of Economic Advisers, for delivering a fascinating set of luncheon remarks at the conference at which these papers were presented. Greg’s remarks focused on how current analyses of the estate tax overstate the progressivity of the tax. He noted that assigning tax burdens to decedents rather than to those who receive bequests is likely to overstate the wealth and income of those who pay the estate tax. He also called for more systematic analysis of the general equilibrium incidence of the estate tax, and in particular for greater recognition that the tax may discourage capital accumulation and therefore burden workers as well as those who accumulate and bequeath wealth. Finally, I wish to thank each of the authors whose papers are included in this volume for striving to communicate their important research findings in a readable and clear fashion. I appreciate their efforts and willingness to participate in this very important opportunity for interchange between the research community and policy-makers.

THE CHARACTER AND DETERMINANTS OF CORPORATE CAPITAL GAINS Mihir A. Desai Harvard University and NBER

William M. Gentry Williams College and NBER

EXECUTIVE SUMMARY This paper analyzes how corporate capital gains taxes affect the capital gains realization decisions of firms. The paper outlines the tax treatment of corporate capital gains, the consequent incentives for firms with gains and losses, the efficiency consequences of these taxes in the context of other taxes and capital market distortions, and the response of firms to these incentives. Despite receiving limited attention, corporate capital gains realizations have averaged 30 percent of individual capital gains realizations over the last 50 years and have increased dramatically in importance over the last decade. By 1999, the ratio of net long-term capital gains to income subject to tax was 21 percent and was distributed across various industries, which suggests the importance of realization behavior to corporate financing decisions. Time-series analysis of aggregate realization behavior demonstrates that corporate capital gains taxes affect realization behavior significantly. Similarly, an analysis of firm-level investment and property, plant, and equipment (PPE) disposal decisions This paper was prepared for the Tax Policy and the Economy Conference in November 2003 in Washington DC. We thank John Graham, Jim Hines, and Jim Poterba for helpful comments and Sean Lubens for helpful research assistance. Mihir A. Desai thanks the Division of Research of Harvard Business School for generous funding.

2

Desai & Gentry

and gains recognition behavior also suggests an important role for these taxes in determining when firms raise money by disposing of assets and realizing gains.

1. INTRODUCTION Analyses of the impact of the tax system on corporate behavior typically emphasize the role of the corporate income tax in altering firm financing and investment decisions. These financing and investment decisions, in turn, have been shown to depend critically on the wedge between the costs of internal and external finance. One obvious and important source of internal finance, aside from retained earnings, is the disposal of assets and investments. The role of taxes in influencing these types of financing decisions may be nontrivial given the system of taxing corporate capital gains and the distortions that arise from costly external finance. Despite the potential importance of asset sales as a source of financing corporate investment, relatively little research has been done on how corporate capital gains taxes might affect asset sales. Analyses of capital gains taxes have focused almost exclusively on the realization behavior of individuals, with particular attention on the revenue consequences of changing capital gains tax rates and on the impact on risk taking and expected asset returns. The relative oversight of the corporate capital gains tax system is surprising given the substantial volume of corporate capital gains—U.S. corporations realized $146.5 billion of net long-term capital gains, or 21 percent of their income subject to tax, in 1999—and the potentially distortionary impact of these taxes stemming from interactions with capital market imperfections. In this paper, we address this oversight by detailing U.S. tax policy toward corporate capital gains, characterizing the nature and distribution of corporate capital gains activity, and examining the effect of these taxes on the financing and investment decisions of firms. There are several important reasons for studying the taxation of corporate capital gains. First, while many of the economic issues regarding the tax effects of corporate and individual capital gains are similar, the possible distortions in the corporate and individual settings differ along some dimensions. For example, taxing corporate capital gains can impede asset sales and reorganizations that reallocate capital between firms. If such reallocations raise the productivity of assets, then discouraging these transactions reduces the pretax rate of return. In contrast, for most assets held by individuals, the identity of the owner is unlikely to affect asset

The Character and Determinants of Corporate Capital Gains

3

returns.1 More generally, the increased emphasis on the role of corporate governance in determining economic performance suggests that tax policy that alters the incentives for cross-holding or corporate venture capital can have important economic consequences.2 Finally, if firms are deterred from disposing of assets as a result of the capital gains tax, corporate capital gains taxes potentially exacerbate pre-existing distortions arising from capital market imperfections that make external finance more costly. Second, President G. W. Bush’s recent proposal to eliminate the double taxation of corporate income brought to the forefront the question of the appropriate structure of capital income taxation. With regard to corporate investment in other corporations, U.S. tax policy provides corporations some relief from multiple layers of taxation on intercorporate dividends through the dividends received deduction (DRD), which allows the exclusion of the majority of intercorporate dividends from the corporate income tax. The logic behind the DRD is to avoid having a full third layer of taxation on capital income in the U.S. tax system that already taxes corporate income twice (i.e., corporate income is taxed at the corporate level and the dividends are taxed at the shareholder level). By this logic, one might expect that capital gains earned on intercorporate investments would similarly be provided some relief, but the tax code does not provide a preferential corporate tax rate for capital gains. Understanding how corporate capital gains taxes influence the holding behavior of firms provides a first step in understanding the consequences of this policy as an element in the overall system of capital taxation. Third, the volume of corporate capital gains is substantial, and increasingly so, when compared to either individual capital gains or other metrics of corporate activity. From 1954–1999, corporations reported realized 1 Edwards, Lang, Maydew, and Shackelford (2003) consider the effect of such potential reallocations by examining the stock market reaction to the German tax reform of 2000, which eliminated the capital gains tax on corporate cross-holdings. Given Germany’s history of substantial corporate cross-holdings, the reform was predicted to have sweeping effects in the level of merger and acquisition activity in Germany. Edwards et al. find a substantial stock market impact of the reform, but it is concentrated among a small number of banks and insurance companies with substantial cross-holdings. They report that the early evidence on the amount of corporate restructuring after the tax reform does not support the idea of widespread restructuring; however, the implementation of the tax reform corresponded to a worldwide slowdown in merger activity, so it is difficult to measure the tax effect. 2 See Morck (2003) on the interaction of cross-holdings and intercompany dividend tax policy; Wolfenzon (1999) on the consequences of pyramidal ownership; and Gompers, Lerner, and Scharfstein (2003) on corporate venture capital activity.

4

Desai & Gentry

net long-term capital gains that averaged 30 percent of the realizations reported by individuals. By 1999, corporate net long-term gains were more than 20 percent of corporate income subject to tax and averaged 16 percent through the 1990s. From a tax policy perspective, given that corporations face a tax rate of up to 35 percent on realized net capital gains while individuals face a maximum tax rate of 15 percent on capital gains, the taxation of corporate capital gains has substantial revenue consequences. We examine several aspects of corporate capital gains taxation. The incentives for realization are fairly complex, and we begin with a discussion of tax policy, with particular reference to the effects of taxes on net long-term gains. With the incentives and potential economic effects established, we outline the scope of this activity and distinguish among the types of capital gains realized and characterize their distribution relative to several benchmarks. We employ two empirical approaches to examine the responsiveness of corporate capital gains to variation in marginal tax rates. Following Plesko (2002), we study the time-series behavior of aggregate corporate capital gains realizations. In this analysis, as in the studies of individual capital gains taxes (see Auerbach, 1988, and Eichner and Sinai, 2000), we rely on time-series variation in tax policy to identify possible tax effects on realizations. We add several additional controls for possible determinants of corporate capital gains—including proxies for sentiment, consolidation activity, and capital market activity—and find statistically significant tax price elasticities of approximately −1.3 for corporations with respect to realization behavior. Such a time-series analysis is problematic for several reasons, so we turn to firm-level financial reporting data to examine whether the propensity to sell assets or realize gains is related to firm-specific variation in estimated marginal tax rates. For the firm-level analysis, the key variation in effective tax rates arises due to the rules related to operating losses. Using proxies for the marginal tax rate provided via the methodology in Graham (1996) and controlling for firm characteristics and timevarying investment opportunities, we find that the sales of investments and property, plant, and equipment (PPE) are more likely and considerably larger in low-tax years. In addition to this evidence on disposal behavior, the likelihood and volume of gains is particularly guided by tax considerations. In the next section, we review the basic tax rules governing corporate capital gains. In section 3, we discuss the various incentive effects of corporate capital gains taxation, including both the efficiency costs to such taxes and how these taxes affect corporate tax planning efforts. In section 4, we provide an overview of the general features of corporate capital


5

gains realizations and broad time-series trends in those same realizations. Section 5 provides the results of our time-series analysis of corporate capital gains realizations. Section 6 presents analysis that examines firmlevel variation in realization behavior. In section 7, we conclude with directions for further research.

2. U.S. TAXATION OF CORPORATE CAPITAL GAINS In determining the tax burden on corporate capital gains, three elements are critical: the definition of capital gains income for corporations; the applicable tax rate on corporate capital gains income; and the rules for netting capital gains with other sources of income, including how capital gains and losses interact with loss carryforward rules. In this section, we address each of these elements in turn and then frame all of them in historical and international perspective.3

2.1 Definition of Capital Gains Capital gains or losses arise from the sale of capital assets. Capital assets are defined as all assets except: (1) inventory; (2) accounts or notes receivable through the ordinary course of business; (3) real or depreciable property used in a business; (4) copyright, literary, musical, or artistic compositions held by the creator; and (5) certain publications of the U.S. government.4 The major categories of capital assets include: (1) investment assets, such as stocks and bonds; (2) assets (including land) held for long-term investment rather than commercial purposes; (3) self-created patents (see Internal Revenue Code, section 1235); and (4) goodwill and going-concern value created by a firm. In addition to the sale of capital assets, capital gains can arise from the sale of real or depreciable property (so-called section 1231 assets) under some circumstances. If these assets are sold for a loss (e.g., the sales price is less than the basis after adjusting for depreciation), then the loss is considered ordinary in character. If such assets are sold for a gain relative to the adjusted basis, then the character of the income depends on the recapture rules. To the extent that the gain arises from deductions for previous 3 Our discussion of the tax rules for corporate capital gains focuses on the regular corporate income tax without considering the effects of the alternative minimum tax (AMT). In general, under current tax rules, capital gains realizations do not generate preference items for the AMT. For the sale of depreciable assets, however, the AMT uses slower depreciation schedules, which tend to result in smaller gains (or larger losses) from the sale of such assets. This difference in depreciation schedules tends to reduce the tentative AMT tax liability for a corporation that sells depreciable assets. 4

Section 1221 of the Internal Revenue Code.

6

Desai & Gentry

depreciation, the gain is considered ordinary income; however, for a gain in excess of the amount of previous depreciation, the gain is considered capital in character. The logic behind the recapture rules that classify gains associated with previous depreciation as ordinary income is that the firm has previously deducted the depreciation allowance from ordinary income but is selling the asset for more than its adjusted basis, which suggests that the depreciation allowances accrued faster than the asset actually depreciated.5 A critical element of the definition of a capital gain is that it depends on an observable transaction, typically the sale of an asset. The realizationbased nature of capital gains taxation creates numerous tax planning incentives, as discussed below. It also complicates measuring the annual effective tax rate on capital gains because the holding period influences the present value of the tax liability associated with owning the asset. When statutory tax rates do not increase over time, the ability to defer the realization of gains reduces the tax burden on the investment.

2.2 Tax Rates on Corporate Capital Gains Unlike individuals who face lower tax rates on capital gains income than on ordinary income, U.S. corporations do not receive preferential tax rates on realized capital gains. Net realized capital gains are added to ordinary income when computing the firm’s taxable income.6 Because corporations do not receive a preferential tax rate on capital gains income relative to ordinary income, the distinction between capital income and ordinary income is often not critical for a firm’s tax liability. As discussed below, however, the character of income affects which types of income can be netted against other types of income and the rules for how firms with net losses can use losses to offset previous or future income.

2.3 Combining Capital Gains, Losses, and Ordinary Income Much of the complexity of taxing corporate capital gains arises from the rules associated with matching different types of capital gains and losses (e.g., short-term versus long-term), pooling different types of income, and carrying losses forward and backward. The general rule is that ordinary 5 The recapture rules are especially important when depreciation allowances for tax purposes are accelerated relative to economic depreciation and when capital gains income faces a lower tax rate than ordinary income. Both of these conditions held before 1986 and created incentives for firms to churn assets by depreciating new assets and then selling them for a gain. For an analysis of these incentives and the role of the recapture rules, see Gordon, Hines, and Summers (1987). 6 For historical reasons, corporations technically have the option of adding capital gains to ordinary income or facing an alternative tax rate of 35 percent, which is the same as the current top corporate marginal tax rate.


7

income and losses, capital gains and losses, and gains or losses on section 1231 assets are aggregated separately. Within capital gains, taxpayers separately aggregate short-term (defined as having a holding period of less than one year) capital gains and losses and long-term capital gains and losses (including any capital gains from the disposition of section 1231 assets). If one of the holding period baskets results in a net gain and the other holding period basket results in a net loss, then the net loss in one basket can be used to offset the net gain in the other basket. After completing this two-step netting process, a net capital gain is included in taxable income; however, corporations are not allowed to use a net capital loss to offset ordinary income.7 Instead, corporations with net capital losses must apply the carryback and carryforward rules. Current law allows capital losses to be carried back to offset net capital gains in the previous three years or carried forward to offset net capital gains in the subsequent five years.8 Because the tax law does not allow for an interest calculation to compensate for the time value of money, carrying losses forward is less valuable than an immediate tax refund or deduction against ordinary income (assuming that the firm’s statutory tax rate is constant over time). In general, the netting rules give corporations a preference for capital gains income over ordinary income but ordinary losses over capital losses. Capital gains have an advantage over ordinary income in their ability to offset capital losses. In contrast, ordinary losses are preferable to capital losses because they can offset ordinary income or capital gains income, while capital losses can offset capital gains only via the netting rules for capital gains.

2.4 Tax Policy Toward Corporate Capital Gains over Time Tax rules governing corporate capital gains have changed over time in a variety of ways. One major change over time is whether corporate capital gains face a preferential rate relative to ordinary income. Before the Tax Reform Act of 1986, corporations could base their tax liability on having net capital gains (i.e., net long-term capital gains in excess of net shortterm losses) taxed at an alternative tax rate. The corporation would pay the minimum of its tax liability, including net capital gains as ordinary 7 In contrast, individuals have a limited opportunity to offset ordinary income with capital losses. 8 In contrast, individuals who exceed the annual limit on using capital losses to offset ordinary income have an unlimited number of years to carry forward capital losses to offset future capital gains. In addition, the time limits on corporate carryovers for capital gains differ from those for operating losses. Operating losses can be carried back by two years or carried forward for 20 years.

8

Desai & Gentry

income and using the alternative tax rate. In 1986, this alternative tax rate was 28 percent, while the maximum tax rate on ordinary income was 46 percent. Between 1954 and 1986, the alternative tax rate varied between 25 and 30 percent. It is worth noting two features of the alternative tax rate system. First, because the same alternative rate is applied to all firms, corporations with relatively low income might prefer the ordinary income tax rate over the alternative tax rate due to the graduated corporate tax rate schedule. Second, because the definition of net capital gains uses the distinction between long- and short-term capital gains, the holding period distinction was more important before 1986 than it was after 1986, with corporations preferring to realize long-term capital gains rather than short-term capital gains, thus qualifying for the lower alternative tax rate.9

2.5 An International Perspective on the Taxation of Corporate Capital Gains The taxation of capital gains for both individuals and corporations varies substantially across countries. Policies can differ along several dimensions. First, how are capital gains taxed relative to other forms of income? Second, do the tax rules for corporate capital gains differ from the tax rules for the capital gains of individuals? One difference is whether capital gains are taxed at a different tax rate than the tax rate for ordinary income, including the possibility of exemption or exclusion from taxation. This rate differential can be targeted toward specific types of assets (e.g., shares in publicly traded firms) or require specific holding periods (e.g., a lower tax rate on long-term capital gains than on short-term capital gains). Indexing of cost basis is another policy option, although it is somewhat rare. Some countries allow for exemptions for individuals of some threshold amount of gains in each year. For example, in 1998, France allowed $8,315 of gains to be excluded from personal income taxation.10 However, these policies do not typically extend to corporate shareholders. 9 As evidence that the holding-period distinction affected behavior, consider the relationship between (1) the difference between the top ordinary income tax rate and the long-term capital gains tax rate and (2) the ratio of net short-term gains to net long-term gains (taken from the Corporate Statistics of Income data described below). From 1954 to 1986, the difference between the ordinary income tax rate and the capital gains tax rate for corporations ranged from 18 to 27.8 percentage points, and the annual ratio of short-term to long-term gains averaged 0.057. From 1988 to 1998, there was no difference in the tax rates, and the ratio of short-term to long-term gains was 0.20. Thus, when it was more advantageous to recognize long-term capital gains instead of short-term gains (i.e., the earlier years), short-term gains were a much smaller percentage of total realizations than when firms were indifferent to the holding period. 10

See American Council on Capital Formation (1998).


9

In addition, countries can allow for rollover provisions in which gains continue to be deferred, provided the proceeds are invested in specific types of assets. The American Council on Capital Formation (1998) surveyed capital gains taxation across 24 countries for 1998.11 Of the 24 countries, six (Belgium, Denmark, Hong Kong, the Netherlands, Poland, and Singapore) exempted long-term and (except for Denmark) short-term capital gains income for both individuals and corporations. Three countries (China, Korea, and Taiwan) exempted gains associated with local companies or companies traded on the major stock exchange, but they taxed gains on other equities (and presumably gains associated with other assets). Another five countries (Canada, France, India, Indonesia, and Italy) had long-term capital gains tax rates below the top marginal income tax rate for both individuals and corporations. In eight countries (Argentina, Brazil, Germany, Japan, Mexico, Sweden, the United Kingdom, and the United States), individuals faced preferential tax rates (relative to ordinary income) for long-term capital gains, but corporations faced the top marginal income tax rate on capital gains.12 In three of these countries (Argentina, Germany, and Mexico), individual shareholders were exempt from capital gains taxes.13 In two countries (Australia and Chile), capital gains were taxed at the ordinary income tax rates for both individuals and corporations; however, Australia allowed for indexing of cost basis for both individuals and corporations, while Chile allowed corporations to index their cost basis and provided individuals an exemption for the first $6,600 of capital gains. Even this cursory review of capital gains taxes in other countries reveals substantial heterogeneity in tax policies toward capital gains. The U.S. tax system of preferential capital gains tax rates for individuals but 11 The survey, conducted by Arthur Andersen, focuses on the tax treatment of investment in equities. The countries included in the survey are Argentina, Australia, Belgium, Brazil, Canada, Chile, China, Denmark, France, Germany, Hong Kong, India, Indonesia, Italy, Japan, Korea, Mexico, the Netherlands, Poland, Singapore, Sweden, Taiwan, the United Kingdom, and the United States. Presumably, many countries have special tax rules pertaining to the gains or losses on specific assets. 12 Among these countries, Japan and the United Kingdom are somewhat different than the others. In Japan, individual taxpayers have a choice between a 20.0 percent tax rate on the net gain (which is lower than the top marginal tax rate on ordinary income) or a tax of 1.25 percent on the sales price. In the United Kingdom, individuals faced a sliding scale of tax rates so that the tax rate fell as the holding period increased; with a holding period of 10 years, an individual would include only 25 percent of the capital gains in the tax base. For corporations in the United Kingdom in 1998, the tax system allowed corporations to index their cost basis in calculating the gain. 13 As discussed above, Germany subsequently eliminated the corporate capital gains tax on corporate cross-holdings.

10

Desai & Gentry

not for corporations is a relatively common approach, but even more countries provide preferential tax rates for corporate capital gains income relative to ordinary income (either in the form of lower tax rates or exemption). The issue of the relative taxation of dividends and capital gains is also likely to differ across countries given the variation in the extent to which different countries have integrated their personal and corporate income tax systems. An open empirical question is whether this heterogeneity in tax policy affects asset allocation and investment decisions and the level of the cross-ownership of corporate shares across countries.

3. CORPORATE CAPITAL GAINS TAXES AND INCENTIVES The taxation of corporate capital gains affects incentives in three broad categories. First, it affects real decisions related to investment and financing and the allocation of capital across firms and throughout the economy. Second, taxes can affect the timing of corporate decisions. Third, tax policy toward corporate capital gains can affect corporate tax planning activities. In this section, we discuss each of these types of possible behavioral responses.

3.1 The Allocation of Capital and Corporate Capital Gains Taxes The allocation effects of capital gains taxation have primarily been discussed in the context of individual investors, and it is useful to anchor a discussion of the allocation effects for corporations in this literature. For individuals, capital gains taxes can affect investment decisions in two ways. First, for deciding among assets in which to invest, assuming that the statutory tax rate is held constant, the effective tax rate on an asset that is taxed on realization is lower than the effective tax rate on an asset whose return is taxed annually.14 Thus, assets with returns that are taxed on realization have a tax advantage relative to assets that face annual taxation. In addition to affecting capital allocation by pushing more capital into assets that produce capital gains, this differential taxation can also affect asset prices and future returns. In response to their favorable tax 14 We focus on the effects of realization-based taxation. In addition, policymakers frequently debate whether capital gains should be taxed at a lower tax rate than other types of income, under the common assertion that a lower tax rate on risky investment promotes risk taking. Despite the common claim that lower tax rates on capital gains promote risk taking, the theoretical relationship between the tax rate and the amount of risk taking is ambiguous because the tax rate affects both the expected return and the variability of returns.


11

treatment, assets that yield capital gains may offer lower pretax rates of return (after adjusting for risk), which might offset the tax advantages of capital gains generating investments. Second, once investors have an appreciated position, realization-based taxation provides an incentive for investors to defer their tax liability by delaying the sale of their assets. This incentive to delay realization is known as the lock-in effect, which is commonly analyzed in the context of individuals who own a portfolio of financial assets.15 By deferring realization, investors effectively receive an interest-free loan in the amount that they would pay in taxes if they sold the asset and paid taxes. The lock-in effect distorts investors’ portfolio choices because it creates a friction for reallocating capital across investments. Investors may retain an appreciated position even when another investment would provide a superior expected return after controlling for the riskiness of the position. In contrast, if an asset falls in value, then investors may have an incentive to accelerate selling the asset to benefit from deducting the loss against other income (when allowed). Thus, realization-based taxation provides incentives for selective realizations by which investors typically minimize taxes by selling their losers and holding their winners. In the extreme, these optimal trading strategies create opportunities to eliminate income taxation (see Stiglitz, 1983); however, the combination of transaction costs and tax restrictions (e.g., loss limitation rules) prevent these strategies from abusing the capital gains tax rules to the point of eliminating overall income taxation. Poterba (2002) reviews the efficiency consequences of capital gains taxation on individuals. One of the challenges for modeling the deadweight loss of capital gains taxation is that a complete model requires understanding investors’ trading behavior in the absence of taxation. Trading behavior depends in part on heterogeneous beliefs about future returns, an aspect of trading behavior that has proven to be a difficult feature to include in a model with taxation. Part of the deadweight loss arises because individuals hold suboptimal portfolios in terms of riskiness or in terms of expectations about future returns. This distortion is greater if individuals’ risk preferences change with age or if the risk characteristics of an investment change over time. Also, the distortion is probably smaller when investors have relatively similar beliefs about future returns or have the ability to undertake investment trading strategies that allow investors to reap the benefits of a sale (e.g., liquidity and disposition of risk) without triggering a capital gain. 15 For recent analyses of the lock-in effect, see Klein (1999, 2001) and Dammon, Spatt, and Zhang (2001).

12

Desai & Gentry

For evaluating the efficiency costs of corporate capital gains taxation, some of the issues that are pertinent for analyzing individuals are less important for corporations. For example, life-cycle concerns about savings and portfolio choice are not critical issues for corporations. Likewise, concerns about a mismatch between risk preferences and the riskiness of a portfolio of assets are less likely to be concerns for corporate investors because the individuals who own the corporation can diversify such risk issues. However, the distortions from capital gains taxes may have effects on corporations that are less relevant for individuals. Corporate investment often differs from the types of investments made by individuals. For individuals, the identity of the owner of an asset rarely affects the asset’s rate of return, at least for the portfolio investments often considered in discussing the lock-in effect. While this may be true for the liquid investments of corporations, the identity of the owner of corporate assets often affects the return on the assets. Returns generated from matching specific assets with specific owners add another dimension to the deadweight loss from capital gains taxation. For example, consider a corporation that is considering selling a division to another firm. If the incumbent owner has an unrealized capital gain on the division, then the capital gains tax might impede the transaction, even when the potential acquirer has a relatively high rate of return from owning the division. When the realization-based capital gains tax discourages transactions, the social cost is the difference in the returns that could be earned by the two different owners. In addition to the possibility of mismatching in the asset market, the patterns of corporate cross-holdings and the accompanying governance issues associated with those cross-holdings could be influenced by corporate capital gains taxation. La Porta, Lopez-de-Silanes, and Shleifer (1999) document the wide variety of corporate cross-ownership in the world and the prevalence of pyramidal ownership, to which the U.S. experience is a notable exception. Morck (2003) suggests that tax rules on intercorporate dividends (for which, as Morck shows, the United States is exceptional compared to other countries in levying an income tax) and corporate dividends to shareholders interact in the United States so that pyramidal ownership and the associated potential governance abuses are prevented. Presumably, the tax on corporate capital gains is an even more important deterrent to cross-holdings given the DRD.16 In addition to these effects on the patterns of ownership, corporate capital gains taxes may shape corporate venture capital 16 Paul (2003) argues that the triple taxation embedded in corporate capital gains taxation has grown more important recently because U.S. corporations have entered into more relationships that involve intercorporate equity holdings. In addition, she discusses how the repeal of the General Utilities doctrine as part of the Tax Reform Act of 1986 has increased the importance of corporate capital gains taxation.


13

activity and the overall venture capital environment, given the interactions between corporate venture capital and venture capital more generally.17 Finally, capital market imperfections may exacerbate the efficiency costs of the taxation of corporate capital gains taxation. Without capital market imperfections, corporations could finance new projects by attracting new investors; with capital market imperfections, asset sales can be a source of financing for new projects. The possibility of selling existing assets to finance new projects has received relatively little attention in the corporate finance literature. As elaborated below, corporate capital gains realizations constitute a significant fraction of corporate cash flow. As a transaction-based tax, the capital gains tax extracts a toll from firms that want to sell one set of assets to invest in another set of assets. Overall, the magnitude of the economic distortions caused by capital gains taxation depends on the elasticity of behavior along the various relevant margins. For individuals, understanding the elasticity of capital gains realizations to the tax rate is a starting point for measuring the efficiency cost; however, the realizations elasticity does not measure the extent to which the capital gains tax distorts portfolio composition. For corporations, the efficiency cost depends on the heterogeneity in asset returns across different owners and the extent to which the capital gains tax reduces capital reallocation across firms. However, measuring the elasticity of realizations with respect to the tax rate may capture only a small part of the efficiency cost of corporate capital gains taxation.

3.2 Fluctuations in Tax Rates and Timing Incentives for Corporate Capital Gains In addition to the relationship between the levels of realization and tax rates, when tax rates change over time, either due to legislated changes in the tax code or due to changes in firm-specific characteristics, firms have an incentive to time their realizations of capital gains and capital losses. The simple adage is: realize losses when the marginal tax rate is high and realize gains when the tax rate is low. If firms anticipate changes in future tax rates, then anticipated tax rate increases can induce firms to sell assets with appreciated values and to defer the sale of assets with unrealized capital losses. A standard issue in the debate over the realizations elasticity of individual capital gains is separating the responsiveness of realizations into the responsiveness to permanent changes in tax rates versus transitory changes in tax rates.18 A response to transitory changes in tax rates is more likely to involve a pure shift in the timing of asset sales rather than an 17

See Gompers and Lerner (2002) and Gompers, Lerner, and Scharfstein (2003).

18

See Burman and Randolph (1994).

14

Desai & Gentry

increase in the long-run amount of realizations. This same issue arises when considering corporate capital gains. Changes in tax policy can lead to anticipated changes in tax rates that affect behavior. Given the graduated corporate tax rate system and the loss carryforward regime, variability in corporate profitability can generate firm-specific variation in effective marginal tax rates. For example, a firm with net operating loss carryforwards can recognize capital gains without paying taxes in the current year. Instead, the realized gain offsets part of the net operating loss carryforward and reduces the stock of carryforwards taken into the future. The realized capital gain will most likely increase the firm’s future tax liability when it returns to paying taxes.

3.3 Tax Planning and Corporate Capital Gains Taxes The taxation of capital gains also affects the tax-planning efforts of U.S. corporations. In general, these tax-planning responses lead to financial consequences without greatly changing the real activity of the firm. A general rule for corporate tax planning is that firms prefer to have capital gains income but ordinary losses because capital gains can be used to offset either capital losses or ordinary losses. We now turn to several examples of corporate tax planning that are affected by realization-based taxation of gains, especially capital gains. Our first example is how the realization-based nature of capital gains taxation affects the design of merger and acquisition transactions. We already discussed how capital gains taxes can inhibit some asset sales. In addition, the tax rules influence the form of asset transfers. In some cases, it is possible to structure an acquisition so that it defers the realization of capital gains taxes; a common feature in deferring the capital gains tax is that the seller accepts stock instead of cash from the acquirer.19 For example, instead of selling a division for cash and realizing a gain, a corporation can exchange its equity in the division for stock of the acquirer and defer the realization of the gain. Early empirical research on the effects of taxes on merger and acquisition activity found a limited role for taxes. Auerbach and Reishus (1988) find that only a minority of mergers from 1968 to 1983 had large enough tax benefits that the taxes may have been a motivating factor in the reorganization.20 They also find little evidence of taxes affecting the form of 19 See Scholes et al. (2002) for an overview of how capital gains taxes affect mergers and acquisitions. 20 Auerbach and Reishus (1988) focused on elements of the tax code that potentially made mergers more attractive, such as allowing firms to use tax losses and credits (rather than carry them forward), a step up in asset basis, and increased interest deductions; Franks, Harris, and Mayer (1988) also conclude that taxes do not seem to affect the form of the transaction for mergers.


15

acquisitions, but they recognize that their measures of tax benefits are imprecisely measured. In contrast, more recent research has documented a role for taxes in corporate reorganizations. For example, Maydew, Schipper, and Vincent (1999) examine a sample of divestitures in which the parent corporation could choose either an asset sale that would trigger the realization of gains or a tax-free spin-off that would not trigger taxation. They find that the size of the tax differential between divestiture methods affects the form of divestiture; firms with the largest potential tax benefits from using a spin-off opt for spin-offs.21 Thus, realizationbased taxation affects the form of the transaction. The realization-based taxes on the sellers in a corporate reorganization can also affect the price paid in the reorganization. Ayers, Lefanowicz, and Robinson (2003) report that the acquisition premium associated with taxable stock acquisitions increases with the capital gains taxes of the target shareholders (although their sample is based on individual rather than corporate shareholders). Erickson and Wang (2000) examine the acquisition prices when one corporation buys a subsidiary from another corporation and find that the price paid depends on the tax on the gain realized by the selling corporation. A second tax-planning example is that the realization-based nature of capital gains taxation provides an incentive for investors to seek alternatives to selling their investments. One possibility is to enter into a hedging transaction that can reduce the risk of the position and possibly raise cash. Such hedging transactions may be relevant when a corporation obtains shares in another corporation as payment in a corporate reorganization.22 Corporations can execute these hedging transactions through either private deals with investment banks or by issuing exchangeable securities. Gentry and Schizer (2003) examine a sample of corporations that issue public securities as a way of hedging an appreciated position, raising cash, and deferring capital gains taxation. While the volume of public transactions has been relatively modest (roughly $25 billion between 1993 and 2001), private transactions may actually have lower transaction costs and thus may be the predominant form of hedging. Again, the government designs tax rules, such as rules that treat the transaction as a sale if the issuer has eliminated all the risk of the position, to 21 In general, Maydew, Schipper, and Vincent (1999) report that the potential nontax benefits of taxable transactions (e.g., raising capital from asset sales may be a relatively inexpensive source of funds) lead many firms to use taxable asset sales and forego the tax benefits of a spin-off. 22 For example, in 1996, Kerr-McGee acquired stock of Devon Energy in exchange for some oil fields, and in 1999 Kerr-McGee issued securities that hedged some of the risk of holding Devon Energy stock. See Gentry and Schizer (2003) for more examples of such transactions.

16

Desai & Gentry

make this sort of tax planning difficult. Nonetheless, corporate capital gains taxation plays an important role in securities market innovations aimed at allowing firms to avoid the realization-based tax. A third example of tax-planning incentives arises from the differential taxation of corporate capital gains income and intercorporate dividend income. As mentioned in the introduction, the dividends received deduction (DRD) for intercorporate dividends provides some relief from the potential triple taxation of corporate income when corporations own shares in other corporations. The DRD typically reduces the tax rate on intercorporate dividends by 70 percent. Thus, a corporation facing a 35 percent tax rate on ordinary income faces a 10.5 percent tax rate on dividend income; however, capital losses are deductible against capital gains so that the tax rate on capital losses may still be 35 percent. The DRD makes corporations a natural clientele to invest in high-dividend yield stocks, such as preferred stock.23 In addition, these clientele effects can affect when a corporation sells stock. For a corporation considering selling shares near the ex-dividend day, the DRD provides an incentive for the corporation to delay the sale until after the ex-dividend day because this delay increases the dividend portion of the return and increases the after-tax return. At the extreme, these clientele effects give corporations an incentive to engage in short-term trading strategies known as dividend capture or dividend stripping. Dividend stripping is an investment strategy aimed at earning dividend income even if the dividend income is offset by an equal amount of capital losses. Suppose a corporation can invest in a high-yield stock just before its ex-dividend day and that on the exdividend day, the stock price drops one for one with the amount of the dividend. A short-term position in this stock might yield $100,000 of dividend income but would also result in $100,000 in capital losses, so that the economic income on the position is zero. However, the dividend income might result in a tax liability of $10,500, while the capital loss (assuming that it can be used to offset capital gains) creates a tax benefit of $35,000. Thus, the tax-rate differential creates a net benefit of $24,500. Given the size of this potential tax arbitrage, it is not surprising that the tax code includes various restrictions limiting this type of strategy, primarily minimum holding period requirements to qualify for the DRD. Evidence about the importance of dividend stripping is scarce. The most direct evidence is from Koski and Scruggs’s (1998) examination of New York Stock Exchange (NYSE) trading audit data from the early 23 For evidence consistent with the formation of investor clienteles for dividends, see Dhaliwal, Erickson, and Trezevant (1999).


17

1990s. The data includes the buying and selling volume of different types of traders. Consistent with corporate dividend capture, they find increased trading volume for corporations just before the ex-dividend day. The strength of the result is tempered, however, by this increase in volume being uncorrelated with the dividend yield, whereas one would expect a positive correlation because the profitability of dividend stripping increases with dividend yield. In addition, Naranjo, Nimalendran, and Ryngaert (2000) examine the ex-dividend stock returns of high yield stocks. They find that after May 1975, when brokered commissions were introduced on the New York Stock Exchange, the ex-dividend day returns for these stocks were negative (i.e., on the ex-dividend day, share prices fall by more than one for one with the dividend) in every year except 1994, which is consistent with corporate dividend capture affecting the stock returns due to corporations bidding up the price of the shares before the exdividend day. Furthermore, these negative returns are correlated with the corporate tax differential on dividend and capital gains income, consistent with the tax differential driving trading behavior. Grammatikos (1989) focuses on the 1984 tax changes that increased the holding period required to qualify for the DRD, which presumably increased the cost of dividend stripping by increasing the associated risk. Consistent with the increased holding period reducing dividend stripping, Grammatikos reports that abnormal returns on the ex-dividend day increased after the tax change. Overall, these studies suggest that corporations engage in some amount of dividend capture trading but that both transaction costs associated with trading and tax restrictions play important roles in limiting this behavior.

4. THE SCOPE OF CORPORATE CAPITAL GAINS Before turning to the effects of corporate capital gains taxes on the realization behavior of firms, it is useful to get a sense of the magnitude and distribution of corporate capital gains. Data on corporate capital gains is available from 1954 to 1999 through the Statistics of Income (SOI) reports on corporation income tax returns. The top panel of Table 1 provides summary numbers on the aggregate amount of gains for 1999 distinguished by type of gain. In 1999, net long-term capital gains realized by all corporate entities amounted to $146.5 billion, net short-term capital gains realized amounted to $94.9 billion, and net gains on all noncapital assets amounted to $64.7 billion. Given the proliferation of pass-through corporate entities and the possible concentration of capital gains in particular sectors, it is useful to isolate the volume of gains according to these distinctions. While net long-term capital gains and net gains on noncapital assets are largely in non-pass-through entities, the vast majority of net

Net short-term capital gains, 1999 Net gains on noncapital assets, 1999 The distribution of economic activity, 1990–1999 Share of total assets, 1990–1999 Share of total receipts, 1990–1999 Share of income subject to tax, 1990–1999 The distribution of gain activity, 1990–1999 Share of net long-term gains, 1990–1999 Share of net short-term gains, 1990–1999 Share of net gains on noncapital assets, 1990–1999

All industries

Manufacturing

FIRE

Other industries

146,520.147 144,547.184 94,913.405 14,314.130 64,698.446 58,284.289

32,817.607 32,817.607 3,590.537 3,590.537 19,372.208 18,309.051

40,658.231 38,685.268 86,208.324 5,609.048 8,027.470 6,820.218

73,044.309 73,044.309 5,114.544 5,114.545 37,298.768 33,155.020

All entities Non-pass-through entities All entities Non-pass-through entities All entities Non-pass-through entities

18.7% 21.4% 30.8% 33.6% 40.8% 39.9%

56.3% 49.2% 16.0% 16.8% 23.3% 22.8%

25.0% 29.4% 53.2% 49.6% 35.9% 37.3%

All entities Non-pass-through entities All entities Non-pass-through entities All entities Non-past-through entities

25.7% 25.7% 3.3% 15.1% 34.4% 38.1%

39.6% 36.6% 91.7% 60.4% 20.4% 18.1%

34.7% 37.7% 4.9% 24.5% 45.3% 43.8%

All entities Non-pass-through entities All entities Non-pass-through entities All entities Non-pass-through entities

Note: The top panel characterizes the magnitude of capital gain activity in 1999 (in thousands of dollars) for all entities and only non-pass-through entities in varied industrial groupings. The middle panel characterizes the distribution of economic activity for the 1990s for all entities and only non-pass-through entities in varied industrial groupings. The ratios presented in the rows of the middle panel sum to 100 percent. The bottom panel characterizes the distribution of capital gain activity for the 1990s for all entities and only non-pass-through entities in varied industrial groupings. The ratios presented in the rows of the bottom panel sum to 100 percent.

Desai & Gentry

The magnitude of gain activity, 1999 Net long-term capital gains, 1999

18

TABLE 1 The Magnitude and Distribution of Corporate Capital Gains, 1990–1999


19

short-term capital gains in 1999 were in pass-through entities, most of which were in the finance, insurance, and real estate (FIRE) sector. To isolate further the sectoral distribution of gains and to ensure that they do not reflect a peculiarity associated with 1999, the bottom two panels of Table 1 provide the share of overall economic activity and of gain activity through the 1990s for manufacturing, FIRE, and other industries for all entities and for only non-pass-through entities. Throughout the 1990s, net short-term gains were concentrated in FIRE. In contrast, net long-term gains were distributed across all three industrial groupings in a manner that accords with the underlying distribution of assets across those groupings. Finally, net gains on noncapital assets are disproportionately in manufacturing and other industries (relative to their shares of total assets), which suggests that these gains correspond to the section 1231 assets described above. Given the distribution of gain activity across sectors and types, the analysis and descriptive statistics that follow emphasize net long-term gains from the SOI data. The data in Table 1 distinguish between short-term and long-term gains and separate the gains on noncapital assets, but they do not address what specific types of assets are being sold. While such a breakdown is not readily available for the United States, Inland Revenue reports what types of assets underlie corporate capital gains realizations for the United Kingdom.24 For the accounting period ending 2000–2001, 51 percent of the gains of non-lifeinsurance companies were from financial assets; more specifically, 16 percent were from shares listed on the London exchange, 16 percent were from unquoted shares, 17 percent were from selling subsidiaries, and 2 percent were from other financial assets. The remaining 49 percent of capital gains were from nonfinancial assets that were concentrated in intangible assets (27 percent of total gains) and commercial assets (13 percent of total gains). In terms of holding periods, for gains realized in 2000–2001, 70 percent of the gains on financial assets and 58 percent of the gains on nonfinancial assets were on assets held for over 10 years, which highlights the importance of tax deferral for gains taxation.25 To the extent that U.S. and U.K. corporations are similar, these data provide a general picture of the types of assets that generate corporate capital gains in the United States. Given the familiarity with capital gains realized by individuals, it is useful to frame the volume of net long-term corporate capital gains 24

Inland Revenue, National Statistics, http://www.inlandrevenue.gov.uk/stats/capital_ gains/menu.htm. It is worth noting that Inland Revenue states that annual data varies substantially across years.

25 The pattern of holding periods varies over time (even more than the variation in the sources of gains); for 1999–2000, 37 percent of realized gains on financial assets and 54 percent of gains on nonfinancial assets were from assets held over 10 years.

20

Desai & Gentry

FIGURE 1. Ratio of Corporate to Individual Capital Gains, 1954 –1997 Note: The figure plots the ratio of corporate capital gains to individual capital gains from 1954 to 1997. Corporate capital gains are defined as net long-term capital gains reduced by net short-term losses for all active corporations. Ratios are expressed in percentage terms.

realizations relative to those gains. As Figure 1 demonstrates, the ratio of realized corporate capital gains to realized individual capital gains has averaged approximately 0.30 from 1954 to 1998. It is useful to note that this ratio has evolved significantly over that period. Until the late 1970s, this ratio was both relatively lower and more consistent than the period after the late 1970s. Specifically, from 1980 on, this ratio averaged 0.36, and it ranged from a high of 0.45 in 1987 to a low of 0.28 in 1984. Overall, the relative magnitude of corporate and individual capital gains suggests further research on corporate capital gains is warranted, especially because corporations face higher tax rates on capital gains than individuals face. To give a sense of how important capital gains are for corporate behavior, it useful to frame the magnitude and trends in corporate capital gains realizations relative to corporate cash flow and assets. To do so, Table 2 provides the ratio of net long-term capital gains realizations to income subject to tax and assets and the ratio of all gains to assets for the 1990s by industrial grouping. Of course, the gain may be much smaller than the proceeds raised by selling an asset so that merely measuring gains understates the importance of asset sales for cash flow.26 This ratio may be subject to cyclical effects because it was at its highest value (33.5 percent) 26 If corporations more readily sell assets with losses than those with gains (a pattern encouraged by the tax rules), then the net gain may be a very small fraction of the cash proceeds from aggregate sales.


21

TABLE 2 Scope of Corporate Capital Gains, Non-Pass-Through Entities, 1990–1999 Year

All industries

Manufacturing

Ratio of net long-term gains to income subject to tax 1990 33.5% 21.0% 1991 11.7% 7.4% 1992 11.9% 7.3% 1993 12.2% 8.5% 1994 9.6% 8.3% 1995 10.8% 7.1% 1996 11.7% 8.3% 1997 14.7% 10.0% 1998 18.7% 9.8% 1999 20.9% 13.2%

FIRE

Other industries

62.9% 19.6% 22.5% 22.2% 12.3% 17.3% 21.1% 22.8% 38.9% 31.9%

27.4% 9.2% 8.6% 8.0% 7.8% 8.5% 8.4% 8.4% 9.9% 10.3%

Ratio of net long-term gains to assets 1990 0.27% 1991 0.25% 1992 0.26% 1993 0.29% 1994 0.24% 1995 0.28% 1996 0.32% 1997 0.38% 1998 0.42% 1999 0.45%

0.36% 0.29% 0.28% 0.36% 0.40% 0.35% 0.43% 0.50% 0.38% 0.50%

0.16% 0.17% 0.24% 0.27% 0.13% 0.21% 0.27% 0.27% 0.43% 0.31%

0.42% 0.37% 0.28% 0.25% 0.36% 0.39% 0.33% 0.57% 0.43% 0.55%

Ratio of all gains to assets 1990 0.46% 1991 0.46% 1992 0.47% 1993 0.50% 1994 0.41% 1995 0.50% 1996 0.54% 1997 0.62% 1998 0.67% 1999 0.67%

0.64% 0.54% 0.52% 0.62% 0.67% 0.66% 0.83% 0.92% 0.80% 0.83%

0.26% 0.33% 0.41% 0.44% 0.20% 0.34% 0.37% 0.40% 0.56% 0.41%

0.76% 0.69% 0.58% 0.54% 0.68% 0.75% 0.69% 0.94% 0.70% 0.84%

Note: The top panel provides the ratios of net long-term gains to income subject to tax for all non-passthrough entities through the 1990s for all industries and selected subindustries. The middle panel provides the ratios of net long-term gains to total assets for all non-pass-through entities through the 1990s for all industries and selected subindustries. The bottom panel provides the ratios of all gains to assets for all non-pass-through entities through the 1990s for all industries and selected subindustries.

22

Desai & Gentry

during the one economic downturn during this period (1990). In general, for all industries, there seems to be an upward trend in the second half of the sample, with the ratio of net long-term capital gains to income subject to tax increasing from 9.6 percent in 1994 to 20.9 percent in 1999. This same ratio for the basic industrial grouping of manufacturing, FIRE, and all other industries suggests that FIRE is particularly reliant on net long-term capital gains but that the cyclical nature and recent increase is also evident for manufacturing and other industries. By 1999, 13.2 percent of income subject to tax for manufacturing firms was net long-term gains. Given that the cyclical nature and upward trend in this ratio may reflect the dynamics of income subject to tax rather than the dynamics of net long-term capital gains, the second panel of Table 2 demonstrates that those same trends hold when scaling net long-term gains by total assets. In 1999, firms across all industries realized net long-term capital gains equal to 0.45 percent of total assets, which represented a sharp increase over the decade. The bottom panel of Table 2 aggregates all gains, compares them to total assets, and finds largely similar results. The upward trend in gains realizations appears to be particularly significant for the manufacturing sector.

5. EVIDENCE FROM TIME-SERIES ANALYSIS Analyses of capital gains realization behavior for individuals have employed the responsiveness of aggregate realizations to time-series variation in tax rates. For example, Eichner and Sinai (2000) follow Auerbach (1988) in estimating long-run elasticities on the basis of such a time-series analysis. While limited in several ways, such an analysis of corporate capital gains realization is a useful starting point prior to turning to firm-level data. Before turning to the time-series regressions, it useful to get a sense of the general pattern in realization behavior and to compare it to the timeseries pattern of individual realization behavior. Figure 2 traces the relationship between realization behavior at the individual level and the corporate level. It also plots the ratio of individual capital gains to household financial assets and the ratio of net long-term corporate capital gains to total assets. Burman and Plesko (2002) provide a version of this figure but deflate the two nominal series and conclude that a correlation of 0.97 exists over the period. Scaling realization amounts as in Figure 2 provides a similar conclusion regarding the high level of correlation between these series. Plesko (2002) interprets the high correlation between corporate and individual realization in the time series as evidence of some omitted


23

FIGURE 2. The Importance of Capital Gains for Individuals and Corporations, 1954–1997 Note: The figure plots the ratio of realized individual capital gains to household financial assets (on the left axis) and the ratio of realized corporate capital gains to total corporate assets (on the right axis) from 1954 to 1997. Corporate capital gains are defined as net long-term capital gains reduced by net short-term losses for all active corporations. Ratios are expressed in percentage terms.

variable in realization behavior that may bias estimated tax effects for individuals’ capital gains realization behavior.27 Our goal is to identify variables that may be omitted from the standard set of variables employed by Eichner and Sinai (2000) for individuals and by Plesko (2002) for both individuals and corporations. In searching for these omitted variables, we limit our analysis to corporate capital gains, which is the general focus of our paper, rather than estimating models for both individuals and corporations. Column (1) of Table 3 provides the baseline specification that follows Plesko’s analysis of realization behavior for corporations, which in turn follows Eichner and Sinai and others in choosing explanatory variables. This specification employs the log value of aggregate corporate capital gains realization as a dependent variable and, in addition to the top corporate capital gains tax rate, controls for the price level (as measured by the gross national product [GNP] deflator), the value of corporate equities (as measured by the level of the Standard & Poor [S&P] 500), and GNP and the first difference of GNP. As with Auerbach (1988) and Eichner and Sinai (2000), all values 27 Plesko (2002) jointly estimates individual and corporate realizations and concludes that single-equation models of individual realization behavior overstate tax sensitivities for individual realization behavior.

24

Desai & Gentry

are first-differenced to accommodate concerns regarding the presence of a unit root in these series. The −9.20 coefficient on the corporate capital gains rate in column (1) translates into a tax elasticity of −2.6. Starting with this baseline set of variables, we consider other variables that might capture factors influencing realization behavior. Because these variables are available only after 1963, the specification in column (2) provides an alternative baseline specification for this shortened period, with estimated coefficients on the corporate capital gains rate that are approximately the same as for the longer period examined in column (1). Columns (3) to (6) of Table 3 consider several proxies for sentiment and capital market activity that could influence these series. Baker and Wurgler (2003) provide a review of various measures of sentiment and their interrelationship. If managers can opportunistically time sales and capitalize on market sentiment, these variables may explain realization activity. Similarly, if these realizations are related to trends in merger activity or equity offerings, then measured responsiveness to taxes could be mismeasured. Column (3) adds an additional control for the closed-end fund discount, which has been used to proxy for market sentiment. The coefficient on the capital gains rate moves modestly and retains its high level of significance, while variation in the closed-end fund discount appears to be associated with realization behavior to some limited degree. Column (4) tests for the role of merger activity by controlling for the share of value of the University of Chicago’s Center for Research in Securities Prices (CRSP) file that is acquired in a given year.28 Again, the coefficient on the capital gains tax rate is largely unchanged and remains significant, while this proxy for merger activity does not appear to determine realization behavior significantly. Finally, the inclusion of the level of initial public offering (IPO) activity in column (5) does appear to play a significant positive role in determining realization behavior and reduces the level and significance of the coefficient on capital gains tax rate. Jointly controlling for merger and IPO activity, as in column (6), produces a marginally significant coefficient that translates into a tax elasticity of −1.3, which is at the upper end of Eichner and Sinai’s estimated elasticities for individuals. Such a time-series analysis provides indicative evidence that measures of capital market activity may shape realization behavior either because they provide an opportunity for corporations to disgorge capital gains or because they measure sentiment in a way that might shape realization behavior. Further investigation of the role of these measures of sentiment 28 Andrade, Mitchell, and Stafford (2001) provide background on the construction of this series.

TABLE 3 Determinants of Corporate Capital Gains Realizations, 1954–1998 (Dependent Variable: Deflated Corporate Capital Gain Realizations)

Corporate capital gains rate Log real GNP Lagged log real GNP Log S&P 500 index Log GNP deflator

(2) 1963–1997

(3) 1963–1997

(4) 1963–1997

(5) 1963–1997

(6) 1963–1997

−0.0973 (0.1087) −9.2577 (2.6969) 3.5385 (2.0995) −0.8293 (1.3772) 0.7231 (0.2230) 0.8356 (1.4820)

−0.1411 (0.1260) −9.0111 (2.8258) 4.4797 (2.1710) −0.1636 (1.5463) 0.5973 (0.2283) 1.1678 (1.6930)

−0.1464 (0.1316) −7.8719 (4.0530) 4.6545 (2.1636) −0.0614 (1.5283) 0.5549 (0.2304) 1.0589 (1.8423) −0.0048 (0.0066)

−0.1710 (0.1391) −8.9049 (3.0853) 5.0439 (2.3573) −0.2418 (1.5204) 0.6322 (0.2201) 1.4632 (1.7788)

−0.2692 (0.1357) −4.1537 (2.9736) 6.7226 (2.0242) −2.0807 (1.4875) 0.6268 (0.2189) 2.007 (1.9352)

−0.2617 (0.1393) −3.9293 (2.9881) 6.5702 (2.0375) −2.1513 (1.5956) 0.6111 (0.2304) 1.9014 (2.0358)

0.0004 (0.0001) 35

1.6430 (5.1817) 0.0004 (0.0002) 35

Closed-end fund discount Percentage of CRSP value acquired

−3.3062 (3.9924)

Number of IPOs Number of observations

42

35

35

35

25

Note: The columns present specifications where the dependent variable is the log deflated value of net long-term corporate capital gain realizations described in the paper. All variables are in first differences. “Corporate capital gains rate” is the applicable corporate capital gains rate over the sample period. “Log S&P 500 index” is the log value of the S&P 500 index over the sample period. “Log GNP deflator” is the log value of the GNP deflator over the sample period. “Closed end fund discount” is the average closed end fund discount over the sample period. “Percentage of CRSP value acquired” is the share of CRSP value acquired in a given year, as presented in Andrade, Mitchell, and Stafford (2001). “Number of IPOs” is the number of initial public offerings in a given year.


Constant

(1) 1956–1997

26

Desai & Gentry

and capital market activity in determining individual and corporate realizations seems warranted. Inclusion of these measures of capital market activity still produces large, if only marginally statistically significant, tax elasticities for corporate realization behavior. Obviously, while this timeseries analysis has the advantage of capitalizing on significant variation in capital gains tax rates, it suffers from well-known econometric problems. To investigate the effect of capital gains taxes on realization behavior further, we turn to analysis of firm-level data.

6. EVIDENCE FROM CORPORATE FINANCIAL REPORTS Much as Burman and Randolph (1994) and Auten and Clotfelter (1982) investigate individual realization behavior using micro data, this section employs firm-level data to investigate whether a firm’s tax position influences the decision to dispose of assets and the nature of gain and loss recognition. We use financial reporting data from Standard & Poor’s Compustat industrial database to shed light on corporate capital gains behavior. Financial statements include several items related to taxable capital gains. First, firms report their proceeds from the sale of investments and their proceeds from the sale of property, plant, and equipment (PPE). Presumably, the sale of investments captures many assets defined by the tax code as capital assets and the sale of PPE captures so-called section 1231 assets (which can create a combination of capital gains and ordinary income). For these variables, we use data from 1980 to 2002. Obviously, observing sales proceeds does not necessarily inform us about the recognition of gain or loss. For the years 1987 to 2002, Compustat reports the gain (or loss) on the sale of assets. This variable, however, may not match taxable capital gains (or losses) perfectly for several reasons. First, financial reporting does not isolate assets using the tax code’s definition of capital asset, so the financial reporting gain may include some ordinary income. Second, for depreciable assets, the depreciation rules differ between financial and tax accounting. Despite these measurement issues, we believe that financial reporting data can shed light on corporate capital gains behavior. Unlike studies of individual capital gains realization behavior that use tax return data, financial reporting data has two distinct limitations. First, using such sources implicitly relies on the decision of a firm to disclose specific actions in public documents. Reporting decisions are mediated presumably by auditor advice and managerial motives. Thus, it is unclear a priori why reporting the presence or volume of asset sales or gain/loss activity would be subject to anything but possibly materiality concerns. In


27

addition, we should emphasize that financial accounting differs from tax accounting, so the measured gain or loss from selling assets differs across accounting systems. Second, public financial documents do not allow one to infer precisely the tax position of a firm, which forces us to rely on proxies devised by Graham (1996) that largely identify probabilities of having net operating losses.29 Before turning to our regression analysis of when firms sell assets and the associated gains or losses, the descriptive statistics in panel B of Table 4 provide some useful information. To measure the propensity of different events, we create dummy variables for whether a firm reports (1) some proceeds from the sale of investments, (2) some proceeds from the sale of PPE, (3) a gain from the sale of assets, and (4) a loss from the sale of assets.30 For the period from 1980 to 2002, 26 percent of firm-year observations contain positive values for the sale of investments, and 50 percent report the sale of PPE, indicating that asset disposal is fairly common. With regard to gains and losses, the sample is limited to the period from 1987 to 2002, but the implications are similar. Over this period, 45 percent of firm-year observations contain either a net gain or loss, with 26 percent of those being gains and 19 percent characterized as losses. As such, gain or loss recognition appears to be fairly common over the sample period. The two panels of Table 5 analyze the determinants of disposal decisions by examining how taxes influence the decision to sell investments (panel A) and PPE (panel B). The first three columns of each panel employ a dummy variable equal to 1 if there is a nonzero value for the sale of either investments (in panel A) or PPE (in panel B) as the dependent variable. We estimate linear probability models for the various extensive margins that we examine so that we can easily incorporate firm fixed effects in the econometric specification. The remaining three columns use the log of the value of those sales as the dependent variable. In examining the size of the sales, we include only observations that have the particular type of sale; hence, the regressions are conditional on having a sale and do not combine the effects of the explanatory variables on the extensive and intensive decisions regarding asset sales. As such, the latter three columns of both panels analyze whether taxes influence the magnitude of these sales conditional on the presence of a recorded sale. All specifications 29 We thank John Graham for making his tax-rate variables available via his Web site at http://www.duke.edu/~jgraham/. Additional discussion of the methodology underlying his tax-rate measure is available in Graham (1996). 30 The dummy variables for recognizing a gain or a loss are created from a single continuous measure of the net gain from asset sales; hence, we cannot infer whether a firm simultaneously recognizes gains and losses.

28

Desai & Gentry TABLE 4 Descriptive Statistics for Regression Analysis Mean

Panel A: times series analysis Log deflated net long-term capital gains realizations Corporate capital gains rate Log real GNP Log S&P 500 index Log GNP deflator Closed end fund discount Percentage of CRSP value acquired Number of IPOs Panel B: panel analysis Sale of investment dummy Log proceeds from sale of investments Sale of PPE dummy Log proceeds from sale of PPE Gain dummy Log value of gain Loss dummy Log value of loss Marginal tax rates Log total assets Log q ratio

Standard Median deviation

Number of observations

−1.1533

−1.2068

0.6601

44

0.2882 8.4173 1.2663 2.9447 8.8669 1.22%

0.2800 8.4775 1.1682 2.9313 9.3820 1.09%

0.0367 0.4640 0.5286 0.4036 8.1197 0.85%

44 49 49 44 39 36

352

351

263

40

0.2587 1.3975

0.0000 1.2834

0.4379 3.2458

91,325 23,626

0.5016 −0.7323 0.2627 −0.2014 0.1851 −1.9489 0.2074 4.6701 0.4024

1.0000 −0.7012 0.0000 −0.2231 0.0000 −2.0636 0.2943 4.5446 0.2312

0.5000 2.7451 0.4401 2.8163 0.3884 2.5255 0.1839 2.3820 0.6546

76,325 38,284 67,741 17,797 67,741 12,541 100,646 100,646 100,646

Note: Panel A provides descriptive statistics for the sample employed in the time series analysis presented in Table 4. Panel B provides descriptive statistics for the sample employed in the panel analysis presented in Tables 5 and 6. “Log deflated net long-term capital gains realizations” is the log value of net long-term corporate capital gain realizations described in the paper. “Corporate capital gains rate” is the applicable corporate capital gains rate over the sample period. “Log S&P 500 index” is the log value of the S&P 500 index over the sample period. “Log GNP deflator” is the log value of the GNP deflator over the sample period. “Closed end fund discount” is the average closed end fund discount over the sample period. “Percentage of CRSP value acquired” is the share of CRSP value acquired in a given year as presented in Andrade, Mitchell, and Stafford (2001). “Number of IPOs” is the number of initial public offerings in a given year. “Sale of investment dummy” is a dummy variable equal to 1 if a corporation reports the sale of investments in a given year. “Log proceeds from sale of investments” is the log value of those sale proceeds. “Sale of PPE dummy” is a dummy variable equal to 1 if a corporation reports the sale of PPE in a given year. “Log proceeds from Sale of PPE” is the log value of those sale proceeds. “Gain dummy” is a dummy variable equal to 1 if a corporation reports the gain on the sale of investments and PPE in a given year. “Log value of gain” is the log value of that gain value. “Loss dummy” is a dummy variable equal to 1 if a corporation reports the loss on the sale of investments and PPE in a given year. “Log value of loss” is the log value of that loss value. “Marginal tax rates” are calculated via the methodology described in Graham (1996). “Log total assets” is the log value of total firm assets. “Log q ratio” is the log of the q ratio calculated from Compustat data as described in the paper.

TABLE 5 Determinants of Corporate Asset Disposal Behavior, 1980–2002 Sale dummy −0.0459 (0.0090)

Log total assets Log q Number of observations Number of firms Firm fixed effects? Year fixed effects?

88,102 12,788 Y N

Panel B: disposal of PPE Marginal tax rate

−0.0402 (0.0097) 0.0396 (0.0021) −0.0044 (0.0032) 84,122 12,645 Y Y

−0.5111 (0.0991)

22,612 5,997 Y N

Sale dummy −0.0747 (0.0107)


−0.0400 (0.0092) 0.0410 (0.0018) −0.0028 (0.0031) 84,122 12,645 Y N

73,530 12,033 Y N

−0.0669 (0.0109) −0.0164 (0.0021) −0.0387 (0.0036) 70,023 11,878 Y N

−0.2778 (0.0949) 0.0618 (0.0372) 0.9378 (0.0191) 21,749 5,861 Y N

−0.0452 (0.0985) 0.8460 (0.0228) 0.0271 (0.0379) 21,749 5,861 Y Y

Log sale value −0.1223 (0.0114) 0.0072 (0.0025) −0.0290 (0.0036) 70,023 11,878 Y Y

−0.8025 (0.0679)

36,724 7,756 Y N

−0.5995 (0.0680) 0.6160 (0.0160) −0.1248 (0.0285) 35,213 7,606 Y N

−0.7698 (0.0711) 0.6643 (0.0193) −0.1184 (0.0291) 35,213 7,606 Y Y

29

Note: Panel A presents specifications that analyze investment disposal behavior, and Panel B presents specifications that analyze PPE disposal behavior. In both panels, the first three columns employ a dummy variable set equal to 1 if there is a sale as a dependent variable. In both panels, the second three columns employ the log value of that sale amount as a dependent variable. All columns employ firm fixed effects, and columns 3 and 6 of both panels also employ year fixed effects. “Marginal tax rates” are calculated via the methodology described in Graham (1996). “Log total assets” is the log value of total firm assets. “Log q ratio” is the log of the q ratio calculated from Compustat data as described in the paper.


Panel A: disposal of investments Marginal tax rate

Log sale value

30

Desai & Gentry

employ firm fixed effects and, as a consequence, are identifying tax effects from within-firm variation in a firm’s tax status. The first column of both panels employs only firm-level variation in tax rates as an explanatory variable and finds that firms time their disposal of investments and PPE to occur in years associated with low tax rates. The estimated coefficients imply that when a firm’s effective marginal tax rate is 10 percentage points higher, the probability of the firm selling investments decreases by 0.46 percentage points and the probability of it selling PPE decreases by 0.75 percentage points. Both of these estimated effects are statistically different from zero at the 99 percent confidence level. While firm fixed effects control for a host of unobservable characteristics, it is useful also to control for firm size and investment opportunities by including the natural logarithm of total assets and the natural logarithm of a proxy for a q ratio.31 Because we add these control variables to the specification with firm fixed effects, the econometric identification for the size and investment opportunity variables arises from each firm’s size and investment opportunities changing over time. The second column of each panel includes these additional control variables and finds that the tax effects identified previously remain statistically significant and are diminished only slightly. Turning to the estimated coefficients on the control variables, we find that when a firm is larger, it is more likely to sell investments but less likely to sell PPE. When firms have better investment opportunities (i.e., higher values of q), they are less likely to sell PPE (and this estimated effect is statistically different from zero), which is not consistent with the idea that firms sell assets as a source of financing when opportunities are good but outside sources of finance are limited. We do not find the relationship between investment opportunities and the sale of investments to be statistically different from zero. By including firm fixed effects, our econometric identification strategy focuses on within-firm variation over time. Some of this intertemporal variation arises from legislated changes in the tax schedule, and some of the variation comes from changes in each firm’s tax position for a given tax code (e.g., how it is affected by loss offset rules). In part, to separate these sources of variation, we include year fixed effects in the specification reported in the third column of each panel in Table 5. The inclusion of year fixed effects in the regressions has a quite small effect on estimated effects in the sale of investment regressions, but the estimated effects in the sale of PPE regressions change considerably (e.g., the estimated coefficient on the marginal tax rate shifts from −0.0669 in the second column 31 This q ratio is the ratio of total assets plus the difference between the market value and book value of equity to total assets.


31

to −0.1223 in the third column). One possible explanation for this result is that the time-series changes in the level of the corporate marginal tax rate are correlated with changes in depreciation rules (which are not as relevant for selling investments that are not typically depreciable assets). The year fixed effects may be capturing how changing depreciation rules affect the propensity to sell PPE, and these differences may be correlated with the level of the tax rate. The final three columns of each panel of Table 5 provide a similar analysis using the log value of proceeds from the sale of investments (in panel A) and PPE (in panel B) as the dependent variable. Conditional on selling investments or PPE, a firm’s tax rate is negatively related to the volume of its investment or PPE sales. These estimated effects are statistically different from zero at the 99 percent confidence level. Including controls for firm size and investment opportunities diminishes the magnitude of the estimated effect of the tax rate, but it retains its high level of significance. While larger firms tend to sell a larger volume of investments and PPE (conditional on selling some assets), the estimated effect of investment opportunities (as measured by q) on the size of investment sales is positive, suggesting that firms with better investment opportunities sell more investment assets possibly as a source of financing investment in new projects. But the estimated effect of investment opportunities on the size of PPE sales is negative. Finally, the inclusion of year effects has little effect on the size or statistical significance of the estimated coefficients in the PPE regression; however, for the size of investment sales, the estimated effect of the marginal tax rate is much smaller and not statistically different from zero, and the coefficients on the other explanatory variables also change dramatically. The two panels of Table 6 provide a similar empirical framework for investigating the presence and magnitude of gains and losses on the sale of assets as reported by firms. In contrast to the two panels of Table 5, the two panels of Table 6, examine gain behavior (in panel A) and loss behavior (in panel B) separately, but we cannot distinguish between investments and PPE. As noted above, 45 percent of all firm-year observations are associated with either a gain or loss, with the majority of these nonzero observations being gains. Again, the latter three columns measure the effect of taxes on the size of gains and losses conditional on the existence of either gains or losses. In panel A of Table 6, the estimated effects of tax rates are broadly consistent with expectations. When a firm has a high marginal tax rate, it is less likely to report a gain from the sale of assets. The estimates imply that a 10-percentage-point increase in the marginal tax rate decreases the propensity to realize a gain by between 0.397 to 0.757 percentage points,

Panel A: gain recognition behavior Marginal tax rate

−0.0709 (0.0146)


65,809 10,815 Y N

Panel B: loss recognition behavior Marginal tax rate

−0.0397 (0.0151) 0.0128 (0.0031) −0.0223 (0.0043) 62,994 10,658 Y Y

−0.1480 (0.1409)

17,268 5,757 Y N

Loss dummy −0.1120 (0.0131)


−0.0757 (0.0151) 0.0376 (0.0027) −0.0106 (0.0043) 62,994 10,658 Y N

Log gain value

65,809 10,815 Y N

−0.1041 (0.0135) 0.0073 (0.0024) −0.0215 (0.0038) 62,994 10,658 Y N

− 0.3283 (0.1416) 0.7021 (0.0289) −0.0416 (0.0533) 16,713 5,618 Y N

−0.2674 (0.1423) 0.5690 (0.0332) −0.0760 (0.0542) 16,713 5,618 Y Y

Log loss value −0.0687 (0.0136) −0.0075 (0.0028) −0.0243 (0.0039) 62,994 10,658 Y Y

−1.2982 (0.1718)

12,000 5,099 Y N

−1.3574 (0.1750) 0.4345 (0.0354) −0.1956 (0.0497) 11,581 4,980 Y N

−1.2516 (0.1761) 0.3272 (0.0409) −0.2003 (0.0505) 11,581 4,980 Y Y

Note: Panel A presents specifications that analyze gain recognition behavior, and Panel B presents specifications that analyze loss recognition behavior. In both panels, the first three columns employ a dummy variable set equal to 1 if there is a gain or loss as a dependent variable. In both panels, the second three columns employ the log value of that gain or loss amount as a dependent variable. All columns employ firm fixed effects; columns 3 and 6 of both panels also employ year fixed effects. “Marginal tax rates” are calculated via the methodology described in Graham (1996). “Log total assets” is the log value of total firm assets. “Log q ratio” is the log of the q ratio calculated from Compustat data as described in the paper.

Desai & Gentry

Gain dummy

32

TABLE 6 Determinants of Corporate Gain/Loss Realization Behavior, 1986–2002


33

and these estimated effects are statistically different from zero at the 95 percent confidence level. Furthermore, conditional on reporting a gain, the gains are smaller when the firm faces a higher marginal tax rate, and this effect is statistically different from zero in the specifications with controls for firm size and investment opportunities. In addition, when a firm is larger, it is more likely to report gains and, conditional on having a gain, the gain is larger. The behavior of reported losses, shown in panel B of Table 6, is somewhat puzzling. Contrary to the prediction that firms with high tax rates will value reporting losses, the estimated coefficients on the marginal tax rate are negative for both the extensive margin of reporting a loss and the intensive margin of the size of the loss (conditional on having a loss). Two issues complicate the analysis of loss behavior. First, as mentioned above, much of the variation in Graham’s (1996) estimates of marginal tax rates is driven by the presence of operating losses. However, the firm’s operating performance is not completely divorced from whether it has gains or losses on its existing investments. For example, poor operating performance may lead to both a low tax rate and a large stock of potential losses that the firm can recognize, which would be consistent with the estimated coefficients.32 Second, the netting rules for capital losses complicate the predicted relationship between tax rates and observed net losses. Suppose a firm has a high tax rate due to having substantial operating income. Because capital losses cannot be netted against positive operating income, it would not be surprising if we found no relationship between the observed tax rate and reporting a net loss.33

7. CONCLUSIONS Corporate capital gains realizations are an increasingly significant component of corporate cash flow. Net long-term capital gains are significant compared to individual capital gains, and they are increasingly important. As this paper outlines, the distortionary effects of such taxes largely subsume those associated with individual capital gains. Specifically, lockin effects at the corporate level may alter productivity levels by changing the patterns of corporate and asset ownership in a manner that taxes on individual capital gains do not. 32 This endogeneity between firm performance and the effective marginal tax rate would bias against finding that high tax rates are associated with the lower propensity to recognize gains that we report in panel A of Table 6. 33 To sort through these issues, it would be helpful to have separate data on gains and losses so that one can observe how firms match capital losses with capital gains; unfortunately, we have data only on the net gain or loss.

34

Desai & Gentry

The time-series analysis of this paper suggests that the elasticities of corporate realizations to tax costs is higher than those derived in similar equations used to estimate the elasticities of individual capital gains. Micro analysis also suggests that firms time their sales and magnitudes of investments and PPE opportunistically and that the realization of gains appears to be shaped particularly by tax incentives. In sum, the corporate capital gains tax regime appears to influence significantly the decisions of firms to dispose of assets and realize gains and losses. Our empirical evidence captures only one dimension—realization behavior—of the effect of corporate capital gains taxes. More generally, these taxes are likely to influence business planning on various margins, including merger activity, the initiation and termination of lines of business, and the patterns of cross-holdings. In combination with the curious distinction between the treatment of intercorporate dividend payments and intercorporate capital gains, the results in this paper and these broader consequences suggest that tax policy for corporate capital gains may be ripe for reevaluation and that much more needs to be understood about how corporate capital gains taxes influence firm behavior.

REFERENCES American Council on Capital Formation (1998). “An International Comparison of Capital Gains Tax Rates.” Special Report. Washington DC: ACCF Center for Policy Research. August. Andrade, Gregor, Mark Mitchell, and Erik Stafford (2001). “New Evidence and Perspectives on Mergers.” Journal of Economic Perspectives, Spring:103–120. Auerbach, Alan (1988). “Capital Gains Taxation in the United States: Realizations, Revenue, and Rhetoric.” Brookings Papers on Economic Activity no. 2:595–631. Auerbach, Alan J., and David Reishus (1988). “The Impact of Taxation on Mergers and Acquisitions.” In Mergers and Acquisitions, Alan J. Auerbach (ed.). Chicago, IL: University of Chicago Press. Auten, Gerald E., and Charles T. Clotfelter (1982). “Permanent Versus Transitory Tax Effects and the Realization of Capital Gains.” Quarterly Journal of Economics 97(4):613–632. Ayers, Benjamin C., Craig E. Lefanowicz, and John R. Robinson (2003). “Shareholder Taxes in Acquisition Premiums: The Effect of Capital Gains Taxation.” Journal of Finance, 58(6):2783–2801. Baker, Malcolm, and Jeffrey Wurgler (2003). “Investor Sentiment and the CrossSection of Stock Returns.” Harvard University, Working Paper. Burman, Leonard E., and George A. Plesko (2002). “Individual and Corporate Capital Gains Are Highly Interrelated.” Tax Notes 553. October 28. Burman, Leonard E., and William C. Randolph (1994). “Measuring Permanent Responses to Capital-Gains Tax Changes in Panel Data.” The American Economic Review 84(4):794–809. Dammon, Robert M., Chester S. Spatt, and Harold H. Zhang (2001). “Optimal


35

Consumption and Investment with Capital Gains Taxes.” Review of Financial Studies 14(3):583–616. Dhaliwal, Dan S., Merle Erickson, and Robert Trezevant (1999). “A Test of the Theory of Tax Clienteles for Dividend Policies.” National Tax Journal 52(2):179–194. Edwards, Courtney, Mark H. Lang, Edward L. Maydew, and Douglas A. Shackelford (2003). “Germany’s Repeal of the Corporate Capital Gains Tax: The Equity Market Response.” University of North Carolina Working Paper. Eichner, Matthew, and Todd Sinai (2000). “Capital Gains Tax Realizations and Tax Rates: New Evidence from Time Series.” National Tax Journal 53(3, part 2):663–682. Erickson, Merle, and Shiing-wu Wang (2000). “The Effect of Transaction Structure on Price: Evidence from Subsidiary Sales.” Journal of Accounting and Economics 30:59–97. Franks, Julian R., Robert S. Harris, and Colin Mayer (1988). “Means of Payment in Takeovers: Results for the United Kingdom and the United States.” In Corporate Takeovers, Alan J. Auerbach (ed.). Chicago, IL: University of Chicago Press. Gentry, William M., and David M. Schizer (2003). “Frictions and Tax-Motivated Hedging: An Empirical Exploration of Publicly-Traded Exchangeable Securities.” National Tax Journal 56(1):167–195. Gompers, Paul, and Josh Lerner (2002). “The Determinants of Corporate Venture Capital Success: Organizational Structure, Incentives and Complementarities.” In Concentrated Corporate Ownership, R. Morck (ed.). Chicago, IL: University of Chicago Press. Gompers, Paul, Josh Lerner, and David Scharfstein (2003). “Entreprenurial Spawning: Public Corporations and the Genesis of New Ventures, 1986–1999.” NBER Working Paper no. 9816. July. Gordon, Roger H., James R. Hines, Jr., and Lawrence H. Summers (1987). “Notes on the Tax Treatment of Structures.” In The Effects of Taxation on Capital Accumulation, Martin Feldstein (ed.). Chicago, IL: University of Chicago Press. Graham, John R. (1996). “Proxies for the Corporate Marginal Tax Rate.” Journal of Financial Economics 42:187–221. Grammatikos, Theoharry (1989). “Dividend Stripping, Risk Exposure, and the Effect of the 1984 Tax Reform Act on the Ex-Dividend Day Behavior.” Journal of Business 62(2):157–173. Klein, Peter (1999). “The Capital Gain Lock-In Effect and Equilibrium Returns.” Journal of Public Economics 71:355–378. Klein, Peter (2001). “The Capital Gain Lock-In Effect and Long-Horizon Return Reversal.” Journal of Financial Economics 59:33–62. Koski, Jennifer L., and John T. Scruggs (1998). “Who Trades Around the ExDividend Day? Evidence from NYSE Audit File Data.” Financial Management 27:58–72. La Porta, Rafael, Florencio Lopez-de-Silanes, and Andrei Shleifer (1999). “Corporate Ownership Around the World.” Journal of Finance 54:471–517. Maydew, Edward L., Katherine Schipper, and Linda Vincent (1999). “The Impact of Taxes on the Choice of Divestiture Method.” Journal of Accounting and Economics 28:117–150. Morck, Randall (2003). “Why Some Double Taxation Might Make Sense: The Special Case of Inter-Corporate Dividends.” NBER Working Paper no. 9651. April.

36

Desai & Gentry

Naranjo, Andy, M. Nimalendran, and Mike Ryngaert (2000). “Time Variation of Ex-Dividend Day Stock Returns and Corporate Dividend Capture: A Reexamination.” Journal of Finance 55(5):2357–2372. Paul, Deborah L. (2003). “Triple Taxation.” The Tax Lawyer 56:571–610. Plesko, George A. (2002). “Omitted Variable Bias in Time Series Estimates of Capital Gains Realizations.” MIT Working Paper. November. Poterba, James M. (2002). “Taxation, Risk-Taking, and Household Portfolio Behavior.” In Handbook of Public Economics, vol. 3, A. Auerbach and M. Feldstein (eds.). Amsterdam: North Holland. Scholes, Myron S., Mark A. Wolfson, Merle Erickson, Edward L. Maydew, and Terry Shevlin (2002). Taxes and Business Strategy: A Planning Approach. Upper Saddle River, NJ: Prentice Hall. Stiglitz, J. E. (1983). “Some Aspects of the Taxation of Capital Gains.” Journal of Public Economics 21(2):257–294. Wolfenzon, Daniel (1999). “A Theory of Pyramidal Ownership.” Harvard University Working Paper.

HOW FAST DO PERSONAL COMPUTERS DEPRECIATE? CONCEPTS AND NEW ESTIMATES Mark E. Doms Federal Reserve Bank of San Francisco

Wendy E. Dunn Federal Reserve Board

Stephen D. Oliner Federal Reserve Board

Daniel E. Sichel Federal Reserve Board

EXECUTIVE SUMMARY This paper provides new estimates of depreciation rates for personal computers (PCs) using an extensive database on prices of used PCs. Our results show that PCs lose roughly half their remaining value, on average, with each additional year of use. We decompose that decline into agerelated depreciation and a revaluation effect driven by the steep ongoing drop in the constant-quality prices of newly introduced PCs. Our results are directly applicable for measuring the depreciation of PCs in the We wish to thank Ana Aizcorbe, Darrel Cohen, Barbara Fraumeni, Jane Gravelle, Michael Kiley, Brent Moulton, Jim Poterba, and Brian Sliker for helpful comments on an earlier draft of the paper. We also thank Chuck Hulten, Dale Jorgenson, and Nick Oulton for useful discussions concerning the measurement of capital. Michael Hurlbut, Robert Little, and Brian Rowe provided excellent research assistance. The views expressed in the paper are those of the authors and should not be attributed to the Federal Reserve Board, the Federal Reserve Bank of San Francisco, or other staff members at either institution.

38

Doms, Dunn, Oliner, & Sichel

National Income and Product Accounts (NIPAs) and were incorporated into the December 2003 comprehensive NIPA revision. Regarding tax policy, our estimates suggest that the current tax depreciation schedule for PCs closely tracks their actual loss of value in a zero-inflation environment. However, because the tax code is not indexed for inflation, the tax allowances would be too small in present value for inflation rates above the very low level now prevailing.

1. INTRODUCTION The measurement of depreciation is a complex—and, to some, obscure— area of economics. Getting the numbers right is critical, however, for important issues in tax policy and capital measurement. This is particularly true for computers and other high-tech capital goods, which have assumed an increasingly central role in U.S. business activity over the past decade. In this regard, the recent debate about the contribution of information technology to economic growth has focused attention on the measurement of high-tech capital goods and consequently on the rate at which they depreciate. Over the years, economists have devoted considerable effort to the measurement of depreciation. Notable early studies include Griliches (1960), who estimated depreciation rates for farm tractors, and Hall’s (1971) work on pickup trucks.1 Somewhat later, Hulten and Wykoff (1981a, 1981b) estimated depreciation rates for many different types of equipment and structures, and the Bureau of Economic Analysis (BEA) has adopted their figures for use in the U.S. National Income and Product Accounts (NIPAs). For high-tech assets, however, the literature on depreciation is remarkably sparse given their importance in the economy. Hulten and Wykoff’s pioneering research more than two decades ago predated the explosion in demand for information technology capital; thus, their studies did not include computing equipment, and they treated quality change in a relatively limited way. Oliner (1993, 1994) estimated depreciation rates for mainframe computers and computer peripheral equipment, but these results are somewhat dated at this point, and there has been no follow-up research for these assets. To our knowledge, only two prior studies— Geske, Ramey, and Shapiro (2003) and Wykoff (2003)—have estimated depreciation for personal computers. Wykoff’s paper mainly concerns methodology, and his empirical work uses a very small sample of computer 1 See Jorgenson (1974) for an excellent overview of this early literature, along with Fraumeni (1997) and Jorgenson (1996) for more recent surveys of the literature.

How Fast Do Personal Computers Depreciate?

39

prices merely to illustrate his approach. Geske, Ramey, and Shapiro have a richer data set, which they employ to estimate depreciation for PCs and to highlight the role of obsolescence in driving depreciation for these assets. Our paper builds on the work of Geske, Ramey, and Shapiro to narrow further the knowledge gap surrounding depreciation rates for PCs. Like their work and most of the earlier literature, we rely on prices of used assets to estimate depreciation. We construct a large data set of prices and model characteristics for used PCs listed in bluebooks. This data set includes nearly 13,000 observations and spans the period from 1985 to 2002, covering almost the entire era of personal computers. With these data in hand, we followed the empirical approach in Oliner (1993, 1994), which regressed used asset prices on product characteristics and functions of age and time. This approach—which relates closely to Hall’s (1971) framework—effectively adds age to a standard hedonic price regression. This framework allows us to decompose the total decline in a PC’s price into two parts. The first is the revaluation of existing units over time as new models are introduced at lower constant-quality prices. The second part is the decline in a PC’s value as it ages because of reduced efficiency and eventual scrapping. Although PCs suffer some wear and tear, we believe that this aging effect arises mainly because older PCs become unable to run the latest software or lack features (like a CD-ROM drive) that become standard. As discussed in the next section, these two components of price change— depreciation and revaluation—have direct applications to tax policy and capital measurement. Not surprisingly, our empirical results indicate that PCs lose value at a rapid pace. Over our full sample period, the value of a PC declines roughly 50 percent, on average, with each year of use, implying that a newly installed PC can be expected to be nearly worthless after five or six years of service. In addition, our results suggest that both depreciation and revaluation contribute to the sharp drop in the value of installed PCs, although revaluation plays the dominant role, especially in the early years of a PC’s life. Our results have important implications for tax policy and for the measurement of depreciation and capital stocks in the NIPAs. Regarding tax policy, we show that in a world with changing relative prices, firms should be permitted to deduct both the age-related decline in an asset’s value (i.e., depreciation) and the revaluation of that asset relative to the general price level. The combined effect of depreciation and real revaluation measures the asset’s real loss of value. Allowing firms to deduct this loss of value equalizes the effective tax rate across assets, a

40


standard prescription for capital taxation. We then evaluate the current tax rules for depreciating PCs against this benchmark. The Internal Revenue Code allows firms to depreciate PCs and other computing equipment over a five-year service life, with annual deductions that equal or exceed 40 percent of the undepreciated value. Our estimates imply that these tax allowances closely approximate the real loss of value that PCs experience when inflation is very low; however, higher rates of inflation induce some distortion because the tax code is not indexed for inflation. Turning to capital measurement, our preferred specification generates a nongeometric schedule of depreciation that averages about 22 percent annually over the first five years of a PC’s service life. However, this schedule cannot be applied directly to the NIPAs because the constantquality PC prices generated by our data set trend down more rapidly than the corresponding NIPA price series, which relies on the producer price index for PCs. As we will show, the combined effect of depreciation and constant-quality price change is tightly pinned down in our data set, which means that altering the rate of constant-quality price decline implies an opposite change in the measured rate of depreciation. To generate a depreciation schedule suitable for use in the NIPAs, we estimate a version of our regression that constrains the path for constant-quality prices to conform with the NIPA series. When we do so, the estimated depreciation rate rises to an average annual pace slightly above 34 percent. This paper is organized as follows. The next section provides a conceptual discussion of depreciation. We highlight the concepts appropriate for tax policy and for capital accounting and then link these concepts to our empirical framework. Section 3 describes our data, and section 4 presents the empirical results, including a comparison to those in Geske, Ramey, and Shapiro (2003). Section 5 draws out the implications of our results for tax policy and for capital measurement. Conclusions are presented in Section 6, which is followed by an appendix that proves two propositions cited in section 2 of the text.

2. MEASUREMENT OF DEPRECIATION, USER COST, AND CAPITAL STOCKS The literature on measuring capital is vast, complex, and often confusing. This section summarizes—with a minimum of technical detail—the concepts that guide our empirical work. We discuss the measures of depreciation that are relevant for tax policy, for constructing the user cost of capital, and for calculating capital stocks and capital consumption


41

allowances in the NIPAs. For broader overviews of the literature on capital measurement, see Diewert (2002) and OECD (2001).

2.1 Background As the starting point for our discussion, let p zk, t, a denote the price of a capital good of type k that has the set of embodied characteristics z. For a personal computer, z specifies the speed of the processor, the size of the hard drive, the amount of memory, and so on. The other subscripts for p indicate that the price is observed at time t for a unit that is a years old. One year later, when time has moved forward to t + 1 and the unit is a + 1 years old, the price becomes p zk, t + 1, a + 1 , and the percentage decline in the asset’s price over the year can be written as: 1-

p zk, t + 1, a + 1 p zk, t, a

J pk p zk, t + 1, a NO z, t + 1, a + 1 =1-K k # K p z, t + 1, a p zk, t, a O P L

(1)

The first term in parentheses compares the price of an (a + 1)-year-old unit with an a-year-old unit at a fixed point in time. We denote this price ratio by 1 − δ k, where δ k is the depreciation in asset value from an additional year of age. The second term compares the price of an a-year-old unit at times t and t + 1. We denote this second price ratio by 1 + π k, where π k represents the percentage change in asset value between t and t + 1, with age held fixed.2 If we substitute 1 − δ k and 1 + π k into equation (1), we obtain: 1-

p zk, t + 1, a + 1 p zk, t, a

= 1 - (1 - d k )( 1 + r k ) . d k - r k

(2)

where the final expression omits the cross-product, δ kπ k, which is small compared to δ k and π k. It is important to note that the two components of price change in equation (2), δ k and π k, are measured conditional on a fixed set of performance characteristics, z. This is evident from the notation on the left side of equation (2), which shows that z does not vary as time and age each move forward by one year. Thus, the time-related element of the price change, π k, represents the change in price when holding quality fixed, which makes π k a constant-quality price measure. For personal computers, constant-quality prices have trended down at a rapid rate, mostly because of advances in semiconductor chip technology. Under standard assumptions 2 By writing δ k and π k without subscripts, we have implicitly assumed that both dimensions of price change are constant. We have done this only to simplify the notation; the entire discussion would remain valid if we allowed δ k and π k to vary with age and time.

42


of competitive equilibrium, each introduction of a more powerful model at a lower constant-quality price forces down the prices of older models. The term π k captures this ongoing revaluation. The age-related element in the equation, δ k, measures the additional decline in asset value that stems from wear and tear, reductions in the efficiency of older models, and the approach of the asset’s retirement. To avoid double-counting, δ k cannot include the revaluation effect. To illustrate these points, consider the following example involving PCs. Assume that the price of a new PC is $1,000 every year and that the value of a PC drops 45 percent over the course of a year (the value is $550 after one year, $302.50 after two years, and so on). The combined effect of depreciation and revaluation, δ k − π k, is 45 percent in this example. Now, to allocate this loss of value between δ k and π k, assume that the quality of new PCs increases 25 percent each year. Given the fixed $1,000 price of a new PC, the 25 percent annual increase in quality implies that new PC prices drop 25 percent per year in constant-quality terms; in a competitive equilibrium, the prices of existing models are pushed down by the same amount. Thus, π k (which captures this revaluation effect) would equal negative 25 percent, while δ k (which represents the annual loss of value over and above the revaluation effect) would be 20 percent. For personal computers, δ k largely reflects the influence of obsolescence rather than physical decay. Even units that continue to function like new lose value as they age because they become too slow to perform some tasks efficiently, can no longer run the latest software, or lack features that come to be considered essential. The literature on capital measurement has used different terms to describe δ k and π k, as shown in Table 1. Consistent with our discussion, Fraumeni (1997) labeled δ k as depreciation, π k as the revaluation term, and δ k − π k as the combined effect of depreciation and revaluation. Several other terms can be found in the literature for each of these concepts. One can make a case for each of these differing sets of terms, but we will use depreciation and revaluation, which strike us as the most intuitive among the alternatives.

2.2 User Cost of Capital Depreciation and revaluation are important elements of the user cost of capital, which measures the implicit cost of using a capital good for a given period of time. The user cost is the equivalent of the wage rate for capital and, as such, it plays a key role in a wide range of economic analyses, including studies of business investment, tax policy, and productivity growth.3 Using a capital good for a given period generates two types of 3

See Hall and Jorgenson (1967) for the classic discussion of the user cost of capital.


43

TABLE 1 Alternative Terms in the Literature Terms for δk

Authors

Terms for δk − πk

Fraumeni (1997)*

Depreciation

Depreciation and revaluation

Oliner (1993, 1994)

Partial depreciation

Full depreciation

Wykoff (2003)

Economic depreciation

Economic depreciation and revaluation

Diewert (2002); Hill (1999, 2000)

Cross-section depreciation

Time-series depreciation

Hulten and Wykoff (1981a, 1981b)

Economic depreciation

Economic depreciation and asset inflation

*We use Fraumeni’s (1997) terms in this paper.

costs. The first is the cost of financing the acquisition of the capital good. Assuming that the purchase is debt financed (a similar analysis holds for equity financing), this cost equals the prevailing interest rate (i) multiplied by the purchase price of the capital good. The second cost is the change in the value of the capital good over the period of use, which—as discussed above—equals the combined effect of depreciation and revaluation [pk(δ k − πk)]. The user cost of capital also reflects several features of the tax code. As described by Hall and Jorgenson (1967), these tax parameters include the statutory tax rate on corporate profits (which we denote by τ), the present value of the depreciation deductions for capital goods of type k (denoted by xk), and any investment tax credit for such capital goods (denoted by θk). Given these tax parameters, we show in the appendix that the user cost of capital can be written as: k k c zk, t, a = 1 - i - xx p zk, t, a (i + d k - r k ) 1-x k k = 1 - i - xx p zk, t, a 7 (i - r) + d k - (r k - r) A 1-x

(3)

where the final part of the equation adds and subtracts the aggregate rate of inflation, π, to express the interest rate and the revaluation effect in real terms. As would be expected, equation (3) indicates that a larger investment tax credit and more accelerated depreciation allowances act to reduce the cost of capital. Equation (3) will prove useful in our discussion of the appropriate tax allowances for depreciation, to which we now turn.

44


2.3 Tax Deductions for Depreciation The U.S. tax code allows businesses to deduct depreciation expenses when figuring their taxable income. These deductions influence the after-tax cost of investing in plant and equipment and thus affect both the overall size and composition of the business capital stock. When designing tax rules for depreciation, a standard prescription is to equalize the effective tax rate across assets to avoid distorting the composition of the capital stock. Numerous studies have shown that effective tax rates can be equalized by setting tax allowances for depreciation to match the actual decline in the asset’s value.4 In the appendix, we derive the implications of this rule when we allow for price inflation in the economy’s aggregate basket of goods and services (π) and asset revaluation relative to this aggregate price index (πk − π). As we show, the implied tax allowance is: DTAXzk, t, a = p zk, t, a 7 d k - (r k - r) A = p zk, t, a (d k - r k ) + p zk, t, a r

(4)

In words, the firm is allowed to deduct [δ k − (π k − π)] percent of the asset’s remaining value in each period. This deduction covers the actual loss of value in nominal terms [pk(δ k − π k)] plus the amount needed to maintain the asset’s real value in the face of aggregate price inflation (pkπ). The total deduction in equation (4) is exactly the sum of depreciation and (real) revaluation in the user cost of capital [see equation (3)]. If this depreciation policy were paired with a deduction for real interest expenses, the combination would grant a deduction for the full user cost of capital. Because the user cost represents the one-period charge that would prevail in a competitive rental market, allowing firms that own (rather than rent) capital to deduct the equivalent of the full user cost means that tax policy would be neutral with respect to the choice of renting versus purchasing capital.5 An example may help clarify how the allowances specified by equation (4) work in practice. As in the previous example, assume that a new PC costs $1,000 and that its value declines 45 percent with each year of service. In addition, assume that the overall inflation rate is 5 percent annually. With these assumptions, δ k − π k is 45 percent and δ k − (πk − π) is 50 percent. As shown in the first line of Table 2, the PC’s initial value [column (1)] falls to $550 at the end of the first year [column (2)]. With an inflation rate of 5 percent, the firm needs $1,050 of PC capital at the end 4 See Gravelle (1982, 1994) for a derivation of this result, Bradford and Fullerton (1981) for an in-depth treatment of effective tax rates, and Auerbach (1982) for a discussion of the connection between effective tax rates and neutral business taxation. 5

This statement assumes that all firms face the same tax rate. If they don’t, a firm with a low tax rate would still have an incentive to lease capital from a firm with a higher rate because the higher-rate firm would realize greater value from the depreciation deductions.


45

TABLE 2 Example of Tax Deductions for a Personal Computer

Year of use 1 2 3 4 5

Beginningof-year value (1)

End-ofyear value (2)

End-of-year amount needed to maintain real value (3)

Tax deduction (4) = (3) − (2)

$1,000.00 550.00 302.50 166.38 91.51

$550.00 302.50 166.38 91.51 50.33

$1,050.00 577.50 317.63 174.69 96.08

$500.00 275.00 151.25 83.18 45.75

Notes: In this example, the personal computer loses 45 percent of its value each year, while the aggregate price level rises 5 percent annually. We assume that the firm disposes of the personal computer at the end of year 5. All amounts in dollars, rounded to the nearest cent.

of the first year [column (3)] to maintain the real value of its initial $1,000 stock in terms of other goods and services in the economy. The difference between $1,050 and $550 gives rise to the $500 tax deduction [column (4)]. This $500 deduction equals [δ k − (π k − π)] percent of the PC’s initial value of $1,000. Lines 2 through 5 of the table repeat this calculation for subsequent years, with the PC assumed to be scrapped at the end of year five.

2.4 Consumption of Fixed Capital and Capital Stocks in National Accounts The Bureau of Economic Analysis (BEA) publishes estimates of capital consumption for the U.S. economy in the NIPAs. In concept, the national accounts measure capital consumption as the outlay required to keep the capital stock intact. While quite intuitive, this notion has generated a surprising amount of controversy in the literature on capital measurement. At the risk of oversimplifying the debate, the key issue is whether the change in asset value that we have labeled “revaluation” should be included in the consumption of fixed capital. An important source of guidance on this issue is the 1993 System of National Accounts (SNA), a comprehensive set of macroeconomic accounts prepared jointly by the World Bank, the International Monetary Fund, the Organisation for Economic Co-operation and Development (OECD), the Commission of the European Communities, and the United Nations. Unfortunately, the standards set out in the 1993 SNA are contradictory. On the one hand, the SNA states that: The value of a fixed asset . . . is determined by the present value of the future rentals . . . that can be expected over its remaining service life. Consumption of

46


fixed capital is therefore measured by the decrease, between the beginning and the end of the current accounting period, in the present value of the remaining sequence of rentals. (section 6.182) This definition of capital consumption includes what we have called the revaluation effect, which is one of several factors that influence the value of an asset’s future rental income. However, the SNA also constructs a revaluation account—separate from the measurement of capital consumption allowances—that records the gain or loss in value that “accrue[s] purely as a result of holding assets over time without transforming them in any way” (section 12.67). The revaluation effect for PCs is clearly a holding loss under this definition. Given these conflicting definitions, the SNA does not settle whether revaluation should be part of the consumption of fixed capital in national accounts. Central to this controversy are different interpretations of what it means to hold capital intact for the purpose of measuring net income. One interpretation states that capital has been held intact if the physical quantity of capital has been maintained. By this definition, capital consumption consists only of what we have called depreciation, which represents the outlay needed to cover the loss of value associated with wear and tear, declines in efficiency, and asset retirements. A contrasting point of view is that capital has been held intact if the ability of the capital stock to produce future income has been maintained. In this case, capital consumption includes not only depreciation but also the revaluation of existing assets. This second view underlies the tax allowances for depreciation described above. These competing interpretations of what it means to hold capital intact date back at least to the debate among Hayek (1941), Pigou (1941), and Hicks (1942), and economists have yet to settle the issue.6 It will be important for researchers and statistical agencies to reach a consensus regarding the appropriate measurement of capital consumption in national accounts. This debate notwithstanding, the NIPAs currently exclude revaluation effects from the consumption of fixed capital, with these effects appearing in a separate revaluation account. Accordingly, if we let Wtk denote the total value of all type k capital goods in year t, the NIPA consumption of fixed capital for this asset is simply: CFCtk = Wtk d k

(5)

6 For more recent views on this question, see Christensen and Jorgenson (1995), who construct an integrated set of national accounts using the narrower notion of capital consumption, and Hill (1999, 2000), who argues for the broader definition that includes revaluation.


47

Because this expression excludes revaluation effects, the NIPA concept of capital consumption measures the outlay needed to maintain the physical quantity of capital, not the future income stream from this capital. The capital stock in equation (5) represents what is known in the literature as a wealth stock (which motivates our use of the symbol W). This wealth stock equals the sum of the current year’s investment in capital goods of type k plus the remaining value of the investment done in previous years. If we let Itk denote this investment in year t and assume that the asset has a service life of T years, the wealth stock in current dollars can be expressed as: T-1

Wtk =

冱I

i=0

k t - i 71

- (d k - r k ) A

i

(6)

In our previous example involving personal computers, we assumed that δ k − π k (the combined effect of depreciation and revaluation) is 45 percent annually and that PCs are scrapped after five years. With these assumptions, the wealth stock of PCs in year t would be: ` I tk + I tk - 1 0.55 + I tk - 2 0.55 2 + I tk - 3 0.55 3 + I tk - 4 0.55 4 j

These expressions for the consumption of fixed capital and the wealth stock are in current dollars. That is, the wealth stock in equation (6) represents the actual dollar value of personal computers (or any other type of asset) in year t. Similarly, the measure of capital consumption in equation (5) represents the actual dollar outlay on personal computers required in year t to keep the stock of PCs intact (according to the NIPA concept). These series can also be expressed in terms of constant dollars. Of particular importance, a constant-dollar measure of capital consumption is needed to convert the economy’s constant-dollar gross product into a figure for its constant-dollar net output. The capital consumption allowance for asset k in constant dollars is simply the current-dollar measure divided by the price deflator for that asset. Dividing both sides of equation (5) by this price deflator yields the expression for capital consumption in constant dollars: CFCtk, 1996$ = Wtk, 1996$ d k

(7)

where we assume that constant-dollar series are measured in 1996 dollars.7 7 This was the NIPA convention when we completed the paper in November 2003. However, the comprehensive NIPA revision one month later shifted the base year for

48


The constant-dollar wealth stock in equation (7) can be computed directly from constant-dollar investment flows.8 As we demonstrate in the appendix: T-1

Wtk, 1996$ =

冱I

t=0

k, 1996$ (1 t-i

- d k )i

(8)

Note that the investment weights here are (1 − δ k)i rather than [1 − (δ k − π k)]i, as in the expression for the current-dollar stock. The weights in equation (8) represent the value in year t of one constant dollar of past investment. As the algebra in the appendix shows, the price deflator used to construct constant-dollar investment in equation (8) embeds the revaluation effect (πk) for existing assets. Hence, one constant dollar of investment from, say, year t − 2 would be worth less than a constant dollar from year t solely because of aging effects (δ k); including π k in the weight double-counts the revaluation effect. This explains why the investment weights for the constant-dollar stock in equation (8) differ from those for the current-dollar stock in equation (6).9

2.5 Summary and Implications for Our Empirical Framework This section has covered a lot of ground. To review the key results, Table 3 summarizes the implications of our discussion for calculating the user cost of capital, for specifying tax allowances for depreciation, and for measuring capital consumption allowances and wealth stocks in the constant-dollar series from 1996 to 2000. This change in base year has no effect on the results in the paper, all of which could be re-expressed in year-2000 dollars. Note also that equation (7) shows the constant-dollar consumption of fixed capital (CFC) for a single type of capital good. BEA calculates an aggregate real CFC by Fisher chain-weighting the constant-dollar CFCs for individual assets. In this paper, we focus on measuring the CFC (and wealth stock) for a single asset type—personal computers—and abstract from the calculation of chainweighted aggregates. 8 One can also arrive at the constant-dollar wealth stock by deflating the current-dollar stock shown in equation (6). We focus on the direct calculation of the constant-dollar stock from constant-dollar investment spending because that is the procedure used by BEA. 9 A related point concerns the distinction between wealth stocks, which measure the value of existing assets, and so-called productive stocks, which measure the services provided by these assets in a given period. Productive stocks are the appropriate concept of capital to use when estimating production functions or when measuring the contribution of capital accumulation to the growth of output or productivity. The constant-dollar productive stock for a given asset, like the constant-dollar wealth stock, is calculated as a weighted sum of current and previous constant-dollar investment flows. The weights generally differ, however, because the weights for the wealth stock reflect the remaining value of each investment cohort, while those for the productive stock reflect its remaining efficiency. This paper deals with the measurement of wealth stocks, although the information we develop on retirement patterns for PCs is also relevant for measuring productive stocks.


49

TABLE 3 Summary of Measurement Results Concept

Notation

Measurement

User cost of capital*

C zk, t, a

p zk, t, a 7 (i - r) + d k - (r k - r) A

Tax allowances for actual loss of value

DTAXtk

p zk, t, a 7 d k - (r k - r) A

Wealth stock T - 1

Current dollars

Wtk

Constant dollars

Wtk, 1996$

冱I

i=0

k t - i 71

- (d k - r k ) A

i

T - 1

冱I

i=0

k, 1996$ 71 t-i

- dkA

i

NIPA consumption of fixed capital Current dollars

CFCtk

Wtk d k

Constant dollars

CFCtk, 1996$

Wtk, 1996$ d k

*We omit the tax parameters in the expression for the user cost to highlight the common elements across the rows of the table.

NIPAs. As shown in the table, the combined effect of depreciation and real revaluation, δ k − (πk − π), plays a central role in the analysis. It represents the percentage loss of value, in real terms, for an asset over the course of a year. This loss of value appears in the user cost of capital, and it constitutes the appropriate percentage deduction for tax purposes. We also need to measure (δ k − πk) and δ k to calculate certain items in Table 3. The term δ k − πk represents the rate of decline in the asset’s nominal value, which enters directly into the calculation of the current-dollar wealth stock and also determines the asset price (pk) that appears elsewhere in the table. The term δ k measures the pure effect of aging on asset value and is needed to compute the constant-dollar wealth stock from data on constant-dollar investment outlays and to calculate the consumption of fixed capital in the NIPAs. Our empirical strategy is to estimate δ k and δ k − (πk − π) directly from data on used PC prices and then to calculate δ k − πk by subtracting the rate of aggregate price inflation, π, from the estimate of δ k − (πk − π). To see how we estimate these parameters, consider a simple log-linear expression for the price of a PC in year t scaled by the aggregate price index in that year: J pk N ln KK zp, t, a OO = a + b ln z + c t + z a (9) t P L For purposes of illustration, we have written this equation with only a single performance characteristic (z), and we have assumed that the

50


effects of age and time on price are constant. (We do not impose these constraints in our actual estimation.) Because this regression equation controls for the PC’s quality, the coefficient on time, γ, measures the rate at which constant-quality PC prices fall relative to the aggregate price level. This is exactly (πk − π) in our notation. Similarly, the coefficient on age, φ, measures the pure effect of aging on PC prices, which is minus δ k in our notation (the minus sign appears because we have expressed the rate of depreciation, δ k, as a positive number). Thus, −φ provides the estimate of δ k, and −(φ + γ) provides the estimate of δ k − (πk − π).

3. DATA The data used in this paper were obtained from Orion Research, which publishes bluebooks for many different used goods, including computers. The computer bluebooks contain prices for various types of used computer equipment and peripherals, based on surveys of used-equipment dealers. The bluebook entry for each PC includes information on the manufacturer, the model name and number, and the year or years in which the model was sold new by the manufacturer. Each bluebook entry also includes detailed information about the characteristics of the PC, including the processor type (Pentium III, for example), the processor speed, the amount of random access memory, and the size of the hard disk.10 Orion’s survey form asks dealers to show the amount paid to the seller of the PC (i.e., the wholesale price) as well as the dealer’s retail selling price and the number of days it took the dealer to sell the PC. Using this information, Orion constructs three used prices for each bluebook entry: the wholesale prices for units in mint and average condition, and the current used price, which measures the average retail price for units sold in 30 days or less. We focus on the current used price for our empirical work, but our results would be nearly the same if we used either wholesale price instead. We collected data on prices and characteristics of desktop PCs from the Orion bluebooks published from 1985 through 2003. Four major computer makers—Compaq, IBM, Dell, and Packard Bell—were included in the sample. From 1995 to 2003, we use the winter edition of the bluebook, published in January of each year. According to Orion, the prices in the winter edition are based on survey data collected in the fourth quarter of the previous year. In 1993 and 1994, only the fall edition was available; 10 In many cases, several additional characteristics were also listed, including whether the PC has a DVD player, a fax/modem, a video card, a sound card, or a network card. However, the reporting of these features appears to be less consistent across models and across years.


51

these prices reflect data collected in the third quarter of each year. Prior to 1993, the bluebooks were published annually, usually at about midyear, so we assume that the prices in these books were observed in the first half of the year. The concept of age in our data set warrants some discussion. The bluebook description of a PC for sale never includes its age. This omission reflects the fact that the value of a PC depends on its characteristics and its general condition; given this information, the date that the manufacturer shipped it from the factory is unimportant. However, a second concept of age—which we use in our empirical work—is relevant for pricing. We define model age as the amount of time that has elapsed since the first shipment of a given model. For example, the Dell Dimension 8100, with a Pentium IV processor, was first sold in 2001; an earlier model, the Dimension V350, with a Pentium II processor, was first sold in 1998. When measured in terms of model age, the V350 units are three years older than the 8100 units. The older model would be expected to sell at a lower price both because it is a less powerful computer and because it likely has fewer remaining years of use before obsolescence. Because PCs depreciate quickly, it is important to be as precise as possible about the timing of the observed prices. Table 4 lists each edition of TABLE 4 Timing of Price Observations in Each Bluebook Edition Bluebook edition 1985 1986 1987 1988 1989 1990 1991 1992 Fall 1993 Fall 1994 Winter 1995 Winter 1996 Winter 1997 Winter 1998 Winter 1999 Winter 2000 Winter 2001 Winter 2002 Winter 2003

Survey period January–June 1985 January–June 1986 January–June 1987 January–June 1988 January–June 1989 January–June 1990 January–June 1991 January–June 1992 July–September 1993 July–September 1994 October–December 1994 October–December 1995 October–December 1996 October–December 1997 October–December 1998 October–December 1999 October–December 2000 October–December 2001 October–December 2002

52


the bluebooks included in our sample, along with the time period in which, to the best of our knowledge, the prices were observed. We construct the time and age variables for our empirical work at the monthly frequency using the midpoint of the date range for the survey period. For example, we assign the price observations from the 2001 bluebook to November of the prior year. To calculate model age, we assume that a given model was first shipped in June of the year it was introduced. Model age is then defined as the number of months between the (assumed) first-shipment date and the survey date. For instance, the 2001 bluebook price for a PC listed as first sold in 1998 is associated with a model age of 29 months (June 1998 to November 2000). Figure 1 illustrates the resulting distribution of observations by model age. Our data set contains a large number of observations in each age group, as the figure shows. In previous work, Dulberger (1989) and Oliner (1993) found that, even after controlling for performance characteristics, the prices of semiconductors and mainframe computers varied significantly depending on whether they were near the frontier of the technologies available at the time a price was observed. This finding was taken as evidence of disequilibrium in these markets. We allow for the possibility of a similar pattern in the market for PCs. Accordingly, we construct a dummy variable (denoted FAST) that distinguishes models with best-practice technology from all other models. FAST equals 1 if the PC’s processor speed is in the

FIGURE 1. Number of Price Observations by Model Age in Years


53

TABLE 5 Characteristics of Used PCs by Year Sold*

Year sold 1985 1986 1987 1988 1989 1990 1991 1992 1993 1994 1995 1996 1997 1998 1999 2000 2001 2002

Used Number of price observations (dollars) 4 6 18 36 77 175 250 312 478 1,368 970 1,191 1,319 1,445 1,548 1,818 970 911

$2,439.0 2,661.2 2,549.7 3,084.0 2,154.9 1,980.4 2,094.0 1,213.5 642.1 567.9 473.2 471.0 346.0 299.3 242.0 335.7 289.5 272.9

Model age (months) 12.0 25.0 29.0 32.3 38.0 28.0 34.3 39.2 41.8 42.4 44.5 43.8 49.0 43.3 43.2 45.9 33.9 40.5

Central Amount of processing random Size of unit (CPU) access hard disk speed memory (Mb) (MHz) (RAM) (Mb) 12.5 15.0 31.1 33.4 45.9 89.8 106.7 110.3 167.7 316.7 366.8 598.4 920.9 2,063.8 3,911.6 6,411.5 11,139.1 14,046.2

6.1 6.9 8.9 9.0 10.1 15.7 17.2 18.6 25.0 31.1 42.1 60.4 76.9 131.1 197.5 276.4 476.6 579.6

0.6 0.5 0.7 0.8 0.8 1.3 1.6 1.8 2.8 4.7 5.8 8.3 11.0 21.9 36.4 50.2 80.7 96.3

*Mean values shown for price and characteristics.

highest 10th percentile of chip speeds available at the time of the price observation. For the regressions that follow, we exclude observations that are missing data on price or the major performance characteristics. In addition, we exclude used PCs with prices greater than $10,000 because they are likely to be servers rather than personal computers, the desired focus of our study. The final sample contains 12,896 observations. Table 5 lists the mean values of various characteristics by the year in which the used PC was sold. The table shows that large advances in chip speeds and memory have been accompanied by rapid declines in the price of used PCs.

4. ESTIMATION AND RESULTS This section presents the results of the hedonic regression that we use to estimate various aspects of price change for personal computers. The

54


dependent variable is the log of the used PC price deflated by the chainweight price index for gross domestic product (GDP), while the explanatory variables include product characteristics and functions of both time and the PC’s model age. We allow the effects of time to vary in the usual way by including separate dummy variables for all but one of the 19 periods in which prices were observed. However, the dummy variable approach does not work as well for model age because of the large number of different ages in our sample. To simplify the regression but still allow for a wide range of age-related price movements, we employ a fourth-order polynomial function of model age. Thus, the general form of our regression equation is: J pk N z, t, a ln KK GDP OO = a + pt L P

18

冱 b ln z + 冱 c i

i

i

n=1

4

n tn

+

冱z

j=1

j

aj

(10)

The vector z of product characteristics includes the log of the CPU speed (denoted by MHZ), the log of the amount of random access memory (RAM), the log of hard disk size (HD), a dummy variable for whether the price for the PC includes a monitor (MON), a dummy variable for whether the PC has a CD-ROM or DVD (CDROM), and the FAST chip dummy variable. We also include brand dummies for Compaq, IBM, and Packard Bell (with Dell as the excluded dummy) to control for differences among the four brands in the sample. With these controls for quality, the equation becomes: J pk N z, t, a ln KK GDP OO = a + b 1 ln MHZ + b 2 ln RAM + b 3 ln HD + b 4 ln MON pt P L + b 5 ln CDROM + b 6 FAST + b 7 COMPAQ + b 8 IBM 18

+ b 9 PBELL +

冱c

n=1

(11)

4

n tn

+

冱z

j=1

j

aj

4.1 Regression Results The columns labeled “Baseline” in Table 6 presents the results from OLS estimation of equation (11). As shown in the table, the regression fits the data quite well, with an adjusted R-squared of 0.92. The coefficients on all of the performance characteristics except the FAST dummy variable are strongly significant and have the expected positive signs. Among these characteristics, differences in processor speed have the largest price effects, consistent with previous findings by Oliner (1993), Dulberger (1989), and Cartwright (1986). The brand effects are also significant,


55

TABLE 6 Regression Results* Variable

Imposed Baseline PC price

Variable

Imposed Baseline PC price

2.045 (.174) .463 (.013) .393 (.009) .174 (.007) .182 (.009) .093 (.012) .014 (.018) −.027 (.010) .102 (.012) −.269 (.018)

5.234 (.045) −.039 (.013) .308 (.012) .147 (.009) .227 (.011) .105 (.015) .268 (.021) −.065 (.013) .040 (.014) −.437 (.022)

−.251 (.183) Mar. 1989 dummy −.578 (.176) Mar. 1990 dummy −1.225 (.173) Mar. 1991 dummy −1.494 (.172) Mar. 1992 dummy −2.097 (.172) Aug. 1993 dummy −3.073 (.171) Aug. 1994 dummy −3.278 (.171) Nov. 1994 dummy −3.422 (.171) Nov. 1995 dummy −3.846 (.171) Nov. 1996 dummy −4.173 (.172)

.495 (.208)

.425 (.254)

Nov. 1997 dummy

−4.721 (.172)

NA

(Model age/100)2

−5.362 (.650)

−9.836 (.800)

Nov. 1998 dummy

−5.987 (.173)

NA

(Model age/100)3

6.979 (.767) −2.753 (.298) −.126 (.268)

11.354 (.948) −3.918 (.370) NA

Nov. 1999 dummy

NA

Nov. 2000 dummy

−6.784 (.174) −6.777 (.175) −7.620 (.176)

NA

Nov. 2002 dummy

−7.774 (.177)

NA

Number of observations

12,896

12,896

Constant Ln MHZ Ln RAM Ln HD Monitor dummy CD-ROM dummy FAST chip dummy Compaq dummy IBM dummy Packard Bell dummy Model age/100

(Model age/100)4 Mar. 1985 dummy Mar. 1986 dummy Mar. 1987 dummy Adjusted R2

0 (omitted) −.324 (.195) .92

Mar. 1988 dummy

Nov. 2001 dummy

NA NA NA NA NA NA NA NA NA NA

NA NA

NA .84

*The dependent variable is the natural log of the price for the used personal computer divided by the gross domestic product chain-weight price index; standard errors in parentheses. NA indicates that this coefficient was not estimated in the regression.

56


revealing a large price discount for Packard Bell PCs relative to other brands with similar product characteristics. The coefficients on model age are all significant, especially those on the higher-order terms. As a result, we can strongly reject the hypothesis that the coefficients on a2, a3, and a4 are jointly zero, which means that the used PC prices do not decline at a constant rate with model age. That is, the estimated age-price profile is not geometric.11 In addition, the coefficients on the time dummies fall sharply over the sample period. As noted earlier, these coefficients represent the rate of decline in constant-quality PC prices relative to GDP prices. The coefficients on the time dummies for March 1985 and November 2002 imply that constant-quality PC prices dropped at an average annual rate of 35.1 percent in real terms between these dates.12 After accounting for the rise in GDP prices over this period, this figure implies that constant-quality PC prices fell in nominal terms at an average annual rate of 32.7 percent. 11 Age-price profiles estimated from used asset prices—as in this paper—can be affected by the lemons problem first identified by Akerlof (1970). Akerlof showed that prices on secondhand markets may embed a lemons discount when buyers cannot assess the quality of the goods offered for sale and thus presume that sellers are attempting to pass off inferior goods. In this case, the observed prices for units of a given age provide a downward biased estimate of the average price for all units of that age. Although we cannot rule out a lemons bias in our data, we doubt this bias is a serious problem. The condition of a used PC can be assessed rather easily, which limits the information asymmetry that lies behind the lemons issue. Even someone with minimal knowledge of computers can detect whether a PC has significant defects by visually inspecting the unit and checking that its key components operate properly. In addition, we performed a simple empirical test that failed to turn up evidence of a lemons problem. In particular, we reran the baseline regression using the wholesale prices for units in mint condition instead of the retail prices. These wholesale prices measure what the dealers—who tend to be sophisticated buyers—paid for the PCs that they assessed to be in excellent condition. If our baseline regression were affected by a lemons problem, we might expect the age-price profile based on wholesale mint prices to differ from that based on retail prices. However, the two profiles were nearly the same. 12

We measure this average annual rate of real price decline as: R V N1/17.67 GDP SJ pk W z, t = 1 1/02, a pt = 1 1/02 O K 100 # S K k - 1W GDP O S K p z, t = 03/85, a pt = 03/85 O W SL W P T X where 17.67 years elapse between the two pricing dates, and we use the values of the GDP chain-weight price index for the quarters containing these pricing dates. We then calculate the price ratio in parentheses as follows, where the second equality is based on equation (11): R J k J pk NV p zk, t = 1 1/02, a ptGDP S K p z, t = 1 1/02, a NO = 1 1/02 K z, t = 03/85, a OWW exp ln ln / S GDP K ptGDP OW p zk, t = 03/85, a ptGDP S KL pt = 1 1/02 OP = 3/85 = 03/85 L PX T = exp [c 1 1/02 - c 03/85 ] The terms γ 11/02 and γ03/85 are the coefficients on the time dummies for those dates. All other coefficients in equation (11) drop out from the calculation.


57

This pace is similar to estimates of quality-adjusted price change for PCs in Berndt and Rappaport (2003) and in Geske, Ramey, and Shapiro (2003).13 The rate of price decline that we estimate is more rapid than the drop in BEA’s constant-quality price index for personal computers, which fell at an average annual rate of 21.5 percent over 1985–2002. Diagnosing the source of this gap requires an analysis beyond the scope of this paper. A host of possible reasons for the gap—including differences in source data, hedonic techniques, and the construction of price indexes—should be explored in future research. For the time being, however, we focus on the implications of this difference for the measurement of depreciation. Recall that the total change in a capital good’s price over a given time period is the sum of depreciation and constant-quality price change (the revaluation effect). In the various specifications that we tested, we found that the estimate of the overall price change was nailed down tightly, while the individual components were less so; as a result, speeding up the decline in constant-quality prices had the effect of reducing the depreciation rate by roughly the same amount. Given this negative correlation between the components, the depreciation rate we estimated conditional on the constant-quality price change in our data set would be inappropriate for use in conjunction with the NIPA measure of constant-quality prices for PCs, which falls more slowly than our measure. To produce a depreciation estimate suitable for use in the NIPAs, we reestimate the baseline regression after constraining the path for constant-quality prices in our data to conform with the BEA series. We impose this constraint by replacing the time dummies in equation (11) with the natural log of (ptPC, BEA /ptGDP ), which is forced to have a coefficient equal to 1. The results of this regression are shown in the columns labeled “Imposed PC price” in Table 6. As can be seen from the drop in the adjusted R-squared, from 0.92 to 0.84, the overall fit of this regression is not as good as the baseline regression, which reflects the imposition of a trend rate of constant-quality price change that conflicts with the pattern in the data. This constraint affects the estimates of some other coefficients in the regression. In particular, the coefficient on ln(MHZ), which was strongly positive in the baseline regression, turns slightly negative in the constrained regression. The FAST chip dummy becomes positive and 13 However, Pakes (2003) found a notably slower pace of quality-adjusted price decline for PCs—roughly 15 to 20 percent per year on average. See Landefeld and Grimm (2000) for a comparison of results from earlier studies, and Berndt and Rappaport (2001) for additional background on the estimation of hedonic indexes for PCs.

58


highly significant. The coefficients on model age also change quite a bit, and we examine the effects of these changes on the estimated age-price profile for PCs in the next subsection.

4.2 Price Profiles Tables 7 and 8 present the estimated price profiles for PCs that we need to explore the implications of our results for tax policy and for capital measurement in the national accounts. Table 7 uses the coefficient estimates from the baseline regression; Table 8 employs those from the constrained regression. Both tables have the same structure, allowing an easy comparison of results. All of the price profiles in both tables have been normalized to equal 100 for new models. Column (1) in Table 7 shows the age-price profile implied by the coefficients on model age in the baseline regression.14 These coefficients capture the age-related decline in price across models at a given time, controlling for differences in performance characteristics. As can be seen, the age-price profile in column (1) is essentially flat for the first 12 months of model age before it declines steadily to about 56 percent of initial value at the 78-month mark. This age-price profile is based on prices for PCs that are still in use and does not account for the units that have been removed from service. Because these retired PCs presumably had a low implicit price relative to those that remained in use, the age-price profile in column (1) provides an upward biased estimate of the expected profile for an initial cohort of PCs. To correct this bias, we follow the procedure in Hulten and Wykoff (1981a, 1981b) and Oliner (1993). On the assumption that the salvage value of retired PCs is zero, we multiply the age-price profile in column (1) by the survival probabilities from an assumed retirement distribution. In particular, let g(a) be the age-price profile from our regression, and let S(a) be the survival function representing the probability that a PC will remain in 14

Each entry in this column equals:

J pk N K z, t, a = a * O 100 # K k K p z, t, a = 0 OO L P where a* varies from zero months to 78 months in six-month increments. We calculate this price ratio as follows, where the second equality is based on equation (11): R J k V N NW J pk S Kp pzk, t, a = a * z, t, a = a * O z, t, a = 0 O K S W / exp ln K - ln K ptGDP OW S K ptGDP OO pzk, t, a = 0 S L W P L P T X = exp 8 z 1 a* + z 2 (a* ) 2 + z 3 (a* ) 3 + z 4 (a* ) 4 - ( z 1 # 0 + z 2 # 0 2 + z3 # 0 3 + z 4 # 0 4 ) B = exp 8 z 1 a* + z 2 (a* ) 2 + z 3 (a* ) 3 + z 4 (a* ) 4 B

TABLE 7 Price Profiles and Survival Probabilities for Personal Computers: Baseline Regression

Model age (months)

Survival probability (2)

Survival-adjusted age-price profile (3) = (2)*(1)

Real revaluation (4)

No inflation (5) = (4)*(3)

1% inflation (6)

4% inflation (7)

100.0 101.2 99.4 95.4 90.2 84.5 78.9 73.6 68.9 65.0 61.7 59.2 57.2 55.8

100.0 99.2 96.9 93.2 88.2 82.2 75.4 68.1 60.6 53.0 45.7 38.8 32.4 26.6

100.0 100.4 96.3 88.9 79.6 69.5 59.5 50.1 41.7 34.5 28.2 23.0 18.5 14.9

100.0 80.5 64.9 52.2 42.1 33.9 27.3 22.0 17.7 14.3 11.5 9.2 7.4 6.0

100.0 80.9 62.5 46.4 33.5 23.5 16.2 11.0 7.4 4.9 3.2 2.1 1.4 0.9

100.0 81.5 63.4 47.5 34.5 24.5 17.0 11.6 7.9 5.3 3.5 2.3 1.5 1.0

100.0 83.3 66.3 50.8 37.7 27.3 19.4 13.6 9.4 6.4 4.4 2.9 2.0 1.3


0 6 12 18 24 30 36 42 48 54 60 66 72 78

Total effect on price

Age-price profile (1)

59

60

Model age (months) 0 6 12 18 24 30 36 42 48 54 60 66 72 78

Total effect on price

Age-price profile (1)

Survival probability (2)

Survival-adjusted age-price profile (3) = (2)*(1)

Real revaluation (4)

No inflation (5) = (4)*(3)

1% inflation (6)

4% inflation (7)

100.0 99.3 93.1 83.5 72.6 61.7 51.8 43.3 36.3 30.6 26.2 22.7 20.0 18.0

100.0 99.2 96.9 93.2 88.2 82.2 75.4 68.1 60.6 53.0 45.7 38.8 32.4 26.6

100.0 98.5 90.2 77.8 64.0 50.7 39.1 29.5 22.0 16.2 12.0 8.8 6.5 4.8

100.0 87.2 76.1 66.4 57.9 50.5 44.1 38.4 33.5 29.2 25.5 22.2 19.4 16.9

100.0 85.9 68.7 51.6 37.1 25.6 17.2 11.3 7.4 4.7 3.1 2.0 1.3 0.8

100.0 86.5 69.6 52.7 38.1 26.5 17.9 11.9 7.8 5.0 3.3 2.1 1.4 0.9

100.0 88.2 72.3 55.8 41.1 29.1 20.1 13.6 9.0 6.0 3.9 2.6 1.7 1.1


TABLE 8 Price Profiles and Survival Probabilities for Personal Computers: Imposed PC Prices


61

service at age a. Then the age-price profile corrected for retirements is S(a)g(a).15 Little is known about retirement patterns for PCs. The limited evidence suggests that the modal age at retirement is roughly four years.16 Because many PCs probably continue to be used in lower-value applications for a number of years, we assume that the mean age of retirement is five years, one year longer than the modal age. We then select an asymmetric retirement distribution that matches these parameter values. Although the Winfrey distributions have a long history in economics for portraying retirement patterns, we use the Weibull distribution instead because of its convenient parametric form.17 The survival probabilities generated by our retirement distribution are shown in column (2). As can be seen, this distribution implies that 88 percent of PCs remain in service after 24 months of use, after which the pace of retirements increases. The probability of retirement is highest over the range from 36 to 60 months of use and then slows, leaving a long righthand tail to the distribution. Column (3) shows our estimated age-price profile adjusted for retirements, calculated as the product of columns (1) and (2). Column (3) represents our estimate of depreciation (δk), taking account of the implicit zero price for retired units. The next column in Table 7 presents the real revaluation effect (πk − π), which equals the estimated rate of decline in constant-quality PC prices relative to the path for GDP prices. As noted above, this real decline in PC prices averaged 35.1 percent annually over our full sample period. Column (4) shows the cumulative effect of this real revaluation for successively older models. Columns (5) through (7) bring together the separate influences on PC prices. Column (5) presents the combined effect of depreciation and real revaluation, calculated as the product of the profiles in columns (3) and (4). The rate of decline shown in column (5) is the estimate of δ k − (πk − π), 15 An alternative way to adjust for retirements would be to multiply each price observation in our sample by the survival probability associated with that observation and then to run the regression with the adjusted data. We tried both methods and found that the results were nearly the same either way. 16 A recent story on the Bloomberg News Service (2003) cited an industry analyst who suggested that, in the recent past, large firms have been replacing PCs every four years. In addition, Richards (2002) estimated that the replacement cycle for PCs is 3.9 years, based on spectral analysis of computer investment flows. 17 Outside economics, the Weibull distribution has been used extensively to model survival patterns. See Johnson, Kotz, and Balakrishnan (1994). Among applications in economics, Sliker (2003) has used the Weibull to model retirement patterns for motor vehicles. The survival function implied by our Weibull retirement distribution is exp[−(age/β)η], where β = 67.8, η = 2, and age is measured in months.

62


which figures so prominently in our measurement system. If aggregate price inflation were zero, δ k − (π k − π) would equal δ k − π k, and this column would represent the actual drop in the value of a personal computer with each additional period of use—hence the label “No inflation” for this column. As can be seen, the value of a PC declines quickly after it enters service. Twelve months after installation, the PC’s value has fallen to 62.5 percent of its initial price, almost entirely reflecting revaluation. After 24 months, only 33.5 percent of the initial value remains, and after 60 months the PC is nearly worthless. Columns (6) and (7) show the analogous schedules when aggregate price inflation is 1 percent and 4 percent, respectively—the range observed over our sample period. As is evident from the similarity of the three columns, the strong downward pressure on PC prices from revaluation and age-related factors overwhelms the effect of aggregate price inflation. The price profiles in Table 7 are those implied by our unconstrained baseline regression. We will use these profiles in section 5 to assess the implications of our results for tax depreciation schedules for PCs. As we discussed above, however, the baseline regression implies that constantquality PC prices have declined considerably faster than is indicated by the BEA series. To obtain price profiles that mesh with BEA’s constantquality price index, Table 8 recalculates all the profiles using the results from the constrained regression. The age-price profile in column (1) of Table 8 declines more rapidly than its counterpart in Table 7. This difference is a direct result of imposing the BEA price index in the constrained regression. That is, the slower rate of constant-quality price decline in the constrained regression forces adjustments in other coefficients to fit the sharp drop in bluebook prices for a given PC over its service life. One such adjustment, shown in column (1), is a speed-up in the estimated rate of age-related price decline, which carries through to the survival-adjusted age-price profile in column (3) (after applying the unchanged survival function). This age-price profile differs substantially across the two tables. Indeed, over the first 60 months of the PC’s service life, the average annual rate of depreciation is 34.6 percent in Table 8, well above the 22.4 percent rate in Table 7. This difference offsets almost all of the gap in the estimated revaluation rate between the constrained and unconstrained regressions, as can be seen by comparing column (5) across the two tables. After 36 months of use, the PC’s remaining value (taking account of both depreciation and revaluation) is 16.2 percent in Table 7, very similar to the 17.2 percent figure in Table 8. The difference becomes even smaller with additional periods of use. Thus, the data enforce a strong negative correlation between the estimated rates of depreciation and constant-quality price


63

change, leaving their sum largely invariant to constraints imposed on either component.

4.3 Alternative Regressions Table 9 summarizes the main results from our empirical work. It also presents several tests of robustness and briefly compares our results to those in Geske, Ramey, and Shapiro (2003), abbreviated henceforth as GRS. Column (2) of the table shows the combined effect of depreciation and revaluation for various specifications of our regression, while columns (3) and (4) display these two components of price change. Column (5) shows the cross-product term that arises when combining the effects of depreciation and revaluation.18 We present these price measures from: ● ● ●

●

●

The baseline regression (line 1); The constrained regression that imposes BEA’s PC price index (line 2); Alternative versions of the baseline regression that enlarge the set of performance characteristics (line 3), that allow the coefficients on the characteristics to vary over time (line 4), and that use wholesale PC prices rather than retail prices as the dependent variable (line 5); The baseline regression estimated over the 1990–2000 sample period used by GRS (line 6); Two sets of results from GRS (lines 7 and 8).

Perhaps the key point to take away from the table is shown in column (2): namely, the various regression specifications all imply that the value of a PC falls roughly 50 percent on average over the course of a year. The estimates of this annual price decline are tightly clustered in a range from 46.9 percent to 51.9 percent, despite substantial differences in the form of the regression, the presence or absence of constraints, and the estimation period. Thus, as noted above, the data yield a robust estimate of the combined effect of depreciation and revaluation on PC prices. However, the decomposition of this total price change between depreciation and revaluation is less certain. The baseline specification using the full sample (line 1) implies an annual (survival-adjusted) depreciation rate of 22.4 percent and a (nominal) revaluation rate of 32.7 percent. Imposing the BEA constant-quality price index for PCs essentially reverses the relative magnitudes of depreciation and revaluation. This shift 18 Specifically, column (2) of Table 9 displays the value of ω k from the equation (1 − ω k) = (1 − δ k)(1 + π k), with δ k shown in column (3), negative π k in column (4), and δ kπ k in column (5). The term ω k equals (δ k − π k) + δ kπ k [i.e., the sum of columns (3), (4), and (5)]. We measure δ k as the average annual decline in the survival-adjusted age-price profile over the initial 60 months of the PC’s service life. Also, we measure −π k as the average annual rate of constant-quality price decline for PCs over the time period shown in column (1).

64

Sample period (1)

Total price decline (2)

Depreciation (3)

Revaluation (4)

Crossproduct (5)

This paper 1. Baseline specification 2. Imposing BEA PC prices 3. Adding CPU dummies 4. Allowing time-varying effects of characteristics 5. Using wholesale average-condition prices 6. Baseline specification, shorter sample

1985–2002 1985–2002 1985–2002 1985–2002 1985–2002 1990–2000

47.8 48.7 46.9 49.2 47.5 51.9

22.4 34.6 18.7 22.4 24.2 21.6

32.7 21.5 34.7 34.6 30.7 38.7

−7.3 −7.4 −6.5 −7.8 −7.4 −8.4

Geske, Ramey, and Shapiro (2003) 7. Unconstrained 8. Imposing BEA computer prices

1990–2000 1990–2000

49.8 50.3

27.3 41.1

30.9 15.7

−8.4 −6.5

Notes: Revaluation is shown here as a rate of decline and thus without the negative sign that appears elsewhere in this paper. All figures at annual rates.


TABLE 9 Alternative Estimates of Price Declines for Personal Computers


65

highlights the fact that the estimate of depreciation depends on the assumed rate of constant-quality price change. Line 3 shows that the estimates of depreciation and revaluation also depend somewhat on the control variables included in the regression. To obtain the result reported on line 3, we augment the baseline set of performance characteristics with a set of dummy variables for the type of central processor chip in the PC. These dummies indicate whether the processor is a 286, 386, 486, Pentium I, Pentium II, Pentium III, or Pentium IV. The CPU dummies can be viewed as capturing some unmeasured dimensions of quality to the extent that processor speed and memory—the standard measures—do not fully determine the processor’s capabilities. We find that these CPU dummies are significant in the regression and, as shown on line 3, their presence tends to slow the rate of depreciation while increasing the rate of revaluation.19 We also examined whether the coefficients on the characteristics change over time and whether any such variation affects the estimates of depreciation and revaluation. This issue is particularly important in light of Pakes’s (2003) critique of standard hedonic procedures. Pakes argued that the coefficients in hedonic regressions may change over time in response to changes in market structure or preferences. For similar reasons, he also argued that caution is required in interpreting these coefficients and that they need not have the expected signs. To examine these issues, we allow the coefficient on each characteristic to differ across three subperiods: 1985–1995, 1996–1999, and 2000–2002. The coefficients on characteristics in this regression did vary somewhat over time but, as line 4 of the table shows, the implied depreciation rate is the same as in the baseline regression and the revaluation rate is only a bit faster. Lines 5 and 6 of Table 9 show the results of two other tests of the baseline regression. To explore robustness with respect to our price measures, we estimate the baseline regression using the wholesale price for PCs in average condition rather than the retail price. As can be seen on line 5, this change had very little effect on the estimated rates of depreciation and revaluation.20 Line 6 shows that using the GRS sample period 19 The interpretation of these CPU dummies is subject to some ambiguity. As we noted, they could be significant because they proxy for unmeasured elements of quality. However, the CPU dummies are also correlated with a PC’s model age. Indeed, GRS use a similar variable to account for the depreciation that they estimate in a regression similar to our baseline specification. If the CPU dummies function mainly as proxies for model age rather than as proxies for unmeasured quality, the baseline specification would provide a more accurate measure of age-related depreciation. 20 The gap between wholesale and retail prices represents the dealer’s margin, and in a competitive market, it measures the transaction cost of selling used PCs. One could be concerned that swings in dealer margins might influence our estimated price profiles; however, our nearly identical results using either wholesale or retail prices indicate that margins have not varied systematically over time or with the age of the PC.

66


(1990–2000) reduced the estimated depreciation rate only slightly relative to the baseline regression, while it increased the revaluation rate more substantially. The final lines of Table 9 present the results from GRS that most closely resemble our baseline and constrained regressions. The regression associated with line 7 allows the data to determine the rate of constant-quality price change, as in our baseline specification, while the regression associated with line 8 imposes BEA’s price index for computers and peripheral equipment.21 As we noted above, the total price decline shown on lines 7 and 8 closely approximates the pace that we estimate. The GRS depreciation rates are somewhat faster than ours, however, which highlights the sizable confidence band around the estimates from our study and theirs concerning this element of price change.

5. IMPLICATIONS This section explores the implications of our empirical results for tax policy, for capital accounting in the NIPAs, and for measuring the user cost of capital.

5.1 Tax Depreciation Allowances for Personal Computers Under current tax rules (the modified accelerated cost recovery system), PCs and other types of computing equipment are depreciated over a fiveyear period. The annual deductions are calculated using the doubledeclining-balance (DDB) method, with a switch to the straight-line method at the point that maximizes the present value of the deductions. The double-declining-balance method specifies an annual percentage 21 Lines 7 and 8 of our table reflect the results shown in GRS, Table 6, columns (8) and (3), respectively. Several points should be noted about the GRS regression results. First, the BEA price series imposed on their regression covers all computers and peripheral equipment, not just PCs. This broader price index has tended to fall somewhat less rapidly than the index for PCs alone, which accounts for the relatively small revaluation effect on line 8 of our table. Second, GRS’s results make no adjustment for retirements. We adjusted their age-price profiles with the survival function shown in our Tables 7 and 8, which places their depreciation estimates on the same conceptual footing as ours. Third, GRS allow for what they call instantaneous depreciation, defined as the loss of value that occurs when a buyer opens the box containing a new PC. They attempt to identify this effect from the new list prices shown in the Orion bluebooks. Their estimates imply a large instantaneous loss of value, ranging from about 20 to 25 percent of the new PC price in the regression specifications that most resemble ours. However, this apparent loss of value could arise, at least in part, from unmeasured price discounts. That is, if new PCs actually sell at a discount to list prices, the regression would overstate the price drop when a new PC leaves the store. Given this identification issue, we chose to exclude the new list prices from our data set, and we present the GRS depreciation rates excluding their estimate of the instantaneous price decline.


67

TABLE 10 Tax Depreciation Schedules for Personal Computers (Percentage of Initial Value) Current law (1)

Current law, adjusted for retirements (2)

π = 0% (3)

π = 1% (4)

π = 4% (5)

1 2 3 4 5 6 7

20.0 32.0 19.2 11.5 11.5 5.8 0.0

20.6 34.6 21.1 11.9 8.7 3.1 0.0

19.1 34.4 22.9 12.5 6.1 2.8 2.1

19.1 34.7 23.4 13.0 6.4 3.0 2.3

19.1 35.5 25.1 14.5 7.5 3.7 3.0

PDV, π = 0% PDV, π = 1% PDV, π = 4%

92.5 90.6 85.2

93.0 91.2 86.1

92.9

Year

Covering full loss of value

92.7 92.3

Note: The figures in the table are based on the assumption that the personal computer (PC) is installed in the middle of the year. Given this half-year convention, the entries for year 1 show the depreciation over the first six months of the PC’s life, the entries for year 2 show the depreciation between six and 18 months, and so on. PDV stands for present discounted value.

deduction that is twice the straight-line rate. For an asset with a five-year recovery period, the DDB deduction rate is 40 percent annually. Column (1) of Table 10 shows the stream of tax allowances for a PC under current law, with each year’s deduction expressed as a percentage of the asset’s initial value. Note that the first-year deduction—20 percent—is only half of the full-year amount, reflecting a half-year convention that assumes the asset was put in place at midyear. After this deduction, 80 percent of the PC’s initial value remains to be depreciated. Applying the 40 percent rate to this remaining value yields the 32 percent deduction for the second year. The third-year deduction is calculated in the same way. The schedule then switches to the straight-line pattern, with the undepreciated part of the PC’s initial value written off over the remaining 21⁄2 year recovery period. Given our assumed retirement distribution for personal computers, a substantial fraction of PCs would be retired before being fully depreciated under current tax rules. In such cases, the tax code allows a firm to deduct the full amount of the remaining allowances in the year of retirement.22 Column (2) of the table adjusts the statutory allowance in column (1) to 22 See CCH (2002, p. 337), “Abandonment and Obsolescence Losses.” The deduction would be reduced by the amount of any sale proceeds or insurance recovery. Implicitly, we have assumed that the asset is uninsured and has a salvage value of zero.

68


account for these early retirements. To make this adjustment, we use our estimated retirement distribution to divide a cohort of newly installed PCs into those that are retired in the first six months of service (to reflect the half-year convention), the next full year, the year after that, and so on. We then calculate the appropriate depreciation schedule for each subcohort. For example, the small fraction of PCs retired within the first six months of service would receive a 100 percent deduction in the first tax year; those retired between six months and 18 months would receive the usual 20 percent deduction in the first tax year and the remaining 80 percent deduction in the second tax year. We proceed in this fashion for successive annual slices of the retirement distribution, and then we aggregate the depreciation schedules for each slice using weights that equal the probability of retirement within that slice. A comparison of columns (1) and (2) of Table 10 shows that this adjustment results in a small acceleration of the statutory schedule of deductions. During the first two tax years, the adjusted allowances total 55.2 percent of the initial value of the PC cohort, up from 52 percent in the statutory schedule. This adjustment—while conceptually necessary—is fairly small because the early retirements in our distribution are concentrated in years four and five, after the bulk of the tax allowances have been taken. We now compare the retirement-adjusted schedule in column (2) to the allowances implied by our empirical results. As discussed above, the allowance in a given period equals the PC’s loss of value in real terms, which we calculate as the product of the PC’s value at the beginning of a period and the real percentage decline in value that it experiences over the period. Both terms in this product were shown in Tables 7 and 8. We use the figures in Table 7, which reflect the baseline (unconstrained) regression. Columns (5) through (7) in that table display the first term in the product—the PC’s remaining value as it ages— under different rates of general price inflation. For the purpose of this exercise, we measure the PC’s value at ages six months, 18 months, 30 months, and so forth, to be consistent with the half-year convention in the tax code. The second term in the product, the real percentage decline in a PC’s value during a given period [δ k − (π k − π)], is calculated from column (5) of Table 7. Moving down that column gives the periodby-period values for δ k − (π k − π). For example, the real decline in value over the initial six months of use is 19.1 percent [1 − (80.9/100)]. To conform to the half-year convention, we use the rate of decline between zero and six months, six and 18 months, 18 and 30 months, and the successive 12-month intervals.


69

Columns (3) through (5) of Table 10 show the resulting schedule for depreciation allowances under different rates of aggregate inflation. Focus first on column (3), the schedule of allowances when the aggregate inflation rate is zero. This schedule is remarkably similar to the deductions allowed under current law after adjusting for early retirements, column (2). The first-year deductions under both schedules are close to 20 percent of the PC’s initial value, and the second-year deductions are both a shade less than 35 percent. When we allow for general price inflation [columns (4) and (5)], the deductions become slightly larger than in column (3) because the nominal value of the PC—its tax basis for our calculations—declines less rapidly in the higher inflation environment. The bottom part of the table compares the present value of the deductions under the various schedules. To calculate these present values, we discount the annual deductions with a nominal after-tax interest rate of 31⁄2 percent in the case of no inflation, 4 1⁄2 percent when inflation is 1 percent, and 7 1⁄2 percent when inflation is 4 percent.23 With no inflation, the present value of the current-law deductions (adjusted for retirements) is $93.0 per $100 of initial asset value, almost identical to the $92.9 figure for deductions that cover the PC’s full loss of value. The gap widens considerably, however, when we introduce inflation. At 4 percent inflation, the present value of current-law deductions (again adjusted for retirements) is $86.1, a fair amount less than the $92.3 figure in column (5) because the higher inflation erodes the present value of the unindexed deductions under current law.24 Thus, the current-law deductions do an excellent job of approximating the full loss of value for personal computers under zero or very low inflation, but the lack of 23 Over our sample period, the real (pretax) interest rate on BAA-rated corporate bonds averaged about 51⁄2 percent (where we compute the real rate as the nominal rate minus the expected ten-year inflation rate from the Philadelphia Federal Reserve Bank’s survey of professional forecasters). Under Fisher’s law (modified to account for taxation), each percentage point of inflation adds 1/(1 − τ) percentage points to the nominal pretax interest rate, where τ represents the corporate tax rate, which we take to be 35 percent. The resulting nominal after-tax interest rate is

r ; 5 1⁄2 + 1 - x E * (1 - x) = 0.65 * 5 1⁄2 + r which equals approximately 31⁄2 percent when π = 0, 4 1⁄2 percent when π = 1, and 7 1⁄2 percent when π = 4. 24

Note that the present value differs slightly across columns (3)–(5), even though the PC’s remaining value is adjusted for inflation in each case. The difference arises because, for simplicity, we have ignored the cross-product in Fisher’s law between the real interest rate and the inflation rate as well as the cross-product between the real decline in PC prices and the aggregate inflation rate in the inflation-adjusted tax basis for depreciation.

70


indexation causes the tax deductions to fall short of this benchmark when inflation moves higher.

5.2 Capital Accounting 5.2.1 NIPA Wealth Stocks for Personal Computers As described in Bureau of Economic Analysis (1999), real NIPA wealth stocks are calculated by summing past real investment flows with weights generally based on the geometric depreciation rates estimated by Hulten and Wykoff (1981b). However, Hulten and Wykoff did their work prior to the widespread introduction of personal computers. Thus, BEA must look beyond Hulten and Wykoff’s results for estimates of depreciation for PCs. Prior to the December 2003 comprehensive revision of the NIPAs, BEA used a depreciation schedule for PCs based on Lane (1999). This schedule is nearly geometric and assumes that the value of a PC declines to 10 percent of its original value after five years. Importantly, this schedule incorporates the full loss in a PC’s value as it ages and thus captures both depreciation and revaluation. As we discussed in section 2, BEA’s calculation of the real wealth stock should rely on weights that exclude revaluation.25 Based on a preliminary version of this paper, BEA decided to adopt a geometric depreciation rate of 34 percent for PCs for the comprehensive NIPA revision in December 2003. This figure is close to the average depreciation rate in column (3) of Table 8, which is calculated from the regression in which we impose BEA’s price index for PCs.26 5.2.2 NIPA Consumption of Fixed Capital As indicated in equations (5) and (7), BEA’s estimate of the consumption of fixed capital (CFC) for an asset can be calculated as the product of the wealth stock and the depreciation rate for that asset. Our estimate of the depreciation rate for PCs (conditional on BEA’s constant-quality price index) is lower than the 39 percent rate that the agency used prior to the December 2003 revision. By itself, the move to a lower rate would reduce BEA’s estimate of the CFC for personal computers. A rough calculation suggests, however, that this effect is approximately offset by the upward revision to the wealth stock that results from using a lower depreciation rate to construct the stock. 25 Cummins and Violante (2002) also discussed this difficulty with Lane’s depreciation schedule for use in the NIPAs. 26 The figures in Table 8 imply an average depreciation rate of 34.6 percent over the first five years of a PC’s life; the difference between 34 percent and 34.6 percent reflects assorted small changes to our data set and specification between the time we provided BEA with preliminary results and the completion of the paper. Although our results suggest that depreciation is not geometric, time constraints prevented BEA from considering nongeometric depreciation for this revision of the NIPAs.


71

Thus, we believe that BEA’s switch to a lower depreciation rate implies little change to its estimate of the CFC for personal computers.

5.3 User Cost of Capital The user cost of capital in equation (3) depends on an asset’s total loss of value in real terms, δ k − (π k − π). Our estimates of δ k − (π k − π) are always in the neighborhood of 50 percent annually, and we would argue that analysts calculating a user cost for PCs for growth accounting or investment analyses ought to use such a figure. For the purpose of constructing the user cost, uncertainty about the precise split between depreciation and real revaluation does not matter.

6. CONCLUSION This paper provides new estimates of depreciation rates for personal computers using an extensive database on used prices. The approach in the paper most closely follows that in Oliner (1993, 1994), and it is very much in the spirit of Hall (1971). Essentially, we regress prices of used PCs (adjusted for the overall GDP price deflator) on a set of performance characteristics, flexible functions of time and age, and other controls. After adjusting for retirements—to avoid the censoring bias from unobserved prices for retired PCs—the coefficients on the age variables provide estimates of age-related depreciation, while the coefficients on time provide a constant-quality price index for PCs. To map our results into the concepts needed for tax policy and capital measurement, we develop a conceptual framework laying out how depreciation should be measured for these purposes. Our results show that PCs lose roughly half their remaining value, on average, with each additional year of use. The bulk of that decline reflects the downward revaluation of existing PCs, which is driven by the steep ongoing drop in the constant-quality prices of newly introduced models. In addition, PCs experience age-related declines in value that stem from the inability of older models to perform the full range of desired tasks and from the decision to retire installed units. We estimate that the resulting depreciation proceeds slowly during the early part of the PC’s lifetime but then picks up. In our preferred specification, the depreciation rate averages about 22 percent annually over the first five years of service. As we discussed, however, this figure is sensitive to the estimated rate of constant-quality price change. When we constrain our regression to follow the NIPA constant-quality price series, the depreciation rate increases to an average pace a bit above 34 percent. This estimate of depreciation is suitable for use in the NIPAs, and BEA

72


decided to adopt this geometric approximation in the December 2003 NIPA revision. Regarding tax policy, our conceptual framework describes the depreciation allowances that would equalize effective tax rates across assets in the face of both general price inflation and changes in relative asset prices. Given this benchmark, our empirical estimates suggest that the current tax depreciation schedule for PCs is about right in a zero inflation environment. Because the tax code is not indexed for inflation, however, the tax allowances would be too small in present value for inflation rates above the very low level now prevailing.

APPENDIX This appendix proves two propositions that are cited in section 2 of the text. The first derives the tax depreciation allowances that equalize effective tax rates across assets when one allows for both general price inflation and changes in relative asset prices. This proposition generalizes the well-known result for equalizing effective tax rates in a world with constant relative prices. The second proposition shows how to calculate constant-dollar wealth stocks, again allowing for changes in relative prices.

Proposition 1: Specifying Depreciation Allowances That Equalize Effective Tax Rates Let p zk, t, a denote the price of a type k capital good with the set of embodied characteristics z; this price is observed in year t when the capital good is a years old. In addition, let DTAXzk, t, a denote the schedule of tax depreciation allowances for type k capital goods, let θk denote the investment tax credit (ITC) for these goods, let τ denote the statutory tax rate on corporate profits, and let τk,e denote the effective tax rate on the income generated by a type k capital good (taking account of depreciation allowances and any ITC). Numerous studies (Jorgenson and Sullivan, 1981, and Gravelle, 1982, for example) have shown that the effective tax rate for every type of capital good equals the statutory corporate tax rate if the tax allowances for depreciation reflect the asset’s actual loss of value and there is no ITC. In the context of our model with changes in relative asset prices, this allowance includes both the age-related loss of value and the revaluation of the asset in real terms. Thus, we show that τk,e = τ for all k if θk = 0 and: DTAXzk, t, a = (d k - (r k - r)) p zk, t, a

(12)


73

where δk and πk are, respectively, the rate of depreciation and the rate of constant-quality price change for type k capital goods, and π is the general rate of inflation.27

Proof The proof proceeds in three steps. We begin by deriving the expression for the user cost of capital. Next, we express the effective tax rate as a function of the tax parameters and other terms in the user cost. The final step is to show that the effective tax rate equals the statutory tax rate for every type of capital good if θk = 0 and the tax allowance for depreciation accords with equation (12). User Cost of Capital Our derivation of the user cost of capital follows the standard method in the literature (see Hall and Jorgenson, 1967, for example). We begin by expressing the current price of the capital good as the discounted value of its future after-tax rental income, plus the present value of its tax allowances for depreciation and any investment tax credit it receives. Let xk denote the present value of the depreciation allowances for type k capital goods; xk and θk are both measured per dollar of the capital good’s value. Also, let c zk, t, a denote the pretax user cost for type k capital. In equilibrium, the user cost equals the pretax rental income generated by the capital good, allowing its price to be written as: p zk, t, a =

3

#0 ^1 - x h c zk, t + s, a + s e - is ds + _ i k + xx k i p zk, t, a

(13)

Equation (13) adopts the usual assumption that the asset has an infinite service life within a continuous time framework; this setup simplifies the algebra while preserving the key economic results. With the asset assumed to depreciate at a constant rate of δk percent and to experience a constant-quality price change of πk percent per period, the user cost (and hence the asset’s rental income) declines at a rate of (δk − πk) percent. Thus, equation (13) can be expressed as: p zk, t, a =

3

#0^1 - x h c zk, t, a e - (d

k-

rk ) s

e - is ds + _ i k + xx k i p zk, t, a

(14)

Solving equation (14) for c zk, t, a yields: k k c zk, t, a = 1 - i - xx p zk, t, a _ i + d k - r k i 1-x k k = 1 - i - xx p zk, t, a 9^ i - r h + d k - _ r k - r i C 1-x

(15)

27 Gravelle (1982) and others have demonstrated that an investment tax credit of θ percent for all capital goods reduces the effective tax rate more for short-lived assets than for

74


where the expression on the second line adds and subtracts the general rate of inflation (π). For the remainder of the derivation, we use the ratio of the asset’s user cost to its price, as shown in equation (16): k k (c/p) kz, t, a = 1 - i - xx 7 (i - r) + d k - (r k - r) A 1-x

(16)

Effective Tax Rate The effective tax rate (τ k,e) typically has been defined in the literature as the asset’s pretax return (ρ k) minus its after-tax return (r k), expressed as a percentage of its pretax return; that is, τ k,e = (ρ k − r k)/ρk. Both the pretax return and the after-tax return are measured net of depreciation and general price inflation. Abstracting from relative price changes (the standard approach in the tax literature), the real pretax return net of depreciation is ρ k = (c/p)k − δ k, where (c/p) k is calculated from equation (16) with π k = π. To specify after-tax returns, the usual assumption is that competitive forces equalize the real after-tax return on all assets net of depreciation, so that r k = i − π. We generalize this framework to allow for changes in relative asset prices. Only two modifications are required, both affecting the measurement of the pretax return. First, we calculate (c/p)k from equation (16) without forcing πk to equal π. Second, we subtract both depreciation (δk) and the real revaluation term (πk − π) from (c/p)k. The intuition is that, with changing relative prices, firms must cover both depreciation and revaluation effects to maintain the real value of their capital stocks. In this general case, the real pretax return net of depreciation and revaluation is ρk = (c/p)k − [δk − (πk − π)]. Using this expression for ρk and recalling that r k = i − π, the effective tax rate can be written as: x

k, e

k k k t k - r k 9 (c/p) z, t, a - _ d - (r - r)i C - (i - r) = = tk 9 (c/p) kz, t, a - _ d k - (r k - r)i C

(i - r) =1k 9 (c/p) z, t, a - _ d k - (r k - r)i C

(17)

Next, substitute the expression for (c/p)k from equation (16) into equation (17), which yields: x k, e = 1 -

(1 - x)( i - r) (1 - i k - xx k ) 7 (i - r) + d k - (r k - r) A - (1 - x) 7 d k - (r k - r) A

(18)

long-lived assets. One could counteract this effect by granting a progressively larger credit to longer-lived assets, but given our focus on depreciation allowances, we derive the proposition with the ITC set to 0.


75

Equalizing Effective Tax Rates Across Assets The final step is to derive the conditions under which the effective tax rate for every type of capital equals the statutory corporate tax rate. We will show that τ k,e = τ for all k if θk = 0 and the schedule of tax depreciation allowances matches equation (12), which is repeated here as equation (19): DTAXzk, t, a = (d k - (r k - r)) p zk, t, a

(19)

The present value of these allowances per dollar of asset value is: 3

xk = = =

#0 DTAXzk, t + s, a + s e - is ds p zk, t, a 3

#0

3

L

(d - (r - r)) e k

k

N

J pk

#0 (d k - (r k - r)) KK

z, t + s, a + s O - is e ds p zk, t, a O

(20)

P

- (d k - r k ) s

e

- is

ds

where the second line substitutes for DTAX from equation (19), and the third line makes use of the assumption that the asset’s value declines at a constant rate of (δk − πk) percent per period. Equation (20) implies that: xk =

d k - (r k - r) d k - (r k - r) k k = (i - r) + d k - (r k - r) i+d -r

(21)

Finally, substitute the expression for xk from equation (21) and θk = 0 into equation (18). After some algebra, the right-hand side of equation (18) reduces to τ, completing the derivation.

Proposition 2: Constructing Constant-Dollar Wealth Stocks Let Wtk and Wtk,1996$ denote, respectively, the current-dollar and constantdollar wealth stocks for capital of type k. In addition, let Itk and Itk,1996$ denote, respectively, current-dollar and constant-dollar investment outlays for this type of capital. Following the NIPA convention at the time we were writing, we assume that constant-dollar series are expressed in 1996 dollars. We show that the constant-dollar wealth stock can be calculated in two equivalent ways: T-1

Wtk, 1996$ =

Wtk = I k, 1996$ (1 - d k ) i (1 + r k ) t - 1996 i = 0 t - i

冱

76


That is, the constant-dollar wealth stock can be calculated by deflating the current-dollar stock or by constructing an appropriately weighted sum of constant-dollar investment flows.28

Proof To begin, recall that the current-dollar wealth stock for type k capital equals the sum of current-year investment plus the remaining value of the investment done in previous years: Wtk =

T-1

冱I

i=0

k t - i ((1

+ r k )( 1 - d k )) i

(22)

Next, multiply and divide the right-hand side of equation (22) by (1 + π k)t-i-1996 to obtain: T-1

Wtk =

k t-i k t - i - 1996

冱 (1 + r I )

i=0

(1 + r k ) t - i - 1996 (1 + r k ) i (1 - d k ) i

T-1

= (1 + r )

k t - 1996

I tk - i k i k t - i - 1996 (1 - d ) i = 0 (1 + r )

冱

(23)

Note that I tk - i/(1 + πk)t-i-1996 equals the constant-dollar investment done in $ year t-i, which we have denoted by I tk,-1996 . Hence, equation (23) can be i written as: T-1

Wtk = (1 + r k ) t - 1996

冱I

i=0

k, 1996$ (1 t-i

- d k )i

(24)

To complete the derivation, note that the constant-dollar wealth stock, Wtk,1996$, equals the current-dollar stock, Wtk, divided by the price deflator for year t, (1 + πk)t − 1996. Thus, equation (24) yields: Wtk, 1996$ =

Wtk = (1 + r k ) t - 1996

T-1

冱I

i=0

k, 1996$ (1 t-i

- d k )i

(25)

28 For this derivation, we shift back to a discrete-time framework and assume that the asset has a finite service life; the discrete-time framework conforms more closely with the actual data on investment and wealth stocks, and there is no algebraic advantage in this case from using continuous time.


77

REFERENCES Akerlof, George A. (1970). “The Market for ‘Lemons’: Quality Uncertainty and the Market Mechanism.” Quarterly Journal of Economics 84(August):488–500. Auerbach, Alan J. (1982). “Tax Neutrality and the Social Discount Rate: A Suggested Framework.” Journal of Public Economics 17(April):355–372. Berndt, Ernst R., and Neal J. Rappaport (2001). “Price and Quality of Desktop and Mobile Personal Computers: A Quarter-Century Historical Overview.” American Economic Review 91(May):268–273. Berndt, Ernst R., and Neal J. Rappaport (2003). “Hedonics for Personal Computers: A Reexamination of Selected Econometric Issues.” Massachusetts Institute of Technology. Mimeo. August 21. Bloomberg News Service (2003). “U.S. Economy: Computer Replacement Delays May Damp Recovery.” May 23. Bradford, David F., and Don Fullerton (1981). “Pitfalls in the Construction and Use of Effective Tax Rates.” In Depreciation, Inflation, and the Taxation of Income from Capital, Charles R. Hulten (ed.). Washington, DC: The Urban Institute Press, 251–278. Bureau of Economic Analysis (1999). Fixed Reproducible Tangible Wealth in the United States, 1925–94. Washington DC: U.S. Government Printing Office. Cartwright, David W. (1986). “Improved Deflation of Purchases of Computers.”Survey of Current Business 66(March):7–10. CCH (2002). 2003 U.S. Master Tax Guide. Chicago, IL: CCH Incorporated. Christensen, Laurits R., and Dale W. Jorgenson (1995). “Measuring Economic Performance in the Private Sector.” In Productivity, Volume 1: Postwar U.S. Economic Growth, Dale W. Jorgenson (ed.). Cambridge, MA: The MIT Press, 175–272. Cummins, Jason G., and Giovanni L. Violante (2002). “Investment-Specific Technical Change in the United States (1947–2000): Measurement and Macroeconomic Consequences.” Review of Economic Dynamics 5(April):243–284. Diewert, Erwin (2002). “Measuring Capital.” Working Paper, April 18. Available at http://www.econ.ubc.ca/diewert/594cap.pdf (accessed March 1, 2004). Dulberger, Ellen R. (1989). “The Application of a Hedonic Model to a QualityAdjusted Price Index for Computer Processors.” In Technology and Capital Formation, Dale W. Jorgenson and Ralph Landau (eds.). Cambridge MA: The MIT Press, 37–75. Fraumeni, Barbara M. (1997). “The Measurement of Depreciation in the U.S. National Income and Product Accounts.” Survey of Current Business 77(July):7–23. Available at http://www.bea.doc.gov/bea/an1.htm (accessed March 1, 2004). Geske, Michael J., Valerie A. Ramey, and Matthew D. Shapiro (2003). “Why Do Computers Depreciate?” University of Michigan. Mimeo. September 2. Gravelle, Jane G. (1982). “Effects of the 1981 Depreciation Revisions on the Taxation of Income from Business Capital.” National Tax Journal 35(March):1–20. Gravelle, Jane G. (1994). The Economic Effects of Taxing Capital Income. Cambridge, MA: The MIT Press. Griliches, Zvi (1960). “The Demand for a Durable Input: U.S. Farm Tractors, 1929–57.” In The Demand for Durable Goods, Arnold C. Harberger (ed.). Chicago, IL: University of Chicago Press, 181–207.

78


Hall, Robert E. (1971). “The Measurement of Quality Change from Vintage Price Data.” In Price Indexes and Quality Change: Studies in New Methods of Measurement, Zvi Griliches (ed.). Cambridge, MA: Harvard University Press, 240–271. Hall, Robert E., and Dale W. Jorgenson (1967). “Tax Policy and Investment Behavior.” American Economic Review 57(June):391–414. Hayek, F. A. (1941). “Maintaining Capital Intact: A Reply.” Economica 8(August):276–280. Hicks, J. R. (1942). “Maintaining Capital Intact: A Further Suggestion.” Economica 9(May):174–179. Hill, Peter (1999). “Capital Stocks, Capital Services and Depreciation.” Presented at the meeting of the Canberra Group on Capital Stock Statistics, November 8–10, 1999, Washington DC. Available at http://www.oecd.org/dataoecd/12/ 47/2549891.pdf (accessed March 1, 2004). Hill, Peter (2000). “Economic Depreciation and the SNA.” Paper prepared for the 26th General Conference of the International Association for Research in Income and Wealth, August 27–September 2, Cracow, Poland. Available at http://www.iariw.org/prog2000.htm (accessed March 1, 2004). Hulten, Charles R., and Frank C. Wykoff (1981a). “The Estimation of Economic Depreciation Using Vintage Asset Prices: An Application of the Box-Cox Power Transformation” Journal of Econometrics 15(April):367–396. Hulten, Charles R., and Frank C. Wykoff (1981b). “The Measurement of Economic Depreciation.” In Depreciation, Inflation, and the Taxation of Income from Capital, Charles R. Hulten (ed.). Washington, DC: The Urban Institute Press, 81–125. Johnson, Norman L., Samuel Kotz, and N. Balakrishnan (1994). Continuous Univariate Distributions, Volume 1. New York: John Wiley & Sons. Jorgenson, Dale W. (1974). “The Economic Theory of Replacement and Depreciation.” In Econometrics and Economic Theory: Essays in Honour of Jan Tinbergen, Willy Sellekaerts (ed.). White Plains, NY: International Arts and Sciences Press, 189–221. Jorgenson, Dale W. (1996). “Empirical Studies of Depreciation.” Economic Inquiry 34(January):24–42. Jorgenson, Dale W., and Martin A. Sullivan (1981). “Inflation and Corporate Capital Recovery.” In Depreciation, Inflation, and the Taxation of Income from Capital, Charles R. Hulten (ed.). Washington, DC: The Urban Institute Press, 171–237. Landefeld, J. Steven, and Bruce T. Grimm (2000). “A Note on the Impact of Hedonics and Computers on Real GDP.” Survey of Current Business 80(December):17–22. Lane, Richard N. (1999). “Appraisal Report ‘Large Aerospace Firm’ Personal Property.” Los Angeles County. Revised February 2, 1999. OECD (2001). Measuring Capital—OECD Manual: Measurement of Capital Stocks, Consumption of Fixed Capital and Capital Services. Paris: OECD. Oliner, Stephen D. (1993). “Constant-Quality Price Change, Depreciation, and Retirement of Mainframe Computers.” In Price Measurements and Their Uses, Murray F. Foss, Marilyn E. Manser, and Allan H. Young (eds.). Chicago, IL: University of Chicago Press, 19–61. Oliner, Stephen D. (1994). “Measuring Stocks of Computer Peripheral Equipment: Theory and Application.” Federal Reserve Board. Mimeo. May.


79

Pakes, Ariel (2003). “A Reconsideration of Hedonic Price Indices with an Application to PCs.” American Economic Review 93(December):1578–1596. Pigou, A. C. (1941). “Maintaining Capital Intact.” Economica 8(August):271–275. Richards, Gordon R. (2002). “Nonlinear Technical Advance in the Aggregate Production Function: An Econometric Model.” Intel Corp. Mimeo. Sliker, Brian K. (2003). “Steps Toward Modeling the Distribution of Automobile Retirements.” Bureau of Economic Analysis. Mimeo. July 10. Wykoff, Frank C. (2003). “Obsolescence in Economic Depreciation from the Point of View of the Revaluation Term.” Pomona College. Mimeo. February 28.

TAX POLICY AND EDUCATION POLICY: COLLISION OR COORDINATION? A CASE STUDY OF THE 529 AND COVERDELL SAVING INCENTIVES Susan Dynarski Harvard University and NBER

EXECUTIVE SUMMARY Coverdell Educational Savings Accounts and 529 saving plans are marketed as attractive vehicles for college savings. The main finding of this paper is that college savings plans can actually harm some families. The joint treatment by the income tax code and financial aid system of college savings creates tax rates that exceed 100 percent for those families on the margin of receiving additional financial aid. Because even families with incomes above $100,000 receive need-based aid, the impact of these very high taxes is quite broad. I find that an aid-marginal family with funds in a Coverdell is worse off than if it did not save at all. Simulations show that $1,000 of pretax income placed in a Coverdell for a newborn and left to accumulate until college will face income and aid taxes that consume all of the principal, all of the earnings, and an additional several hundred dollars. This perverse outcome is the product of

I gratefully acknowledge support from the Russell Sage Foundation and NBER National Institute on Aging Grants P30-AG12810 and K12-AG000983. Naomi Calvo, Joe Ciesla, Betsy Kent, and Juan Saavedra provided excellent research assistance. I thank Julie-Anne Cronin, Martin Feldstein, Andrew Samwick, James Stedman, and participants in the NBER Taxation and Saving conference and the Tax Policy and the Economy conference for helpful comments.

82

Dynarski

poor coordination between the income tax code and the financial aid system.

1. INTRODUCTION In the past few years, a new breed of tax-advantaged savings vehicle has emerged. The federal Coverdell Education Savings Account (ESA) allows annual, after-tax deposits of up to $2,000 a year, with asset earnings untaxed as long as withdrawals are used for educational expenses. At the state level, nearly every state offers a tax-advantaged 529 savings plan. These accounts allow participants to make annual, after-tax deposits of up to $11,000 a year per child, comparable to the annual ceilings on the 401(k).1 The tax treatment is like that of the ESA: earnings are untaxed by the federal government, and by almost every state, when the funds are used for postsecondary education. In about half the states, deposits are exempt from state taxation, further increasing the income tax advantages of the 529. Politicians and financial advisers aggressively market 529 saving plans and the ESA as attractive vehicles for college savings. For many families, the favorable tax treatment of these savings vehicles does make them more attractive than other methods of saving for college. As I show in this paper, however, some families are worse off saving in an ESA than they would be in an alternative savings vehicle, such as an IRA or even a non-tax-advantaged account. For families on the margin of getting more financial aid, holding funds in an education savings account results in substantial decreases in aid eligibility. In the case of the ESA, more than a dollar in aid is lost for each dollar held in the account, more than undoing its tax incentive for saving and in fact leaving a family worse off than if it had not saved at all. One might dismiss the results of the paper as irrelevant by observing that the poor get aid but do not save, and the rich save but do not get aid. This common wisdom is wrong. As I show in the next section, a substantial proportion of families with incomes above $70,000, and even $100,000, receive need-based aid in the form of both grants and loans. Upperincome students at expensive, four-year private colleges often qualify for need-based grant aid from their schools, while even those at lessexpensive four-year public colleges often qualify for subsidized, need-based 1 Each parent can deposit $11,000 per child in a given year without triggering a gift tax. A two-parent family with three children could therefore move $66,000 per year into a taxadvantaged 529 account. Grandparents can also make deposits up to these limits, further expanding the amount of assets that can be shielded from taxation. A five-year averaging option allows a participant to contribute $55,000 in a single year without triggering a gift tax.

Tax Policy and Education Policy

83

federal loans. These families are therefore subject to the aid policies I describe in the paper. Of course, such families also save and so have assets that are affected by the intersection of tax policy and aid policy described in this paper. The fact that a tension exists between policies intended to increase saving and distribute aid according to need is unsurprising. The intent of the 529 and ESA is to increase saving by increasing after-tax returns. The intent of the need-based aid system is to give less aid to those with greater assets. These two sets of policies inevitably work at cross-purposes because the aid system taxes away part of the increase in assets and asset returns that the savings incentives create.2 This tension between targeting funds to those who are most needy and discouraging desirable behaviors is an inherent characteristic of all means-tested programs. For example, the old welfare system had an earned income test: welfare benefits were reduced proportionally for each dollar earned. This acted as a tax on labor supply and theoretically discouraged work effort by welfare recipients. Similarly, the need-based aid system taxes increases in income and assets, thereby potentially discouraging saving. Unless assets and asset income are completely disregarded in the distribution of need-based aid, the aid determination process inevitably reduces asset returns and perhaps saving rates. The conclusion of this paper is that the tension between targeting aid and discouraging saving can be managed well or poorly. For example, I find that the aid system assesses different assets at highly variable rates, with the drop in aid associated with a dollar in assets ranging from 0.50 to nearly $2.00. This variation in asset treatment has a cost because it distorts decisions about the composition of savings. There is no concomitant benefit, however, because these wildly varying policies do not improve the targeting of aid toward needy students. If anything, such arbitrary policy variation undermines the goals of need-based aid because families with identical financial positions receive very different levels of aid, depending on whether they are savvy enough to steer their savings toward the right vehicles. It now appears that the Department of Education is moving to improve the treatment of the ESA documented in this paper. In early November 2003, the department posted revisions to the online version of the Student Financial Aid Handbook, its reference manual of aid rules. These revisions indicate that, in the future, the ESA will be given the treatment currently applied to the 529 savings plans. This treatment will eliminate the socalled aid tax of over 100 percent that is currently applied to the ESA. 2

The aid tax was first discussed by Edlin (1993) and Feldstein (1995).

84

Dynarski

It is not clear when this new policy will become effective. Given how the department collects asset data from applicants, a necessary step in implementing the new policy is revision of the Free Application for Federal Student Aid (FAFSA). The FAFSA does not collect separate data on each type of asset; if it did, the department could change the formula that calculates aid eligibility without altering the FAFSA. Rather, the 2003–2004 FAFSA, which has not been revised, instructs families to add ESA balances to other miscellaneous student assets, while 529 balances are added to other parental assets. Parental and student assets are then run separately through the aid formula, with $1.00 in student assets leading to a reduction in aid of more than $1.00 over the course of a college career. Note that all student assets are subject to this treatment; changing the treatment of the ESA will still leave other student assets subject to the very high taxes, that are the subject of this paper. The discussion in section 6 addresses this point. The paper is organized as follows. In section 2, I show that families quite high in the income distribution are affected by aid policy. In section 3, I provide background on the tax-advantaged college savings plans. I calculate returns on various savings vehicles net of income taxes in section 4. In section 5, I explain the aid determination process and calculate returns that account for both income taxes and the reductions in aid caused by holding savings in various vehicles. Section 6 discusses the results, and section 7 concludes.

2. WHO IS AFFECTED BY AID POLICY? What kind of family is affected by the aid system and its treatment of assets? Given the historically high level of tuition prices, relatively welloff families qualify for need-based aid and so are affected by the aid rules. This scenario is particularly true if the student attends a private college or if a family has multiple students in college at the same time.3 As this section will show, families all along the income distribution are affected by the need-based aid system and its treatment of assets and asset returns. For two kinds of families, however, the aid system’s treatment of assets is irrelevant. The first type of family is extremely needy (as defined by the need-based aid system) and receives the maximum aid allowed.4 For this family, a marginal decrease in assets does not increase its aid, nor does a marginal increase in its assets decrease its aid. Because no link exists 3

A family that has multiple children in college at a given point in time is eligible for more need-based aid than if those children attended college in sequence.

4

Total aid is capped by a student’s actual schooling costs, which includes tuition and fees plus an allowance for items such as food, rent, and other living expenses.


85

between assets and aid for this family, its net asset returns are unaffected by the aid system. The second type of family is at the other end of the spectrum: this family is well off (again, as defined by the need-based aid system) and receives no aid. Marginal changes in assets do not affect this family’s aid eligibility. For any family that is not at one of these two extremes of need, asset returns are affected by the rules discussed in this paper.

2.1 Who Gets Aid? Families all along the income distribution get financial aid. Table 1 shows the probability that a student with a given family income will receive need-based aid. The table also shows the average amount of aid received among aid recipients. These data are for nonforeign, full-time, dependent

TABLE 1 Need-Based Aid Receipt, by Income, for Dependent, Full-Time Undergraduates, Academic Year 1999–2000 Household income 0

85% $6,859

62% $5,937

37% $5,371

22% $4,975

Received Mean if > 0 Received Mean if > 0 Received Mean if > 0

68% $2,259 26% $4,074 52% $3,835

9% $1,056 24% $5,060 49% $3,491

0% — 18% $4,793 28% $3,322

0% — 12% $4,617 12% $3,518

0%

6%

14%

8%

4%

39%

39%

19%

4%

45%

53%

27%

Note: Data are from NPSAS 2000 (National Center for Education Statistics, 2000)

86

Dynarski

undergraduates attending a single college in academic year 1999–2000 and are taken from the 2000 National Postsecondary Aid Survey (NPSAS). I show results separately for four categories of need-based aid: (1) all types, including grants, loans, and work study; (2) federal Pell Grants only; (3) need-based grants provided by colleges; and (4) subsidized federal loans. Low-income families are most likely to receive aid and get the largest aid packages.5 Among students with family incomes below $40,000, 85 percent receive need-based aid, with their total aid packages averaging $6,859. However, middle- and even upper-income families are quite likely to receive substantial amounts of aid. Of students from families with incomes of $40,000 to $70,000, 62 percent receive need-based aid in the form of grants, loans, or work study, with the aid of recipients averaging $5,937. Moving up the income distribution, we see that 37 percent of students from families with incomes of $70,000 to $100,000 receive needbased aid averaging $5,371. Even in the highest income group, 22 percent of students receive some form of need-based aid, averaging $4,975. The composition of this need-based aid varies considerably across the income groups. Pell Grant distribution is highly progressive. While 68 percent of students from families with income below $40,000 receive a Pell Grant, only 9 percent of students from families with incomes of $40,000 to $70,000 receive a Pell Grant, and no students in higher income categories receive one. While the Pell Grant is heavily concentrated among lowincome students, the story is quite different for other forms of need-based aid. Colleges and universities, especially the more expensive private schools, distribute their own need-based scholarships. The more expensive the school, the more likely that a student of a given income level will qualify for need-based aid from that school. Among students with family income below $40,000, 26 percent receive need-based grants from their schools, with the grant of recipients averaging $4,074. In the next higher income category, the share receiving a need-based grant drops barely, to 24 percent, while the average grant received rises to $5,060. This reflects the tendency of these higher-income families to send their children to expensive schools. Even among families with incomes above $100,000, 12 percent receive need-based grants from their schools averaging $4,617 per grant.6 5 Note that the average amount of need-based aid does not drop very rapidly with income. This situation arises because higher-income students are more likely to attend expensive private institutions, and need is a function of both ability to pay and actual schooling costs. 6 Most schools follow the federal formulas described in this paper in distributing their own need-based grant. Eighty-seven percent of four-year public schools and 57 percent of fouryear private schools use the federal formula in distributing their own need-based grants (see National Association of Student Financial Aid Administrators and the College Board, 2002).


87

Many middle- and upper-income students also qualify for need-based, subsidized federal loans. While loans are obviously less valuable than grants, the need-based Perkins and Stafford loans have very attractive terms, with all interest paid while the child is in school and a low rate of interest paid thereafter. The subsidy value of a need-based Stafford loan is currently about 30 cents on the dollar.7 In the $40,000 to $70,000 income range, 49 percent of students receive one of these federal loans, with annual borrowing averaging $3,419, not very different from the borrowing patterns in the lowest-income group (52 percent borrowing, with loans averaging $3,835). Even in the highest income category, the figures are 12 percent and $3,518, respectively.

2.2 Who Is on the Margin of Getting More Aid? Many of the families who receive need-based aid are on the margin of getting more aid—that is, an increase (decrease) in their financial resources decreases (increases) the amount of aid for which they are eligible. So too are those families who currently get no aid at all but would if their financial resources, as defined by the aid system, were to decrease. We can learn how many students are on the aid margin by examining the population of current students, and in particular those who apply for financial aid. Note that who applies for aid is almost certainly influenced by individuals’ expectations about whether they will qualify for aid and how much they might receive. For example, an upper-income family with substantial funds in an ESA might not apply for aid under the current policy regime, but that same family would if the aid system treated ESAs differently. By using data from those students who apply for aid to estimate the share of all students who are on the margin of aid, I underestimate the share of the student population that would be affected by a change in the aid formula because I do not account for such endogenous changes in the extensive aid margin.8 Describing who is on the aid margin requires some understanding of how the need-based aid system defines need. As I will describe in greater detail later in the paper, need is determined by comparing a student’s projected schooling costs with the amount that the aid formula determines 7 See my previous work in Dynarski, 2002. The bulk of the subsidy arises from the government paying the interest on the loan while the student is in school. The subsidy value on the Stafford is at a historical low because market interest rates are quite low. As market interest rates rise, so too does the subsidy value. The subsidy value rises especially rapidly when market rates exceed the statutory rate cap of 8.25 percent because above this rate, the government assumes all interest rate risk. 8 Note that in the following calculations, when a student does not have EFC information, as is the case for anyone who has not applied for aid, I have assumed that she or he is not on the margin of getting aid.

88

Dynarski

that her or his family can afford to pay toward college. This latter amount is referred to as the expected family contribution (EFC). Need is defined as schooling costs minus the EFC. Two types of students are on the aid margin: (1) those receiving no aid but who would if their financial situation changed marginally (these students have nonpositive need) and (2) those receiving some aid who would receive more or less aid if their financial situation changed marginally (these students have positive need). I treat these two cases in turn. A student with nonpositive need is not eligible for need-based aid because the aid formula calculates that he and his family can handle the full cost of college. Those with very negative need (EFC >> schooling cost) are far from the aid margin because the aid formula indicates that they can contribute an amount well above schooling costs; such families are not on the aid margin. But for those whose need is relatively small and negative, marginal decreases in their financial resources push them over the margin into aid eligibility. At the bottom of Table 1, I show the share of students in each income category whose need lies between 0 and −$5,000. For these students, changing the formula so that their expected family contribution drops by $5,000 or less pushes them over the margin into receiving aid. To get a sense of the magnitude of this change in EFC, note that a high school senior whose family has $15,000 of college savings in a Coverdell ESA or Uniform Transfer to Minors Act (UTMA) account has a freshman-year EFC about $5,000 higher than a senior whose family has no college savings.9 The lowest-income families (less than $40,000) are always eligible for some form of need-based aid, so none of them are on this aid margin. However, 6 percent of students from families in the $40,000 to $70,000 income range, and 14 percent of those in the $70,000 to $100,000 income range, would be pushed into aid eligibility by a decrease in their EFC of $5,000 or less. In the highest-income group, 8 percent are on this aid margin.10 Another type of family is getting some need-based aid but would get more if their EFC dropped. These families have positive need, but they are not so needy that changes in their financial situation cannot increase or decrease their aid package. I define these families as those whose EFCs are sufficiently far from zero (at least $5,000) that they will see substantial changes in need if their financial resources (as defined by the aid system) 9 As I will show later in the paper, the aid eligibility of subsequent years of college is also negatively affected by this ESA account, so the ultimate impact of ESA and UTMA holdings on aid is substantially larger than that described in this sentence. 10 Note that the share of students on this margin is likely to be underestimated using data on aid applicants, as discussed above.


89

alter. For such families, dollar decreases in need translate into dollar increases in aid. For example, their need can increase by at least $5,000 without bumping up against the ceiling of the student’s total schooling costs.11 Again, few low-income students are on this aid margin—just 4 percent. However, about 40 percent of students in the $40,000 to $100,000 income group have room for their need to grow by at least $5,000; 19 percent of students in the top income group fall on this aid margin. As the bottom row of Table 1 shows, a substantial share of families fall on one of these two aid margins. Roughly half of students from families with income between $40,000 and $100,000 are on the margin of getting more aid, as are one-quarter of those from families with incomes above $100,000. The interaction of aid policy and tax policy described in this paper therefore affects a large number of families.

3. INCOME TAX INCENTIVES FOR COLLEGE SAVING 3.1 Legislative History In 1997, the Education IRA was established. The Education IRA was structured much like the then-new Roth IRA. In both types of vehicles, aftertax dollars grow tax-free. Earnings are never taxed if Education IRA withdrawals are used for postsecondary expenses or if Roth funds are withdrawn after age 591⁄2. Annual contributions to the Education IRA were capped at $500 per child until 2001, when the contribution limit was raised to $2,000. The same year, eligible educational expenses were expanded to include primary and secondary education, and the name of the Education IRA was changed to Coverdell Education Savings Account (ESA). While the ESA is a product of federal legislation, the 529 savings plans are innovations of the states. The 529 savings plans have their roots in prepaid tuition plans, the first of which was introduced by Michigan in 1986. Those who purchased shares in Michigan’s plan were guaranteed that their investment would cover the cost of a certain number of semesters at Michigan schools. Essentially, Michigan created a savings plan whose rate of return was linked to tuition costs at the state’s public postsecondary schools, thereby allowing parents to insure against the risk of rising tuition prices.12 Michigan exempted investment returns in its prepaid 11 This assumes that all need is met by some combination of loans, grants, and work study provided by government and schools. 12 A key drawback of the prepaid plans is that the tuition guarantee is only for in-state schools. Funds can be used at out-of-state schools, but the implied rate of return on funds used in this way is quite low.

90

Dynarski

plan from state taxes, and the state argued to the Internal Revenue Service (IRS) that returns should also be exempted from federal taxes. The IRS disagreed, but Michigan went forward with the plan and sued the IRS for a refund of taxes paid, winning its case in 1994. While the Michigan case was wending through the courts, several other states introduced their own prepaid tuition plans. In 1997, Congress codified in Internal Revenue Code (IRC), section 529, the federal tax treatment of the tuition plans, which was to tax earnings in these accounts only at withdrawal. IRC 529 also contains language that recognized a variant on the prepaid plans that had been introduced by a handful of states: the tax-advantaged college savings plan. Like the Education IRA, these new savings plans allowed after-tax investments to grow free of federal and state taxes; however, withdrawals used for postsecondary costs were exempt only from state taxation. With the passage of tax reform in 2001, the federal tax on withdrawals from 529 savings plans was eliminated.13 Every state except Washington now has a 529 savings plan, as does the District of Columbia. The growth of the 529 savings plans has far outstripped that of the prepaid plans likely because of their greater fungibility and potentially higher returns.14 In this paper, I focus on the 529 savings plans.

3.2 Eligibility for and Tax Advantages of the 529 and ESA The tax treatments of the ESA and 529 are quite similar: after-tax dollars put into savings and earnings are not taxed as they accrue, nor are they taxed at withdrawal if the withdrawal is used for educational expenses.15 There are some key differences, however, between the two savings vehicles. First, there is an income limit on participation in the ESA. Joint-filer households with incomes above $220,000 and single-filer households with incomes above $110,000 cannot contribute to an ESA; eligibility begins to phase out at $190,000 and $95,000, respectively. There is no income limit on contributions to a 529 savings plan.16 A second distinguishing characteristic of the 529 is that its contribution limits are much higher than the limit on the ESA. Each account owner (a 13 This federal tax treatment of the 529 savings plans sunsets in 2010. The present analysis assumes that the provision will be extended indefinitely. 14 The bull market of the 1990s made the tuition plans appear stodgy to investors accustomed to double-digit returns. Also, the plans substantially constrained the college choices of beneficiaries, who could use the funds at out-of-state schools only at unattractive terms. 15 As discussed below, some states exempt contributions to the 529 from state taxable income, thereby increasing the tax advantages. 16 In some states, the exclusion of contributions from state taxable income phases out as income rises. The exclusion of earnings from taxable income is not linked to income in any state.


91

parent or grandparent, for example) can put $11,000 in after-tax income per beneficiary, per year, into a 529.17 A two-parent family with three children can put $66,000 a year into 529 savings plans for their children, but just $6,000 into ESAs. Each state has a lifetime limit on the account balance that can be reached in an account held in a given beneficiary’s name. When the account reaches this limit, no additional contributions can be made. The limit averages $241,000, and it ranges from $182,000 in Louisiana to $305,000 in South Dakota.18 Third, while families can invest their ESAs as they wish, they are constrained in their ability to allocate assets in a 529. Each state determines the investment options open to investors in its plan, and by federal law, assets can be reallocated by the investor only once a year. Until recently, most 529 savings plans provided only a single investment option, an agebased portfolio that grew less aggressive as the child neared college age. Most plans now offer several investment options. Finally, the 529s are creatures of state government, with each state sponsoring its own plan. Therefore, heterogeneity in 529 characteristics, including portfolio choice, tax treatment, and net returns, exists across the states. Each state contracts with a mutual fund company to run its plan, chooses the mutual funds that will be available to investors, decides on the treatment of deposits and earnings for the purposes of state taxation, and negotiates fees that will be paid by the investor to the state and fund company. Individuals are free to participate in any state’s plan. Many of the states encourage their residents to invest in the local plan by allowing them to deduct contributions to its 529 savings plan from state taxable income. Some states also tax withdrawals from other states’ 529 plans, further encouraging investors to choose their home state’s plan. There is considerable cross-state variation in fees charged on the 529 accounts. Fees for 529 accounts also appear to be somewhat higher, on average, than fees on ESAs or retail mutual funds. For the purposes of this paper, I ignore this source of variation in net returns across states and savings vehicles. By assuming that pretax returns on the various savings vehicles are identical, I can focus on variation in returns driven by the income tax code and the financial aid system. In ongoing work, I explicitly focus on sources of cross-state heterogeneity in 529 returns and its impact on savings decisions.

17 A total of $55,000 per account owner, per year, can be deposited in a single year for a beneficiary if no deposits are made for the next four years. 18

See Cerulli Associates, 2003.

92

Dynarski

4. CALCULATION OF AFTER-TAX RETURNS ON THE ESA, 529, AND ALTERNATIVE SAVINGS VEHICLES In this section, I calculate returns, net of the income tax, on the 529 and ESA, in absolute terms and relative to other vehicles. First, I show variation in net returns across vehicles for a single household type, with household income of $100,000 and two dependent children. Because the benefits of tax-advantaged accounts vary with marginal tax rates, I then calculate returns for a range of household incomes.

4.1 Assumptions For the purposes of assigning tax rates, I consider a household that consists of a married couple, filing jointly, with two dependent children. All earned income is assumed to come from one earner.19 The children are assumed to have no income other than that produced by any college savings held in their name. The marginal federal and state tax rates on earned income, capital gains, and interest for this household, as well as for the other income groups I will be analyzing, are shown in Table 2. The state tax rates in Table 2 are the average of the states’ 2002 marginal tax rates for each income group, as calculated by the TAXSIM program from the National Bureau of Economic Research.20 Table 2 shows (and the paper’s calculations use) federal tax rates effective as of the Jobs and Growth Tax Relief Reconciliation Act (JGTRRA) of 2003. Some of these rates are scheduled to revert to pre-2003 rates in a few years. It is difficult to forecast which, if any, of these provisions will be allowed to sunset, so I calculate the effect of making the current provisions permanent in this paper. For each savings vehicle, I calculate the return to $1,000 of pretax income placed in an account at the time of a child’s birth. A family saving for college will likely start with a portfolio heavily weighted toward stocks and move toward a more conservative mix as the start of college nears. Every state’s 529 savings plan offers an age-based portfolio that follows this pattern. I use a portfolio mix typical of state 529s in calculating returns; this portfolio is shown in Table 3. I assume an identical portfolio 19 Some assumption about the distribution of earned income within the household must be made before FICA rates can be assigned. For each earner, the FICA rate is 7.65 percent up to $87,000 and 1.45 percent thereafter. 20 The average is taken over the states that have an income tax. I use effective marginal state tax rates calculated by TAXSIM rather than the bracket rates. The effective marginal rates account for the interaction of state and federal taxes as well as the phaseout of various credits and deductions.


93

TABLE 2 Marginal Tax Rates Used in Calculations Household income $35,000 $50,000 $100,000 $150,000 $200,000 $335,000+

Earned income

Capital gains

Interest income

Federal

State

FICA

Federal

State

Federal

State

10% 15% 25% 28% 33% 35%

5.08% 5.65% 6.29% 6.43% 6.38% 6.40%

7.65% 7.65% 1.45% 1.45% 1.45% 1.45%

5% 5% 15% 15% 15% 15%

4.41% 4.83% 5.22% 5.61% 5.48% 5.56%

10% 15% 25% 28% 33% 35%

5.08% 5.65% 6.29% 6.43% 6.38% 6.40%

Notes: Federal rates are 2003 bracket rates. State rates are average of effective 2002 marginal rates calculated by NBER TAXSIM (Feenberg and Coutts, 1993 and NBER, 2004). State averages are taken across states that have an income tax.

TABLE 3 Age-Based Portfolio Used in Return Calculations

Year

1–3

Stock share 90% Bond share 10%

4–6

7–8

9

10

85% 74% 68% 59% 15% 26% 32% 41%

11–12

13

58% 42%

45% 55%

Nominal rate of 14–15 16–22 return 42% 58%

25% 75%

9% 4%

Note: Values reflect typical age-based 529 portfolio.

mix for the other savings vehicles, so that any variation in returns across the vehicles will be induced by variation in their treatment by the income tax and financial aid systems. Stocks are assumed to earn a nominal rate of return of 9 percent and bonds to earn a rate of 4 percent. To simplify the analysis, I assume that all stock returns take the form of long-term capital gains. Capital gains are realized when the funds are withdrawn from the account to pay for college; these withdrawals begin at the end of the account’s eighteenth year.21 After any relevant taxes on asset earnings are paid, earnings are reinvested. In about half the states, deposits to the 529 are excluded from state taxable income. I calculate returns for 529s both with and without this upfront deduction. I also calculate returns for a non-tax-advantaged mutual fund account in the name of the parent, a Uniform Transfer to Minors Act (UTMA) account in the name of the student, and a traditional 21 The family withdraws 1/nth of the remaining balance each year, with n representing the number of years remaining until college completion. For the calculations in the paper, I assume four years of college.

94

Dynarski

IRA. Table 4 summarizes the income tax treatment of these savings vehicles. For all vehicles, I assume that all capital gains realizations are put off until the account is drawn down and that no dividends are earned. Therefore, the only relevant taxes on the inside buildup are those on bond interest.22 Note that the IRAs can be used for higher education expenses without the 10 percent penalty usually assessed on withdrawals before retirement age. However, the earnings portion of such early withdrawals from a Roth is subject to taxation as ordinary income. As a result, the Roth is not an advantageous vehicle for college savings if its use requires early withdrawal.

4.2 Calculation of Returns Net of Income Taxes: Example First, I calculate the nominal returns for a family with household income of $100,000, using the assumptions laid out above. The return for a nonadvantaged mutual fund account, held in the name of the parent, forms the benchmark used to gauge the financial benefits of the tax-advantaged vehicles. After paying social security and Medicare taxes (FICA), as well as federal and state income taxes, on $1,000 of pretax income, this household has $673 to deposit. The family uses the portfolio allocation shown in Table 3, putting 90 percent of the funds into stocks and the balance into bonds. Interest on the bonds is taxed as ordinary income; the interest net of taxes is reinvested in the account. After 18 years, the account will have grown to $1,135, with 55 percent of the account’s value consisting of unrealized capital gains. At the end of year 18, one-quarter of the account balance is withdrawn to pay for college. Capital gains taxes are paid on the portion of this withdrawal that represents unrealized capital gains. After four years of withdrawals, the account is empty. Accounting for income and payroll taxes, as well as taxes on interest and capital gains, a family following the investment path just described nets $1,113 on its $1,000 in pretax saving, as shown in Table 5 and in the first bar in Figure 1. The tax-advantaged vehicles, including the 529 and ESA, increase returns by reducing or eliminating the taxes assessed before the initial deposit, during the inside buildup, and/or at withdrawal. The return for each of these vehicles is shown in Figure 1. The second column of Table 5 shows the returns on assets held in these vehicles relative to returns for a nonadvantaged account in the name of the parent. I briefly discuss the tax advantages conferred by each of these vehicles below. 22 I assume that current tax law will persist despite the scheduled sunset of the exclusion of 529 earnings from federal taxable income.

TABLE 4 Tax Treatment of College Saving Alternatives Investment option

Income limit, married, filing jointly

Nonadvantaged account, parent Traditional IRA

Federal and state, plus FICA. $70,000 No income limit if no work-related retirement plan.

$220,000

Taxed paid on inside buildup Federal and state.

Federal and state on realized capital gains. Federal and state on entire withdrawal.

Federal and state: First $750 untaxed. Child 14+: earnings >$750 at child’s rate. Child $1,500 at parent’s rate.

Federal and state on realized capital gains, child’s rate.

FICA.

Federal and state, FICA. No state taxes if 529 with deduction. Federal and state, FICA. Federal and state. FICA.

Note: Unless otherwise indicated, applicable tax rate is that applied to parent’s income.

Taxes paid at withdrawal


529

Coverdell UTMA

Taxes paid on income, pre-deposit

95

96

Dynarski TABLE 5 After-Tax Return to College Savings Alternatives Nominal return

Non-advantaged account, parent $35K $50K $100K $150K $200K $335K+ UTMA $35K $50K $100K $150K $200K $335K+ 529 plan (deduction) $35K $50K $100K $150K $200K $335K+ 529 plan (no deduction) $35K $50K $100K $150K $200K $335K+ ESA $35K $50K $100K $150K $200K $335K+ Traditional IRA $35K $50K $100K $150K $200K $335K+

Return relative to parental account

$1,735 1,485 1,113 987 803 728

1.00 1.00 1.00 1.00 1.00 1.00

$1,824 1,618 1,453 1,338 1,157 1,084

1.05 1.09 1.31 1.36 1.44 1.49

$2,188 1,976 1,811 1,683 1,475 1,391

1.26 1.33 1.63 1.71 1.84 1.91

$2,026 1,808 1,634 1,511 1,317 1,238

1.17 1.22 1.47 1.53 1.64 1.70

$2,026 1,808 1,634 1,511 1,317 1,238

1.17 1.22 1.47 1.53 1.64 1.70

$2,026 1,808 1,634 1,511 1,317 1,238

1.17 1.22 1.47 1.53 1.64 1.70

Notes: Assumes portfolio mix of Table 3, with stock returns of 9 percent and bond returns of 4 percent. One-time investment of $1,000 of pretax income with all earnings reinvested. Funds drawn down over the final four years of the investment horizon.


97

FIGURE 1. After-Tax Return to College Savings Options (Nominal Return to $1,000 of Pretax Savings, Household Taxable Income of $100,000) Notes: Assumes the portfolio mix of Table 3, with stock returns of 9 percent and bond returns of 4 percent. One-time investment of $1,000 of pretax income with all earnings reinvested. Funds are drawn down over the final four years of the investment horizon.

The UTMA account shifts assets into the child’s name and, thereby, the child’s lower tax bracket.23 The initial pretax savings are taxed at the parent’s rate, and so $673 is deposited into the UTMA, as was true for the parental account discussed above. For a family with taxable income of $100,000, these tax advantages translate into a substantially higher return on the UTMA account rather than on a parental account. This family yields $1,453 in an UTMA account nearly one-third more than in a parental account. A 529 savings account confers even greater tax advantages than the UTMA because the taxes on the inside buildup and withdrawals are not just reduced; they are eliminated. In a state that does not allow families to deduct 529 deposits from taxable income, $1,000 of pretax income translates into the same $673 deposit that was placed in the parental account and UTMA account. Because no taxes are levied against the inside 23 In an UTMA, annual asset earnings up to $750 are untaxed. For a child younger than 14, the next $750 is taxed at the child’s rate and the remaining earnings are taxed at the parents’ rate. For children 14 and over, all earnings over $750 are taxed at the child’s rate. Note that the tax advantages of the UTMA drop as asset holdings (and earnings) grow because an ever-smaller share of earnings are taxed at a zero rate.

98

Dynarski

buildup, by the time the child enters college, the family has a slightly higher balance in a 529 than it would in a parental account or UTMA ($2,314 as compared to $2,135 and $2,277, respectively). The relative advantage of the 529 grows as the family begins to draw down the funds and is exempted from any taxes on the resulting capital gains realizations. Accounting for these taxes, the family nets a $1,634 return on its $1,000 in pretax savings, 47 percent more than in a parental account and 12 percent more than with an UTMA. The ESA confers the same tax advantages as the 529 without an upfront deduction and therefore yields the same return.24 The return on these two college savings vehicles is also identical to that on the traditional IRA. The traditional IRA is the mirror image of the college savings account because there are no upfront taxes on the $1,000 deposit and no taxes on the inside buildup, but withdrawals are taxed as ordinary income. Note that there is no penalty for early withdrawal (before age 591⁄2) from the traditional IRA if the funds are used for higher education expenses. The traditional IRA therefore yields the same return as the ESA and 529, producing a return 47 percent greater than a nonadvantaged parental account. The option with the highest return is a 529 in a state that allows deposits to be deducted from state taxable income. For a given $1,000 in pretax income, more can be deposited into this account than is true for a nondeductible 529 or ESA. With the typical state tax rate on earned income of 5.95 percent, the initial deposit is $718 rather than the $673. Going forward, the tax treatment is the same as for a standard 529 or ESA. The 529 with an upfront deduction yields a return of $1,811, or 63 percent more than a nonadvantaged account in the parent’s name. As these calculations make clear, the education savings accounts provide new and substantial tax advantages. The 529, with the upfront reduction, offers a higher return than any existing investment option. The 529 and ESA, while yielding the same after-tax return as the traditional IRA, substantially expand the assets that can be shielded from taxation. Finally, because the 529 has no eligibility requirements, it provides the first opportunity for tax-advantaged saving for those families ineligible for the IRA or ESA because of their incomes or their access to a pension program at work.

4.3 Calculation of Returns Net of Income Taxes: All Income Groups In this section, I examine the advantages of the education savings accounts for a range of household incomes, ranging from the lowest 24 A key difference, however, is that much larger amounts can be deposited in a 529 than can be deposited in an ESA.


99

federal tax bracket (household income of $35,000) to the highest (household income of over $335,000). The groups and their associated state and federal tax rates on earned income, capital gains, and interest are shown in Table 2. First, I show how returns vary by income in the benchmark, a non advantaged account held in the name of the parent. In Figure 2 and Table 5, we see that the lowest-income household has the highest absolute returns. This situation is due to this group’s relatively low tax rates on two types of income. First, this group’s lower marginal tax rates on earned income produce a larger deposit for a given $1,000 of pretax income: they start with $773 in principal, compared to $572 for the highestincome family. This difference in the upfront taxation of income accounts for most of the variation across income groups in net returns. Second, the lowest-income household faces the lowest marginal tax rates on capital gains and interest. As a result of these two aspects of the tax code, the highest-income household earns an after-tax return of $728 on its pretax savings of $1,000, while the lowest-income household earns 2.4 times as much, or $1,735. By eliminating some forms of taxation, the tax-advantaged vehicles flatten this income gradient in after-tax returns. Figures 3 and 4 show the after-tax return on the ESA and 529 for each income group. Figure 4 shows the returns in dollar terms, while Figure 3 scales the returns relative to the return in the nonadvantaged account for that income group. Note that

FIGURE 2. After-Tax Return to Non-advantaged Account Held in Name of Parent (Nominal Return to $1,000 of Pretax Savings)

100

Dynarski

FIGURE 3. After-Tax Return to College Saving Options (Relative to Nonadvantaged Account Held in Name of Parent) Notes: Assumes the portfolio mix of Table 3, with stock returns of 9 percent and bond returns of 4 percent. One-time investment of $1,000 of pretax income with all earnings reinvested. Funds are drawn down over the final four years of the investment horizon.

FIGURE 4. After-Tax Return to College Saving Options Notes: Assumes the portfolio mix of Table 3, with stock returns of 9 percent and bond returns of 4 percent. One-time investment of $1,000 of pretax income with all earnings reinvested. Funds are drawn down over the final four years of the investment horizon.


101

because their returns for the investment scenario laid out earlier are identical, I have collapsed the ESA, 529 without an upfront deduction, and the traditional IRA into one category. Recall, however, that the contribution limits are far higher for the 529 than the ESA or IRA, making the 529 particularly advantageous to those who save above the ESA or IRA limits or to those participating in a retirement plan at work and above the associated IRA income limits. Note also that the top two income groups do not qualify for the ESA but do qualify for the 529. The largest increases in returns accrue to the highest income group, both in dollar terms (Figure 4) and relative terms (Figure 3). For those in the top federal tax bracket, the 529 with an upfront deduction delivers a net return almost twice as high as that on a nonadvantaged account. The 529 without an upfront deduction and the ESA net an after-tax return 70 percent higher than funds held in a nonadvantaged account. For those in the lowest bracket, the proportional increases are much lower: the return on a 529 with an upfront deduction is 26 percent. The corresponding figure is 17 percent for the ESA and 529 with no upfront deduction. Note that the UTMA is of almost no benefit for this lowest-income household because the child and parent are in the same low tax bracket. These calculations make clear that both the relative and absolute advantages of the education savings accounts rise steeply with income. At the bottom of the income distribution, where marginal tax rates are the lowest, the new accounts offer after-tax returns 17 to 26 percent higher than returns on a nonadvantaged account. For an initial pretax investment of $1,000, this translates into an additional return of $291 to $453. At the top of the income distribution, the new accounts offer after-tax returns 70 to 91 percent higher than returns on a nonadvantaged account. For an initial pretax investment of $1,000, this translates into an additional return of $511 to $663.

5. THE TREATMENT OF ASSETS BY THE FINANCIAL AID SYSTEM This section turns to the financial aid system. First, I discuss in general terms the aspects of aid determination that affect net returns to savings. Next, I calculate the impact of the aid system’s treatment of assets on returns to various savings vehicles.

5.1 Overview of the Financial Aid Determination Process The federal government distributes need-based aid according to a formula called the federal methodology, which I describe in this

102

Dynarski

section.25 Most schools use the same formula when distributing their own need-based aid. Eighty-seven percent of four-year public schools and 57 percent of four-year private schools use the federal methodology to distribute need-based institutional grants.26 The financial aid determination process I describe here is used for dependent students in academic year 2002–2003.27 Families applying for aid fill out the Free Application for Federal Student Aid (FAFSA), which collects detailed information on family income, assets, and expenses. A new FAFSA, with current data, must be filled out previous to every academic year for which a student wants aid. Financial data from the FAFSA is put through an algorithm that calculates the expected contribution of the family and of the student toward schooling costs. If the sum of the expected contributions from the family and student is less than anticipated schooling costs, the student is eligible for aid. In the calculation of the expected contribution, savings are “taxed” because both assets and asset income are considered resources for paying for college. The resources of the family and the student are calculated separately and are assessed at differing rates in the determination of aid. In the calculation of the family’s contribution, an algorithm sums parental income from all sources. Asset income, in the form of dividends, interest, and capital gains, is included.28 In particular, the earnings portion of withdrawal from some asset accounts is counted as income by the aid formula. After summing income, the aid algorithm subtracts allowable expenses, including taxes, an allowance based on family size, tuition paid for primary and secondary school, and unusually high medical costs. To this net income figure is added 12 percent of certain family assets.29 From the perspective of the financial aid system, assets fall into three categories. A first class of assets, notably home equity, pensions, and other 25 The aid determination process is described in detail by the U.S. Department of Education (2002, 2003) in annual releases of The Student Financial Aid Handbook. The 2002–2003 version used in this paper was downloaded from http://ifap.ed.gov/IFAPWebApp/ currentSFAHandbooksPag.jsp on October 17, 2003. 26 See National Association of Student Financial Aid Administrators and the College Board (2002). 27 In the past, students gamed their dependency status because, for an independent student, family income is not counted in the need determination process. Today, almost all collegeage students are considered dependents because rule changes have made it much more difficult for young people to declare themselves independent. 28 I discuss below changes the idea for which types of assets this source of income is counted and not counted. 29 If a family is not required to file a 1040 and has an adjusted gross income (AGI) below $50,000, then no assets are added at this point.


103

retirement vehicles, is completely sheltered from consideration by the aid formula. These assets are not considered available for college expenses. A second class of assets, including 529 savings accounts and ESAs, is considered fully available for college expenses. The first dollar of these assets is assessed in the determination of aid. A final class consists of any assets that do not fall into these first two categories. These assets are partially sheltered from consideration by an asset protection allowance. Each family is allowed a certain level of savings, based on the age of the oldest parent; the assumption is that older parents need a higher level of savings for their approaching retirement. Below this allowance, assets in this class are assumed to be unavailable for schooling costs. The highest allowance is $70,000; for a family in which the oldest parent is 50, the allowance is $44,000. Above the allowance, 12 percent of assets is added to the net income figure. The resulting weighted sum of income, expenses, and assets is the family’s adjusted available income (AAI). A progressive schedule, with rates ranging from 22 percent to 47 percent, is applied to AAI to determine the expected family contribution (EFC).30 The schedule is quite steep: an AAI of $11,000 is marginally assessed at 22 percent, while the schedule tops out at an AAI of $24,000, which is marginally assessed at 47 percent. In the calculations below, I assume that families are at the top of this schedule. The expected contribution of the student is calculated analogously to the process just described, with fewer protections for income and assets. All student income above $1,750 is assumed to be available for college and is assessed at a rate of 50 percent. There is no asset protection allowance for students. For each year of college, students are expected to contribute 35 percent of their assets.

5.2 The Treatment of Assets and Asset Income in the Aid Determination Process Assets returns are affected twice in the process just described because both asset balances and asset income are considered available for college expenses. First, I will describe the treatment of an asset balance. Consider an entering freshman whose parents have $45,000 in financial assets that are not held in retirement accounts or college savings plans. These assets fall into the third category described above and so are partially sheltered by the asset protection allowance. Assuming the older parent is 50, $44,000 is protected from consideration by the aid formula and $1,000 is subject to assessment. Twelve percent of the $1,000 is added to adjusted 30 Families who are not required to file a 1040 and whose incomes are below $13,000 are automatically assigned an EFC of zero.

104

Dynarski

available income, which is then assessed at 47 percent. Thus, 5.64 (0.12 × 0.47) percent of the $1,000 is considered available for the first year of college. Freshman-year financial aid is reduced by $56.40 as a result of this aspect of the aid formula. If the child goes on for another year of college and applies again for aid, sophomore-year financial aid is again reduced by 5.64 percent of the remaining asset balance. As a result, the total impact of this $1,000 asset on aid received throughout college is a function of the annual assessment on the asset balance (ta, 5.64 percent, in this case) and the number of years spent in college. If a family draws down an equal share of the initial asset balance for each year of college (and for simplicity, we assume no asset earnings once the child enters college), we can summarize the reduction in aid received over the college career as a result of owning a dollar in assets as of the senior year of high school as follows: T

xa =

冱 Tt t

t=1

a

Here, t indexes each year of college for which aid is requested, T is the total number years of college for which aid is requested; a indexes different types of assets. Assume that the high school senior we are considering ends up spending four years in college, drawing down equal increments of the asset per year for expenses. For this family, each dollar of assets held as of the senior year leads to a reduction in aid over the four years of college of 14.1 cents: 4

xa =

冱 4t t = 25 0.0564 = 0.141

t=1

a

or 14.1 percent. Note that this result does not show the full impact of the asset on aid because we have not allowed the asset to grow while the child is in college, which produces income that goes into the aid formula, nor have we considered that some portion of the withdrawals may consist of earnings, which again produces income that goes into the aid formula. I will consider both of these issues below and in the main calculations of the paper. I have just described the financial aid system’s treatment of a parental asset. Some savings are considered assets of the child in the calculation of aid eligibility, which changes the annual assessment rate from 5.64 to 35 percent. The second column of Table 6 shows the annual assessment on asset balances for the savings vehicles we have been considering throughout the paper. Balances in a 529 savings plan are treated as an asset of the parent in the determination of the aid tax on asset balances. Balances in retirement vehicles are ignored by the aid formula. An ESA is considered


105

TABLE 6 Aid System’s Treatment of Saving Alternatives in Aid Determination

Investment option Non-advantaged account, parent Traditional IRA 529 Coverdell UTMA

Annual assessment on asset balance (%)

Annual assessment on earnings net of income tax (%)

5.64

47

0

0

5.64 35 35

0 0 50

Assessment of withdrawals 47% of realized earnings net of income tax 47% of withdrawal net of income tax None 50% of realized earnings 50% of realized earnings net of income tax

by the financial aid system to be owned by the potential student, as is an UTMA or any other asset in the child’s name.31 For such assets, the relevant annual assessment on asset balances is 35 percent rather than 5.64 percent. The first dollar of such assets is assessed at this rate because there is no asset protection allowance for the student. Over four years of college, the assessment on these asset balances amounts to 87.5 percent: 4

xa=

冱 4t t = 25 0.35 = 0.875

t=1

a

As the table and these calculations make clear, the aid tax on asset balances varies widely across savings vehicles. Next, I describe the aid system’s treatment of asset earnings. Like all other sources of income, asset income is considered a financial resource that a family can apply toward college costs, and so increases in asset 31 Department of Education documents for the 2003–2004 school year clearly state that the ESA is to be treated as an asset of the child, which is assessed at the 35 percent rate: “The Education IRA is counted as an asset of the beneficiary,” and “Education IRAs have been appropriately renamed education savings accounts; they are considered an investment asset for the student beneficiary (pp. AVG-20 and AVG-19, respectively, in The Student Financial Aid Handbook 2003–2004). This document can be accessed at http://ifap.ed.gov/ sfahandbooks/attachments/0304AVGMaster.pdf. This same information is contained in the dozens of financial advising documents, news articles, and financial aid resources for parents, schools, and aid professionals that I have consulted. It is also contained in the instructions for completing the online FAFSA. As of November 2003, it appears that the Department of Education is moving to improve the treatment of the ESA documented in this paper. The department has now posted revisions to the online version of the Student Financial Aid Handbook, its reference manual of aid rules. These revisions indicate that, in the future, the ESA will be given the treatment currently applied to the 529 savings plans. This will eliminate the aid tax of over 100 percent that is currently applied to the ESA. See additional discussion of implementation of this policy shift in the introduction to this paper.

106

Dynarski

income lead to decreases in aid. Asset income is considered only if it is realized during a year when income is considered in the determination of aid. Asset income is assessed with a one-year lag because it is based on income reported on the previous year’s 1040. Freshman-year aid, for example, is based on the FAFSA filed when the student was a high school senior. This FAFSA contains tax return data on asset income for the calendar year that spans the spring of the junior year and fall of the senior year of high school. Any earnings received during that period count as income in the determination of financial aid for freshman year. These earnings might take the form of interest, dividends, or capital gains realizations stemming from the sale of stock or liquidation of a mutual fund. Note that in any account that has been building value for 18 years, a substantial portion of the balance consists of unrealized gains. As the account is drawn down for college, these earnings are realized and assessed by the aid formula. In a nonadvantaged account, given the investment scenario assumed throughout the paper, unrealized gains represent about 55 percent of account value. When withdrawals are made to pay for college, 55 percent of each withdrawal is treated as income. Any income taxes paid in a given year offset the income taxed by the financial aid formula. For example, interest earned in a nonadvantaged account is taxed by the state and federal governments. Interest adds to adjusted available income, and taxes paid on the interest subtract from it. The financial aid system therefore assesses asset income net of any income taxes paid on that income. The last two columns of Table 6 show the assessment rate on asset earnings for the different savings vehicles. I show separately the treatment of earnings accruals and of withdrawals. For the 529 and ESA, earnings are ignored by the aid system as they accrue; these earnings do not appear on the FAFSA. The earnings portion of withdrawals from the 529 is also ignored by the aid system. However, the earnings portion of withdrawals from the ESA are assessed at 50 percent.32 For the traditional IRA, earnings are ignored by the aid system as they accrue, and the entirety of any withdrawal is treated as income and is assessed at the parents’ rate of 47 percent. However, any income taxes paid on these withdrawals reduce the amount of income that goes into the aid formula.33 32 33

See page AVG-17 of U.S. Department of Education (2003).

Personal communication with Anthony Jones, U.S. Department of Education. See also Chapter 6 of the 2003–2004 Federal Student Aid Handbook, which contains the worksheets detailing the treatment of various assets and income.


107

5.3 Returns Net of Income Taxes and Aid Reductions Table 7 shows how reductions in aid affect after-tax returns for various savings vehicles. The first column shows the returns on a given savings vehicle for a household that is unaffected by aid policy; we have seen these returns in earlier tables. As discussed earlier, asset returns are unaffected by aid policy for two types of families. The first type is extremely needy (with very low financial resources and/or very high schooling costs) and receiving the maximum aid allowed. 34 The second type of family is not at all needy (with very high financial resources and/or very low schooling costs) and receiving zero aid. For neither family does a marginal change in assets affect aid. The second column of Table 7 shows returns net of reductions in financial aid induced by asset holdings. I assume, as I have throughout the paper, that the account funds are drawn down over the four years of college. These results are not shown for the top two tax brackets, in which I assume household income is sufficiently high (above $150,000) that the child is beyond the margin of eligibility of financial aid at even the most expensive institutions. Columns (3) and (4) express the loss in aid as a percentage of the asset balance at the start of college and as a percentage of the after-tax return, respectively. The impact on returns is enormous, especially for the UTMA and ESA, for which returns are negative once losses in aid are considered. Each of these assets is considered by the financial aid system to belong to the child. As a result, the annual assessment on asset balances held in either of these vehicles is 35 percent rather than the 5.64 percent applied to the other savings vehicles. When we consider only income taxes, an aidmarginal family with taxable income of $50,000 who puts $1,000 pretax in an ESA nets a return of $1,808. This return is 22 percent higher than if the funds were invested in a nonadvantaged account (see Table 4). But once we consider losses of need-based aid, the financial advantage of the ESA disappears. The final return on the $1,000 pretax investment, net of income and aid taxes, is −$1,194. This family loses all principal and all earnings, plus an additional $194, to income taxes and foregone aid. The aid lost due to owning assets in the ESA, expressed as a percentage of the return net of income tax, ranges from 160 percent for the family with $35,000 of income to 172 percent for the family with $100,000 of income. A similar story holds for the UTMA, with the reduction in returns ranging from 178 to 194 percent. 34 As already noted, total aid is capped by a student’s actual schooling costs, which includes tuition and fees plus an allowance for items such as food, rent, and other living expenses.

108

Dynarski

TABLE 7 After-Tax Return to College Savings Alternatives, Net of Financial Aid Losses

After-tax After-tax return, net return of aid loss (1) (2) Non-advantaged account, parent $35K $1,735 $635 $50K $1,485 $490 $100K $1,113 $267 $150K $987 $193 $200K $803 — $335K+ $728 — UTMA $35K $50K $100K $150K $200K $335K+

Return net of aid loss relative to non-advantaged account parent (3)

Aid loss as a percentage of asset Aid loss as balance at a percentstart of age of aftercollege tax return (4) (5)

1.00 1.00 1.00 1.00

43% 43% 40% 39% — —

63% 67% 76% 80% — —

$1,824 $1,618 $1,453 $1,338 $1,157 $1,084

−$1,422 −1,391 −$1,366 −$1,349 — —

−2.24 −2.84 −5.12 −6.98

124% 124% 124% 124% — —

178% 186% 194% 201% — —

529 plan (deduction) $35K $2,188 $50K $1,976 $100K $1,811 $150K $1,683 $200K $1,475 $335K+ $1,391

$1,772 $1,587 $1,444 $1,333 — —

2.79 3.24 5.41 6.90

15% 15% 15% 15% — —

19% 20% 20% 21% — —

529 plan (no deduction) $35K $2,026 $50K $1,808 $100K $1,634 $150K $1,511 $200K $1,317 $335K+ $1,238

$1,631 $1,441 $1,290 $1,183 — —

2.57 2.94 4.84 6.13

15% 15% 15% 15% — —

19% 20% 21% 22% — —

−$1,209 −$1,194 −$1,182 −$1,174 — —

−1.90 −2.44 −4.43 −6.08

122% 122% 122% 122% — —

160% 166% 172% 178% — —

ESA $35K $50K $100K $150K $200K $335K+

$2,026 $1,808 $1,634 $1,511 $1,317 $1,238


109

TABLE 7—Continued

After-tax After-tax return, net return of aid loss (1) (2) Traditional IRA $35K $2,026 $50K $1,808 $100K $1,634 $150K $1,511 $200K $1,317 $335K+ $1,238

$987 $844 $730 $649 — —

Return net of aid loss relative to non-advantaged account parent (3) 1.55 1.72 2.74 3.36

Aid loss as a percentage of asset Aid loss as balance at a percentstart of age of aftercollege tax return (4) (5) 33% 31% 27% 26% — —

51% 53% 55% 57% — —

Because this result is so extraordinary, I will lay out in detail the losses in aid associated with holding funds in an ESA. Table 8 shows the calculation in detail for the ESA, for a 529 without an upfront deduction, and for an UTMA. As an example, I examine a family with $100,000 in income.35 In January of year 18, when the child is a high school senior, this family files a financial aid form. At this time the ESA account, which has been gaining value since it was established at year 0 with an after-tax contribution of $673, contains $2,314; this balance is shown in column (1) of Table 8. Thirty-five percent of this balance, or $810, is considered available for college costs; this is shown in column (2). At the end of the year, when the child is a freshman in college, the family draws down $609, onequarter of the end-of-year balance [column (4)]. Seventy-one percent of this amount ($432) consists of earnings, which is considered income of the child in the calculation of aid and is assessed at 50 percent. Aid is therefore reduced by $216 because of this withdrawal. The remaining rows repeat these calculations for the three subsequent years until the account is emptied. Asset balances are assessed four times, once for each FAFSA that is filed. Withdrawals are assessed just three times because income is recorded on the FAFSA with a one-year lag, and so the final, senior-year withdrawal does not appear on a filed FAFSA. The total reduction in aid is $2,816, while the balance at the start of year 18 was $2,314. The ratio of these two numbers is the 122 percent shown in Table 7. 35 Family income does not affect the aid process depicted in Table 8, but it is necessary to choose an income tax bracket to pin down the dollar amounts shown therein.

110

Dynarski TABLE 8 Detailed Aid Calculations: Loss in Aid for Family Holding Various Assets Coverdell ESA Loss of aid due to asset balance, Balance, 35% of start of start-of-year year balance (1) (2)

Balance, end of year (3)

Loss of aid due to asset income, 50% of after-tax Withdrawal, realized end of year earnings (4) (5)

Year 18 $2,314 $ 810 $2,435 Year 19 $1,826 $ 639 $1,922 Year 20 $1,282 $ 449 $1,349 Year 21 $ 674 $ 236 $ 710 Total $2,134 As share of year 18 starting balance: 122%

$609 $641 $674 $710

Total loss of aid (6)

$216 $227 $239 $682

$2,816

529, no upfront deduction

Balance, start of year (1)

Loss of aid due to asset balance, 5.64% of start-of-year balance (2)


Loss of aid Withdrawal, due to asset end of year income, none (4) (5)

Year 18 $2,314 $130 $2,435 Year 19 $1,826 $103 $1,922 Year 20 $1,282 $ 72 $1,349 Year 21 $ 674 $ 38 $ 710 Total $343 As share of year 18 starting balance: 15%

$609 $641 $674 $710


$— $— $— $—

$343

UTMA Loss of aid due to asset balance, Balance, 35% of start of start-of-year year balance (1) (2) Year 18 Year 19

$2,277 $1,794

$797 $628

Balance, end of year (3) $2,392 $1,884

Loss of aid due to asset income, 50% of after-tax Withdrawal, realized end of year earnings (4) (5) $598 $628


$226 $238 Continued


111

UTMA Loss of aid due to asset balance, Balance, 35% of start of start-of-year year balance (1) (2)


Year 20 $1,256 $ 440 $1,320 Year 21 $ 660 $ 231 $ 693 Total $2,096 As share of year 18 starting balance: 123%

Loss of aid due to asset income, 50% of after-tax Withdrawal, realized end of year earnings (4) (5) $660 $693


$241 $705

$2,801

Notes: Unrealized earnings account for 68–71% of withdrawals. Family has household income of $100,000.

The 529 savings plans are not as hard hit by the aid tax because the financial aid system considers this asset as belonging to the parent rather than to the child. The aid tax on net of income tax returns for the 529 is 57 to 63 percent, lower than that on a nonadvantaged account in the name of the parent (63 percent to 81 percent). Once we consider aid taxes, the 529, with or without an upfront deduction, nets higher returns than the nonadvantaged account, the UTMA, or the ESA. In the case of the nonadvantaged account, the 529 performs better because its inside buildup is not taxed by the financial aid system. In the case of the UTMA and ESA, the 529 performs better because the tax on the asset balance is 5.64 percent rather than 35 percent.

6. DISCUSSION The intent of the savings incentives is to increase saving by increasing net returns. The intent of the financial aid system is to give less aid to those with higher income and assets. These two sets of policies inevitably work at cross-purposes because the aid system attempts to tax away any increase in assets and income that the savings incentives create. Unless assets and asset income are completely disregarded, asset returns for aid-marginal families are reduced by the aid determination process. Given this constraint, we can make the aid tax as non-arbitrary as possible. Here, I discuss the results of the paper’s analysis in the context of this goal.

112

Dynarski

6.1 Asset “Taxes” Greater Than 100 Percent It is difficult to infer any reasonable policy goal that is consistent with the aid system’s current treatment of the Coverdell, the UTMA, and all assets held in the name of the student. Families that put funds in these vehicles lose all their assets to income taxes and aid reductions; that is, these vehicles face income taxes and “aid taxes” that sum to well over 100 percent. The paper’s simulations show that a middle-income family who puts $1,000 into a Coverdell loses all of the principal and earnings, plus an additional $194, to income and aid taxes. A family that puts funds into the name of the student in an UTMA is even worse off, losing principal and earnings plus an additional $391. Any asset held in the name of the child faces similar treatment. Fully taxing away principal and earnings—a tax of 100 percent—is consistent with a very strict, narrow formulation of need: at the time of college attendance, it puts a saving family in the same position vis-à-vis the aid system as a nonsaving family.36 However, taxing away more than principal and earnings is certainly not consistent with this strict formulation of need because it places the saving family in a worse-off position than the nonsaving family—by thousands of dollars if they save at the rate recommended by financial counselors.

6.2 Sharply Differing Tax Rates on Parents’ and Students’ Assets The differing treatment of assets held by the parent and the student has a large impact on aid received and net returns, as shown in Figure 5. This operates counter to the aid system’s goal of treating equally families with equal resources because two families with the same asset levels face vastly divergent tax rates depending on whose name is on the account. A middle-income family (income of $50,000) who puts funds in the child’s name in an UTMA yields a small income tax advantage—a 9 percent increase in the lifetime return (not annualized return), as shown in Table 4. For a family who deposits $1,000 of pretax income in an account and leaves it to accrue for 18 years, this translates into a savings of $133. However, the associated loss in aid more than erases this small gain from gaming the income tax system. Once we consider both the income tax and losses in aid, this family loses $1,881 by having the funds in an UTMA rather than in the parent’s name [netting a return of −$1,391 versus $490, see column (2) of Table 7]. They also end up with far less than they would 36 The saving family has forgone consumption to save, and so it is worse off in a lifetime sense than if it had not saved at all when principal and earnings are fully taxed away. Edlin (1993) discusses this and other aspects of the equity of the aid tax.


113

FIGURE 5. Return to College Saving Options, Net of Aid Lost and Income Tax Notes: Assumes those in the bottom four brackets are on the aid margin. The assumed aid tax is zero in the top two brackets.

have had they not saved at all, having lost their principal, their earnings, and an additional $391. The first dollar of funds held in the child’s name results in aid reductions, while assets held in the parents’ name are protected by an asset allowance. As a result, the average aid tax rate on the parents’ assets is well below that on children’s assets.

6.3 Sharply Differing Tax Rates on Different Savings Vehicles As Table 7 makes clear, the impact of aid policy on asset returns varies wildly, depending on the savings vehicle. A dollar in assets held by the family of a high school senior produces, over a four-year college career, a reduction in need-based aid of $0.15 if the funds are held in a 529 savings plan, $0.26 to $0.39 if the funds are held in an IRA, about $0.40 if they are held in a typical mutual fund account in the parent’s name, $1.22 if they are held in a Coverdell ESA, and $1.24 if they are held in an UTMA. We can express these aid reductions relative to after-tax returns on these various savings vehicles. The reduction in aid caused by holding funds in a given vehicle ranges from 19 percent of after-tax returns for the 529 to 200 percent for the UTMA.

114

Dynarski

As discussed above, in the context of asset ownership by parents and children, such wildly varying treatment of assets does not advance the goals of the need-based aid system. It induces an efficiency loss because it encourages shifting of assets toward those vehicles that are treated preferentially by the aid system. And it induces a loss in equity because it imposes significant losses on those who do not know how to game the system. This leads to a loss of horizontal equity because the aid system treats unequally those who have the same asset levels but have made differing strategic choices about where to put the funds.

6.4 Policy Alternatives Two key points emerge from this discussion. First, some assets are treated extremely punitively by the aid system, resulting in those who save losing more than one dollar in aid for each dollar they hold in assets. A second and distinct point is that the treatment of assets is highly inconsistent. I have already explained the efficiency and equity losses induced by these aspects of the aid system. Here I lay out and critique several policy options that address these two points. There are two main sources of the variable treatment of assets by the aid system: the differential treatment of parents’ and children’s assets and the differential treatment of different asset types—e.g., retirement accounts, home equity, college savings plans, and nonadvantaged accounts. The differential treatment of parents’ and children’s assets accounts for most of the variance in the treatment of assets; for example, because the Coverdell is defined as an asset of the child, it faces an annual assessment rate on asset balances of 35 percent rather than the 5.64 percent imposed on the 529 savings plans, which are defined as an asset of the parent. Were the Coverdell instead defined as an asset of the parent, $1.00 held in a Coverdell would lead to a reduction in aid of $0.15, the same as the 529 savings plan, rather than $1.22. The fact that the first dollar of children’s assets is assessed while many parental assets are protected by an allowance that goes as high as $70,000 also contributes to the divergence in the treatment of parents’ and children’s assets.37 Pooling parents’ and children’s assets in aid determination will therefore go far in reducing the variability in the treatment of assets. All children’s assets, and not just those of the student applicant, would be included in this pool. Applying what is currently the treatment of parents’ assets to children’s assets is the simplest solution, but this approach would obviously lead to a higher level of aid expenditures 37 The paper’s calculations use marginal tax rates rather, so than average rates, so they ignore this aspect of the divergence in the treatment of parents’ and children’s assets.


115

because children’s assets would be assessed at a much lower rate. A revenue-neutral alternative would be a new assessment rate on the pooled assets that is the asset-weighted average of the current assessment rates. Note that this pooling of assets would bring children’s assets under what is currently the parents’ asset protection allowance. The second issue is how different types of assets are treated. Currently, the value of retirement assets and home equity are completely excluded from aid determination.38 All other assets contribute to the net worth considered available for college costs. An alternative is to pool all assets— regular accounts, Coverdells, 529s, UTMAs, retirement assets, home equity—and tax them uniformly in the aid determination process. Under such a system, the aid tax rate on assets would be the same across savings vehicles. Unifying assets in this way would reduce the deadweight loss caused by families shifting assets to avoid the aid tax. It would also eliminate several sources of horizontal inequity. For example, homeowners in areas with high real estate values (the East and West coasts) have greater opportunity to shield assets from the aid system than do renters or those in areas with lower real estate values (the middle of the country). Similarly, those who work in jobs that provide access to a 401(k) plan have a greater ability to shield assets than do other workers. If other aspects of the aid determination process were unchanged, the main effect of pooling all types of assets would be to decrease aid because it would add retirement assets and home equity to the net worth considered available for paying for college. To maintain the current level of aid spending, the assessment rate on all assets could be reduced below its current maximum of 5.64 percent. Alternatively, the asset protection allowances could be increased so that the total net worth assessed by the aid system remains unchanged.

7. Conclusion This paper has examined the income tax code’s most recent experimentation with education policy, in the form of the Coverdell Education Savings Account and the 529 savings plans. Tax incentives for college saving were designed to increase savings by increasing after-tax returns. From the narrow perspective of the income tax code, they have succeeded in increasing after-tax returns. But if we broaden our perspective to include the interaction of the new tax incentives with existing educational policy—in the form of the financial aid system—these policies fail. Families who save for college are potentially subject to taxation not only by federal and 38

After-tax withdrawals from retirement funds are treated as available income, however.

116

Dynarski

state taxing authorities but also by the federal, state, and college financial aid systems. As I have shown, the aid tax on savings can extend well up the income distribution because fairly well-off families can qualify for aid at expensive private institutions. For families caught in the cross fire between aid policy and tax policy, the impact on the bottom line is not pretty. A family that heeds advice to save for college in one of the new college savings vehicles can find itself far worse off than if it had simply placed funds in a non-advantaged account in the parents’ name. Further, those who put funds in a Coverdell can find themselves worse off than if they had not saved at all. These perverse outcomes indicate that greater attention to the interaction of aid and taxes is required if the tax code is to succeed as an instrument for education policy.

References Cerulli Associates (2003). “A Competitive Outlook for 529 Savings Plans.” The Cerulli Report. Boston, MA: Cerulli Associates, Inc. Dynarski, Susan (2002). “Loans, Liquidity and Schooling Decisions.” Kennedy School of Government. Working Paper. Edlin, Aaron (1993). “Is College Financial Aid Equitable?” Journal of Economic Perspectives 7(2): 143–158. Feenberg, Daniel Richard and Elisabeth Coutts (1993). “An Introduction to the TAXSIM Model.” Journal of Policy Analysis and Management 12(1): 189–194. Feldstein, Martin (1995). “Scholarship Rules and Private Savings.” American Economic Review 85(3): 552–566. National Association of Student Financial Aid Administrators and The College Board (2002). “Financial Aid Professionals at work in 1999–2000, Results from the 2001 Survey of Undergraduate Financial Aid Policies, Practices, and Procedures.” Washington, DC: NASFAA. National Bureau of Economic Research (2004). “TAXSIM Related Files at the NBER,” www.nber.org/~taxsim (accessed March 8, 2004). National Center for Education Statistics (2002). National Postsecondary Student Aid Survey 2000. Washington, DC. U.S. Department of Education (2002). Student Financial Aid Handbook 2002–2003. Washington, DC: U.S. Government Printing Office. U.S. Department of Education (2003). Student Financial Aid Handbook 2003–2004. Washington, DC: U.S. Government Printing Office.

REPORTED INCOMES AND MARGINAL TAX RATES, 1960–2000: EVIDENCE AND POLICY IMPLICATIONS Emmanuel Saez University of California and NBER

EXECUTIVE SUMMARY This paper uses income tax return data from 1960 to 2000 to analyze the link between reported incomes and marginal tax rates. Only the top 1 percent of income earners show evidence of behavioral responses to taxation. The data display striking heterogeneity in the size of responses to tax changes over time, with no response either short-term or long-term for the very large Kennedy top income tax cuts in the early 1960s, and striking evidence of responses, at least in the short term, to the tax changes since the 1980s. The 1980s tax cuts generated a surge in business income reported by high-income individual taxpayers, due to a shift away from the corporate sector, and the disappearance of business losses for tax avoidance. The Tax Reform Act of 1986 and the recent 1993 tax increase generated large short-term responses of wages and salaries reported by top income earners most likely because of retiming in compensation to take advantage of the tax changes. It is unlikely, however, that the extraordinary trend upward of the shares of total wages accruing to top wage income earners, which started in the 1970s and accelerated in the 1980s and especially the late 1990s, can be explained solely by the evolution of marginal tax rates. I am very grateful to Dan Feenberg for his help using the micro tax return data and the TAXSIM calculator. I thank Alan Auerbach, Gerald Auten, Robert Carroll, Dan Feenberg, Martin Feldstein, Wojciech Kopczuk, Andrew Leigh, Thomas Piketty, James Poterba, and especially Joel Slemrod for very helpful comments and discussions. Financial support from the Sloan Foundation and NSF Grant SES-0134946 is gratefully acknowledged.

118

Saez

1. INTRODUCTION Over the last 40 years, the U.S. federal income tax system has undergone large changes. Perhaps the most striking change has been the dramatic decrease in top marginal income tax rates. From 1950 to the early 1960s, the statutory top marginal income tax rate was 91 percent. This rate was reduced to 70 percent by the Kennedy tax cuts in the mid-1960s. During the Reagan administrations of the 1980s, the top income tax rate was further reduced to 50 percent in 1982 by the Economic and Recovery Tax Act (ERTA) of 1981, and was reduced again to 28 percent in 1988 by the Tax Reform Act (TRA) of 1986. The top income tax rate was then increased to 31 percent in 1991 and further to 39.6 percent in 1993 by the Omnibus Budget Reconciliation Act (OBRA) of 1993. The top rate has been reduced to 35 percent in 2003 by the 2001 tax reform. Only about 500 taxpayers were subject to the top marginal tax rate of 91 percent in the early 1960s, but by 2000, more than half a million taxpayers were subject to the top rate.1 Thus, the continuous and drastic progressivity of the federal income tax system up to the very highest income taxpayers has been replaced by a much flatter tax structure, where an upper-middle-class family can face the same marginal tax rate as the highest-income earners in the United States. In addition to the redistributive effects, the dramatic reductions in top income tax rates might have generated large behavioral responses: the net-of-tax value of an additional dollar of pretax income (excluding state and local taxes) for those in the highest income bracket has experienced enormous variations over the period, from less than $0.10 in the early 1960s to more than $0.70 by the late 1980s and around $0.60 by 2000. It is plausible to think that such variations might have had substantial effects on the economic activity of high-income earners, such as labor supply decisions, career choices, and savings decisions, as well as on the form of compensation (salary versus untaxed fringe benefits, for example). Indeed, the intellectual weight behind the dramatic reduction in marginal income tax rates in the 1980s was the logic of supply-side economics, which argued that lower tax rates would generate important increases in economic activity and perhaps even tax revenues. As documented by Feenberg and Poterba (1993, 2000) and Piketty and Saez (2003), there has indeed been an extraordinary increase in the share of total income accruing to upper-income groups in the income distribution over the last 25 years. For example, the income share of the top 1 percent of taxpayers 1 The statistics on the number of taxpayers in each tax bracket have been reported regularly since 1961 in the Internal Revenue Service (IRS) annual publication Statistics of Income.

Reported Incomes and Marginal Tax Rates, 1960–2000

119

(excluding capital gains from the analysis) has surged from less than 8 percent in the early 1970s to almost 17 percent in 2000 (Piketty and Saez, 2003). Feenberg and Poterba (1993) pointed out that the timing of the increase in top income shares, and most notably the surge in top income from 1986 to 1988 around TRA of 1986, appears to be closely related to the cuts in top income tax rates. Slemrod and Bakija (2001) and Piketty and Saez (2003) note, however, that the surge in top incomes accelerated in the late 1990s, although top income tax rates increased substantially in 1993. The goal of this paper is to understand the effects of marginal income tax rates on reported incomes by analyzing the shares and composition of incomes accruing to various groups in the top tail of the income distribution, and the marginal income tax rates faced by those groups. The analysis will focus on the 1960–2000 period because it spans all the important tax changes since World War II.2 This same period allows me to use the large and stratified public-use tax return microfiles released by the IRS since 1960, as well as the TAXSIM tax calculator created and maintained by the National Bureau of Economic Research (NBER) to estimate marginal and average tax rates.3 Many researchers have tried to estimate the effects of taxes on decisions such as those involving the labor supply, savings, and retirement. Over the past decade, researchers have pointed out that these standard behavioral responses are only components of what drives reported incomes; other responses (such as the form of compensation, tax-deductible activities, unmeasured effort, and compliance) also ultimately determine reported incomes, and these responses may be more elastic with respect to taxation. Feldstein (1999) shows that, under certain conditions, the overall elasticity of taxable income with respect to the net-of-tax rate (1 minus the marginal tax rate) is relevant for assessing the implications of tax changes for revenue raising and welfare. The influential studies of Lindsey (1987) and Feldstein (1995), which examined the 1980s tax cuts, estimated very large elasticities, in excess of 1. This striking conclusion has generated a substantial body of work on this central elasticity parameter and generated a wide range of estimated elasticities, ranging from Feldstein’s (1995) and Lindsey’s (1987) separate estimates at the high end to close to zero at the low end, depending on the estimation methodology and the tax reforms considered.4

2 There are few studies on behavioral responses to taxation in the United States in the prewar era. Goolsbee (1999) provides a simple analysis of the most important episodes. 3

See Feenberg and Coutts (1993) for a description of the TAXSIM calculator.

4

See Gruber and Saez (2002) for a survey.

120

Saez

It is important to note that, in contrast to most previous studies, my analysis focuses on reported incomes before deductions, such as adjustments to gross income, personal exemptions, and standard and itemized deductions. Therefore, my income concept is market income rather than taxable income. Because taxable income is a smaller base than gross income, and because some components of deductions such as charitable giving or mortgage interest deductions are also responsive to marginal tax rates, the elasticities of taxable income are likely to be larger than the elasticities of reported incomes that I analyze here.5 My analysis shows that only the reported incomes of taxpayers within the top 1 percent of the income distribution appear to be responsive to changes in tax rates over the 1960–2000 period. Even upper-middleincome taxpayers (within the top decile but below the top 1 percent), who experienced substantial changes in marginal tax rates, show no evidence of responses to taxation, either in the short-run or the long-run. Attributing all the gains of the top 1 percent relative to the average to the changes in tax rates produces large elasticities of income with respect to net-of-tax rates, in excess of 1. However, allowing for simple secular and non-tax-related time trends in the top income share reduces the elasticity drastically (to about 0.5). Top income shares within the top 1 percent show striking evidence of large and immediate responses to the tax cuts of the 1980s, and the size of those responses is largest for the topmost income groups. In contrast, top incomes display no evidence of short- or longterm response to the extremely large changes in the net-of-tax rates following the Kennedy tax cuts in the early 1960s. Data on the composition of income show that part of the response to the 1980s tax cuts has been due to a sudden and permanent shift of corporate income toward the individual income sector using partnerships and Subchapter S corporations, legal entities taxed only at the individual level. However, most of the surge in top incomes since the 1970s has been due to a smooth and extraordinary increase in the wages and salary component (which includes stock-option exercises). This wage income surge started slowly in the early 1970s and has accelerated over the period, and especially during the last decade, and does not seem to be closely related to the timing of the tax cuts. There is evidence of short-term responses of the wage income component around TRA 1986 and OBRA 1993: top wage 5 Gruber and Saez (2002) indeed find larger elasticities for taxable income than for adjusted gross income. Here, I focus on gross income because the nature and size of deductions has changed considerably over time so that, in contrast to gross income, it is not possible to construct consistent time series of taxable income. A large part of the literature has analyzed the response of the main components of itemized deductions such as charitable contributions and interest deductions.


121

income shares spike just after the tax reduction of 1986 and just before the tax increase of 1993, suggesting that highly paid employees were able to retime their compensation to take advantage of the tax changes. It is difficult, however, to tell apart a long-term effect of tax cuts from a non-taxrelated secular widening of the disparity of earnings. The paper is organized as follows. Section 2 describes the key identification issues in estimating behavioral elasticities of income with respect to marginal tax rates and shows how such elasticity estimates can be used for tax policy analysis. Section 3 presents the results on income shares and marginal tax rates, as well as the evolution of the composition of top incomes. Section 4 concludes by contrasting the U.S. experience with evidence from other countries.

2. CONCEPTUAL FRAMEWORK AND METHODOLOGY 2.1 Estimating Elasticities The economic model underlying the estimation of behavioral responses to income taxation is a simple extension of the static labor supply model. Individuals maximize a utility function u(c, z) increasing in after-tax income c (available, for example, for consumption) and decreasing in beforetax income z (earning income is costly, for example). The budget constraint takes the form c = (1 − τ) z + R, where τ is the marginal tax rate and R is virtual income. Such maximization generates an individual “reported income” function of z(1 − τ, R) which depends on the net-of-tax rate 1 − τ and virtual income R.6 Each individual has a particular income supply function reflecting his or her skills, taste for labor, etc. Income effects are ignored, so the income function z is independent of R and depends only on the net-of-tax rate.7 The key point is that, in contrast to the standard labor supply model, changes in hours of work isn’t the only factor that can affect earnings z; intensity of work on the job, career choices, form of compensation, tax-deductible activities, etc., can also affect earnings. The analysis below will show that it is indeed the full response of reported incomes that is relevant for tax policy (a point made by Feldstein, 1999). The literature on behavioral responses to taxation has attempted to use tax reforms to identify the elasticity of reported incomes with respect to 6 This reported income supply function remains valid in the case of nonlinear tax schedules; c = (1 − τ) z + R then represents the linearized budget constraint at the utility maximizing point. 7 Labor supply studies in general estimate modest income effects. See Blundell and Macurdy (1999) for a survey. Gruber and Saez (2002) try to estimate both income and substitution effects in the case of reported incomes and find small and insignificant income effects.

122

Saez

the net-of-tax rate defined as e = [(1 − τ)/z] ∂z/∂(1 − τ) in the notation used above. To isolate the effects of the net-of-tax rate, one compares observed reported incomes after the tax rate change to the incomes that would have been reported had the tax change not taken place. Obviously, the latter are not observed and must be estimated. The simplest method consists in using as proxy reported incomes before the reform, and hence in relating changes in reported incomes before and after the reform to changes in tax rates. Lindsey (1987) and Feldstein (1995) applied this methodology to the ERTA 1981 and TRA 1986 tax changes and found that top income groups, which experienced the largest marginal tax cuts, also experienced the largest gains in reported incomes. As a result, Lindsey (1987) and Feldstein (1995) obtain extremely large elasticities, between 1 and 3, with preferred estimates around 1.5. Several important issues surround those estimates. First, as pointed out by Slemrod (1996, 1998) and Goolsbee (2000a), these elasticities are upward biased if, for non-tax-related reasons, top incomes increased more rapidly than average incomes during that period. A large body of work has suggested that nontax factors, such as skillbiased technical progress, the development of international trade, or the decline of unions, might have led to a substantial increase in earnings disparity in the 1980s [see Katz and Autor (1999) for a survey]. To overcome this issue, it would be preferable to compare taxpayers with similar incomes rather than comparing high incomes to middle incomes. In the case of income taxation, this approach is difficult for two reasons. First, for most reforms, taxpayers with similar incomes face very similar tax changes.8 Second, although the discontinuity in marginal tax rates due to the progressive bracket structure creates sharp changes in marginal incentives for taxpayers with very similar incomes, this situation cannot be satisfactorily exploited to estimate elasticities because it appears that taxpayers either control their incomes imperfectly or are not well aware of the details of the tax code and their precise location on the tax schedule.9,10 Therefore, it is conceivable that only large or salient tax changes are likely to generate behavioral responses, which raises some interesting and 8 In contrast, for redistributive programs (such as the Earned Income Tax Credit, which is targeted to taxpayers with children) taxpayers with no children but similar income can be used as a plausibly better control group for identifying the effects of the program (see, for example, Eissa and Liebman, 1996). 9 In an earlier study (Saez, 2003), I tried to exploit this feature and the bracket creep from 1979 to 1981 to identify behavioral responses. 10 In an earlier study (Saez, 2002), I documented in detail the fact that bunching, as predicted by theory, does not occur at the kink points of the tax schedule.


123

complicated issues about the estimation of behavioral responses and the design of tax policy [see Liebman and Zeckhauser (2003) for an analysis along these lines]. Second, comparing years just before and just after the reform might reveal a short-term elasticity, which can be quite different from the longterm elasticity, the relevant parameter for tax policy. Slemrod (1995) discusses this point, and Goolsbee (2000b) shows convincingly that executives exercised numerous stock options in 1992 to avoid the higher tax rate starting in 1993, which created a large short-term elasticity of reported income around OBRA 1993; the longer-term elasticity was much smaller and possibly equal to zero.11 Looking at times series spanning several years before and after the reform, as in Feenberg and Poterba (1993), can be helpful for making progress on these two issues. Slemrod (1996) proposes an aggregate time-series regression framework, for the period 1954 to 1990, to try and disentangle tax and nontax influences on the share and composition of income accruing to the top 0.5 percent taxpayers. Third, the Lindsey (1987) and Feldstein (1995) studies assume implicitly that reported income elasticities are the same for all income groups and, as we will see, the data strongly suggest that those taxpayers with very high incomes are much more responsive to changes in taxation than taxpayers in the middle or upper-middle class. More precisely, instead of adopting the simple difference method just described, they compare changes in the incomes of the very high incomes (experiencing the largest tax rate changes), to changes in incomes of the middle and upper-middle class (experiencing more modest tax changes). This difference-in-differences of (log) incomes is then divided by the corresponding difference-in-differences of (log) net-of-tax rates to obtain an elasticity estimate of the following form: et =

∆ log (z H ) - ∆ log (z M ) ∆ log (1 - x H ) - ∆ log (1 - x M )

where zH, zM and τ H, τ M denote the incomes and marginal tax rates of the high (H) and middle (M) income groups, respectively, and ∆ denotes the changes from before to after the tax change. But suppose that the middle class has a zero elasticity, so that ∆ log(zM) = 0, and that high-income individuals have an elasticity of e, so that ∆ log(zH) = e∆log(1 − τ H). Assume further that the middle class experiences an increase in its net-of-tax rates that is half as large as that experienced by the high-income taxpayers, so 11 Feldstein and Feenberg (1996) note a decrease in top reported incomes from 1992 to 1993 and interpret this finding as evidence of large behavioral elasticities. As compensation of executives continued to soar throughout the late 1990s, negative long-run elasticity estimates would be obtained by repeating Goolsbee’s (2000a) analysis and comparing incomes in 1992 to those of the late 1990s.

124

Saez

that ∆ log(1 − τ M) = 0.5 $ ∆ log(1 − τ H). Then the estimated elasticity ê will be twice the true elasticity e of the high-income group, a dramatic upward bias in the estimate. This simple but realistic example shows that it is not appropriate to rely on comparisons of the responsiveness of the reported incomes of the middle- and upper-income groups when there is a strong suspicion that the behavioral elasticities for the two groups are quite different. Fourth, the increases in top incomes following the 1980s tax changes might have been due partly to income shifting rather than the creation of new income. As I show below, the critical distinction for policy and welfare analysis is whether the increase in reported incomes comes at the expense of untaxed activities (for example, leisure, fringe benefits, and perquisites) or taxed activities (for example, profits in the corporate sector, future capital gains, and deferred compensation such as pensions). Slemrod (1996) points out that part of the surge in top incomes following TRA 1986 was due to a dramatic increase in S-corporation income, suggesting that many businessowners switched the legal form of their corporations from Subchapter C (which faces the corporate income tax on profits) toward Subchapter S (which does not face the corporate tax and whose profits are taxed directly at the individual level) because the top individual income tax rate became lower than the corporate income tax rate by 1988.12 Carroll and Joulfaian (1997) explore this issue in more detail using a panel of corporations from 1985 to 1990, and they confirm Slemrod’s (1996) earlier findings. Gordon and Slemrod (2000) perform a systematic study of income shifting by analyzing simultaneously tax changes and reported incomes at the corporate and personal level. In this paper, I analyze in detail the composition of reported individual incomes to cast light on the source of the changes in reported incomes following tax reforms. The early studies by Lindsey (1987) and Feenberg and Poterba (1993) used the large and stratified annual cross-sectional public-use tax return data to document the evolution of top reported incomes. Following Feldstein’s (1995) influential analysis of the TRA 1986, several studies have used panel data to estimate elasticities. The main justification for 12 A C-corporation faces the corporate tax on its profits. Profits are then taxed again at the individual level if they are paid out as dividends. If profits are retained in the corporation, they may generate capital gains that are taxed at the individual level, but in general would be taxed more favorably than dividends, when they are realized. Profits from S-corporations (or partnerships and sole proprietorships) are taxed directly and solely at the individual level. Distributions from S-corporations to individual owners generate no additional tax. Thus, an S-corporation is fiscally more advantageous than a C-corporation the lower the individual tax rate, the higher the corporate tax rate, and the higher the capital gains tax rate. See Scholes and Wolfson (1992, Chapter 4) for extensive details and examples. A business can switch to and from the C and S status, but an S-corporation cannot have more than a limited number of stockholders (75 currently), issue more than one class of stock, or be a subsidiary of other corporations.


125

using panel data instead of repeated cross-sections was that they might alleviate the issue of non-tax-related changes in income inequality because the same individuals are followed before and after the reform. It is plausible to think, however, that an increase in income inequality might be due mostly to high-income individuals experiencing larger gains than do lower-income individuals; in which case, a panel analysis does not solve the issue. Furthermore, a tax cut might induce middle-income people to try harder to become rich, and this behavioral response will be missed by a Feldstein-type panel data analysis. The use of panel data has two additional important drawbacks. First, the publicly available panel of tax returns is not stratified and hence does not allow nearly as precise a study of the evolution of top incomes as does the large, stratified cross-sections.13 Second, comparing groups ranked according to pre-reform incomes generates a mean reversion problem: if there is mobility in incomes from year to year, then it can cause highincome taxpayers in one year to appear in low-income brackets in the next, aside from any true behavioral response.14 Eliminating this mobility bias requires control of pre-reform income in the estimation, but this approach will weaken and possibly destroy identification because the size of net-of-tax-rates changes is closely correlated with income.15 Many authors, including Lindsey (1987) himself, have argued that comparing income groups using repeated cross-sections is a valid strategy only if taxpayers stay in the same groups from year to year. Following a tax rate cut such as ERTA 1981 or TRA 1986, however, one would like to know how the distribution of reported income has changed relative to a scenario where the tax change does not take place. Whether there is mobility in incomes from year to year is independent of this question as long as the income distribution is stationary (without the tax change). In contrast, mobility in incomes is precisely what complicates the panel data analysis. Panel data have key advantages, however, for studying some questions more subtle than the overall response of reported incomes. For example, if one wants to study how a tax change affects income mobility

13 Auten and Carroll (1999) have used a larger panel available only at the U.S. Treasury to compare years 1985 and 1989. It is difficult, however, to create longer panels to analyze longer-term time series because of attrition issues. 14 This would generate a downward bias in the elasticity estimates in the case of a tax rate decrease, such as TRA 1986, and an upward bias in the case of a tax rate increase, such as OBRA 1993. 15 This point is discussed in Gruber and Saez (2002), who overcome this problem by using many years instead of just two in the analysis. The implicit assumption they make, however, is that mobility remains stable from year to year.

126

Saez

(i.e., do more middle-income taxpayers become successful entrepreneurs following a tax rate cut?), panel data is clearly necessary. Measuring the tax-induced change in the income distribution is exactly what is needed to derive the tax revenue consequences of the tax change. Because we do not observe the counterfactual income distribution when no tax change takes place, we have to rely on income distributions from previous years, and there is no systematic bias in the repeated crosssection analysis as long as the income distribution remains stationary, without the tax change. The direct focus on the income distribution series over time allows a much more concrete and simple grasp of the evolution of incomes for different groups than does panel analysis because it is straightforward to divide the population into various percentiles for each year and to analyze simultaneously the evolution of the incomes and the marginal tax rates of these groups. By relating the changes in incomes to the changes in net-of-tax rates, we can obtain elasticity estimates. Finally, Slemrod (1998) and Slemrod and Kopczuk (2002) make the important point that the elasticity of reported incomes with respect to tax rates might not be a fixed parameter, and it depends on the legal details and the enforcement of the tax system. For example, if it is easy for corporations to switch from Subchapter C to Subchapter S to avoid taxes, the individual tax base might be much more elastic than in a setting where Subchapter S corporations do not exist. Kopczuk (2003) performs an empirical analysis of this issue for the United States from 1979 to 1990 and shows that taxable income elasticities are negatively related to the base of incomes subject to taxes. This result suggests that introducing additional deductions increases the responsiveness of taxable incomes. Goolsbee (1999) studies the key tax changes in the United States since the 1920s and finds enormous heterogeneity in the observed responses from episode to episode, although he does not try to explain the discrepancies. The present analysis of the period 1960–2000 also displays significant heterogeneity in responses over time.

2.2 Using Elasticities for Tax Policy The empirical analysis that follows will show that evidence of behavioral responses to changes in marginal tax rates is concentrated in the top of the income distribution, with little evidence of any response for the middleincome and upper-middle-income class.16 Therefore, it is useful to focus 16 The low end of the income distribution is beyond the scope of this paper because many low-income families and individuals do not file income tax returns. The large amount of research on responses to welfare and income transfer programs targeted toward low-income earners has displayed evidence, however, of significant labor supply responses. See Meyer and Rosenbaum (2001), for example, for a recent analysis.


127

on the analysis of the effects of increasing the marginal tax rate on the upper end of the income distribution. Therefore, let us assume that incomes in the top bracket, above a given threshold zr , face a constant marginal tax rate τ.17 N is the number of taxpayers in the top bracket. Assume that incomes reported in the top bracket depend on the net-oftax rate 1 − τ, and z (1 − τ) denotes the average income reported by taxpayers in the top income bracket. As discussed above, income effects in the analysis are ignored, and thus the net-of-tax rate is the only relevant parameter. The elasticity (compensated or uncompensated because there are no income effects) of income in the top bracket with respect to the netof-tax rate is therefore defined as e = [(1 − τ)/z]∂z/∂(1 − τ). Suppose that the government increases the top income tax rate τ by a small amount dτ (with no change in the tax schedule for incomes below zr ). This small tax reform has two effects on tax revenue. First, there is a mechanical increase in tax revenue because taxpayers face a higher tax rate on their incomes above zr . Hence, the total mechanical effect is: dM = N [z - zr ] dx This mechanical effect is the projected increase in tax revenue, without any behavioral response. Second, the increase in the tax rate triggers a behavioral response that reduces the average reported income in the top bracket by dz = −e $ z $ dτ / (1 − τ) on average, and hence it produces a loss in tax revenue equal to: dB =- N $ e $ z $

x dx 1-x

Summing the mechanical and the behavioral effect, I obtain the total change in tax revenue due to the tax change: dR = dM + dB = Ndx (z - zr ) $ ; 1 - e $

z $ x E z - zr 1 - x

Let us use a to denote the ratio z/(z − zr ). Note that a ≥ 1 and that a = 1 when zr = 0, that is, when there is a single flat tax rate applying to all incomes. If the top tail of the distribution is Pareto distributed, then the parameter a does not vary with zr and is exactly equal to the Pareto parameter.18 Because the tails of actual income distributions are closely approximated by Pareto distributions, it turns out that the coefficient a is 17 In the case of the 2003 tax law, for example, taxable incomes above zr = $311,950 are taxed at the top marginal tax rate of r = 35 percent. 18 A Pareto distribution has a density function of the form f(z) = C/z1 + α, where C and α are constant parameters; α is called the Pareto parameter.

128

Saez

extremely stable for zr above $200,000. Saez (2001) provides such an empirical analysis for 1992 and 1993 incomes using tax return data. The parameter a measures the thinness of the top tail of the income distribution: the thicker the tail of the distribution, the larger z is relative to zr , and hence the smaller is a. Feenberg and Poterba (1993) provide estimates of the Pareto parameter a from 1951 to 1990 for the distribution of adjusted gross income (AGI) in the United States using income tax returns. They show that a has decreased from about 2.5 in the early 1970s to around 1.5 in the late 1980s.19 We can rewrite the effect of the small reform on tax revenue dR simply as: dR = dM ; 1 -

x $e$a E 1-x

(1)

Equation (1) is of central importance. It shows that the fraction of tax revenue lost through behavioral responses—the second term in the square bracket expression—is a simple function increasing in the tax rate τ, the elasticity e, and the Pareto parameter a. This expression is also equal to the marginal deadweight burden created by the increase in the tax rate. More precisely, because of the envelope theorem, the behavioral response creates no additional welfare loss because individuals are maximizing utility, and thus the utility loss (in dollar terms) created by the tax increase is exactly equal to the mechanical effect dM. However, tax revenue collected is only dR = dM + dB, with dB < 0. Thus, −dB represents indeed the extra amount lost in utility over and above the tax revenue collected, dR. The marginal excess burden expressed in terms of extra taxes collected is simply: e$a$x - dB = dR 1 - x - e $ a $ x

(2)

These formulas are valid for any tax rate τ and income distribution, even if individuals have heterogeneous utility functions and behavioral elasticities, as long as income effects are assumed away.20 Thus, this formula should be preferred to the Harberger triangle approximations, which require small tax rates to be valid. The parameters τ and a are straightforward to obtain; the elasticity parameter e is thus the central nontrivial parameter necessary to make use of equations (1) and (2). For example, in 2000, for the top 1 percent income cutoff (corresponding 19 Piketty and Saez (2003) provide estimates of thresholds zr and average incomes z corresponding to various fractiles within the top decile of the U.S. income distribution from 1913 to 2000. This approach allows a straightforward estimation of the parameter a for any year and income threshold. 20

The elasticity e is the average (income weighted) of individual elasticities.


129

approximately to the top 39.6 percent federal income tax bracket in that year), Piketty and Saez (2003) estimate that a = 1.6. For an elasticity estimate e = 0.5, corresponding to the mid- to upper range of the estimates from the literature, the fraction of tax revenue lost through behavioral responses (dB/dM), should the top tax rate be increased slightly, would be 52.5 percent, more than half of the mechanical projected increase in tax revenue. In terms of marginal excess burden, increasing tax revenue by $1 requires the creation of a utility loss of 1/(1 − .525) = $2.11 for taxpayers, and hence a marginal excess burden of $1.11, or 111 percent of the extra $1 tax collected. Following the supply-side debates of the early 1980s, much attention has been focused on the tax rate which maximizes tax revenue, the so-called Laffer rate. The Laffer rate τ* maximizes tax revenue; hence, the bracketed expression in equation (1) is exactly zero when τ = τ*. Rearranging the equation, we obtain the following simple formula for the Laffer tax rate τ* for the top bracket: x* =

1 1+a$e

(3)

A top tax rate above the Laffer rate is an inefficient situation because decreasing the tax rate would increase both government revenue and the utility of high-income taxpayers.21 At the Laffer rate, the excess burden becomes infinite because raising more tax revenue becomes impossible. Using our previous example with e = 0.5 and a = 1.6, the Laffer rate τ* would be 55.6 percent, not much higher than the combined maximum federal, state, Medicare, and sales tax rate. Note that when zr = 0 and the tax system has a single tax rate, the Laffer rate becomes the well-known expression τ* = 1/(1 + e). Because a ≥ 1, the flat rate maximizing tax revenue is always larger than the Laffer rate for high incomes only. Increasing the top tax rate collects extra taxes only on the portion of incomes above the bracket threshold zr but produces a behavioral response for high income taxpayers as large as an across-the-board increase in marginal tax rates. The analysis has assumed so far that the reduction in incomes due to the tax rate increase has no other effect on tax revenue. This assumption 21 When the government has strong redistributive tastes and does not value the marginal consumption of high-income individuals relative to the average individual, the optimal income tax rate for high-income individuals is exactly equal to the Laffer rate in equation (3). When the government generally values the marginal consumption of high-income individuals at 0 ≤ g < 1, the optimal tax rate for the high-income individuals is such that the bracketed expression in equation (1) is equal to g. See my earlier work (Saez, 2001) for a more detailed exposition following the classical optimal income tax theory of Mirrlees (1971).

130

Saez

is reasonable if the reduction in incomes is due to reduced labor supply (and hence an increase in untaxed leisure time) or to a shift from cash compensation toward untaxed fringe benefits or perquisites (more generous health insurance, better offices, company cars, etc.). In many instances, however, the reduction in reported incomes is due in part to a shift away from individual income toward other forms of taxable income such as corporate income, or deferred compensation, that will be taxable to the individual when paid out (see Slemrod, 1998). For example, Slemrod (1996) and Gordon and Slemrod (2000) show convincingly that part of the surge in top incomes after the Tax Reform Act of 1986 was due to a shift of income from the corporate sector toward the individual sector. I will cover this topic in detail later. Therefore, let us assume that the incomes that disappear from the individual income tax base following the tax rate increase dτ are shifted to other bases taxed at rate t on average. For example, if two-thirds of the reduction in individual reported incomes is due to increased leisure and one-third is due to a shift toward the corporate sector, t would be onethird of the corporate tax rate because leisure is untaxed. In that case, it is straightforward to show that equation (1) becomes: dR = dM ; 1 - x - t $ e $ a E 1-x

(4)

The same envelope theorem logic applies for welfare analysis, and the marginal deadweight burden formula is also modified accordingly by replacing e $ a $ τ by e $ a $ (τ − t) in both the numerator and denominator of equation (2). The Laffer rate in equation (3) becomes: x* = 1 + t $ a $ e 1+a$e

(5)

If we assume again that a = 1.6 and e = .5, but that incomes disappearing from the individual base are taxed at t = 20 percent on average, the fraction of revenue lost due to behavioral responses drops from 52.5 to 26 percent, and the marginal excess burden (expressed as a percentage of extra taxes raised) decreases from 111 to 35 percent if the initial top tax rate is τ = 39.6 percent. The Laffer rate increases from 55.6 to 64.5 percent. This simple theoretical analysis shows therefore that, in addition to estimating the elasticity e, it is critical to analyze the source or destination of changes in reported individual incomes.

2.3 Data and Methodology I estimate the level and shares of total income accruing to various upperincome groups using the large cross-sectional individual tax return data


131

annually released by the Internal Revenue Service (IRS) since 1960.22 The data are a stratified sample of tax returns oversampled for high-income taxpayers, which allows an extremely precise analysis of top reported incomes. The top income shares are estimated based on the Piketty and Saez (2003) analysis.23 The unit of analysis is the tax unit defined as a married couple living together (with dependents) or a single adult (with dependents), as in the current tax law. It is important to note that top income shares series measured at the tax unit level, as I do here, might be different from series estimated at the individual level. As displayed in Table 1, since 1960, the average number of individuals per tax unit has decreased from 2.6 to 2.1 because of the decrease in the average number of dependent children per tax unit as well as the decrease in the fraction of married tax units. Those long-term demographic changes imply that real average income growth per tax unit will be substantially smaller than real income growth per capita. These demographic changes can also affect top income shares if the reduction in tax unit size is not uniform across income groups. However, the tax return data show that the reduction in tax unit size has been about the same for high-income taxpayers as it has for the U.S. population as a whole. From 1960 to 2000, the number of individuals per tax unit in the top decile has declined from 3.6 to 2.9, which is the same 20 percent decline as in the general population (from 2.6 to 2.1). From 1960 to 2000, the fraction of married tax units has declined from about 60 to 50 percent for the total population (due to the increased number of single parents and unmarried couples) but only from 90 to 85 percent for the top decile tax units. An increase in single tax units with lower incomes contributes to increasing top income shares. Similarly, an increase in the correlation of earnings between spouses (due, for example, to the increased labor force participation of married women) would also increase top income shares estimated at the tax unit level. Those slow moving demographic changes are small, however, relative to the dramatic trends I document and can explain at best only a small fraction of the changes in the top most income shares. Each upper-income group is defined relative to the total number of potential tax units in the entire U.S. population, estimated from population and family census data as the sum of married men, divorced and widowed men and women, and single adults never married (age 20 and 22 23

There is no micro data for years 1961, 1963, and 1965.

The main (and very minor) difference is that government transfers such as social security benefits and unemployment compensation have been excluded from the income definition in this paper to obtain better consistency in the income definition over the years. The estimates have been extended to year 2000.

1960 1961 1962 1963 1964 1965 1966 1967 1968 1969 1970 1971 1972 1973 1974 1975 1976 1977 1978 1979

Tax Units (000s) (1)

Number of tax returns (000s) (2)

68,681 69,997 71,254 72,464 73,660 74,772 75,831 76,856 77,826 78,793 79,924 81,849 83,670 85,442 87,228 89,127 91,048 93,076 95,213 97,457

61,028 61,499 62,712 63,943 65,376 67,596 70,160 71,652 73,729 75,834 74,280 74,576 77,573 80,693 83,340 82,229 84,670 86,635 89,771 92,694

Total Income

Wage earners and wage income

Total Income Average Average (millions Income marginal (2)/(t) Population 2000 $ (2000 $, tax rate (%) (%) (000s) (4)/(1) CPI-U) CPI-U) (9) (3) (10) (4) (5) (8) 88.9 87.9 88.0 88.2 88.8 90.4 92.5 93.2 94.7 96.2 92.9 91.1 92.7 94.4 95.5 92.3 93.0 93.1 94.3 95.1

180,671 183,691 186,538 189,242 191,889 194,303 196,560 198,712 200,706 202,677 205,052 207,661 209,896 211,909 213,854 215,973 218,035 220,239 222,585 225,055

2.63 2.62 2.62 2.61 2.61 2.60 2.59 2.59 2.58 2.57 2.57 2.54 2.51 2.48 2.45 2.42 2.39 2.37 2.34 2.31

1,850,218 1,907,985 2,011,233 2,099,285 2,236,911 2,361,753 2,500,162 2,600,178 2,719,064 2,794,675 2,845,542 2,905,636 3,093,721 3,225,502 3,195,330 3,093,548 3,235,043 3,339,935 3,480,248 3,503,689

26,939 27,258 28,226 28,970 30,368 31,586 32,970 33,832 34,938 35,469 35,603 35,500 36,975 37,751 36,632 34,709 35,531 35,884 36,552 35,951

22.55 23.32 21.64 21.30 21.62 24.33 25.53 24.11 23.06 23.62 24.77 25.82 25.40 26.04 27.71 29.16 29.19

Tax units with wages (11)

Total wages (millions 2000 $, CPI-U) (12)

52,554 51,946 53,338 53,893 55,216 57,239 60,358 61,571 62,836 64,371 63,778 63,194 64,750 67,614 68,518 66,671 68,459 70,898 74,503 77,038

1,587,214 1,615,622 1,705,361 1,772,347 1,877,056 1,987,572 2,125,707 2,213,824 2,337,364 2,435,448 2,447,144 2,484,179 2,630,468 2,748,251 2,697,802 2,609,012 2,722,938 2,825,066 2,961,075 2,979,812

Average Average wages marginal (2000 $, tax rate CPI-U) (%) (13) (14) 30,201 31,102 31,972 32,886 33,995 34,724 35,219 35,955 37,198 37,834 38,370 39,311 40,625 40,646 39,373 39,132 39,775 39,847 39,745 38,680

22.68 23.35 21.66 21.19 21.59 24.10 25.15 24.20 23.29 23.73 24.68 25.61 25.91 26.53 28.33 29.87 30.04

Inflation indexes

CPI-U (2000 base) (15)

CPIU-RS (2000 base) (16)

17.189 17.361 17.552 17.762 17.993 18.299 18.830 19.376 20.190 21.280 22.535 23.527 24.280 25.785 28.621 31.226 33.037 35.185 37.859 42.137

20.183 20.385 20.609 20.856 21.127 21.486 22.110 22.751 23.662 24.693 25.882 27.031 27.864 29.608 32.541 35.236 37.257 39.635 41.340 45.224

Saez

Tax Units and Population

132

TABLE 1 Reference Totals for Population, Income, and Inflation, 1960-2000

99,625 101,432 103,250 105,067 106,871 108,736 110,684 112,640 114,656 116,759 119,055 120,453 121,944 123,378 124,716 126,023 127,625 129,301 130,945 132,267 133,589

93,902 95,396 95,337 96,321 99,439 101,660 103,045 106,996 109,708 112,136 113,717 114,730 113,605 114,602 115,943 118,218 120,351 122,422 124,771 127,075 129,272

94.3 94.0 92.3 91.7 93.0 93.5 93.1 95.0 95.7 96.0 95.5 95.2 93.2 92.9 93.0 93.8 94.3 94.7 95.3 96.1 96.8

227,726 229,966 232,188 234,307 236,348 238,466 240,651 242,804 245,021 247,342 250,132 253,493 256,894 260,255 263,436 266,557 269,667 272,912 276,115 279,295 282,339

2.29 2.27 2.25 2.23 2.21 2.19 2.17 2.16 2.14 2.12 2.10 2.10 2.11 2.11 2.11 2.12 2.11 2.11 2.11 2.11 2.11

3,412,006 3,419,549 3,405,788 3,466,971 3,637,968 3,760,935 3,876,141 4,046,941 4,305,720 4,350,842 4,377,181 4,286,889 4,356,547 4,320,595 4,424,217 4,581,375 4,730,336 4,974,958 5,268,063 5,522,779 5,705,414

34,248 33,713 32,986 32,998 34,041 34,588 35,020 35,928 37,553 37,263 36,766 35,590 35,726 35,019 35,474 36,353 37,064 38,476 40,231 41,755 42,709

30.66 31.68 29.22 27.36 26.99 27.27 27.26 24.47 22.92 23.06 23.05 23.11 22.99 23.94 24.29 24.58 24.75 25.33 25.56 25.84 26.13

76,913 77,439 75,771 76,260 80,008 81,936 83,340 85,618 88,121 90,145 91,348 89,813 89,883 91,279 93,270 95,388 97,338 100,161 103,069 105,233 107,693

2,880,118 2,876,292 2,844,255 2,913,254 3,075,930 3,193,778 3,321,487 3,442,337 3,572,571 3,609,277 3,632,403 3,574,052 3,645,188 3,687,902 3,783,593 3,891,745 3,986,011 4,170,993 4,429,422 4,626,416 4,836,329

37,446 37,143 37,537 38,202 38,445 38,979 39,855 40,206 40,542 40,039 39,764 39,794 40,555 40,402 40,566 40,799 40,950 41,643 42,975 43,963 44,909

31.77 32.95 30.71 28.90 28.36 28.59 28.77 25.98 24.75 24.65 24.77 24.61 24.91 25.60 25.82 26.29 26.65 27.32 27.79 28.39 28.99

47.825 50.258 52.751 54.974 56.022 58.185 57.814 60.602 60.300 63.020 62.471 65.161 63.658 66.310 65.950 68.569 68.654 71.066 71.949 74.158 75.834 77.883 79.019 80.737 81.390 82.878 83.832 85.018 86.011 86.881 88.419 89.061 91.072 91.478 93.167 93.460 94.657 94.768 96.740 96.750 100.000 100.000

133

Notes: Population and tax unit estimates based on census and current population surveys (Historical Statistics of the United States, and Statistical Abstract of the United States). Tax units estimated as sum of married men, divorced and widowed men and women, and singles men and women aged 20 and over. Income defined as adjusted gross income less realized capital gains, taxable social security and unemployment insurance benefits and adding back all adjustments to gross Income. Income of nonfilers is imputed as 20 percent of average income. Marginal tax rates are weighted by income and estimated using TAXSIM calculator (see Feenberg and Coutts, 1993) and the tax return micro-files and ignoring interactions with state income taxes. Marginal income tax rate in column (10) is a weighted average of marginal tax rate on earned income and other income. Tax units with wages defined as the total number of employees (from National Income and Product Accounts) less number of married women employed (from Statistical Abstract of the United States). Total wages from total compensation of employees from National Income and Product Accounts. Marginal income tax rate in column (14) is the average (wage income weighted) marginal tax rate on wages and salaries. Consumer Price Index (CPI-U) is the official CPI index from Economic Report of the President. CPI-U-RS includes retrospectively improvements on CPI estimation method for the 1967–1998 period.


1980 1981 1982 1983 1984 1985 1986 1987 1988 1989 1990 1991 1992 1993 1994 1995 1996 1997 1998 1999 2000

134

Saez

above).24 The income definition I use is consistent over time and includes all income items except realized capital gains reported on tax returns and before all deductions such as adjustments to gross income, exemptions, and itemized and standard deductions.25 I exclude government transfers such as social security (SS) benefits and unemployment insurance (UI) benefits. Thus, my income measure is defined as adjusted gross income (AGI) less realized capital gains included in AGI, less taxable SS and UI benefits, plus all the adjustments to gross income. Hence, my measure of income is a broader measure than taxable income, on which many previous studies have focused. If deductions to income, such as charitable giving, mortgage interest payments, etc., are also responsive to taxation, taxable income might be more responsive to tax rates than my broader income measure. Because the nature of deductions allowed has changed substantially over the period 1960–2000, however, it is impossible to construct a consistent taxable income definition over the full period. As a result, refer to previous studies analyzing specifically the components of taxable income that I exclude from the analysis. As in Piketty and Saez (2003), I consider various groups within the top decile of the income distribution. To get a more concrete sense of those upper-income groups, Table 2 displays the thresholds, the average income level in each group, and the number of tax units in each group, all for 2000. The median income as well as the average income for the bottom 90 percent of tax units, are quite low, around $25,000. Those numbers are smaller than those reported by the Census Bureau based on the Current Population Survey (CPS) for two reasons. First, my income definition does not include any government transfers. Second, CPS income is reported at the household level, which is a larger unit than the tax unit I consider.26 The groups in the top decile below the top 1 percent (the top 10–5 percent denotes the bottom half of the top decile, and the top 5–1 percent denotes the next four percentiles) have average incomes of $100,000 and $160,000, respectively, which corresponds to the popular view of the middleincome and upper-middle-income class (perhaps surprisingly given how 24 From 1960 to 2000, between 90 and 95 percent of potential tax units actually filed an income tax return because many nontaxable families file to get tax refunds. 25 Realized capital gains are excluded because they form a volatile component of income and face in general a different tax treatment than do other forms of income. Much of the literature focuses on the response of capital gains realizations to tax changes. See Auerbach (1988) for a survey. 26 For example, a cohabiting couple or two roommates form a single household but are two separate taxpayers.


135

TABLE 2 Thresholds and average incomes in top income groups in 2000 Percentile threshold (1) Median Top 10% Top 5% Top 1% Top .5% Top .1% Top .01%

Income threshold (2) $25,076 $87,334 $120,212 $277,983 $397,949 $1,134,849 $5,349,795

Income groups (3)

Number of tax units (4)

Full population Bottom 90%

133,589,000 120,230,100

$42,709 $26,616

6,679,450 5,343,560 667,945 534,356 120,230 13,359

$100,480 $162,366 $327,970 $611,848 $2,047,801 $13,055,242

Top 10–5% Top 5–1% Top 1–0.5% Top 0.5–0.1% Top 0.1–0.01% Top 0.01%

Average income in each group (5)

Notes: Computations are based on income tax return statistics. Income is defined as annual gross income reported on tax returns excluding capital gains and all government transfers (such as social security, unemployment benefits, welfare payments, etc.) and before individual income taxes and employees’ payroll taxes. Amounts are expressed in 2000 dollars. Column (2) reports the income thresholds corresponding to each of the percentiles in column (1). For example, an annual income of at least $87,334 is required to belong to the top 10 percent tax units, etc.

far up the income distribution those groups are). In 2000, an annual family income of at least $280,000 is required to be part of the top 1 percent. Hence, the top 1 percent corresponds perhaps to the popular view of the high-income tax payers. About 140,000 tax units (or slightly more than 0.1 percent of all tax units) report incomes larger than $1 million (the highincome taxpayers). Finally, the top .01 percent, the smallest top group I consider, is formed by the top 13,400 tax units, who reported, on average, $13 million of annual income in 2000. These are the super-high-income American families. I estimate shares of income by dividing the income amounts accruing to each group by reported income, and I have assumed that nonfiling units earn 20 percent of the average income.27 I then estimate the composition of income for each group and consider seven components: salaries and wages (including exercised stock options, bonuses, and private pensions), S-corporation income, sole proprietorship (Schedule C income) and farm income, partnership income, dividends, interest income, and other income (including smaller items such as rents, royalties, and other miscellaneous items). Marginal tax rates are estimated using the TAXSIM tax calculator. For each individual record, I compute a weighted marginal tax rate based on wage income and other income because various provisions in the tax code 27 Because only between 5 and 10 percent of tax units do not file returns, my results are not sensitive to this assumption.

136

Saez

generate differences in the tax treatment of wage income and other forms of income. For each income group, I then estimate an average marginal tax rate weighted by income.28 Note that my marginal tax rate computations ignore state income taxes because the data does not provide state information for high-income earners. My tax measure also ignores other taxes such as social security and Medicare taxes, corporate taxes, and nonincome taxes such as sales and excise taxes. I use the same methodology to compute top wage shares using wages and salaries reported on tax returns. Wages and salaries include exercised stock options and bonuses. In this case, groups are defined relative to the total number of tax units, with positive wage income estimated as the number of part-time and full-time workers from the National Income and Product Accounts less the number of married women who are employees. The sum of total wages in the economy used to compute shares is obtained from the National Income and Product Accounts (total compensation of employees). The marginal tax rates for upper-wage-income groups are, of course, those relevant for wages and salaries and are also weighted by wage income (see Table 1). I propose a simple time-series regression methodology to obtain various elasticity estimates, and illustrate some of the identification difficulties. Because of potential heterogeneity in elasticities across income groups, all regressions are run for a single income group. The simplest specification consists in regressing log real incomes on log net-of-tax rates (and a constant) for a given group. Of course, as real incomes grow over time, time trends can be added in the regression to control for exogenous (i.e., non-tax-related) real income growth. These estimates are unbiased estimates of behavioral elasticities if, absent any tax change, real incomes in that specific group do not change (first specification) or follow a regular time pattern (second specification). These assumptions may not be met. Because many years of data are included, these estimates capture mostly the long-term behavioral elasticities.29 As we will see, the pattern of average incomes for the full population does not appear to be related to the evolution of average marginal tax rates. Therefore, to control for average income growth, most of the regressions are run in terms of log income shares instead of log average incomes.30 These regressions control 28 As we saw above, for tax policy analysis, it is necessary to weight marginal tax rates by income. 29 I leave for future research the regression analysis of the dynamics of tax responses. Such a formal analysis has been attempted in the case of capital gains realizations. See, for example, Auerbach (1988). 30 Slemrod (1996) adopted the same approach, although he controlled for nontax factors explicitly rather than using general time trends controls, as I do here.


137

automatically for overall income growth. Adding time trends in that case amounts to assuming that incomes for the particular group considered may diverge from the average income in the economy. Because timeseries regressions are run and the error terms appear to be correlated over time (according to the standard Durbin-Watson test), Ordinary Least Squares (OLS) standard errors are not correct. Therefore, the Newey-West standard errors are computed, assuming that the error terms can be correlated up to an eight-year lag.31 Because of the progressive structure of the income tax, increases in incomes lead to higher marginal tax rates, or bracket creep. As a result, an increase in top income shares (for non-tax-related reasons) might also induce a mechanical increase in the marginal tax rate faced by those high-income taxpayers, hence potentially biasing downward the elasticity estimates. A simple way to investigate the extent of the problem is to use the statutory top marginal income tax rate (or more precisely, the log of 1 minus the top rate) as an instrument for the effective log net-of-tax-rate variable. The results show that the OLS and Instrumental Variables (IV) estimates are extremely close, suggesting that progressive structure of the income tax system and bracket creep do not create a significant estimation problem.

3. INCOME SHARES AND MARGINAL TAX RATES 3.1 Trends in Average Incomes Figure 1 shows the average federal marginal individual income tax rate (weighted by income) and the average income (per tax unit) reported in real terms for the full population from 1960 to 2000. Incomes are expressed in 2000 dollars using the standard Consumer Price Index–All Urban Consumers (CPI-U) deflator (see Table 1). Figure 1 also shows that real incomes increased quickly from 1960 to 1973 and then increased hardly at all until the early 1990s. From 1993 to 2000, real incomes have increased quickly but are only 13 percent higher than in 1973. Real growth depends critically on the Consumer Price Index (CPI) deflator. Improvements in the CPI estimation have been made over the years, and some of them have been incorporated retrospectively in the so-called Consumer Price Index Research Series using current methods (CPI-U-RS) deflator (see Stewart and Reed, 1999). Using the CPI-U-RS instead of the CPI-U would display about 29 percent real income growth instead of 13 percent from 1973 to 2000 (see Table 1). 31 An eight-year lag is close to maximizing the size of the standard errors and thus should be seen as conservative.

138

Saez

FIGURE 1. Average Real Income, Marginal and Average Tax Rate, All Tax Units, 1960–2000 Note: Based on Table 1.

Average marginal tax rates display significant movements, with a steady increase from 21–22 to 30 percent from the mid-1960s to the early 1980s (with a temporary surge during the Vietnam War surtaxes from 1968 to 1970). In the 1980s, the average marginal tax rate decreased to 23 percent, and it increased slightly to 26 percent during the 1990s. Figure 1 displays no clear relationship between the level of real incomes and the level of marginal tax rates. As displayed in panel A of Table 3, a simple OLS regression of log average incomes on the log of the net-of-tax rate, always displays insignificant elasticity coefficients. Therefore, the aggregate data display no evidence of significant behavioral responses of reported incomes relative to changes in the average marginal tax rate. Figure 2 shows a striking contrast between the bottom 99 percent tax units (panel A) and the top 1 percent (panel B). The average real income of the bottom 99 percent increased steadily from 1960 to 1973 and then stagnated; real incomes in 2000 are hardly higher than in 1973.32 The decline in marginal tax rates faced by the bottom 99 percent, from almost 32 If one uses the CPI-U-RS deflator, the bottom 99 percent of real incomes would have grown by about 13 percent. In any case, it is clear that real growth of incomes has been slow in the last quarter of the twentieth century relative to the 1950–1973 period. It is also important to note that this slow growth is not due to a decrease in the number of adults per tax units (see Table 1).


139

TABLE 3 Elasticities of income with respect to net-of-tax rates in the aggregate, bottom 99%, and top 1% Regression in levels (1) Panel A: all tax units Elasticity

−0.02 (0.38) Yes

0.20 (0.55) Yes Yes

−0.66 (0.70)

−0.41 (0.37) Yes

−0.04 (0.38) Yes Yes

1.83 (0.37)

0.71 (0.22) Yes

0.50 (0.18) Yes Yes

Time trend Time trend square Panel C: top 1% tax units Elasticity Time trend Time trend square

Regression in levels + time controls (3)

−0.44 (0.84)

Time trend Time trend square Panel B: bottom 99% tax units Elasticity

Regression in levels + time control (2)

Notes: Estimates obtained by time-series regression of log(average real income) (using CPI-U deflator) on a constant, log(1 − average marginal tax rate) from 1960 to 2000 (38 observations). In column 1, simple OLS regression is run, standard errors from Newey-West with 8 lags. In column 2, a time trend is added. In column 3, time ^2 trend is added.

30 percent in 1981 to around 23 percent in 2000, does not seem to have noticeably improved the growth of real incomes. Indeed, as shown in panel B of Table 3, regressing the log average incomes on the log net-oftax rate for the bottom 99 percent displays negative (although insignificant) coefficients whether or not a time trend is included. In stark contrast, the average real income of the top 1 percent has increased by 160 percent since the early 1970s (or by 200 percent if one uses the CPI-U-RS), and the average marginal tax rate has also declined substantially, from around 50 percent before 1981 to less than 30 percent by 1988. It is striking to note that the top 1 percent incomes start increasing precisely in 1981, when marginal tax rates start going down. The jump in top incomes from 1986 to 1988 corresponds exactly to the sharp drop in marginal tax rates, from 45 to 29 percent, after the Tax Reform Act of 1986. These points, first noted by Feenberg and Poterba (1993), suggest that high-income taxpayers are indeed quite responsive to taxation. The other striking feature of the figure is the extraordinary increase in top incomes

140

Saez

FIGURE 2. Marginal Tax Rates and Average Real Incomes for the Bottom 99% and the Top 1% Note: Series Based on from Tables 1 and 4.


141

from 1994–2000, in spite of the increase in tax rates, from about 32 percent to almost 40 percent in 1993. Thus, although the marginal tax rates faced by high-income taxpayers in 2000 are hardly lower than in the mid-1980s (39 percent instead of 44–45 percent), top incomes are more than twice as large. Figure 2 illustrates clearly the difficulty of obtaining convincing estimates of the elasticity of reported income with respect to the net-of-tax rate. It seems obvious that the sharp, and unprecedented, increase in incomes from 1986 to 1988 is related to the large decrease in marginal tax rates that happened exactly during those years. The central issue, however, is whether this short-term response persists over time. In particular, how should we interpret the continuing rise in top incomes since 1994? If one thinks that this surge is evidence of diverging trends between highincome taxpayers and the rest of the population independent of tax policy, which started in the 1970s, then it is tempting to consider the response to TRA 1986 as a purely short-term spike followed by lower growth from 1988 to 1993, before getting back to the normal upward trend by 1994. On the other hand, one could argue that the surge in top incomes since the mid-1990s might have been the long-term consequence of the decrease in tax rates in the 1980s and that such a surge would not have occurred had tax rates for high-income taxpayers remained as high as they did in the 1960s and 1970s. I will return to this point later. These issues are illustrated formally in the regression results in panel C of Table 3. When no time trend is included in the regression of log income on log net-of-tax rate, all the growth in top incomes is attributed to the decline in top rates, and the elasticity obtained is extremely large 1.83 (.37). In contrast, including a time trend produces a much smaller, although still sizable, elasticity of .71 (.22) because part of the rise in top incomes is attributed to a secular rise. Adding an additional time square control further reduces the elasticity to 0.5 (0.18). This analysis also shows that comparing two single years by taking the ratio of the difference in log incomes to the difference in log net-of tax rates, as is done in most studies, can produce a wide range of elasticity estimates. Comparing 1981 to 1984, as in Lindsey (1987), produces an elasticity of 0.77.33 Comparing 1985 and 1988, as in Feldstein (1995) and Auten and Carroll (1999), produces an extremely large 1.7 elasticity.34 In contrast, 33 Lindsey (1987) obtains larger estimates because he compares the upper-income to the middle-income groups, creating an upward bias if, as is apparent in the data, elasticities are increasing with income (see discussion in section 2.1). 34 Auten and Carroll (1999) obtain a much smaller 0.6 elasticity because they compare 1985 to 1989 (instead of 1988, as did Feldstein [1995]) and because of the mean reversion issue discussed in Section 2.1, which is difficult to correct with only two years of data.

142

Saez

comparing 1991 to 1994 (as in Goolsbee, 2000b) produces a zero elasticity because top incomes are about constant, while tax rates increase by almost 10 percentage points.35 The elasticity would even become negative if one compares 1991 to the late 1990s because both top incomes and the tax rate have increased.36 The large micro data sets can be used to obtain these simple elasticity estimates directly from regressions at the individual level, as is done in many studies, with small standard errors. The regression counterpart would be to pool the samples of top 1 percent earners for the pre- and postreform years and run a Two Stage Least Squares (2SLS) regression of log incomes on the log net-of-tax rate using as an instrument a postyear dummy.37 To cast additional light on these issues and try to separate tax effects from other effects, I turn to a closer analysis of various upper-income groups, with particular emphasis on the change in the composition of reported incomes.

3.2 Trends in Top Income Shares and Marginal Tax Rates Average real incomes do not seem to respond to average marginal tax rates in the aggregate, and responses seem to be concentrated in the upper 1 percent of the income distribution. From now on, therefore, top incomes are normalized by considering the shares of total income accruing to various upper-income groups (as in Feenberg and Poterba, 1993, 2000, and Piketty and Saez, 2003). This approach has two advantages. First, the income share measures are independent of the CPI deflator used. Second, the top shares are normalized automatically for overall real and nominal growth in incomes. All the top income share series and corresponding average marginal tax rates (income weighted) are reported in Tables 4 and 5, respectively. Table 6 displays several regressions of the (log) top 1 percent income share on the log net-of-tax rate, varying the number of time trend controls and instrumenting or not the tax variable with the log net-of-tax top rate. As discussed above, introducing time trends reduces substantially the elasticity, from 1.6 (with no controls) to about 0.6–0.7 (with many controls). After adding linear and square controls in time, the adjusted 35 In contrast, comparing 1992 to 1993 would produce a significant short-term elasticity of 0.63, as in Feldstein and Feenberg (1996). 36 Carroll (1998) and Sammartino and Wiener (1997) analyze panel tax return data. They also show that short-term responses around OBRA 1992 are much larger than longer-term responses. 37 It is doubtful, however, that these small standard errors would be accurate because random year effects are most likely to be present in the data, making 2SLS standard errors far too low and hence worthless (in addition to creating the identification problems discussed in section 2.1). See Bertrand, Duflo, and Mullainathan (2003) for a detailed discussion of these econometric issues.

TABLE 4 Top Income Shares in the United States, 1960–2000 Top 5% (2)

Top 1% (3)

Top .5% (4)

Top .1% (5)

Top .01% (6)

Top 10–5% (7)

Top 5–1% (8)

Top 1–5% (9)

Top .5–1% (10)

Top .1–.01% (11)

Top .01% (12)

1960 1962 1964 1966 1967 1968 1969 1970 1971 1972 1973 1974 1975 1976 1977 1978 1979 1980 1981

31.70 32.37 32.18 32.01 32.12 32.06 31.86 31.59 31.82 31.70 31.93 32.47 32.74 32.56 32.60 32.63 32.53 33.05 32.96

20.81 21.23 21.04 21.01 21.12 21.03 20.72 20.45 20.54 20.43 20.64 21.12 21.14 20.97 20.99 21.05 21.01 21.36 21.16

8.28 8.42 8.25 8.35 8.42 8.36 8.03 7.81 7.79 7.76 7.75 8.15 8.04 7.92 7.96 8.01 8.09 8.24 8.03

5.53 5.59 5.46 5.56 5.61 5.58 5.30 5.15 5.11 5.09 5.06 5.41 5.32 5.23 5.27 5.32 5.40 5.53 5.38

2.13 2.10 2.05 2.14 2.15 2.13 1.99 1.92 1.90 1.90 1.87 2.09 2.02 2.00 2.03 2.07 2.15 2.22 2.17

0.59 0.57 0.56 0.60 0.59 0.58 0.54 0.52 0.51 0.52 0.49 0.56 0.55 0.56 0.56 0.57 0.61 0.65 0.64

10.89 11.14 11.14 11.00 11.00 11.02 11.14 11.14 11.28 11.27 11.29 11.35 11.61 11.59 11.60 11.59 11.52 11.69 11.80

12.53 12.81 12.78 12.66 12.70 12.67 12.70 12.64 12.76 12.67 12.89 12.98 13.09 13.04 13.04 13.03 12.91 13.11 13.13

2.75 2.83 2.80 2.79 2.80 2.78 2.73 2.66 2.68 2.67 2.69 2.74 2.73 2.69 2.69 2.69 2.69 2.71 2.65

3.40 3.49 3.41 3.42 3.47 3.44 3.31 3.22 3.21 3.19 3.19 3.32 3.29 3.23 3.24 3.25 3.26 3.31 3.21

1.54 1.53 1.49 1.54 1.56 1.56 1.45 1.40 1.39 1.39 1.38 1.53 1.47 1.45 1.47 1.49 1.54 1.57 1.54

0.59 0.57 0.56 0.60 0.59 0.58 0.54 0.52 0.51 0.52 0.49 0.56 0.55 0.56 0.56 0.57 0.61 0.65 0.64


Year

Top 10% (1)

Continued

143

144

TABLE 4—Continued Top 5% (2)

Top 1% (3)

Top .5% (4)

Top .1% (5)

Top .01% (6)

Top 10–5% (7)

Top 5–1% (8)

Top 1–5% (9)

Top .5–1% (10)

Top .1–.01% (11)

Top .01% (12)

1982 1983 1984 1985 1986 1987 1988 1989 1990 1991 1992 1993 1994 1995 1996 1997 1998 1999 2000

33.81 34.37 34.54 34.86 35.20 36.68 38.85 38.70 39.12 39.00 40.36 39.99 39.93 40.54 41.14 41.70 42.06 42.59 43.91

21.83 22.25 22.50 22.81 23.02 24.70 27.17 26.89 27.32 26.98 28.35 27.85 27.85 28.46 29.15 29.83 30.31 30.91 32.15

8.50 8.71 8.98 9.20 9.22 10.87 13.28 12.74 13.12 12.48 13.71 13.03 13.04 13.53 14.10 14.77 15.28 15.85 16.94

5.79 5.99 6.26 6.44 6.41 7.83 10.02 9.45 9.79 9.12 10.25 9.58 9.57 9.99 10.48 11.12 11.60 12.14 13.10

2.45 2.60 2.82 2.94 2.86 3.74 5.22 4.76 4.92 4.44 5.26 4.75 4.74 4.98 5.32 5.80 6.19 6.63 7.37

0.77 0.86 0.97 0.96 0.99 1.30 1.99 1.75 1.83 1.61 2.03 1.75 1.74 1.82 1.97 2.19 2.40 2.63 3.06

11.99 12.12 12.03 12.05 12.18 11.98 11.68 11.82 11.81 12.02 12.01 12.14 12.08 12.08 11.99 11.87 11.74 11.68 11.76

13.32 13.55 13.52 13.61 13.80 13.83 13.89 14.15 14.20 14.50 14.65 14.82 14.81 14.93 15.05 15.07 15.04 15.06 15.21

2.72 2.72 2.72 2.76 2.81 3.04 3.26 3.29 3.32 3.36 3.46 3.45 3.47 3.54 3.62 3.65 3.68 3.71 3.84

3.34 3.39 3.44 3.50 3.55 4.09 4.80 4.69 4.88 4.68 4.99 4.83 4.82 5.00 5.16 5.31 5.41 5.51 5.73

1.68 1.74 1.84 1.98 1.87 2.44 3.23 3.02 3.09 2.83 3.23 3.01 3.00 3.17 3.35 3.61 3.79 4.00 4.32

0.77 0.86 0.97 0.96 0.99 1.30 1.99 1.75 1.83 1.61 2.03 1.75 1.74 1.82 1.97 2.19 2.40 2.63 3.06

Notes: Computations by authors on tax return statistics. Taxpayers are ranked by gross income (excluding capital gains and government transfers). Income of nonfilers is imputed as 20 percent of average income. Groups defined relative to all tax units (filers and nonfilers). The table reports the percentage of total income accruing to each of the top groups. Top 10 percent denotes the top decile, top 10–5% denotes the bottom half of the top decile, etc.

Saez

Year

Top 10% (1)

TABLE 5 Marginal Tax Rates (MTR) for Top Income Groups in the United States, 1960–2000 Top 5% (2)

Top 1% (3)

Top .5% (4)

Top .1% (5)

Top .01% (6)

Top 10–5% (7)

Top 5–1% (8)

Top 1–.5% (9)

Top .5–.1% (10)

Top .1–.01% (11)

Top MTR (12)

1960 1962 1964 1966 1967 1968 1969 1970 1971 1972 1973 1974 1975 1976 1977 1978 1979 1980 1981 1982 1983 1984 1985

32.32 33.17 31.19 30.58 31.05 34.55 35.56 34.29 33.48 34.55 36.19 37.56 36.53 38.32 40.88 42.65 42.57 44.14 45.01 40.60 38.24 37.33 37.74

37.33 38.02 35.72 34.91 35.49 39.38 40.40 39.05 38.30 39.42 41.26 42.73 41.12 43.02 45.90 47.43 47.44 48.46 48.72 43.72 41.27 40.22 40.73

51.47 51.89 48.43 47.13 47.61 52.37 53.37 51.53 50.73 51.19 52.37 53.79 51.38 53.10 54.93 55.45 54.99 54.84 54.12 47.44 46.07 44.65 45.53

57.92 58.05 54.00 52.00 52.29 57.03 58.04 55.76 54.89 54.48 55.36 56.56 54.52 56.04 56.89 57.37 56.53 56.18 55.20 47.45 47.17 45.72 46.81

69.89 69.07 62.78 59.90 59.67 64.31 65.22 61.87 61.06 59.36 60.14 61.20 59.34 60.77 60.12 60.62 58.61 57.79 56.11 46.49 47.48 45.88 47.14

81.30 79.31 70.43 65.22 64.74 67.44 68.62 64.28 63.50 61.40 63.22 63.68 61.87 64.36 61.74 62.75 59.90 58.79 56.30 44.90 47.15 46.56 47.16

22.74 23.92 22.65 22.32 22.53 25.32 26.54 25.57 24.71 25.72 26.91 27.95 28.18 29.82 31.81 33.96 33.70 36.25 38.36 34.92 32.68 31.92 32.09

27.98 28.91 27.51 26.85 27.46 30.82 32.21 31.34 30.72 32.22 34.58 35.78 34.82 36.89 40.39 42.50 42.70 44.46 45.41 41.34 38.18 37.28 37.49

38.50 39.73 37.54 37.42 38.25 43.03 44.30 43.33 42.78 44.94 46.74 48.32 45.28 47.39 51.09 51.67 51.91 52.10 51.92 47.44 43.66 42.18 42.54

50.42 51.41 48.71 47.06 47.72 52.51 53.72 52.12 51.24 51.57 52.55 53.64 51.55 53.10 54.87 55.30 55.15 55.11 54.59 48.15 46.92 45.59 46.53

65.55 65.27 59.89 57.84 57.74 63.15 63.95 60.98 60.16 58.60 59.04 60.30 58.38 59.40 59.50 59.80 58.10 57.37 56.03 47.22 47.65 45.53 47.13

87 87 77 70 70 75.25 77 71.75 70 70 70 70 70 70 70 70 70 70 70 50 50 50 50

145

Continued


Year

Top 10% (1)

146 Saez

TABLE 5—Continued

Year

Top 10% (1)

Top 5% (2)

Top 1% (3)

Top .5% (4)

Top .1% (5)

Top .01% (6)

Top 10–5% (7)

Top 5–1% (8)

Top 1–.5% (9)

Top .5–.1% (10)

Top .1–.01% (11)

Top MTR (12)

1986 1987 1988 1989 1990 1991 1992 1993 1994 1995 1996 1997 1998 1999 2000

37.58 33.88 29.03 29.10 29.20 29.93 29.87 32.34 32.57 32.62 32.39 33.21 33.63 33.79 33.95

40.52 35.85 29.46 29.56 29.74 30.99 30.88 34.38 34.61 34.60 34.17 35.21 35.54 35.78 36.02

45.34 37.31 28.59 28.50 28.91 32.01 31.83 39.01 39.27 38.74 37.74 39.15 39.05 38.94 38.83

46.51 37.07 27.53 27.42 27.90 31.50 31.34 39.55 39.68 38.98 37.90 39.47 39.36 39.32 39.28

47.31 36.93 27.33 27.09 27.65 31.29 31.25 39.99 39.95 39.51 38.42 39.48 39.43 39.32 39.21

46.72 36.53 27.07 26.99 27.57 31.21 31.15 39.83 39.80 39.46 38.38 39.35 39.37 39.19 39.01

32.03 29.82 28.05 28.05 27.96 27.57 27.47 27.66 27.86 27.94 28.05 28.18 28.69 28.53 28.37

37.30 34.69 30.29 30.51 30.50 30.11 30.00 30.31 30.51 30.86 30.83 31.35 31.98 32.45 32.93

42.66 37.92 31.83 31.60 31.90 33.39 33.29 37.50 38.14 38.04 37.26 38.16 38.07 37.67 37.27

45.87 37.21 27.76 27.75 28.15 31.70 31.42 39.12 39.41 38.47 37.37 39.47 39.29 39.32 39.35

47.62 37.14 27.49 27.15 27.70 31.35 31.32 40.08 40.04 39.53 38.44 39.55 39.47 39.41 39.35

50 38.5 28 28 28 31 31 39.6 39.6 39.6 39.6 39.6 39.6 39.6 39.6

Notes: Marginal tax rates computed using microfiles of tax returns and the TAXSIM calculator (Feenberg and Coutts, 1993). Marginal tax rates include only federal income taxes and ignore state income taxes. Marginal tax rates are weighted by income and are a weighted average of marginal tax rates on earnings and other income (excluding capital gains). Column (12) reports the top marginal tax rate. In 1960–1963, the top bracket rate is 91 percent, but there is maximum average tax rate of 87 percent. In 1971–1981, the top marginal tax rate for labor income is lower (see Table D2).

OLS 2SLS OLS (Newey(Top rate (NeweyWest s.e.) instrument) West s.e.) (1) (2) (3) Elasticity Time trend Time trend square Time trend cube Adjusted R-square First-stage t-statistics

2SLS (Top rate instrument) (4)

2SLS OLS OLS (Newey- (Top rate (NeweyWest s.e.) instrument) West s.e.) (6) (7) (5)

1.58 (0.28)

1.70 (0.19)

0.85 (0.21) Yes

−0.02 (0.34) Yes

0.62 (0.12) Yes Yes

0.59 (0.08) Yes Yes

0.72

0.71 10.10

0.86

0.74 5.37

0.98

0.98 10.1

0.68 (0.15) Yes Yes Yes 0.98

2SLS (Top rate instrument) (8) 0.61 (0.09) Yes Yes Yes 0.98 11.7

Notes: Estimates obtained by time-series regression of log(top 1% income share) on a constant, log(1 − average marginal tax rate), and polynomials time controls from 1960 to 2000 (38 observations). In columns 1, 3, 5, and 7, simple OLS regression is run, standard errors (s.e.) from Newey-West with 8 lags. In columns 2, 4, 6, and 8, 2SLS regression is run using log(1 − top marginal tax rate) as an instrument.


TABLE 6 Elasticities of the top 1% income share with respect to net-of-tax rates

147

148

Saez

R-square reaches 98 percent, and the elasticity coefficient is not sensitive to adding additional controls. The IV estimates are close in magnitude to the OLS estimates and have a strong first stage [except in the case of column (4) where the first stage is weak]. This finding suggests that the issue of reverse causality because of the progressive nature of the tax schedule is not an important issue. Figure 3 illustrates these issues by plotting, along with the top 1 percent income share series, the fitted values from the regressions with no time controls (line with triangles) and with two time controls (solid line). The line with triangles shows that the pure tax effects explain quite poorly the evolution of the top 1 percent income share. In contrast, the solid line with two time trends captures extremely well the pattern of the top 1 percent income share (the adjusted R-square of the regression is 98 percent). The line with squares in Figure 3 displays the counterfactual pattern, assuming that the marginal tax rate for the top 1 percent had remained constant since 1960. This curve shows that most of the growth in the top 1 percent income share is due to the time trends and that only two out of the nine-percentage-point increase in the top 1 percent income

FIGURE 3. The Top 1% Income Share and Fitted Values from Elasticity Regressions Source: Series based on regression analysis presented in Table 6, columns (1) and (5). Notes: The diamond line is the top 1 percent income share. The line with triangles is the fitted regression curve, including only the net-of-tax rate. The solid line is the fitted regression curve, including time controls. The line with squares is the same fitted regression curve but the marginal tax rate is frozen at the 1960 value.


149

share from the 1960s to 2000 is due to the decline in marginal tax rates. Therefore, in summary, attributing all the increase in the top income shares to the tax developments generates large elasticities but fits the data poorly. Controlling for time trends fits the data much better and reduces substantially the elasticity as well as the fraction of the increase in top incomes that can be attributed to tax changes. Figure 4 displays the share of income accruing to the bottom half of the top decile (panel A) and to the bottom half of the top percentile (panel B), along with the average marginal tax rate faced by these two groups. The figure shows that the top 10–5 percent income group has experienced moderate gains since 1960, and the pattern of the gains does not appear to be correlated with the pattern of the marginal tax rates that the group faces (rising up to 1981, then declining in the 1980s, then stable in the 1990s). Panels A and B in Table 7 show that regressing the log of the top income shares of the top 10–5 percent and top 5–1 percent on their log netof-tax rates, with or without time trend controls, produces elasticities close to zero. Therefore, upper-middle-income families and individuals (up to the top 1 percent threshold, around $280,000 per year in 2000) do not appear to be sensitive to taxation.38 It is striking, in particular, that these upper-middle-income taxpayer shares increase little during the 1980s; although they experience quite sizable marginal tax rate cuts (about 9 percentage points for the top 10–5 percent, and over 13 points for the top 5–1 percent).39 Note again that IV estimates are also almost identical to OLS estimates. Panel B of Figure 4 shows that the top 1–.5 percent share does not decrease during the 1970s, when the marginal tax rate increases from 40 to 50 percent, and does not increase during ERTA 1981, when the marginal tax rate decreases back to 40 percent. In contrast, TRA 1986, which decreases the rate to around 32 percent (thus a smaller percentage change in the net-of-tax rate relative to the 1970s or ERTA 1981), does produce a sizable increase in the income share, producing a noticeable break in the series. The increase in tax rates, to about 38 percent following OBRA 1992, does not seem to have affected the upward trend following TRA 1986. Thus, although marginal tax rates in the late 1990s are about the same as 38 In principle, the secondary earner labor supply responses should be captured by those elasticities. Thus, my results can be consistent with the large married female labor supply responses obtained by Eissa (1995) only if secondary earners’ income is a small fraction of total reported family incomes. 39 A similar regression analysis for other income groups below the top decile generates small or even negative and always insignificant elasticities. The estimates are not precisely estimated, however, because changes in net-of-tax rates are much smaller below the top decile.

FIGURE 4. Tax Rates and Income Shares for the Medium-High Income Groups Note: Based on Tables 4 and 5.

TABLE 7 Elasticities of income shares with respect to net-of-tax rates for various upper income groups NeweyWest OLS Regression, with time controls (2)

2SLS regression, with time controls (3)

Top 10% 0.77 (0.36)

0.33 (0.08)

0.32 (0.05) 9.94

Top 5% 1.25 (0.30)

0.43 (0.09)

0.39 (0.05) 10.5

Top 1% 1.58 (0.28)

0.62 (0.12)

0.59 (0.08) 10.11

Top 0.5% 1.55 (0.25)

0.72 (0.13)

0.69 (0.09)

First-stage t-statistic of instrument Elasticity First-stage t-statistic of instrument Elasticity First-stage t-statistic of instrument Elasticity

NeweyWest OLS regression, with time controls (5)

2SLS regression with time controls (6)

B. Intermediate income groups

A. Top income groups Elasticity

NeweyWest OLS Regression, no time controls (4)

Top 10–5% −0.44 (0.17)

− 0.11 (0.09)

−0.04 (0.10) 6.5

Top 5–1% 0.14 (0.28)

0.12 (0.04)

0.09 (0.04) 8.16

Top 1–.5% 0.92 (0.21)

0.30 (0.08)

0.29 (0.07) 10.65

Top 0.5–0.1% 1.21 (0.22)

0.52 (0.09)

0.49 (0.08)

151

Continued


NeweyWest OLS regression, no time controls (1)

152

TABLE 7—Continued NeweyWest OLS Regression, with time controls (2)

2SLS regression, with time controls (3)

NeweyWest OLS regression, with time controls (5)

2SLS regression with time controls (6)

B. Intermediate income groups

A. Top income groups First-stage t-statistic of instrument

NeweyWest OLS Regression, no time controls (4)

9.9

9.21

Elasticity Top 0.1% 1.54 (0.27)

0.94 (0.19)

0.89 (0.11) 11.37

Top 0.01% 1.45 (0.36)

1.08 (0.32)

1.09 (0.16)

First-stage t-statistic of instrument

Top 0.1–0.01% 1.44 (0.23)

0.78 (0.16)

0.76 (0.11) 9.69

Top 0.01% 1.45 (0.36)

1.08 (0.32)

1.09 (0.16)

Elasticity

First-stage t-statistic of instrument

18.01

18.01

Notes: Estimates obtained by time-series regression of log(top income share) on a constant, log(1 − average marginal tax rate), time trend, and square of time trend from 1960 to 2000 (38 observations). In columns 1 and 4, OLS regression is run, no time trend included. Newey-West standard errors with 8 lags reported. In columns 2 and 5, OLS regression is run with time and time^2 trend included. Newey-West standard errors with 8 lags reported. In columns 3 and 6, 2SLS regression is run with time and time^2 trend included and instrumented with log (1 − top marginal tax rate).

Saez

NeweyWest OLS regression, no time controls (1)


153

in the 1960s, the income share is 30 percent larger.40 The regressions for the top 1–.5 percent and top .5–.1 percent groups in Table 7 (panels C and D) display significant elasticities, but the size of the elasticity is much smaller when income controls are included. Figure 5 displays the share of income and marginal tax rates for the very top groups: the top .1–.01 percent (panel A), and the top .01 percent (panel B). The responses to ERTA 1981 and TRA 1986 and the shortterm response to OBRA 1993, followed by a surge in income shares since 1995, are even more pronounced than for the groups the top 0.1 percent below. However, the Kennedy tax cuts of the early 1960s provide striking new evidence. For the topmost .01 percent, the progressive tax structure of the early 1960s generated extremely high marginal tax rates (around 80 percent), which were reduced significantly by the Kennedy tax cuts in 1964–1965 (to about 65 percent).41 This implies a 75 percent increase in the net-of-tax rate, a much larger increase than the ERTA 1981 and TRA 1986 tax rate reductions. In spite of this enormous marginal tax rate cut, the topmost income share remains flat in the 1960s and well into the 1970s, which suggests a complete absence of behavioral response in both the short- and the long-run.42 Note that, although the top nominal marginal tax rate was 91 percent, the average marginal tax rate of the top .01 percent is only slightly above 80 percent. This is due to various other provisions of the tax code, such as the maximum average tax of 87 percent on income and charitable gifts by the wealthy.43 Panels E and F of Table 7 show that the regressions for the top .1–.01 percent and the top .01 percent display significant elasticities in all specifications, although pure tax factors can explain only a fraction of the total increase in the top most shares once exogenous time trends are included.

40 These considerations show again that elasticity estimates would be extremely sensitive to the time period considered. The ERTA 1981 and OBRA 1993 episodess would produce 0 elasticity estimates, and TRA 1986 would produce a sizable 0.93 estimate (comparing 1986 and 1988). Comparing 2000 to 1984 and attributing all the large increase in the share to the modest decrease in the marginal tax rate would produce an enormous elasticity estimate of 4.94. 41 These tax cuts were proposed by President Kennedy in the early 1960s but were actually implemented by the Johnson administration after Kennedy’s death in 1963. 42 Lindsey (1990) claimed that the Kennedy tax cuts generated a surge in top incomes, but this erroneous result is due to his casual examination of the tabulations published by the IRS. Goolsbee (1999) makes a more careful use of the same published data (although he does not exclude realized capital gains and does not measure marginal tax rates accurately) and finds no response, as I do here. 43 Considering smaller groups at the very top, such as the top .001 percent, never generates marginal tax rates higher than 80 to 82 percent.

154

Saez

FIGURE 5. Tax Rates and Income Shares for the Top Groups Note: Based on Series obtained from Tables 4 and 5.


155

3.3 Composition In the previous subsection, we saw that the income groups within the top decile display very heterogeneous responses. Groups below the top 1 percent never display evidence of tax responsiveness. Top groups displayed a sharp response to the 1980s tax cuts, especially TRA 1986, but only a short-term response to the tax increase of 1993, and no response for the earlier tax cuts in the 1960s. To cast more light on these findings, I now turn to an analysis of the composition of those incomes.44 The complete composition series of top income groups are reported in Tables D1 and D2 of Saez (2004), a longer version of my work. Figure 6 displays the evolution of the top decile income share from 1960 to 2000 and how those incomes are decomposed into the seven sources

FIGURE 6. The Top 10% Income Share and Composition, 1960–2000 Source: Tables B1 and Table D1 in Saez (2004). Notes: The figure displays the income share of the top 10 percent tax units and shows how the top 10 percent incomes are divided into seven income components: wages and salaries (including exercised stock options), S-corporation profits, partnership profits, sole proprietorship profits, dividends, interest income, and other income.

44 Previous studies have focused mostly on taxable income elasticities. Feenberg and Poterba (1993, 2000) analyze the composition of incomes for the top .5 percent from 1951 to 1990, and Slemrod (1994, 1996) analyzes the composition of top incomes around TRA 1986.

156

Saez

described in section 2. Wage income forms the majority of the top 10 percent of incomes, and its share has increased smoothly from two-thirds to about three-quarters since 1960. The large 12-percentage-point gain in the top 10 percent income share (from 32 to 44 percent) is due almost entirely to a smooth and secular increase in the wage component (from 22 points to 33.5 points), with the size of the other components remaining stable overall (around 10 points, with a squeeze around 7 points in the late 1970s and early 1980s). As depicted in Figure 7, the top 1 percent income share increases from 8.3 percent to almost 17 percent from 1960 to 2000. The striking feature, however, is that 7 out of the 8.7-point increase in the top 1 percent share is due to the wage-income component. As a result, although wages represented only 40 percent of total income for the top 1 percent in the early 1960s, they now represent over 60 percent of top 1 percent incomes. The increase in the wage component appears to have started in the early 1970s and has been fairly regular, with an acceleration in the last two decades

FIGURE 7. The Top 1% Income Share and Composition, 1960–2000 Source: Tables B1 and Table D1 in Saez (2004). Notes: The figure displays the income share of the top 1 percent tax units and shows how the top 1 percent incomes are divided into seven income components: wages and salaries (including exercised stock options), S-corporation profits, partnership profits, sole proprietorship profits, dividends, interest income, and other income.


157

(especially the 1990s). There are two spikes in the wage component series, one in 1988 (just after TRA 1986) and another in 1992 (just before the OBRA 1993 tax increase). However, the short-term nature of those two spikes suggests that they were the consequence of the retiming of wage income to take advantage of lower rates.45 Although the nonwage part stays stable as a whole, the components display interesting patterns. The most striking feature is the emergence of S-corporation income after TRA 1986. Before the 1980s, S-corporation income was extremely small. Indeed, the standard C-corporation form was more advantageous for high-income individual owners because the top individual tax rate was much higher than the corporate tax rate and taxes on capital gains were relatively low. S-corporation income increases sharply from 1986 to 1988 and increases slowly afterward. The sharp increase in S-corporation income just after TRA 1986 certainly reflects in large part a shift in the status from C-corporation to S-corporation status to take advantage of the lower individual rates.46 In contrast, dividends (paid out by C-corporations and foreign corporations) and sole proprietorship income decreased regularly over the period. Partnership income is about the same in the 1960s as in the 1990s; partnership income was very small during the 1980s due to a dramatic increase in partnership losses.47 The dramatic increase of partnership losses from the mid- to late 1970s up to 1986 (during recessions and recoveries alike) is probably due first to the increase in inflation, which might have increased losses because of the deductibility of nominal interest payments.48 Then taxpayers and tax accountants might have realized that partnerships offered an attractive possibility for avoiding taxes. The repeal of the investment tax credit and the passive losses limitations with the TRA 1986, as well as the reduction in top tax rates, have drastically reduced the value of those tax shelters and probably explains the quick and sustained disappearance of most partnership losses just after TRA 1986.49 Sole proprietorship income also displays a similar pattern, with a sharp reduction from the mid-1970s 45 Goolsbee (2000b) showed that many executives exercised their stock options in 1992 to take advantage of the low rate of 31 percent in 1992 before the increase to 39.6 percent in 1993. This retiming explains the large difference between the short-term and long-term elasticity estimates using the OBRA 1993 reform. 46 See Slemrod (1996), Carroll and Joulfaian (1997), and Gordon and Slemrod (2000) for a more precise analysis. 47

Partnership profits have stayed about stable over the full period.

48

Note that interest income (which is not net of interest payment deductions) is also particularly high during that period. 49

See Samwick (1996) for a more detailed analysis.

158

Saez

FIGURE 8. The Top 0.01% Income Share and Composition, 1960–2000 Source: Tables B1 and Table D1 in Saez (2004). Notes: The figure displays the income share of the top .01 percent tax units and shows how the top .01 percent incomes are divided into seven income components: wages and salaries (including exercised stock options), S-corporation profits, partnership profits, sole proprietorship profits, dividends, interest income, and other income.

to the mid-1980s.50 Although the wage income component starts to increase in the early 1970s, the combined effect of sharp reductions in partnership and sole proprietorship incomes from the mid-1970s to 1981 explains why the top 1 percent income share stays almost flat up to 1981. Figure 8 displays the income share and composition of the top .01 percent group. It shows a dramatic shift in the composition of the topmost incomes away from dividends (which represented more than 60 percent of top incomes in the early 1960s) toward wage income (which represents about 60 percent of top incomes in 2000).51 In the early 1960s, the top .01 percent incomes were facing extremely high marginal tax rates of about 50 Sole proprietorship income displays a secular trend downward from 1960 to 2000 most likely because of the secular decline in farming and other traditional small-business activities organized in the form of sole proprietorships. 51 This secular shift from rentiers to the working rich at the top of the U.S. income distribution is described in more detail in Piketty and Saez (2003).


159

80 percent on average (while tax rates on long-term capital gains were around 25 percent). Thus, dividends were a disadvantageous form of income for the rich, which suggests that these top-income earners had little control over the form of payment and thus might have been passive investors. The Kennedy tax cuts did not reduce the top individual rate enough (the top rate became 70 percent) to make the S-corporation form attractive relative to the C-corporation form, which explains perhaps the contrast in behavioral responses between the Kennedy tax cuts and the tax changes of the 1980s. This situation shows, as argued by Slemrod and Kopczuk (2002), that the elasticity of reported incomes is not a constant parameter but may be extremely sensitive to the legal structure and the complete tax environment for corporations and individuals. The share of dividends falls regularly over the period, while the share of wage income starts to increase in 1971. By 1979, the wage component overtakes the dividend component. Figure 8 shows clearly that ERTA 1981 produced a sudden burst of S-corporation income (which was negligible up to 1981) mostly likely because of a shift from C-corporations to S-corporations.52 Note that the increase in S-corporation income is concentrated mostly in the top .01 percent and does not happen at all for groups below the top .1 percent. This situation is consistent with the tax minimization explanation: ERTA 1981 decreased marginal tax rates significantly only for groups above the top .1 percent, for whom the Subchapter S status started to become attractive when the top individual rate was reduced to 50 percent.53 Figure 8 shows that almost all the increase in top incomes from 1981 to 1984, first documented by Lindsey (1987), is also due to the surge in S-corporation income. The wage component increases as well but with no noticeable break in the upward trend around ERTA 1981.54 The Scorporation component increases again sharply from 1986 to 1988 and then stays about stable afterward. The wage component also presents a spike in 1988 and in 1993, but these spikes seem to be short-term responses in a generally upward trending curve. The tax cuts of the 1960s, although extremely large, did not generate any behavioral response perhaps because top individual rates remained substantially higher than the corporate and capital gains tax rate and thus did not induce top-income taxpayers to switch corporate income toward individual income. 52 As discussed in section 2.1, this phenomenon has been well documented in the case of TRA 1986. 53 54

From 1980 to 1986, the corporate tax rate was 42 percent.

Because of the maximum tax of 50 percent on labor income enacted in 1971–1972, the marginal tax rates for top wage incomes actually did not change much with ERTA; see section 3.4.

160

Saez

Therefore, to sum up, the dramatic increase in top income shares is due primarily to a secular increase in the wage income component starting in the early 1970s, and the large tax changes of TRA 1986 and OBRA 1993 seem to have generated only short-term spikes in the overall upward and accelerating trend of the wage component.55 The tax cuts of the 1980s have generated a surge in business income taxed at the individual level. ERTA 1981 created a surge in S-corporation income for the topmost groups of the income distribution. With TRA 1986, S-corporation income surged for all upper-income groups. Partnership income also rose dramatically immediately after TRA 1986 mostly because of the disappearance of partnership losses. These business income components have remained relatively stable after TRA 1986, which suggests they were the consequence of a one-time shift from the corporate sector and the one-time closing of the partnership loss tax shelters. The top tax rate increase of 1993 to 39.6 percent (with a corporate tax rate of 35 percent) was not large enough to induce businessowners to switch back to the C-corporation status. As a result, OBRA 1993 did not produce any long-term income shifting away from the individual sector, and its only effect seems to have been a shortterm retiming of salary income. The surge in business income reported on individual returns in the 1980s cannot be interpreted as a supply-side success because most of these individual income gains came either at the expense of taxable corporate income or could have been obtained from the closing of tax shelters after the imposition of stricter rules on losses from passive businesses.56 Therefore, the success or failure of the tax cuts at generating additional economic activity must be deferred to a more precise analysis of the central wage income component, to which we now turn.

3.4 Top Wage Incomes We have seen that most of the increase in top income shares since the 1970s is actually due to a sharp increase in the wage income component. The time pattern of marginal tax rates for wage income is not the same as the pattern for other forms of income because of the introduction of the maximum tax rate on earned income in 1971, which reduced the top rate 55 Top income shares are flat before 1981, masking the increase in the wage component, because of a large decline in partnership and sole proprietorship income, due in turn perhaps to high interest rates and the development of tax shelters in the 1970s. Partnership income and, to a lesser extent, sole proprietorship income increased back to their early 1970s levels immediately after TRA 1986. 56 It is doubtful that the decrease in tax rates, by reducing the incentives to avoid taxes, was necessary to eliminate abusive partnership losses (as argued, for example, in Samwick, 1996) because partnership losses were almost nonexistent before the late 1970s, a time when tax rates were extremely high.


161

for earned income from 70 percent (the top rate on other income) to 60 percent in 1971 and then 50 percent starting in 1972.57 This provision became irrelevant in 1982, when the top tax rate for any income source was reduced from 70 percent to 50 percent. Therefore, analyzing the wage income component separately is of particular interest. All the top wage income share series and corresponding average marginal tax rates for wage income are reported in Tables 8 and 9, respectively. As for average income, the evolution of average real wage income series (for the full population) does not appear to be correlated with the evolution of marginal tax rates. Figure 9 shows the pattern of real incomes and marginal tax rates for the bottom 99 percent wage earners (panel A) and the top 1 percent wage earners (panel B). The bottom 99 percent have experienced no real growth in wage income since 1972, and the pattern of changes in real wages does not seem to be related to changes in marginal tax rates. In contrast, top 1 percent wage income earners experienced accelerating growth over the 1960 to 2000 period, with almost a tripling in real wage income since the early 1970s. Consistent with the pattern of the wage component for overall income, top wage income earners experienced spikes just after TRA 1986 and just before OBRA 1993, clear evidence of short-term responses (or retiming) of labor income compensation. However, the long-run pattern seems to be an extraordinary and accelerating growth independent of the tax developments because marginal tax rates on these wage income earners were about the same, around 40 percent, in the mid-1960s and in the most recent years. Indeed, the secular growth in top wages starts in the early 1970s, a time when marginal tax rates were actually increasing (due mostly to the progressive nature of the income tax structure and the resulting bracket creep). To understand better this unprecedented increase in top wage incomes, it is useful to consider smaller groups within the top 1 percent, as I did for overall income. Table 10 produces the same regressions as Table 7 but for wage incomes instead of overall income.58 The shares of the bottom groups of the top decile below the top 1 percent (top 10–5 percent and top 5–1 percent) display low elasticities, while all groups within the top 1 percent display significant elasticities when no time trend is included. The elasticities increase sharply from 0.3 to 2.5 as we move up the wage income distribution 57 As described in Slemrod (1994), the marginal income tax rate on labor income could be higher than these limits in several cases because of the interaction of this provision with the regular schedule. 58 I have omitted the IV estimates in the case of wages because the first stage is not as strong as in the case of income and because the estimates are more noisy.

162 Saez

TABLE 8 Top Wage Income Shares in the United States, 1960–2000

Year

Top 10% (1)

Top 5% (2)

Top 1% (3)

Top .5% (4)

Top .1% (5)

Top .01% (6)

Top 10–5% (7)

1960 1962 1964 1966 1967 1968 1969 1970 1971 1972 1973 1974 1975 1976 1977 1978 1979 1980

24.64 24.62 24.98 25.35 25.78 25.60 25.71 25.67 25.67 25.82 26.15 26.63 26.46 26.66 26.94 27.43 27.65 28.06

15.11 15.02 15.25 15.47 15.81 15.66 15.68 15.64 15.67 15.80 16.06 16.48 16.32 16.49 16.70 17.07 17.25 17.60

5.16 5.05 5.12 5.16 5.34 5.24 5.19 5.13 5.18 5.32 5.43 5.66 5.64 5.74 5.85 6.05 6.21 6.43

3.30 3.21 3.24 3.27 3.38 3.32 3.27 3.21 3.25 3.39 3.43 3.63 3.63 3.70 3.79 3.93 4.06 4.23

1.15 1.08 1.07 1.10 1.14 1.12 1.10 1.06 1.08 1.14 1.14 1.26 1.26 1.30 1.35 1.40 1.47 1.57

0.25 0.21 0.21 0.22 0.23 0.23 0.24 0.21 0.22 0.24 0.24 0.27 0.27 0.29 0.30 0.31 0.34 0.38

9.53 9.60 9.73 9.88 9.97 9.94 10.04 10.03 10.00 10.02 10.09 10.15 10.15 10.16 10.25 10.36 10.40 10.46

Top 5–1% (8) 9.95 9.97 10.13 10.31 10.47 10.42 10.49 10.51 10.49 10.48 10.63 10.82 10.67 10.76 10.85 11.02 11.03 11.17

Top 1–.5% (9)

Top .5–.1% (10)

Top .1–.01% (11)

Top .01% (12)

1.86 1.85 1.88 1.89 1.96 1.92 1.92 1.92 1.93 1.94 2.00 2.03 2.01 2.04 2.06 2.13 2.15 2.20

2.15 2.13 2.17 2.16 2.24 2.20 2.17 2.15 2.17 2.24 2.28 2.37 2.38 2.40 2.45 2.53 2.59 2.66

0.91 0.87 0.87 0.88 0.91 0.89 0.87 0.85 0.86 0.90 0.91 0.99 0.98 1.02 1.05 1.09 1.13 1.19

0.25 0.21 0.21 0.22 0.23 0.23 0.24 0.21 0.22 0.24 0.24 0.27 0.27 0.29 0.30 0.31 0.34 0.38

28.15 28.56 29.09 29.61 29.74 29.94 30.60 31.97 31.55 31.81 31.44 32.46 31.85 31.54 32.43 33.16 33.88 34.34 35.11 36.03

17.65 18.02 18.49 18.95 19.05 19.19 19.99 21.37 20.83 21.14 20.77 21.85 21.29 20.95 21.73 22.47 23.19 23.73 24.50 25.42

6.43 6.68 6.96 7.27 7.28 7.33 8.15 9.38 8.70 9.00 8.56 9.63 9.06 8.72 9.26 9.80 10.43 10.98 11.64 12.61

4.24 4.42 4.66 4.96 4.92 4.96 5.69 6.79 6.13 6.41 5.97 6.97 6.41 6.07 6.52 6.97 7.54 8.08 8.71 9.64

1.59 1.67 1.80 1.99 1.98 2.02 2.43 3.16 2.69 2.87 2.57 3.33 2.90 2.63 2.91 3.21 3.67 4.12 4.67 5.44

0.39 0.41 0.47 0.52 0.54 0.58 0.69 1.09 0.82 0.91 0.78 1.22 0.96 0.83 0.94 1.11 1.36 1.65 1.98 2.45

10.50 10.54 10.61 10.66 10.70 10.75 10.61 10.60 10.71 10.67 10.67 10.61 10.56 10.59 10.70 10.69 10.70 10.61 10.61 10.62

11.23 11.34 11.53 11.68 11.77 11.87 11.83 11.99 12.14 12.14 12.21 12.22 12.23 12.22 12.48 12.66 12.75 12.76 12.85 12.84

2.18 2.25 2.30 2.32 2.35 2.36 2.47 2.59 2.57 2.59 2.59 2.66 2.64 2.65 2.73 2.83 2.89 2.89 2.94 2.99

2.66 2.75 2.86 2.97 2.95 2.95 3.25 3.63 3.44 3.54 3.40 3.64 3.51 3.44 3.62 3.77 3.88 3.96 4.04 4.24

1.20 1.26 1.33 1.47 1.44 1.44 1.74 2.07 1.86 1.96 1.79 2.12 1.95 1.80 1.97 2.10 2.31 2.48 2.69 3.03

0.39 0.41 0.47 0.52 0.54 0.58 0.69 1.09 0.82 0.91 0.78 1.22 0.96 0.83 0.94 1.11 1.36 1.65 1.98 2.45

Notes: Computations by authors on tax return statistics. Taxpayers are ranked by wages and salaries (which includes exercise of stock options). Groups are defined relative to all tax units with wage income (filers and nonfilers). The table reports the percentage of total wages and salaries accruing to each of the top groups. Top 10 percent denotes to top decile, top 10–5 percent denotes the bottom half of the top decile, etc.


1981 1982 1983 1984 1985 1986 1987 1988 1989 1990 1991 1992 1993 1994 1995 1996 1997 1998 1999 2000

163

164 Saez

TABLE 9 Marginal Tax Rates (MTR) on Wages for Top Wage Income Groups in the United States, 1960–2000

Year

Top 10% (1)

Top 5% (2)

Top 1% (3)

Top .5% (4)

Top .1% (5)

Top .01% (6)

Top 10–5% (7)

Top 5–1% (8)

Top 1–.5% (9)

Top .5–.1% (10)

Top .1–.01% (11)

Top MTR (12)

1960 1962 1964 1966 1967 1968 1969 1970 1971 1972 1973 1974 1975 1976 1977 1978 1979 1980

28.48 29.44 27.67 26.96 27.54 30.68 31.93 30.96 30.61 31.48 32.78 34.10 34.30 36.04 38.44 40.06 40.21 41.90

32.11 33.06 30.95 30.06 30.83 34.27 35.58 34.57 34.40 35.24 36.70 38.19 37.99 39.76 42.50 43.93 44.19 45.38

43.20 44.39 40.91 39.99 40.93 45.35 46.27 45.36 45.08 45.08 46.01 47.05 46.01 47.27 49.58 49.95 49.92 49.57

48.83 50.08 45.51 45.02 45.61 50.15 50.90 49.72 49.09 48.05 48.43 49.22 48.14 48.94 50.46 50.37 50.31 49.69

60.05 61.05 54.86 53.91 54.10 58.22 57.96 56.25 55.05 50.27 50.04 49.94 49.63 49.30 50.13 50.19 49.10 48.28

67.48 71.97 62.81 60.45 60.52 63.79 60.48 60.53 57.32 50.52 49.97 49.66 49.61 48.10 48.84 49.02 47.63 47.06

22.73 23.78 22.52 22.10 22.31 25.01 26.23 25.32 24.68 25.56 26.55 27.45 28.38 30.01 31.84 33.67 33.62 36.05

26.36 27.31 25.91 25.09 25.69 28.70 30.29 29.32 29.11 30.24 31.95 33.55 33.75 35.76 38.68 40.63 40.97 42.97

33.18 34.53 33.01 31.28 32.86 37.08 38.38 38.07 38.35 39.89 41.86 43.17 42.17 44.24 47.94 49.18 49.17 49.34

42.83 44.53 40.88 40.49 41.31 46.02 47.31 46.51 46.13 46.92 47.63 48.84 47.35 48.74 50.65 50.47 50.99 50.53

58.02 58.41 52.97 52.26 52.46 56.79 57.27 55.20 54.48 50.21 50.06 50.02 49.64 49.64 50.50 50.52 49.55 48.67

87 87 77 70 70 75.25 77 71.75 60 50 50 50 50 50 50 50 50 50

42.87 39.14 37.00 35.94 36.24 36.47 33.32 28.79 28.89 28.97 29.57 29.64 31.78 31.83 31.96 31.75 32.51 32.95 33.13 33.31

45.69 41.61 39.35 38.24 38.62 38.95 35.19 29.28 29.46 29.61 30.62 30.66 33.79 33.84 33.96 33.57 34.56 34.91 35.14 35.37

48.67 44.64 43.14 41.91 42.54 43.20 37.01 28.91 29.09 29.42 32.06 31.88 38.59 38.83 38.52 37.68 39.00 39.02 38.83 38.64

48.49 44.70 44.25 42.95 43.51 44.10 36.82 27.73 27.92 28.32 31.71 31.51 39.46 39.60 39.04 37.98 39.51 39.43 39.33 39.23

47.07 44.12 45.15 42.83 44.80 44.71 36.67 27.10 27.25 27.73 31.35 31.35 40.03 40.10 39.74 38.80 39.71 39.64 39.48 39.32

46.53 43.13 45.33 44.71 44.54 44.37 36.91 26.61 27.33 27.76 31.26 31.24 39.81 40.09 39.88 39.01 39.75 39.60 39.37 39.14

38.12 34.93 32.91 31.84 32.01 32.05 29.79 27.81 27.78 27.71 27.52 27.55 27.74 27.85 27.91 27.92 28.08 28.56 28.50 28.44

43.99 39.83 37.06 35.96 36.19 36.32 33.93 29.56 29.72 29.75 29.60 29.69 30.23 30.28 30.57 30.39 30.93 31.37 31.80 32.24

49.01 44.52 40.89 39.69 40.52 41.30 37.45 32.00 31.89 32.15 32.89 32.85 36.48 37.07 37.29 36.92 37.68 37.90 37.33 36.77

49.34 45.05 43.69 43.03 42.65 43.69 36.93 28.29 28.45 28.79 31.98 31.65 38.99 39.22 38.48 37.28 39.32 39.21 39.17 39.13

47.24 44.44 45.08 42.16 44.90 44.84 36.58 27.36 27.22 27.71 31.39 31.42 40.14 40.11 39.67 38.69 39.68 39.66 39.56 39.46

50 50 50 50 50 50 38.5 28 28 28 31 31 39.6 39.6 39.6 39.6 39.6 39.6 39.6 39.6

Notes: Marginal tax rates on wage income are computed using microfiles of tax returns and the TAXSIM calculator. Marginal tax rates include only federal income taxes and ignore state income taxes, as well as payroll taxes. Marginal tax rates are weighted by wage income. Column (12) reports the top marginal tax rate on labor income. In 1960–1963, the top bracket rate is 91 percent, but there is maximum average tax rate of 87 percent. In 1971–1981, the top marginal tax rate for nonlabor income is 70 percent (see Table 8), and the labor income marginal tax rate can be locally larger than reported.


1981 1982 1983 1984 1985 1986 1987 1988 1989 1990 1991 1992 1993 1994 1995 1996 1997 1998 1999 2000

165

FIGURE 9. Marginal Tax Rates and Average Real Wage Incomes for the Bottom 99% and the Top 1% Source: Based on Series obtained from Tables 1, 8, and 9.


167

TABLE 10 Elasticities of wage income shares with respect to net-of-tax rates for various upper wage income groups Newey-West OLS regression, no time controls (1)

Newey-West OLS regression, with time controls (2)

A. Top wage income groups Elasticity Elasticity Elasticity Elasticity Elasticity Elasticity

Top 10% −0.10 (0.55) Top 5% 0.41 (0.56) Top 1% 1.97 (0.45) Top 0.5% 2.33 (0.54) Top 0.1% 2.44 (0.43) Top 0.01% 2.48 (0.50)

0.10 (0.07) 0.17 (0.09) 0.39 (0.12) 0.51 (0.13) 0.82 (0.17) 0.96 (0.42)

Newey-West OLS regression, no time controls (3)

Newey-West OLS regression, with time controls (4)

B. Intermediate groups Top 10–5% −0.43 (0.18) Top 5–1% −0.17 (0.37) Top 1–.5% 0.31 (0.48) Top 0.5–0.1% 1.50 (0.32) Top 0.1–0.01% 2.16 (0.37) Top 0.01% 2.48 (0.50)

−0.05 (0.02) 0.07 (0.02) 0.15 (0.05) 0.38 (0.08) 0.72 (0.11) 0.96 (0.42)

Notes: Estimates obtained by time-series regression of log (top wage income share) on a constant, log (1 – average marginal tax rate), time trend, and square of time trend from 1960 to 2000 (38 observations). In columns 1 and 3, OLS regression is run, no time trends included. Newey-West standard errors with 8 lags reported. In columns 2 and 4, OLS regression is run with time and time ^2 trend included. Newey-West standard errors with 8 lags reported.

because all the increase in the top wage income shares is attributed to the secular decline in marginal tax rates since the 1960s. Including two time trends reduces significantly the estimated elasticities, which are below 0.4 except for the topmost groups. Even within the top 0.1 percent group, where elasticities are sizable, tax changes can explain only a small fraction of the dramatic surge in top wage incomes. They key point to resolve is whether we should attribute the long-term increase in top wage shares entirely to the long-term decrease in marginal tax rates. Comparing 1960 and 2000, that view seems to be untenable for groups below the top .1 percent because these groups faced comparable marginal tax rates in 1960 and in 2000. As a result, the sizable increase in

168

Saez

the top 1–.5 percent and top .5–.1 percent wage income shares cannot be due entirely to marginal tax rates. The problem is more complicated for the topmost groups (within the top .1 percent) because these groups experienced much larger gains but also experienced a nontrivial decline in marginal tax rates. Undoubtedly, a reason for the huge increase in top wage income shares (the top .01 percent share increased more than tenfold, from .21 percent in 1970 to 2.45 percent in 2000) has been the development of stock options. Stock options also create lumpiness in wage compensation because they are exercised by executives only once every few years. As a result, the top .01 percent might be extremely large in recent years because, in any given year, topmost wage earners are executives who happen to exercise their stock options in that particular year. The stock-option phenomenon, however, has clearly increased the average compensation of top executives because the top 1 percent (which certainly includes almost all the top employees receiving large option grants, even when they do not exercise stock options) more than doubles from 5.1 to 12.6 percent from 1970 to 2000. Thus, the extraordinary increase in top wage incomes, a phenomenon certainly closely related to the explosion in the compensation of chief executive officers (CEOs) and other top executives and sports, movie, and television stars, appears too large to have been solely the direct consequence of the tax reductions through supply-side effects. Furthermore, the surge in top wages is not related closely enough to the timing of the tax cuts to suggest a direct and simple causal link. Particularly surprising is the surge in top wages since 1994, in spite of the significant tax increase in 1993, which makes the secular reduction in marginal tax rates faced by top wage groups appear rather small.59 A more pertinent issue is whether this surge in top wages could have occurred had the tax structure remained the same as in the early 1960s, when the working rich had to pay in taxes more than three-quarters of their compensation. It is plausible to think that the drastic reduction in top marginal tax rates, which started in the 1960s, opened the possibility of the dramatic increase in top wages that started in the 1970s and accelerated in the 1980s and 1990s. Of course, it is impossible to provide a convincing answer to that important issue by looking only at individual income tax statistics in the United States. A promising approach would be to analyze executive compensation data. Many have researched executive 59 Companies might have started granting stock options more aggressively after TRA 1986, however, because of the decrease in individual tax rates. These options can be exercised (and thus appear on individual income tax returns) only several years later. However, Hall and Murphy (2003) show that grants of stock options, valued using the Black-Scholes formula, increased significantly after the tax increase of 1993.


169

compensation; see Murphy (1999) for a survey. Although many studies explain the disparity of CEO pay in cross-sectional data, no convincing explanation for the time-series evidence seems to have been provided.60 If the dramatic surge in top compensation is not fully explained by a comparable surge in the marginal productivity of top executives, then this lack is evidence of a market failure, which would certainly change the welfare and tax policy analysis that I presented above. Perhaps top executive pay may now be aligned with marginal product and was below market value before. Note, however, that the surge in the top 1 percent salaries since the early 1970s has been accompanied by dismal growth for the bottom 99 percent salary earners and thus does not seem to have had a positive impact on the vast majority of working families. An alternative way to make progress in our understanding is by looking at comparable experiences in other countries, a point to which I now turn for the conclusion.

4. CONCLUSION: INTERNATIONAL COMPARISONS No other country offers such a large body of empirical analysis on behavioral responses to individual income taxation as does the United States. Recently, however, several studies have produced series of top income shares using tax return data. Although these studies do not produce corresponding series of marginal tax rates, as I have shown here, interesting findings emerge. First, enormous heterogeneity exists in the behavior of top income shares in recent decades across countries. Some countries, such as the United Kingdom (Atkinson, 2002) or Canada (Saez and Veall, 2003) have experienced notable increases in top income shares, although these increases have not been as pronounced as in the United States. In contrast, countries from continental Europe, such as France (Piketty, 2003), the Netherlands (Atkinson and Salverda, 2003), and Switzerland (Dell, Piketty, and Saez, 2003), have experienced either decline or little change in top income shares since 1960. Second, the U.K. experience seems to be the closest to the U.S. experience. Top income shares in the United Kingdom started increasing exactly in 1979, when the top rate declined from 98 to 75 percent, although the concomitant increase seems modest relative to the size of the net-of-tax 60 It is quite telling to read in the recent survey of Hall and Murphy (2003), two prominent and conservative researchers in this field, that their best explanation for the surge in stockoption compensation was that “boards and managers falsely perceive stock options to be inexpensive because of accounting and cash-flow considerations.”

170

Saez

increase at the top.61 In 1988, the top rate was further decreased to 40 percent and has not changed since then. In contrast to the United States, however, the increase in top share has been relatively smooth since 1979, with no break around the tax changes. Studying the composition and estimating precisely the marginal tax rates faced by top U.K. income-taxpayers seems to be a priority in understanding whether the recent increase in top incomes is due to the tax developments.62 Third, Canada has experienced a surge in top incomes significantly larger than the increase in the United Kingdom (although smaller than that in the United States) and, as in the United States, this increase has been due to a dramatic increase in top salaries since the early 1980s. In contrast to the United States, however, top incomes in Canada have not experienced, large tax cuts since the 1960s.63 Thus, the dramatic increase in top incomes in Canada cannot be attributed solely to fiscal developments in Canada. Saez and Veall (2003) argue that the threat of emigration to the United States has forced Canadian companies to increase the pay of their top employees if they want to retain them, thereby replicating in Canada the dramatic U.S. increase in top employees’ pay. If the migration explanation is correct, it implies that the surge in top wage incomes in the United States is a real phenomenon and not a unique consequence of the repackaging of income to avoid taxes. Last, France, the Netherlands, and Switzerland have experienced relatively small changes in their top tax rates, in contrast to the United States and the United Kingdom. Piketty (1999) shows that the small changes in the French top tax rates generated small shortterm responses from top income taxpayers but that those responses do not seem to persist over time. Switzerland has lower top-income tax rates than does the United States (around 35 percent when adding federal, cantonal, and local income taxes), but has much lower top income shares than does the United States (the top 1 percent share was around 8–9 percent in the 1990s, while it was between 13 and 17 percent in the United States). In sum, high income tax rates do not seem to account for the differences in top income shares across countries, although it is more debatable whether they can account for a substantial part of the time-series pattern within countries. Therefore, a systematic analysis of top incomes in countries 61 It might be the case, however, that for the top .1 percent incomes, the average decline in marginal tax rates has been much more modest. 62 Dilnot and Kell (1988) try to analyze this issue but have access only to a single year of micro tax returns and have to rely on aggregate numbers for their time-series analysis. 63 The top income tax rate in Canada, including provincial taxes, was about 50 percent in 2000.


171

that have experienced drastic cuts in top income tax rates in recent decades, as in the United States and the United Kingdom, would be of most interest. Those results could teach us whether a dramatic cut in top rates is necessarily associated with a rise in top incomes.

REFERENCES Atkinson, Anthony (2002). “Top Incomes in the United Kingdom over the Twentieth Century.” Nuffield College, Oxford. Mimeo. Atkinson, Anthony, and Wiemer Salverda (2003). “Top Incomes in the Netherlands and the United Kingdom over the Twentieth Century.” Nuffield College, Oxford. Mimeo. Auerbach, Alan (1988). “Capital Gains Taxation in the United States.” Brookings Papers on Economic Activity 2:595–631. Auten, Gerald, and Robert Carroll (1999) “The Effect of Income Taxes on Household Behavior.” Review of Economics and Statistics 81(4):681–693. Bertrand, Marianne, Esther Duflo, and Sendhil Mullainathan (2002). “How Much Should We Trust Differences-in-Differences Estimates?” NBER Working Paper no. 8841. March. Forth-coming in Quarterly Journal of Economics. 119(1), 2004. Carroll, Robert (1998). “Tax Rates, Taxpayer Behavior, and the 1993 Tax Act.” Office of Tax Analysis, U.S. Department of the Treasury. Working paper. November Number 79. Carroll, Robert, and David Joulfaian (1997). “Taxes and Corporate Choice of Organizational Form.” Office of Tax Analysis, U.S. Department of the Treasury. Working Paper no. 73. October. Dell, Fabien, Thomas Piketty, and Emmanuel Saez (2003). “The Evolution of Income and Wealth Concentration in Switzerland over the 20th Century.” University of California. Mimeo. Dilnot, Andrew, and Michael Kell (1988). “Top-Rate Tax Cuts and Incentives: Some Empirical Evidence.” Fiscal Studies 9:70–92. Eissa, Nada (1995). “Taxation and Labor Supply of Married Women: The Tax Reform Act of 1986 as a Natural Experiment.” NBER Working Paper no. 5023. February. Eissa, Nada, and Jeffrey Liebman (1996). “Labor Supply Response to the Earned Income Tax Credit.” Quarterly Journal of Economics 111:605–637. Feenberg, Daniel, and Elisabeth Coutts (1993). “An Introduction to the TAXSIM Model.” Journal of Policy Analysis and Management 12(1):189–194. Feenberg, Daniel, and James Poterba (1993). “Income Inequality and the Incomes of Very High Income Taxpayers: Evidence from Tax Returns.” Tax Policy and the Economy 7:145–177. Feenberg, Daniel, and James Poterba (2000). “The Income and Tax Share of Very High Income Households, 1960–1995”. American Economic Review 90(2):264–270. Feldstein, Martin (1995). “The Effect of Marginal Tax Rates on Taxable Income: A Panel Study of the 1986 Tax Reform Act.” Journal of Political Economy 103(3):551–572. Feldstein, Martin (1999). “Tax Avoidance and the Deadweight Loss of the Income Tax”. Review of Economics and Statistics 81(4):674–680.

172

Saez

Feldstein, Martin, and Daniel Feenberg (1996). “The Effect of Increased Tax Rates on Taxable Income and Economic Efficiency: A Preliminary Analysis of the 1993 Tax Rate Increases”. Tax Policy and the Economy 10:89–117. Goolsbee, Austan (1999). “Evidence on the High-Income Laffer Curve from Six Decades of Tax Reform.” Brookings Papers on Economic Activity 2:1–47. Goolsbee, Austan (2000a). “It’s Not About the Money: Why Natural Experiments Don’t Work on the Rich.” In Does Atlas Shrug? The Economic Consequences of Taxing the Rich, J. Slemrod (ed.). Cambridge, MA: Harvard University Press. Goolsbee, Austan (2000b). “What Happens When You Tax the Rich? Evidence from Executive Compensation.” Journal of Political Economy 108(2):352–378. Gordon, Roger, and Joel Slemrod (2000). “Are ‘Real’ Responses to Taxes Simply Income Shifting Between Corporate and Personal Tax Bases?” In Does Atlas Shrug? The Economic Consequences of Taxing the Rich, J. Slemrod (ed.). Cambridge, MA: Harvard University Press. Gruber, Jonathan, and Emmanuel Saez (2002). “The Elasticity of Taxable Income: Evidence and Implications.” Journal of Public Economics 84:1–32. Hall, Brian, and Kevin Murphy (2003). “The Trouble with Stock Options”. NBER Working Paper no. 9784. June. Katz, Lawrence, and David Autor (1999). “Changes in the Wage Structure and Earnings Inequality.” In Handbook of Labor Economics, Volume 3A, O. Ashenfelter and D. Card (eds.). Amsterdam: North-Holland. Kopczuk, Wojciech (2003). “Tax Bases, Tax Rates and the Elasticity of Reported Income.” NBER Working Paper no. 10044. October. Liebman, Jeffrey, and Richard Zeckhauser (2003). “Schmeduling”. Harvard University. Mimeo. Lindsey, Lawrence (1987). “Individual Taxpayer Response to Tax Cuts: 1982–1984, with Implications for the Revenue Maximizing Tax Rate.” Journal of Public Economics 33:173–206. Lindsey, Lawrence (1990). The Growth Experiment: How the New Tax Policy Is Transforming the U.S. Economy. New York: Basic Books, Inc. Meyer, Bruce, and Daniel Rosenbaum (2001). “Welfare, the Earned Income Tax Credit, and the Labor Supply of Single Mothers.” Quarterly Journal of Economics 116(3):1063–1114. Mirrlees, James A. (1971). “An Exploration in the Theory of Optimal Income Taxation.” Review of Economic Studies 38:175–208. Murphy, Kevin J. (1999). “Executive Compensation.” In Handbook of Labor Economics, Volume 3B, O. Ashenfelter and D. Card (eds.). Amsterdam: NorthHolland. Piketty, Thomas (1999). “Les Hauts Revenus Face aux Modifications des Taux Marginaux Supérieurs de l’Impôt sur le Revenu en France, 1970–1996.” Economie et Prévision 138–139:25–60. Piketty, Thomas (2003). “Income Inequality in France, 1901–1998.” Journal of Political Economy 111(5):1004–1042. Piketty, Thomas, and Emmanuel Saez (2003). “Income Inequality in the United States, 1913–1998.” Quarterly Journal of Economics 118(1):1–39. Series updated to year 2000 at http://www.nber.org/papers/w8467 (accessed March, 2003). Saez, Emmanuel (2001). “Using Elasticities to Derive Optimal Income Tax Rates.” Review of Economic Studies 68:205–229. Saez, Emmanuel (2002). “Do Taxpayers Bunch at Kink Points?” University of California. Mimeo.


173

Saez, Emmanuel (2003). “The Effect of Marginal Tax Rates on Income: A Panel Study of ‘Bracket Creep’.” Journal of Public Economics 87:1231–1258. Saez, Emmanuel (2004). “Reported Incomes and Marginal Tax Rates, 1960–2000: Evidence and Policy Implications.” NBER Working Paper no. 10273 February. Saez, Emmanuel, and Michael Veall (2003). “The Evolution of High Incomes in Canada, 1920–2000.” NBER Working Paper no. 9607. April. Sammartino, Frank, and David Weiner (1997). “Recent Evidence on Taxpayers’ Response to the Rate Increases in the 1990s.” National Tax Journal 50(3):683–705. Samwick, Andrew (1996). “Tax Shelters and Passive Losses After the Tax Reform Act of 1986.” In Empirical Foundations of Household Taxation, M. Feldstein and J. Poterba (eds.). Chicago, IL: University of Chicago Press. Scholes, Myron, and Mark A. Wolfson (1992). Taxes and Business Strategy: A Planning Approach. New York: Prentice Hall. Slemrod, Joel (1994). “On the High Income Laffer Curve.” In Tax Progressivity and Income Inequality, J. Slemrod (ed.). Cambridge, England: Cambridge University Press, 177–210. Slemrod, Joel (1995). “Income Creation or Income Shifting? Behavioral Responses to the Tax Reform Act of 1986.” American Economic Review 85(2):175–180. Slemrod, Joel (1996). “High Income Families and the Tax Changes of the 1980s: The Anatomy of Behavioral Response.” In Empirical Foundations of Household Taxation, M. Feldstein and J. Poterba (eds.). Chicago, IL: University of Chicago Press. Slemrod, Joel (1998). “Methodological Issues in Measuring and Interpreting Taxable Income Elasticities.” National Tax Journal 51(4):773–788. Slemrod, Joel, and Jonathan Bakija (2001). “Growing Inequality and Decreased Tax Progressivity.” In Inequality and Tax Policy, K. Hassett (ed.). Washington DC: American Enterprise Institute Press. Slemrod, Joel, and Wojciech Kopczuk (2002). “The Optimal Elasticity of Taxable Income.” Journal of Public Economics 84(1):91–112. Stewart, Kenneth J., and Stephen B. Reed (1999). “CPI Research Series Using Current Methods, 1978–98.” Monthly Labor Review 122(6):29–38.

THE (UN)CHANGING GEOGRAPHICAL DISTRIBUTION OF HOUSING TAX BENEFITS: 1980–2000 Todd Sinai University of Pennsylvania and NBER

Joseph Gyourko University of Pennsylvania

EXECUTIVE SUMMARY Using tract-level data from the 1980, 1990, and 2000 censuses, we estimate how the income-tax-related benefits to owner-occupiers are distributed spatially across the United States. Even though the top marginal tax rate has fallen substantially since 1979 and the tax code more generally has become less progressive, the tax subsidy per household or owner was almost unchanged between 1979 and 1989 and then rose substantially between 1989 and 1999. Geographically, gross program benefits have been and remain spatially targeted. At the state level, California’s owners have received a disproportionate share of the subsidy flows over the past two decades. Their share of the gross benefits nationally has fluctuated from 19 to 22 percent. Depending on the year, these percentages represent from 1.8 to 2.3 times California’s share of the nation’s owners. The median ratio of the share of tax benefits to the share of owners has declined over time, from 0.83 in 1979 to 0.76 in 1999. We are grateful to the National Bureau of Economic Research and the Research Sponsor Program of the Zell/Lurie Real Estate Center at Wharton for supporting this research, and to Daniel Feenberg, Jim Poterba, and Steven Sheffrin for helpful advice and comments. Dan Simundza provided excellent research assistance.

176

Sinai & Gyourko

Examining the data at the metropolitan-area level finds an even more dramatic spatial targeting and a spatial skewness that is increasing over time. Comparing benefit flows in 1979 in the top 20 areas versus those in the bottom 20 areas finds that owners in the highest subsidy areas received from 2.7 to 8.0 times the subsidy reaped by owners in the bottom group. By 1999, the analogous calculation finds owners in the top 20 areas receiving from 3.4 to 17.1 times more benefits than owners in any of the 20 lowest recipient areas. Despite the increasing skewness, the top subsidy recipient areas tend to persist over time. In particular, the highbenefit-per-owner areas are heavily concentrated in California and the New York City–Boston corridor. While taxes are somewhat higher in these places, it is high and rising house prices that appear most responsible for the large and increasing skewness in the spatial distribution of benefits.

1. INTRODUCTION It is generally accepted that the favorable tax subsidy to homeownership in the United States stimulates the demand for housing, raising prices and increasing the homeownership rate.1 The fact that this subsidy comes at a significant cost is also well documented at the national level, with several authors having estimated the tax expenditure associated with the mortgage interest and property tax deductions as well as the untaxed return on housing equity.2 Over time, these marginal incentives for homeownership—and the aggregate cost of these subsidies—have changed considerably. For example, Poterba’s (1992) analysis of the impacts of the various tax reforms of the 1980s reports a significant increase in the marginal cost of owneroccupied housing between 1980 and 1990 across the entire income distribution and particularly for high-income owners, mostly because of a drop in marginal tax rates for high-income households and an overall reduction in the progressivity of the tax code. Even so, we calculate that the real cost of the tax subsidy to homeownership has risen substantially in the last 20 years, from $198 billion (in 1999 dollars) in 1979, to $284 billion in 1989, and to $420 billion in 1999. In addition, recent evidence shows that the value of the subsidy to owner-occupied housing varies dramatically over space. In an earlier 1 See Rosen (1979) for a classic analysis, and see Bruce and Holtz-Eakin (1999); Capozza, Green, and Hendershott (1996); and the report to the Ford Foundation by Green and Reschovsky (2001) for more recent investigations into how the tax code might function in these instances. 2 For example, see Follain and Ling (1991), Follain, Ling, and McGill (1993) and Follain and Melamed (1998).

The (Un)changing Geographical Distribution of Housing Tax Benefits

177

study (Gyourko and Sinai, 2003), using 1990 census data, we found that the benefits of the tax subsidy are highly skewed, with just a handful of metropolitan areas reaping most of the net gains from the favored tax treatment of owner-occupiers. These sets of stylized facts naturally lead one to wonder whether the changes over time in marginal incentives for homeownership and in the aggregate cost of the homeownership subsidy have also affected the geographic distribution of the benefits. Because housing markets are inextricably tied to physical location and are not national in scope, knowing the extent to which the tax benefits vary spatially is important for determining the potential impact of any change in the tax treatment of owner-occupied housing. The nature of the spatial distribution of benefit flows is likely to be important for any consideration of the potential impacts on house prices, the homeownership rate, or the political economy of fundamental tax reform. In addition, knowing how the geographical distribution of program benefits changes is also useful for analysis of the spatial equity of the tax treatment of owner-occupied housing. Every year, for example, the Tax Foundation (Moody, 2003) calculates each state’s ratio of federal spending received to taxes paid and finds substantial variation across states. Our results, that the benefits of the subsidy to owner-occupied housing vary spatially, suggest that this sort of calculation should include implicit tax expenditures and subsidies alongside the observable taxes and spending. Indeed, many of the Tax Foundation’s states with the lowest ratios of spending to actual taxes paid are the same states whose homeowners receive the largest housing-related subsidies. In this paper, we examine how the spatial distribution of the tax subsidy to owner-occupied housing changed over three decades. Using the 1980, 1990, and 2000 censuses, we calculate the value of the tax subsidy to owner-occupied housing as the difference in ordinary state and federal income taxes currently paid by homeowners and the taxes they would pay if the tax code treated them like landlords. In the latter scenario, there is no preference for investing in one’s home relative to other assets. We find that the marginal tax subsidy for homeownership has decreased over the last 20 years on net, but the aggregate value of the tax benefits actually increased. Our analysis indicates that this increase is due to rising house prices and the growth in the number of homeowners more than offsetting the decline in average tax benefit per dollar of house. In particular, the after-tax cost of a dollar of owner-occupied housing rose between 1979 and 1989, before falling slightly by 1999, as the marginal tax rates on housing deductions were reduced and then increased. If all other factors were held constant, one would expect the value of the

178

Sinai & Gyourko

tax benefit to fall with tax rates. However, this scenario does not occur at the per-owner level, where the benefit remained flat during the 1980s before rising by 20 percent during the 1990s. The fact that the aggregate subsidy rose substantially during the 1980s, from $198 billion in 1979 to $284 billion in 1989, is due at least in part to growth in the number of homeowners. Regarding the spatial distribution of the subsidy, these tax changes, increases in house prices, and growth in the number of homeowners were not individually neutral. However, they happen to offset each other so that, at the state level, the spatial distribution of the tax benefits changes little over time. At the metropolitan-area level, however, spatial skewness of the subsidy has been increasing. This phenomenon appears to be driven by the relatively large increases in the price of houses experienced in various coastal areas of California and in the Northeast between New York City and Boston. Even so, the top recipients tend to persist; they just receive a larger fraction of the total subsidy over time. Among states, California always receives the largest gross subsidy flow, but this distribution is not due solely to the fact that it has the most owners. For example, in 2000, it received 18.7 percent of the aggregate subsidy, although it had only 9.4 percent of the nation’s owners. That high ratio of benefits to owners applies to only a small number of other states (such as New York, with 9.5 percent of total benefit flow while being home to only 5.3 percent of the nation’s owners in 2000), indicating that this program has highly spatially targeted beneficiaries. This pattern of spatial skewness related to the flow of program benefits is even more extreme at the metropolitan-area level. Comparing subsidy flows in 1979 in the top 20 areas versus those in the bottom 20 areas finds that owners in the high recipient areas received from 2.7 to 8.0 times the subsidy reaped per owner in the bottom group. By 1999, the analogous calculation finds the typical owner in the top 20 areas receiving from 3.4 to 17.1 times more benefits than owners in any of the 20 lowest recipient areas. The precise economic implications of these results depend on whether or not the subsidy is capitalized into land prices. While such an analysis is well beyond the scope of this paper, the broad range of possible outcomes can be readily understood. If the subsidy were fully capitalized, eliminating it would not affect the user cost of owning, but many owners in a few metropolitan areas would experience significant changes in wealth. While the savings associated with eliminating the subsidy would be redistributed back to homeowners, the net wealth effect could still be significant in many areas, regardless of how one thinks the tax benefits are financed. If the tax subsidy is not capitalized into land prices, then the user cost of ownership must reflect it.


179

The remainder of the paper is organized as follows. In section 2, we describe the tax subsidy to owner-occupied housing and how we measure it. Section 3 reports our results, beginning with an analysis of how benefits flow across states, followed by a description of the distribution across metropolitan areas. Finally, section 4 provides a brief conclusion.

2. MEASURING HOUSING-RELATED TAX BENEFITS The fact that there is a subsidy to owner-occupied housing can be seen most easily by comparing the current tax treatment of homeowners to how they would be taxed if housing were treated like any other asset. In particular, owner-occupied housing gets favorable tax treatment, but housing owned by a landlord is treated like any other income-producing, depreciable asset. Both homeowners and landlords are allowed to deduct mortgage interest and property taxes as expenses (as long as the homeowner itemizes). But a landlord must pay tax on her rental income, while a homeowner does not. The homeowner implicitly pays herself rent to occupy her house, but because she is both landlord and tenant, that transfer is tax-free. If the parties were distinct, however, the rent would be taxed. On the other hand, landlords can deduct depreciation and maintenance, while homeowners cannot. It is apparent from this comparison that the tax subsidy to owneroccupancy arises largely from the nontaxation of the implicit rent on the home. It is not so straightforward, however, to compute the amount of the benefit. Implicit rent is unobserved, and the components of landlords’ tax bills are often difficult to estimate. Instead, as we show below, it is much more straightforward to calculate the difference between the equilibrium taxes paid by homeowners and landlords. Underlying this approach is the same assumption used in the familiar user-cost-of-owning concept developed in Hendershott and Slemrod (1983) and Poterba (1984): the marginal homeowner invests in owner-occupied housing until the point where the annual cost she incurs exactly equals the rent she would have to pay as a tenant in the same property. We begin with the equilibrium annual flow cost of owning. That user cost is described in equation (1) and takes into account the fact that implicit rental income is untaxed, while mortgage interest and property taxes are deductible for itemizers: R H = (1 - x ded ) a i + (1 - x ded ) x p + (1 - x int )( 1 - a) r + (1 - x int ) b + M + d - P H

(1)

180

Sinai & Gyourko

The left-side variable, RH, is the annual cost of owner occupancy per dollar of housing value. This cost includes (1) the after-tax cost of mortgage interest, (1 − τded)αi, where α is the loan-to-value ratio on the house, i is the mortgage interest rate, and τded is the owner-occupier’s marginal tax rate equal to her marginal rate (denoted τint) if she itemizes and zero otherwise; (2) the after-tax cost of property tax payments, (1 − τded)τp, with τp equal to the effective property tax rate; (3) the after-tax opportunity cost of investing equity in the house rather than in some other riskless investment at rate of return, r, given by (1 − τint)(1 − α)r and is a cost to all owners, whether they itemize or not;3 (4) an after-tax risk premium, (1 − τint)β, to account for the difference in risk between bonds and housing, which applies to the entire long position in the house and thus is unaffected by the choice of leverage;4 (5) annual maintenance costs per unit of housing, which are given by M; (6) the cost of true economic depreciation per unit of house, which is assumed to occur at rate δ; and (7) any annual appreciation in the house value, ΠH, which reduces the carrying cost.5 If the homeowner were treated as a landlord, the residence would be taxed just like any other asset. Neutral tax treatment obviously requires taxing the implicit rental income on the home, but if treated like landlords, owner-occupiers would also be able to deduct maintenance expenses and depreciation, not just the mortgage interest and local property taxes presently allowed. In this case, a different annual cost would result, as described in equation (2): R H l = (1 - x) ai + (1 - x) x p + (1 - x)( 1 - a) r + (1 - x) b + xR H l + (1 - x) M + (1 - x) d - (1 - x) P H

(2)

3

We assume that the opportunity cost of tying up equity in a house is foregoing taxable returns. If the homeowner were to invest in a tax-exempt asset instead, we assume the return would be (1 − τ)r rather than r, yielding the same after-tax return.

4 In this framework, the homeowner’s financial position can be thought of as being long one house and short one bond (the mortgage). This approach allows us to decompose the opportunity cost of being long one house as the riskless rate of return plus a premium that reflects the difference in risk between a bond position and an equivalent-risk alternative to investing in housing. The difference between the mortgage interest rate and the equivalentduration riskless rate is reflected in the options to default on or prepay the mortgage. These options have value to the owner, so the premium above the riskless rate for borrowing is rolled into the mortgage rate as a cost. 5 This specification treats capital gains on housing as untaxed and realized every year. Because a $250,000 capital-gains exclusion ($500,000 for married couples filing jointly) can now be applied every other year, this approach is not unrealistic. Even in earlier periods, the assumption of no capital-gains taxation on housing was valid for the vast majority of households.


181

With perfect competition in the rental housing market, rents must equal the annual cost, so τRH′ would be the tax due on imputed rent.6 Grouping the RH′ terms and dividing both sides by (1 − τ) yields the simplified version in equation (3): R H l = ai + x p + (1 - a) r + b + M + d - P H

(3)

One possible strategy for estimating the tax benefits of owner-occupancy is to compute RH′ as the sum of the terms on the right side of equation (3), add that value to the homeowner’s reported income, and then determine the additional tax that would be paid. This approach has two important drawbacks. One is that we do not have good data on maintenance, depreciation, or expected capital gains, so the estimate is likely to be a noisy one. The other is that simply adding the implicit rent to income does not accurately capture the impact of itemization rates because the tax rates on deductions differ for nonitemizers. The alternative strategy we pursue in this paper is to compute the difference between RH′ and RH directly by subtracting equation (1) from equation (3):7 R H l - R H = x ded ai + x dedl (x p ) + x int 7 (1 - a) r + b A

(4)

This approach shows the impact of itemization correctly, and the terms we would have the most problems measuring accurately (M, δ, and Π) difference out in the subtraction. Thus, the tax subsidy to owner-occupancy can be computed as the sum of three components: (1) the tax value of home mortgage interest deductions (τded . α . i), (2) the tax value of local property tax deductions (τded . τp), and (3) the tax that would have been 6 This result also assumes accrual taxation of capital gains that, when combined with statutory ordinary income and with capital-gains rates being equal, allows us to focus on program benefits arising from differential tax treatment of ordinary income. As our 2003 paper (Gyourko and Sinai, 2003) shows, in this setting a dollar of house price appreciation has approximately the same value to owner-occupiers and landlords, so there is no differential impact on user costs. The analysis behind this conclusion is fairly complex, and we refer the interested reader to our 2003 paper for the details. 7 Note that we have abstracted throughout from the amount of housing dollars on which a homeowning family receives a subsidy. A change in the tax treatment of owner-occupied housing might affect house values, but because we measure the subsidy on a per-dollar basis, we abstract from the possibility that there is a second-order effect through changes in house prices. We follow this approach for two reasons. First, determining precisely how a change in the subsidy would be capitalized into house values is beyond the scope of this paper. Second, any change in house price would only increase the magnitudes of our estimates. For example, if the benefit to owner-occupied housing were reduced, house prices might also fall, further decreasing the subsidy.

182

Sinai & Gyourko

FIGURE 1. Calculating the Value of the Tax Subsidy paid on the equity invested in the home had it been invested elsewhere {τint $ [(1 − α) . r + β]}.8 While the sum of these three terms represents total ordinary income tax benefits to owner-occupiers under the current code, we hasten to emphasize that this does not imply that mortgage interest or local property tax deductions themselves are responsible for creating the subsidy. As noted above, the subsidy arises from the nontaxation of imputed rent and merely can be represented algebraically by the three terms on the right side of equation (4). Looking at the deductions alone would underestimate the true subsidy.

2.1 Estimation Strategy and Data The procedure for estimating the tax-code-related subsidy to owneroccupiers is represented graphically in the tax schedule with three marginal tax brackets shown in Figure 1. A homeowning family with no housing-related deductions would have a taxable income (TI) of Y1. If they were not owners, however, they may have invested their housing equity in 8 The depreciation term nets out because we have assumed that landlords can deduct economic depreciation and, after 1986, that assumption is probably not far from the truth. Deloitte and Touche (2000) and Gravelle (2001) conclude that economic lifetimes for rental properties in 1989 (and now) are somewhat shorter than the statutory lifetimes. The statutory depreciable life in 1981 (of 15 years) was shorter than true economic depreciation, so we may overestimate the subsidy to owner-occupiers in 1979.


183

a vehicle that yielded a taxable return that would raise their TI to Y2. Thus, Y2 is the counterfactual TI for a homeowning family if it were to stop being an owner. Starting with that TI, we can compute the tax value of each of the three aforementioned deductions. With a taxable income of Y2, this hypothetical family would have a tax liability of T1. Assume that claiming the home mortgage interest (HMI) deduction would lower TI to Y2 − HMI (presuming for simplicity that all of HMI was above the standard deduction) and the tax liability to T2. Therefore, the tax savings for this family from the mortgage interest deduction is T1 − T2. In this example, the mortgage interest deduction does not move the family into a lower tax bracket, but the property tax deduction does. Beginning with TI equal to Y2 − HMI, we can compute the tax savings from the property tax deduction as the tax bill with only the mortgage interest deduction, T2, minus the tax bill with both the mortgage interest and property tax deductions, T3. In this case, T2 and T3 span a kink in the tax schedule, but they still account for the fact that the average tax rate is less than the marginal tax rate at Y2 − HMI. Finally, we compute the value of the nontaxation of the return on housing equity. Because the return on housing equity is not included in TI, taxable income is measured at Y1 instead of the greater amount Y2. The tax value of not including that income is measured as the change in tax between T3 (the tax bill corresponding to a TI of Y2 − HMI − Tp) and T4 (the tax bill corresponding to a TI of Y1 − HMI − Tp). It is apparent from Figure 1 that the order in which the deductions are taken matters when the tax schedule is not linear. For example, T1 − T2 > T3 − T4, even though HMI < Y1 − Y2. After adding back the implicit return on housing equity, we compute the deductions in the following order: (1) tax savings from the mortgage interest deduction, (2) the tax savings associated with the property tax deduction, and (3) the savings from the return on housing equity not being taxed. We have repeated the estimation using all six possible sequences in which the deductions can be taken. While the relative magnitudes of the categories change, the differences are minor. We calculate each of the tax liabilities T1 through T4 by combining tract level information covering the entire United States from the STF3 files of the 1980, 1990, and 2000 decennial censuses (U.S. Bureau of the Census, 1980, 1990, 2000) with the National Bureau of Economic Research (NBER). TAXSIM program (Feenberg and Coutts, 1993). TAXSIM calculates federal and state tax liabilities from our tax data and allows us to engage in a “what if” calculation to determine what taxes would have been paid had a household not had various housing deductions or had invested in an asset with a taxable income stream. For each year in our

184

Sinai & Gyourko

data, the TAXSIM program incorporates all relevant federal and state tax law, including housing and property tax deductions. To construct representative households to pass through the TAXSIM tax calculator, we start by computing the distribution of household income among homeowners at the tract level.9 For each tract, we divide the household income distribution into deciles and assign the median income for each decile to all the households in that category. Thus, the one-tenth of the households with the lowest-income is assumed to have an income equal to that of the fifth percentile for the tract, the next lowest-income tenth of the households is assigned an income equal to that of the 15th percentile for the tract, and so forth. We then map tract-level information on the distribution of house values, PH, to incomes by assigning to households in each decile of the income distribution the value corresponding to the same decile of the house value distribution. For example, we assume that the household in the 5th percentile of the income distribution for the tract also owns the home in the 5th percentile of the housing price distribution for the same tract.10 The actual value of the tax benefits depends on certain demographic data that are likely to affect the number of exemptions and the overall amount of deductions. Tract-level data that are available in each census year include the distribution of households according to their description as single, married, or single with children; the percentage of households with children; and the percentage of households with at least one member over 65 years of age. We create a representative household for each possible combination of these characteristics and then compute the weighted average estimated tax, where the weights are the tract-level distributions of the demographic characteristics. The census data lack information on most non housing categories of potential tax deductions. We compute mortgage interest, state tax, and property tax deductions, but we do not observe medical expenses, charitable 9 All tax-benefit figures reported in this paper are based on tract-level data that aggregates household income across its various sources. 10 This matching process presumes that owners and renters in a tract have identical income distributions. Fortunately, our spatial results are robust to assuming an extreme case in which all the owners in a tract have a higher income than any of the renters, and houses are matched to owners so that the highest-income owner owns the highest-value house, the next highest-income owner occupies the next highest-value house, and so forth. In reality, any sorting into houses by income would not be perfect, as is suggested by the data in O’Sullivan, Sexton, and Sheffrin (1995), who match tax returns and property tax assessments in California. Unfortunately, those data are no longer available. For the 1989 data, however, we have tried using the mean income and house value in each tract, rather than the full distribution, and it does not make any qualitative difference to the spatial skewness we observe.


185

giving, deductible interest (other than for a home mortgage), and several other miscellaneous categories. Two countervailing problems arise from underestimating possible deductions. First, we would be more likely to assume incorrectly that the family does not itemize. This error would cause us to underestimate the tax value of the mortgage interest and property tax deductions because less would be deducted at the margin. On the other hand, omitting deductions for itemizers could increase the tax value we do measure because the remaining deductions are applied against higher marginal tax rates. Consequently, we impute missing tax deductions to our census data based on data from the Department of the Treasury’s Statistics of Income (SOI) public-use tax microsample. A modified Heckman-style sample selection model is employed to correct for the selective observing of deductions only by itemizers.11 Following the procedure shown in Figure 1, we augment the observed income by an estimate of how much higher the household’s income would have been had its members invested in an equivalently risky taxable asset rather than housing. First, we calculate the opportunity cost of the equity in one’s home, or PH* [(1 − α)*r + β], where r is the riskless yield on seven-year Treasuries in the relevant census year: 9.47, 8.57, and 5.79 percent, respectively. Then we compute β: the risk premium for the whole house.12 The estimates below assume that the expected equivalent-risk opportunity cost of investing in a house is equal to the geometric mean on the value-weighted Standard & Poor’s S&P500 return (including dividends) over a certain time period. For simplicity, we assume that the relevant period always runs from the beginning of 1926 to the end of the census year (i.e., 1926–1979, 1926–1989, and 1926–1999), yielding expected 11 The interested reader should see the appendix to Gyourko and Sinai (2003) for a detailed description of the procedure. The imputation results indicate that, without the correction, we would have underestimated deductions and therefore the number of itemizers. This turns out to be important because the underestimation of itemizers was not random across space. In high-house-value and high-income-tax states such as California, not observing nonhousing deductions only infrequently caused us to miscategorize an owner family as a nonitemizer. Home mortgage interest, local property taxes, and state income taxes generally were sufficient to make California residents itemizers. This scenario was not the case in many states with lower house values and lower state taxes. Hence, the imputation has an important effect on the measured spatial distribution of program benefits. 12 The risk adjustment follows from Poterba (1991), with the calculation effectively assuming that the mortgage rate would be the yield on seven-year Treasuries in the absence of the options to prepay or default. Other assumptions regarding the relative risk of owneroccupied housing obviously could be made because no clear agreement exists on this issue. However, we have repeated all the analyses reported in the paper under widely varying assumptions about the relative risk of owner-occupied housing. While the aggregate subsidy certainly does vary with the presumed opportunity cost of equity in the home, the nature of the spatial distribution of the subsidy across states and metropolitan areas is largely unaffected.

186

Sinai & Gyourko

returns of 8.79, 10.13, and 11.22 percent, respectively. The risk premium is the difference between this yield and the risk-free yield. Thus, for 1989, we define β to be the 10.13 percent S&P500 return minus the 8.57 percent Treasury yield, for a premium of 1.56 percentage points. The opportunity cost of riskless equity and the risk premium are then added to income. We estimate the value of the mortgage interest deduction by computing each tract-decile’s tax value as the weighted average difference in tax bills with and without it. The mortgage interest deduction itself is defined as PH*α*i. Leverage ratios, α, vary by age and are computed from household data in the Survey of Consumer Finances (SCF) closest in time to the relevant census year. A weighted average leverage for each tract was computed based on the tract’s age distribution.13 The mortgage interest rate, i, was calculated by taking an average across households in the same SCFs. From the 1983 SCF, which is the closest in time to 1979, we calculate the average mortgage rate to be 10.21 percent. For 1989, the analogous rate was 9.56 percent, with a rate of 7.85 percent matched from the 1998 SCF to the 1999 census data. The tax value of the mortgage interest deduction can differ from mortgage interest paid times the marginal tax rate for three reasons. First, only families that itemize on their tax returns receive any benefit on the margin from the deductibility of mortgage interest. Also, only the excess of the mortgage interest deduction plus other itemized deductions over the standard deduction has value for a taxpayer. Therefore, we would multiply only the portion of mortgage interest in excess of the standard deduction (after itemizing all other non-housing-related deductions first) by the tax rate. Because the tax schedule is nonlinear, taking the mortgage interest deduction may lower the taxpayer’s marginal and average tax rates. The second component involves the value of the deduction of local property taxes. Property tax payments themselves are defined as PH*τp, where τp is the average effective property tax rate. We were not able to find reliable estimates for this variable over time. Consequently, we use information for an intermediate year—1990.14 This variable is allowed to vary by metropolitan area using data provided by Stephen Malpezzi, who 13 There is considerable heterogeneity in leverage by age in all years. For example, in 1998, loan-to-value ratios by age are as follows: 20- to 24-year-olds: 66.5 percent, 25- to 29-yearolds: 64.2 percent, 30- to 34-year-olds: 62.6 percent, 35- to 39-year-olds: 61.0 percent, 40- to 44-year-olds: 52.3 percent, 45- to 49-year-olds: 44.5 percent, 50- to 54-year-olds: 41.3 percent, 55- to 59-year-olds: 30.9 percent, 60- to 64-year-olds: 21.3 percent, 65- to 69-year-olds: 13.2 percent, 70- to 74-year-olds: 9.6 percent, and 75-year-olds, and older: 4.6 percent. Leverage in previous decades is lower, on average. 14 Property taxes are such a small component of the total subsidy—about 10 percent—that the noise in this measure probably has little qualitative effect on our conclusions.


187

has calculated average property tax rates in 1990 for a large number of areas. Census tracts not located within metropolitan areas covered in the Malpezzi data are assigned the average state-level local property tax rate as reported by the Advisory Commission on Intergovernmental Relations (ACIR) (1987).15 The tax value of the deduction associated with these payments is then computed the same way as for the mortgage interest deduction. The third term we estimate arises from the fact that the government does not tax as income the return homeowners could have earned on their equity had they not invested in their homes. We calculate the reduction in tax liabilities that occurs when we remove the imputed income that we had added in the first step. This approach accounts for the possibility that a family might move into a higher marginal tax bracket if the return on its housing equity were taxed.

3. RESULTS 3.1 Summary Statistics for the Nation The national aggregate gross value to owners of housing-related ordinary income tax benefits, reported in the second column of Table 1, is quite large and has risen over time—from $198 billion in 1979 to $284 billion in 1989, to $420 billion in 1999 (in constant 1999 dollars).16 These subsidies are large and are significantly higher than those typically reported by the Treasury or the Joint Committee on Taxation primarily because those government agencies calculate only the traditional tax expenditures—the tax cost of the mortgage interest and property tax deductions—rather than the failure to tax implicit rent. Because houses are leveraged only partially and the expected return on a house is greater than mortgage rates, those deductions measure only a portion of the true tax expenditure.17 In addition, our figures include state tax subsidies. 15 The ACIR did not report state-by-state breakdowns for 1989, so we use the 1987 data. We have also experimented with assuming a 1 percent and a 1.5 percent national average effective rate. Our findings are not sensitive to these changes. 16 The bulk of the tax-code-related benefits to owners arises from the third of the three components from equation (4). Depending on the census year, from two-thirds to three-quarters of the total benefits are due to not having to pay tax on the return to equity invested in the home plus the difference in expected return on housing versus the cost of the mortgage. Results on the decomposition of the subsidy are available on request. 17 Our estimates of the tax savings from the mortgage interest deduction alone are quite close to, but lower than, what we obtain by looking at actual tax return data. We cannot use the Statistics of Income (SOI) data to compute the full tax expenditure because tax return data do not include information about house values, only itemized deductions. In addition, the SOI data do not report state of residence for taxpayers with adjusted gross income (AGI)

188

Sinai & Gyourko TABLE 1 Aggregate Tax Subsidy, National Level, by Year

Year

Total (billions of $1999)

Per owner ($1999)

Per household ($1999)

1979 1989 1999

$197.9 284.0 420.1

$4,840 4,818 6,024

$3,023 3,121 4,015

The housing subsidy is sizable—and growing—even on a per-owner or per-household basis. While the aggregate real subsidy amount increased 112 percent since 1979, the number of owner-occupied units rose just 70 percent between 1979 and 1999 (from 40.9 million in 1979 to 69.7 million in 1999), so the subsidy per owner-occupied household has been increasing. Gross program benefits per owner-occupied household were $4,840 in 1979, remained constant over the ensuing decade (with the 1989 figure being $4,818), and then rose in the 1990s to $6,024 in 1999. The analogous figures on a per-household basis range from just over $3,000 in 1979 to just over $4,000 in 1999. While it has long been understood that the subsidy is skewed in aggregate toward those with high incomes and high house values, much less is known about the spatial skewness of this aspect of the tax code. We turn now to this issue. We begin by documenting just how the tax subsidy to owner-occupied housing is skewed, describe how that skewness changes over time, and then investigate the factors driving any changes in the distribution of the subsidy across states and metropolitan areas.

3.2 State-Level Results While we will focus most of our analysis on the amount of tax benefits per owner, we begin with the most basic measure of the spatial distribution of the benefits: the aggregate benefit flow for each state by year. Not surprisingly, the most populous state, California, stands out in Table 2, with its owners receiving gross benefits of nearly $40 billion in 1979, well over $60 billion in 1989, and almost $80 billion in 1999. No other state approaches these levels, although the benefit flow to New York has risen dramatically over time. A closer examination shows that, as the national aggregate value of the subsidy increases, the additional benefits appear to be distributed in rough proportion to where they were already going. above a threshold, so our calculations using the SOI are also below the true figure. On the other hand, projected tax expenditure on mortgage interest deductions for 1999 (these do not include state taxes) from the Joint Committee on Taxation’s (1998) is slightly lower than what we calculate.


189

TABLE 2 Aggregate Benefit Flow in Billions of $1999 by State, 1979, 1989, and 1999 State Alabama Alaska Arizona Arkansas California Colorado Connecticut Delaware District of Columbia Florida Georgia Hawaii Idaho Illinois Indiana Iowa Kansas Kentucky Louisiana Maine Maryland Massachusetts Michigan Minnesota Mississippi Missouri Montana Nebraska Nevada New Hampshire New Jersey New Mexico New York North Carolina North Dakota Ohio Oklahoma Oregon Pennsylvania Rhode Island South Carolina South Dakota Tennessee

1979

1989

1999

$1.80 $0.38 $2.69 $0.65 $38.07 $3.37 $4.29 $0.58 $0.99 $8.61 $3.63 $1.81 $0.43 $9.92 $3.01 $1.43 $1.77 $1.28 $2.22 $0.54 $4.53 $5.12 $10.39 $4.11 $1.01 $2.61 $0.43 $0.76 $0.82 $0.64 $8.96 $0.84 $15.20 $2.59 $0.26 $8.09 $1.77 $2.87 $8.80 $0.80 $1.48 $0.23 $2.26

$2.25 $0.40 $3.23 $1.17 $63.73 $3.07 $8.10 $0.89 $1.23 $11.83 $5.30 $2.70 $0.65 $11.87 $3.31 $1.70 $1.94 $1.89 $2.04 $1.37 $7.42 $11.84 $9.92 $4.14 $1.11 $3.64 $0.49 $0.85 $0.93 $1.60 $15.01 $1.12 $32.99 $5.03 $0.27 $7.82 $1.72 $2.50 $10.45 $1.48 $2.48 $0.24 $2.84

$4.18 $0.67 $6.55 $2.09 $78.66 $8.56 $8.23 $1.20 $1.41 $19.62 $10.49 $2.91 $1.55 $19.71 $6.13 $3.07 $2.93 $3.81 $3.49 $1.59 $9.56 $14.03 $17.59 $7.67 $2.00 $6.11 $1.04 $1.67 $2.30 $1.74 $17.60 $2.15 $39.72 $10.54 $0.41 $13.32 $2.67 $6.48 $13.82 $1.49 $4.76 $0.48 $5.61 Continued

190

Sinai & Gyourko TABLE 2—Continued

State

1979

1989

1999

Texas Utah Vermont Virginia Washington West Virginia Wisconsin Wyoming

$9.12 $1.22 $0.11 $5.30 $4.04 $0.87 $4.90 $0.31

$8.88 $1.14 $0.59 $7.82 $4.77 $0.90 $5.11 $0.22

$15.60 $3.21 $0.72 $10.90 $9.52 $1.40 $8.64 $0.46

That is, while the aggregate benefit to California doubles between 1979 and 1999, so does the subsidy to small beneficiaries such as Georgia, Maryland, and North Carolina. Thus, the states tend to maintain their same relative standing, but the absolute (real) dollar difference between the highest and lowest recipient increases substantially. Of course, changes in aggregate subsidy flows are heavily affected by population growth. To net out differential increases in the number of homeowners, Figure 2 reports benefits scaled by the number of owners in each state in 1979 and 1999.18 Even on a per-owner basis, people in only a handful of states, often the most populous states, reap substantially more from tax-code-related housing benefits than the typical owner nationally. For example, while California is no longer the extreme outlier it was in the aggregate data in Table 2, it is still one of only seven states that received at least $6,000 per owner in 1979 and at least $8,000 per owner in 1999. Overall, the per-owner subsidies in the top few states are well over double those received by owners in the vast majority of states. Thus, while the Gini coefficients for the distribution of per-owner benefits across states are relatively low in each decade (0.20 in 1979, 0.32 in 1989, and 0.25 in 1999), it would not be accurate to consider the benefit distribution an especially egalitarian one in spatial terms. Although the subsidy per owned unit has risen over time, the skewness has persisted at least since 1979. Benefit flows are always concentrated in the hands of owners in just a few states, and the top three states have remained at the top for the last 20 years. The spatial distribution has changed some, however, with owners in northeastern states doing better over time. Of course, Figure 2 confounds changes in the national level of subsidy with its distribution across space. However, the typical state receives less than the national average benefit per owner, with a few states receiving 18

Data for all three years—1979, 1989, and 1999—are reported in Appendix Tables A and B, which are available in NBER Working Paper 10322 and at www.nber.org/ data/tpe18.


191

FIGURE 2. Average Tax Benefits per Owned Unit, by State, in 1979 and 1999 about double the average. These disparities rise between 1979 and 1989 but are mitigated somewhat by 1999.19 To isolate the spatial distribution from the dollar value of the subsidy, we have computed the ratio of each state’s share of the subsidy to its share of the nation’s owners. For example, the median state has a ratio of subsidy share to owner share of 0.83 in 1979, 0.71 in 1989, and 0.76 in 1999. These ratios are generally less than half of California’s numbers, which are 1.77 in 1979, 2.29 in 1989, and 2.00 in 1999.20 Figures 3 and 4 provide more detail on the heterogeneity in benefit changes by state over the 1980s and 1990s. Both figures measure each state’s changes relative to the national average change. Figure 3 shows that owners in northeastern and mid-Atlantic states did better than average in the 1980s. California and Hawaii are the only exceptions to that statement. There was less heterogeneity in the 1990s, when owners in the 19 While one cannot compute transfers across states without making assumptions regarding how the program is financed, it seems certain that transfers are flowing from a host of states to owners in California and a select few other states. See our 2003 paper (Gyourko and Sinai, 2003) for transfer estimates assuming lump-sum and proportional financing schemes using 1990 data. In both cases, the outcome is the majority of states transferring resources to owners in the smaller number of other states. 20 While ratios for Hawaii and the District of Columbia are higher in each decade, ratios for California are more relevant empirically because of the state’s large number of owners.

192

Sinai & Gyourko

FIGURE 3. Change in Average Benefits per Owner Relative to National Average, by State, 1979–1989 ($1999)

FIGURE 4. Change in Average Benefits per Owner Relative to National Average, by State, 1989–1999 ($1999) less populous western states of Colorado, Oregon, and Utah experienced significantly greater than average increases. Owners in California and Hawaii received smaller than average benefit flow increases that decade. As suggested in the introduction, many factors have changed over time that could influence the value of the tax benefits associated with owner occupancy. The most obvious is the tax rates themselves. Because owner-


193

occupied housing is a true tax shelter in the sense that one can deduct expenses without declaring any income on the asset, a reduction in tax rates naturally lowers the value of the tax shelter. Figure 5 plots the “average” marginal tax rate (state plus federal) on housing deductions for 1979 and 1999, calculated using the census data and the NBER’s TAXSIM program. While marginal rates do differ across states, these differences have declined over time. Overall, marginal rates fell significantly during the 1980s and then rose modestly during the 1990s because of a series of tax reforms at the federal level.21 However, aggregate benefits rose and benefits per owner did not decline on average between 1979 and 1989; these facts indicate other factors were changing to counterbalance the negative effect that an increase in the tax price of housing would have on the value of the benefit. In addition, the fact that most of the important tax changes were at the federal level may help explain why the nature of the spatial distribution across states was not affected much. Of course, other components of the subsidy, house prices in particular, were changing. Figure 6 graphs mean house price by state in 1979, 1989, and 1999. Figure 7 reports the percentage changes over time for each state. Values in many of the coastal states in particular have skyrocketed over the past 20 years. In California, mean real prices rose from just over $200,000 in 1979 to nearly $300,000 in 1999. The change has been even more dramatic in places like Massachusetts, where the average home was worth a little more than $100,000 in 1979. One decade later, mean prices had doubled (in real terms), and prices held firm in Massachusetts during the 1990s. It seems clear that this type of change has allowed the average subsidy per owner in Massachusetts to rise so much over the past two decades. Indeed, a comparison of Figures 3, 4, and 7 suggests that rising real house prices can help account for the dramatic increases in benefits per owner that have occurred in a small number of states, especially northeastern states, in the 1980s. Of course, other factors, including the rising return in equity markets, which raises the value of the tax shield on home equity in our calculations, are also at work. While a detailed decomposition analysis of changes in the tax benefit over time is beyond the scope of this paper, the data show that the factors that do change did so in a largely offsetting fashion with respect to the spatial distribution across states in the 1980s. The rise in aggregate and per-owner benefits in the 1990s probably reflects 21 Like tax rates, the probability of itemizing declined significantly between 1979 and 1999, reducing the subsidy to owner-occupied housing. Changes in the spatial distribution of itemizers, once one nets out the effect of house prices on the likelihood of itemization, do not seem to determine the changes in the benefits. This result is not surprising because we saw in section 2 that itemization affects the value of only a small portion of the tax subsidy.

194 Sinai & Gyourko

FIGURE 5. Average Marginal Tax Rates, by State, in 1979 and 1999

FIGURE 6. Average House Prices, by State, in 1979, 1989, and 1999

196

Sinai & Gyourko

FIGURE 7. Percentage Change in Mean House Prices, by State, in 1979–1989 and 1989–1999 a growing share of households that are owners, rising real house prices, and increasing tax rates. On net, the spatial distribution of benefits across states is fairly skewed in each census year, with few states experiencing significant changes in their relative status. Whether this holds at the metropolitan-area level is the subject of the following subsection.

3.3 Metropolitan Area-Level Results In this subsection, we disaggregate the data further to examine subsidy flows at the metropolitan-area level and find that the distribution of housing benefits is more skewed than at the state level and that skewness is


197

increasing over time. Results are computed for 380 areas that were identifiable census Core-Based Statistical Areas (CBSAs).22 Aggregate benefit flows at the CBSA level, which are reported for selected areas in appendix tables A and B, document how extremely spatially targeted are the overall benefit flows.23 The vast majority of metropolitan areas receive a relatively modest benefit flow, while a relatively small number of areas receive large aggregate benefit flows. This form of spatial skewness also has increased over time at the metropolitan-area level. For example, if we focus on the three CBSAs that contain the nation’s three largest cities, New York City, Los Angeles, and Chicago, their homeowners received benefit flows equal to $27.3 billion in 1979. While being home to just 10.1 percent of all owners living in designated metropolitan areas in the 1980s, these owners received 14.7 percent of all benefits flowing to metropolitan census tracts. By 1989, the spatial skewness of aggregate tax subsidy flows had become even more extreme. Owners in just these three CBSAs received 17.7 percent of all metropolitan-area benefits while constituting an even smaller share of the nation’s owners, at 9.3 percent. The share of owners in these areas had fallen to 8.5 percent by 1999, but their benefit share was 1.72 times higher, at 14.6 percent. Figure 8 plots benefits scaled by the number of owners in the CBSA. The figure highlights the fact that the subsidy flows disproportionately toward owners in a relatively small number of metropolitan areas and that the skewness is increasing over time. In this figure, CBSAs are ordered by their per-owner subsidy. Thus, the more extreme curvature in the graphs as the decades progress is an indication that spatial skewness, net of population changes, has been on the rise. This scenario is made even more clear in Tables 3 and 4, which report the top and bottom 20 CBSAs in terms of benefits per owner in 1979 and 1999, respectively. (We limit our consideration to the 179 CBSAs that are 22 Benefit flows to census tracts not located within CBSAs are not included in the figures reported in this section. CBSAs are the new (2003) county-based definition of metropolitan areas from the U.S. Bureau of the Census. We apply the same definition in each of the three census files, knowing that the economic relationship among the counties is weaker, of course, in previous decades. By construction, a CBSA must contain at least one urban area of 10,000 or more population. The county (or counties) “in which at least 50 percent of the population resides within urban areas of 10,000 or more population, or that contain at least 5,000 people residing within a single urban area of 10,000 or more population, is identified as a ‘central county’ ” and is included in the CBSA. Additional “outlying counties” are included in the CBSA if they meet specified requirements of commuting to or from the central counties. 23 Appendix Tables A and B, from NBER Working Paper 10322, can be accessed at www.nber.org/data/tpe.

198

Sinai & Gyourko

FIGURE 8. Benefits per Owner, by Metropolitan Area, in 1979, 1989, and 1999

Top 20 areas by per-owner subsidy CBSA name Honolulu, HI Metropolitan Statistical Area San Francisco–San Mateo–Redwood City, CA Metropolitan Division San Jose–Sunnyvale–Santa Clara, CA Metropolitan Statistical Area Santa Barbara–Santa Maria–Goleta, CA Metropolitan Statistical Area Santa Ana–Anaheim–Irvine, CA Metropolitan Division Bethesda–Frederick–Gaithersburg, MD Metropolitan Division Bridgeport–Stamford–Norwalk, CT Metropolitan Statistical Area Los Angeles–Long Beach–Glendale, CA Metropolitan Division San Diego–Carlsbad–San Marcos, CA Metropolitan Statistical Area Lake County–Kenosha County, IL–WI Metropolitan Division Anchorage, AK Metropolitan Statistical Area Santa Cruz–Watsonville, CA Metropolitan Statistical Area Oxnard–Thousand Oaks–Ventura, CA Metropolitan Statistical Area Oakland–Fremont–Hayward, CA Metropolitan Division Washington–Arlington–Alexandria, DC–VA Metropolitan Division Salinas, CA Metropolitan Statistical Area Milwaukee–Waukesha–West Allis. WI Metropolitan Statistical Area Santa Rosa–Petaluma, CA Metropolitan Statistical Area Ann Arbor, MI Metropolitan Statistical Area Warren–Farmington Hills–Troy, MI Metropolitan Division

Subsidy per owner-occupied unit

Subsidy per household

$13,491 $13,126 $11,320 $10,731 $10,719 $10,669 $10,189 $9,585 $8,758 $8,637 $8,616 $8,598 $8,553 $8,427 $8,349 $8,037 $7,738 $7,677 $7,483 $7,387

$7,132 $6,156 $6,830 $5,615 $6,430 $7,080 $6,870 $4,621 $4,813 $6,236 $4,843 $5,332 $5,744 $4,856 $4,425 $4,549 $4,611 $4,857 $4,094 $5,689 Continued


TABLE 3 Benefits per Owner and per Household, Select CBSAs Above Median Population, 1979, in $1999

199

200

TABLE 3—Continued

CBSA name McAllen–Edinburg–Pharr, TX Metropolitan Statistical Area Waco, TX Metropolitan Statistical Area Fort Smith, AR–OK Metropolitan Statistical Area Lakeland–Winter Haven, FL Metropolitan Statistical Area Killeen–Temple–Fort Hood, TX Metropolitan Statistical Area Kingsport–Bristol, TN–VA Metropolitan Statistical Area Pensacola–Ferry Pass–Brent, FL Metropolitan Statistical Area Scranton–Wilkes-Barre, PA Metropolian Statistical Area Columbus, GA–AL Metropolitan Statistical Area Johnstown, PA Metropolitan Statistical Area Jacksonville, FL Metropolitan Statistical Area Deltona–Daytona Beach–Ormond Beach, FL Metropolitan Statistical Area Chattanooga, TN–GA Metropolitan Statistical Area Fayetteville, NC Metropolitan Statistical Area Beaumont–Port Arthur, TX Metropolitan Statistical Area Hickory–Morganton–Lenoir, NC Metropolitan Statistical Area Huntington–Ashland, WV–KY–OH Metropolitan Statistical Area Macon, GA Metropolitan Statistical Area Augusta–Richmond County, GA–SC Metropolitan Statistical Area Springfield, MO Metropolitan Statistical Area Note: Median number of households in 1979 among all 380 CBSAs is 56,664.



$1,687 $2,010 $2,177 $2,180 $2,247 $2,294 $2,303 $2,307 $2,418 $2,418 $2,564 $2,599 $2,602 $2,612 $2,628 $2,652 $2,652 $2,667 $2,733 $2,760

$1,173 $1,252 $1,490 $1,539 $1,332 $1,751 $1,614 $1,564 $1,496 $1,740 $1,670 $1,848 $1,800 $1,619 $1,885 $2,001 $1,896 $1,666 $1,855 $1,851

Sinai & Gyourko

Bottom 20 areas by per-owner subsidy

Top 20 areas by per-owner subsidy CBSA name San Francisco–San Mateo–Redwood City, CA Metropolitan Division San Jose–Sunnyvale–Santa Clara, CA Metropolitan Statistical Area Bridgeport–Stamford–Norwalk, CT Metropolitan Statistical Area Santa Barbara–Santa Maria–Goleta, CA Metropolitan Statistical Area Suffolk County–Nassau County, NY Metropolitan Division Oakland–Fremont–Hayward, CA Metropolitan Division New York–Wayne–White Plains, NY–NJ Metropolitan Division Santa Ana–Anaheim–Irvine, CA Metropolitan Division Salinas, CA Metropolitan Statistical Area Honolulu, HI Metropolitan Statistical Area Santa Rosa–Petaluma, CA Metropolitan Statistical Area Oxnard–Thousand Oaks–Ventura, CA Metropolitan Statistical Area Cambridge–Newton–Framingham, MA Metropolitan Division Los Angeles–Long Beach–Glendale, CA Metropolitan Division Boulder, CO Metropolitan Statistical Area San Diego–Carlsbad–San Marcos, CA Metropolitan Statistical Area Bethesda–Frederick–Gaithersburg, MD Metropolitan Division Boston–Quincy, MA Metropolitan Division Newark–Union, NJ–PA Metropolitan Division Lake County–Kenosha County, IL–WI Metropolitan Division



$26,385 $24,629 $17,418 $16,759 $15,655 $15,151 $14,776 $14,593 $14,554 $14,115 $13,030 $12,895 $12,643 $12,096 $11,855 $11,641 $11,223 $10,941 $10,870 $10,700

$13,327 $14,874 $12,075 $9,593 $12,520 $9,189 $6,123 $8,953 $7,994 $7,944 $8,338 $8,734 $7,804 $5,845 $7,719 $6,476 $7,894 $6,389 $6,823 $8,127 Continued


TABLE 4 Benefits per Owner and per Household, Select CBSAs Above Median Population, 1999, in $1999

201

202

Bottom 20 areas by per-owner subsidy CBSA name McAllen–Edinburg–Pharr, TX Metropolitan Statistical Area Brownsville–Harlingen, TX Metropolitan Statistical Area Beaumont–Port Arthur, TX Metropolitan Statistical Area EI Paso, TX Metropolitan Statistical Area Lubbock, TX Metropolitan Statistical Area Corpus Christi, TX Metropolitan Statistical Area Killeen–Temple–Fort Hood, TX Metropolitan Statistical Area Huntington–Ashland, WV–KY–OH Metropolitan Statistical Area Ocala, FL Metropolitan Statistical Area Lakeland–Winter Haven, FL Metropolitan Satistical Area Fort Smith, AR–OK Metropolitan Statistical Area Kingsport–Bristol. TN–VA Metropolitan Statistical Area Deltona–Daytona Beach–Ormond Beach, FL Metropolitan Statistical Area Shreveport–Bossier City, LA Metropolitan Statistical Area San Antonio, TX Metropolitan Statistical Area Pensacola–Ferry Pass–Brent. FL Metropolitan Statistical Area Youngstown–Warren–Boardman, OH–PA Metropolitan Statistical Area Charleston, WV Metropolitan Statistical Area Mobile, AL Metropolitan Statistical Area Scranton–Wilkes-Barre, PA Metropolitan Statistical Area Note: Median number of households in 1999 among all 380 CBSAs is 92,249.



$1,541 $1,696 $2,027 $2,153 $2,326 $2,341 $2,345 $2,448 $2,466 $2,528 $2,537 $2,789 $2,866 $2,873 $2,931 $3,000 $3,069 $3,071 $3,087 $3,156

$1,126 $1,149 $1,428 $1,380 $1,380 $1,483 $1,329 $1,765 $1,969 $1,855 $1,785 $2,136 $2,162 $1,901 $1,891 $2,134 $2,275 $2,272 $2,158 $2,199

Sinai & Gyourko

TABLE 4—Continued


203

above the median in terms of the number of households.24) The tables also include per-household values of the subsidy, although the sorting is on a per-owner basis. These two tables demonstrate the wide disparities in the size of benefit flows across places. For example, Table 3 documents that, in 1979, an owner in one of the top 20 areas received from three to eight times the benefit flow of an owner in one of the bottom 20 areas.25 The differentials are narrower on a per-household basis, with households in the top 20 areas receiving benefit flows that are from two to four times those in the bottom 20 areas. While differences in ownership rates—which are lower in the top subsidy areas—do account for some of the gap between the top and bottom recipient areas, the disparity is still large, even on a per-household basis. Based on 1999 data, the figures in Table 4 indicate that the differentials widened considerably over the ensuing two decades. For example, a comparison of the per-owner subsidy in the twentieth highest-ranked area (Lake County–Kenosha County, IL–WI, Metropolitan Division) with the same figure for the twentieth lowest-ranked area (Scranton–Wilkes-Barre, PA, Metropolitan Statistical Area [MSA]) finds a ratio of 3.4 to 1—or 1.3 times the ratio for the analogously ranked areas in 1979. Comparing the benefit-per-owner value in the tenth highest-ranked area (Honolulu, HI, MSA) with that for the tenth lowest-ranked area (Fort Smith, AR–OK, MSA) finds a ratio of 5.6 to 1—which is 1.5 times the ratio for similarly ranked areas in 1979. The disparity widens even further when comparing the topranked area (San Francisco–San Mateo–Redwood City, CA, Metropolitan Division) to the bottom-ranked area (McAllen–Edinberg–Pharr, TX, MSA) in terms of benefit per owner, with a ratio of 17.1 to 1 ($26,385 to $1,541). Thus, the top recipient areas were receiving relatively more per area than the bottom-ranked areas in 1999 than in 1979. The benefits flowing to owners in the top areas rose by 50 to 100 percent in real terms, while they were flat or declined slightly in the bottom-ranked areas. An even clearer face can be put on the skewness depicted in Figure 8 by examining who and where the top and bottom recipient areas are on a per-owner basis. Fourteen of the top 20 areas appear in both 1979 and 1999. They include Honolulu, HI; Bridgeport–Stamford–Norwalk, CT; Bethesda–Frederick–Gaithersburg, MD; Lake County–Kenosha County, 24 The top 20 areas in terms of benefits per owner are almost unchanged by restricting the sample to more populous areas containing more than the median number of households. This situation is not the case among the bottom 20 areas. If the full sample of 380 CBSAs is used, Texas is even more overrepresented because it contains a large number of less populous metropolitan areas. 25 These ranges were determined by computing the ratio of benefit per owner in the topranked area versus the bottom-ranked area, from the second- to highest-ranked area versus the second- to lowest-ranked area, and so forth.

204

Sinai & Gyourko

IL–WI; and ten areas spanning the length of California’s coastline. By 1999, a series of areas, primarily located along the New York City–Boston corridor (Suffolk County–Nassau County, NY; New York–Wayne–White Plains, NY–NJ; Cambridge–Newton–Framingham, MA; Boston–Quincy, MA; and Newark–Union, NJ–PA) joined the top-20 list, replacing midwestern areas such as Ann Arbor, MI; Warren–Farmington Hills–Troy, MI; and Milwaukee–Waukesha–West Allis, WI, along with Anchorage, AK, and Washington–Arlington–Alexandria, DC–VA–MD–WV. Thus, the top recipient areas have become even more dominated by coastal areas, with the Northeast being much more heavily represented in the 1999 rankings.26 There is less stability among the 20 bottom ranked areas, with 10 present in both 1979 and 1999. This group always has a strong southern representation (especially, but not exclusively, because of Texas), and the metropolitan areas tend not to be situated along the Atlantic or Pacific coasts. In sum, the spatial skewness of benefit flows per owner has grown over time, with the top areas now receiving large multiples of the subsidy received by the bottom areas. Geographically, this skewness now is a bicoastal phenomenon, with metropolitan areas spanning the state of California and the area between New York and Boston dominating the top 20 benefit-per-owner rankings. Still, strong persistence exists over time in the areas that receive the most benefits, and their share of the total has been rising. Because the most important tax-code changes tend to have occurred at the federal level, plots of tax rates and tax-rate changes at the metropolitan level are not particularly helpful in increasing our understanding of these results. In contrast, examining house prices over time at the local level is illuminating. For example, the plots in Figure 9 show the distribution of mean house values by metropolitan area over time, and they look strikingly similar to the distributions of benefits per owner in Figure 8. While incomes and tax rates are somewhat higher in coastal metropolitan areas, these differences are not nearly as pronounced as they are for house values. Thus, rising real house prices, especially in key coastal metropolitan areas, augmented by generally higher tax rates in those areas, are increasing the absolute and relative benefits flowing to their owners. Because the method of financing for housing has only a secondorder effect (through itemization) on the value of the subsidy, it is not necessary for households to refinance their houses to increase their subsidies. Higher prices reflect higher implicit rental value, so if housing were treated symmetrically, tax revenues would increase with house prices. 26

The only interior area to join the top-20 list in 1999 was Boulder, CO.


205

FIGURE 9. Mean House Value, by Metro Area, in 1979, 1989, and 1999

206

Sinai & Gyourko

4. CONCLUSIONS Estimating the tax subsidy to homeowners by comparing the taxes they now pay with those they would pay if they faced neutral tax treatment— like landlords in our example—shows a substantial increase in the value of the tax benefit over time. While some of the aggregate increase clearly is due to a rise in the number of homeowners, benefits per owner are about 20 percent higher in 1999 than they were in 1979 at the national level. This development is particularly interesting because it occurs despite marginal and average tax rates falling over the past two decades. The evidence suggests that rising house prices, especially in key coastal areas and in certain regions of the country, can help account for the fact that the value of the subsidy has risen, even though the tax subsidy per dollar of housing has declined. We demonstrate that the subsidy flows disproportionately to owners in a relatively small number of states—California, especially. Spatial skewness is even more extreme at the metropolitan level, and the data indicate that skewness there has increased over time, though the top recipient areas tend to remain top recipients. Rising house prices in certain coastal metropolitan areas appear to play a large role in explaining this phenomenon. While the magnitude and skewness of the subsidy are striking, one note of caution is in order when interpreting these results. While it may appear that current homeowners in some parts of the country reap a large tax subsidy, their house prices may be higher. That is, the after-tax annual cost of housing in high-subsidy areas may not differ from low-subsidy areas by the full amount of the tax benefit. In the extreme case, if house prices have fully capitalized the benefit, current homeowners are no better off on a flow basis. Computing the incidence of the tax subsidy to owner-occupied housing—the degree to which the subsidy shows up in higher house prices rather than as a reduced flow cost of homeownership—is beyond the scope of this paper. In addition, no consensus about the issue exists in the economics literature: estimates range from full capitalization to extremely low capitalization.27 Where the incidence lies, however, has crucial implications for public policy. For example, it would be easy to jump to the conclusion that, because of the spatial inequity of the tax subsidy to owner-occupied housing, policymakers should restructure the tax benefit. But if a reduction in benefit is capitalized into house prices, current homeowners may experience a loss of wealth. If those homeowners had been 27 For examples, see Bruce and Holtz-Eakin (1999) and Capozza, Green, and Hendershott (1996)


207

the beneficiaries of the rise in house prices when the tax subsidy increased, such a reduction in asset value might be equitable. It is quite likely, however, that current homeowners purchased their houses with the tax benefit already capitalized into the price, paying more on the expectation of future subsidies. The degree of the capitalization of the subsidy into house prices is also unlikely to be spatially neutral. In places where land is in short supply, an increase in demand for housing is likely to show up more in house prices than it would in cities, where it is easy to add more housing stock. That housing demand can be created by local economic factors or the subsidy to owner-occupied housing. Thus, for the same underlying economic reasons, places where the tax benefit is the greatest are places with high land prices and also places where the subsidy is more likely to be capitalized into the house price. While we cannot say how much of any reduction in the tax benefit would show up as lower house prices, it seems likely that a larger fraction (of a larger benefit) would be reflected in house prices in the high-benefit areas.

References Advisory Commission on Intergovernmental Relations (1987). “Significant Feature of Fiscal Federalism.” Washington, DC: ACIR, 102–103. Bruce, Donald, and Douglas Holtz-Eakin (1999). “Apocalypse Now? Fundamental Tax Reform and Residential Housing Values.” Syracuse University. Mimeo. November. Capozza, Dennis, Richard Green, and Patric Hendershott (1996). “Taxes, Mortgage Borrowing, and Residential Land Prices.” In Economic Effects of Fundamental Tax Reform, Henry Aaron and William Gale (eds.). Washington, DC: Brookings, 171–210. Deloitte & Touche (2000). “Analysis of the Economic and Tax Depreciation of Structures,” Deloitte & Touche. Working Paper. June. Feenberg, Daniel Richard, and Elizabeth Coutts (1993). “An Introduction to the TAXSIM Model.” Journal of Policy Analysis and Management 12(1):189–194. Follain, James, and David Ling (1991). “The Federal Tax Subsidy to Housing and the Reduced Value of the Mortgage Interest Deduction.” National Tax Journal 44(2):147–168. Follain, James, David Ling, and Gary McGill (1993). “The Preferential Income Tax Treatment of Owner-Occupied Housing: Who Really Benefits?” Housing Policy Debate 4(1): 1–24. Follain, James, and Lisa Melamed (1998). “The False Messiah of Tax Policy: What Elimination of the Home Mortgage Interest Deduction Promises and a Careful Look at What It Delivers.” Journal of Housing Research 9(2):179–199. Geolytics, Inc. (1980). CensusCD 1980. (Census Tract-Level Data.) East Brunswick, NJ: Geolytics, Inc. Geolytics, Inc. (1990). CensusCD 1990. (Census Tract-Level Data.) East Brunswick, NJ: Geolytics, Inc.

208

Sinai & Gyourko

Geolytics, Inc. (2000). CensusCD 2000. (Census Tract-Level Data.) East Brunswick, NJ: Geolytics, Inc. Gravelle, Jane (2001). “Whither Depreciation?” National Tax Journal 54(3):513–526. Green, Richard, and Andrew Reschovsky (2001). Using Tax Policy to Increase Homeownership Among Low- and Moderate-Income Households. Final Report to the Ford Foundation. Gyourko, Joseph, and Todd Sinai (2003). “The Spatial Distribution of HousingRelated Ordinary Income Tax Benefits,” Real Estate Economics, 31(4):527–575. Hendershott, Patric, and Joel Slemrod (1983). “Taxes and the User Cost of Capital for Owner-Occupied Housing.” AREUEA Journal 10(4):375–393. Joint Committee on Taxation (1998). “Estimates of Federal Tax Expenditures for Fiscal Years 1999–2003 (JCS-7-98).” Washington, DC: U.S. Government Printing Office, December 14. Moody, J. Scott (2003). “Federal Tax Burdens and Expenditures by State: Which States Gain Most from Federal Fiscal Operations?” Tax Foundation Monograph No. 124. July. Poterba, James (1984). “Tax Subsidies to Owner-Occupied Housing: An Asset Market Approach.” Quarterly Journal of Economics 99(4):729–752. Poterba, James (1991). “House Price Dynamics: The Role of Tax Policy and Demography.” Brookings Papers on Economic Activity 2:143–203. Poterba, James (1992). “Taxation and Housing: Old Questions, New Answers.” American Economic Review 82(2):237–242. Rosen, Harvey (1979). “Housing Decisions and the U.S. Income Tax.” Journal of Public Economics, 36(1):87–109. O’Sullivan, Arthur, Terri A. Sexton, and Steven M. Sheffrin (1995). Property Taxes and Tax Revolts: The Legacy of Proposition 13. New York: Cambridge University Press. Sinai, Todd, and Joseph Gyourko (2004). “The (Un)changing Geographical Distribution of Housing Tax Benefits: 1980 to 2000.” NBER Working Paper no. 10322. February. U. S. Bureau of the Census (1980). Summary Tape File 3 (STF3). (Census Tract-Level Data.) Washington, DC. U. S. Bureau of the Census (1990). Summary Tape File 3 (STF3). (Census Tract-Level Data.) Washington, DC. U. S. Bureau of the Census (2000). Summary File 3 (SF3). (Census Tract-Level Data.) Washington, DC. U.S. Department of the Treasury (1980). Statistics of Income, 1979: Public Use Microsamples. Washington, DC. U.S. Department of the Treasury (1990). Statistics of Income, 1989: Public Use Microsamples. Washington, DC. U.S. Department of the Treasury (2000). Statistics of Income, 1999: Public Use Microsamples. Washington, DC.

Tax Policy and the Economy, Volume 18 (NBER Tax Policy and the Economy)

Tax Policy and the Economy, Volume 20 (NBER Tax Policy and the Economy)

Tax Policy and the Economy, Volume 9 (NBER Tax Policy and the Economy)

Tax Policy and the Economy, Volume 16 (NBER Tax Policy and the Economy)

Tax Policy and the Economy, Volume 12 (NBER Tax Policy and the Economy)

Tax Policy and the Economy, Vol. 21 (NBER Tax Policy and the Economy)

Tax Policy Reform and Economic Growth, OECD Tax Policy Studies

Tax Policy Reform and Economic Growth (OECD Tax Policy Studies)

Tax and the Economy: A Comparative Assessment of Oecd Countries (Oecd Tax Policy Series, 6)

United States Taxes and Tax Policy

Industrial Policy and the International Economy

Tax Policy in the Real World

Sell Globally Tax Locally : Sales Tax Reform for the New Economy (AEI Studies on Tax Reform)

Financing government in the transition--Bulgaria: the political economy of tax policies, tax bases, and tax evasion

OECD Tax Policy Studies No. 02: Tax Burdens: Alternative Measures

Tax Ratios: A Critical Survey (Oecd Tax Policy Studies, 5)

Tax Systems and Tax Reforms in Europe (Routledge Studies in the Modern World Economy, 42)

Tax Systems and Tax Reforms in New EU Members (Routledge Studies in the Modern World Economy)

Using Micro-Data to Assess Average Tax Rates: Oecd Tax Policy Studies No. 8 (Oecd Tax Policy Studies)

Tax policy in sub-Saharan Africa, Volume 19

The X-Tax in the World Economy: Going Global with a Simple, Progressive Tax (AEI Studies on Tax Reform)

Entrepreneurship and the Creative Economy: Process, Practice and Policy

The Political Economy of American Monetary Policy

UK Tax Policy and Applied General Equilibrium Analysis

Tax and Corporate Governance

Aid and the Political Economy of Policy Change

Heaven's Door: Immigration Policy and the American Economy

Migration and the Welfare State: Political-Economy Policy Formation

Tax Policy and the Economy, Volume 18 (NBER Tax Policy and the Economy)

Tax Policy and the Economy, Volume 20 (NBER Tax Policy and the Economy)

Tax Policy and the Economy, Volume 9 (NBER Tax Policy and the Economy)

Tax Policy and the Economy, Volume 16 (NBER Tax Policy and the Economy)

Tax Policy and the Economy, Volume 12 (NBER Tax Policy and the Economy)

Tax Policy and the Economy, Vol. 21 (NBER Tax Policy and the Economy)

Tax Policy Reform and Economic Growth, OECD Tax Policy Studies

Tax Policy Reform and Economic Growth (OECD Tax Policy Studies)

Tax and the Economy: A Comparative Assessment of Oecd Countries (Oecd Tax Policy Series, 6)

United States Taxes and Tax Policy

Industrial Policy and the International Economy

Tax Policy in the Real World

Sell Globally Tax Locally : Sales Tax Reform for the New Economy (AEI Studies on Tax Reform)

Financing government in the transition--Bulgaria: the political economy of tax policies, tax bases, and tax evasion

OECD Tax Policy Studies No. 02: Tax Burdens: Alternative Measures

Tax Ratios: A Critical Survey (Oecd Tax Policy Studies, 5)

Tax Systems and Tax Reforms in Europe (Routledge Studies in the Modern World Economy, 42)

Tax Systems and Tax Reforms in New EU Members (Routledge Studies in the Modern World Economy)

Using Micro-Data to Assess Average Tax Rates: Oecd Tax Policy Studies No. 8 (Oecd Tax Policy Studies)

Tax policy in sub-Saharan Africa, Volume 19

The X-Tax in the World Economy: Going Global with a Simple, Progressive Tax (AEI Studies on Tax Reform)

Entrepreneurship and the Creative Economy: Process, Practice and Policy

The Political Economy of American Monetary Policy

UK Tax Policy and Applied General Equilibrium Analysis

Tax and Corporate Governance

Aid and the Political Economy of Policy Change

Heaven's Door: Immigration Policy and the American Economy

Migration and the Welfare State: Political-Economy Policy Formation

Recommend Documents