IDEALIZATION XII: CORRECTING THE MODEL
POZNAŃ STUDIES IN THE PHILOSOPHY OF THE SCIENCES AND THE HUMANITIES VOLUME 86
EDITORS Jerzy Brzeziński Andrzej Klawiter Piotr Kwieciński (assistant editor) Krzysztof Łastowski Leszek Nowak (editor-in-chief)
Izabella Nowakowa Katarzyna Paprzycka (managing editor) Marcin Paprzycki Piotr Przybysz (assistant editor) Michael J. Shaffer
ADVISORY COMMITTEE Joseph Agassi (Tel-Aviv) Étienne Balibar (Paris) Wolfgang Balzer (München) Mario Bunge (Montreal) Nancy Cartwright (London) Robert S. Cohen (Boston) Francesco Coniglione (Catania) Andrzej Falkiewicz (Wrocław) Dagfinn Føllesdal (Oslo) Bert Hamminga (Tilburg) Jaakko Hintikka (Boston) Jacek J. Jadacki (Warszawa) Jerzy Kmita (Poznań)
Leon Koj (Lublin) Władysław Krajewski (Warszawa) Theo A.F. Kuipers (Groningen) Witold Marciszewski (Warszawa) Ilkka Niiniluoto (Helsinki) Günter Patzig (Göttingen) Jerzy Perzanowski (Toruń) Marian Przełęcki (Warszawa) Jan Such (Poznań) Max Urchs (Konstanz) Jan Woleński (Kraków) Ryszard Wójcicki (Warszawa)
Poznań Studies in the Philosophy of the Sciences and the Humanities is partly sponsored by SWPS and Adam Mickiewicz University
Address:
dr Katarzyna Paprzycka . Instytut Filozofii . SWPS . ul. Chodakowska 19/31 03-815 Warszawa . Poland . fax: ++48 22 517-9625 E-mail:
[email protected] . Website: http://PoznanStudies.swps.edu.pl
IDEALIZATION XII: CORRECTING THE MODEL IDEALIZATION AND ABSTRACTION IN THE SCIENCES
Edited by Martin R. Jones and Nancy Cartwright
Amsterdam - New York, NY 2005
The paper on which this book is printed meets the requirements of "ISO 9706:1994, Information and documentation - Paper for documents Requirements for permanence". ISSN 0303-8157 ISBN: 90-420-1955-7 ©Editions Rodopi B.V., Amsterdam - New York, NY 2005 Printed in The Netherlands
Science at its best seeks most to keep us in this simplified, thoroughly artificial world, suitably constructed and suitably falsified world . . . willy-nilly, it loves error, because, being alive, it loves life. Friedrich Nietzsche, Beyond Good and Evil
This page intentionally left blank
CONTENTS
Preface . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
9
Analytical Table of Contents . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
11
Kevin D. Hoover, Quantitative Evaluation of Idealized Models in the New Classical Macroeconomics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
15
John Pemberton, Why Idealized Models in Economics Have Limited Use. .
35
Amos Funkenstein, The Revival of Aristotle’s Nature . . . . . . . . . . . . . . . . . .
47
James R. Griesemer, The Informational Gene and the Substantial Body: On the Generalization of Evolutionary Theory by Abstraction . . . . . . .
59
Nancy J. Nersessian, Abstraction via Generic Modeling in Concept Formation in Science . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
117
Margaret Morrison, Approximating the Real: The Role of Idealizations in Physical Theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
145
Martin R. Jones, Idealization and Abstraction: A Framework . . . . . . . . . . .
173
David S. Nivison, Standard Time . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
219
James Bogen and James Woodward, Evading the IRS . . . . . . . . . . . . . . . . .
233
M. Norton Wise, Realism Is Dead . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
269
Ronald N. Giere, Is Realism Dead? . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
287
This page intentionally left blank
PREFACE
This volume was conceived over a dozen years ago in Stanford, California when Cartwright and Jones were both grappling – each for their own reasons – with questions about idealization in science. There was already a rich European literature on the subject, much of it represented in this series; and there was a growing interest in the United States, prompted in part by new work on approximation and in part by problems encountered within the American versions of the semantic view of theories about the fit of models to the world. At the time, tuned to questions about idealization and abstraction, we began to see elements of the problem and lessons to be learned about it everywhere, in papers and lectures on vastly different subjects, from Chinese calendars to nineteenth century electrodynamics to options pricing. Each – usually unintentionally and not explicitly – gave some new slant, some new insight, into the problems about idealization, abstraction, approximation, and modeling; and we thought it would be valuable to the philosophical and scientific communities to make these kinds of discussions available to all, specifically earmarked as discussions that can teach us about idealization. This volume is the result. We asked the contributors to write about some aspect of idealization or abstraction in a subject they were studying – “idealization” as they found it useful to think about it, without trying to fit their discussion into some categories already available. As a consequence, most of the authors in this volume are not grappling directly with the standard philosophical literature on the problems; indeed, several are not philosophers – we have a distinguished historian of ideas who specialized in the medieval period, a renowned historian of physics, an eminent economic methodologist, and an investment director for one of the largest insurance companies in the world. The volume has unfortunately been a long time in the making; many of the papers were written a decade ago. We apologize to the authors for this long delay. Still, for thinking about abstraction and idealization, the material remains fresh and original. We hope that readers of this volume will find this diverse collection as rewarding a source of new ideas and new materials as we ourselves have. Two of the papers in particular need to be set in context. Norton Wise’s contribution, “Realism is Dead,” was given at the 1989 Pacific Division Meetings of the American Philosophical Association, in Berkeley, California, as part of an “Author Meets the Critics” session devoted to Ronald Giere’s book Explaining Science, which was hot off the presses at the time. (The other
10
Preface
speaker was Bas van Fraassen.) Giere’s contribution, “Is Realism Dead?”, is based on his response to Wise’s comments on that occasion. Sadly, Amos Funkenstein died during the production of the volume. As a result, we have had to make several decisions during the editing of his paper without being able to confirm our judgements with the author. Some of these decisions concerned the right way to correct apparent typographical error, some the filling-in of bibliographical gaps, and some the fine points of various languages. Thanks in that connection to James Helm of the Classics Department of Oberlin, who helped us to get the Greek right. We hope we have been true to Professor Funkenstein’s original intentions. Any errors which remain are, no doubt, the fault of the editors. Nancy Nersessian’s paper has also been published in Mind and Society, vol. 3, no. 5, and we thank Fondazione Roselli for permission to reprint it here. We also thank the various publishers for their permission to reprint the following figures: figure 2 in James Bogen and James Woodward’s paper, from John Earman and Clark Glymour’s “Relativity and Eclipses: The British Eclipse Expeditions of 1919 and their Predecessors,” Historical Studies in the Physical Sciences, 11:1 (1980), p. 67 (University of California Press); figure 2 in James Griesemer’s paper, reprinted by permission from Nature vol. 227, p. 561 (copyright © 1970 MacMillan Magazines Limited); and figure 3 in James Griesemer’s paper, from John Maynard Smith’s The Theory of Evolution, 3rd ed., 1975 (copyright © John Maynard Smith, 1985, 1966, 1975, reproduced by permission of Penguin Books, Middlesex). In the process of putting the volume together, we have benefited from the help of many people. Particular thanks are due to Jon Ellis, Ursula Coope, Cheryl Chen, Mary Conger, Lan Sasa, George Zouros, and especially to Brendan O’Sullivan, for their careful, thorough, and diligent assistance with various sorts of editorial and bibliographical work. Thanks to Ned Hall for suggesting that we use the phrase “correcting the model” in the title of the book, to Leszek Nowak as series editor, and to Fred van der Zee at Rodopi. We (and especially MRJ) would also like to thank the authors for their patience in the face of delay. Martin Jones’s work on the volume was assisted at various points, and in various ways, by: a President’s Research Fellowship in the Humanities from the University of California; a Humanities Research Fellowship from the University of California, Berkeley; and a Hellman Family Faculty Fund Award, also at the University of California, Berkeley. In addition, H. H. Powers Travel Grants from Oberlin College made it much easier for the editors to meet during two recent summers. M.R.J. N.C.
ANALYTICAL TABLE OF CONTENTS
Kevin D. Hoover, Quantitative Evaluation of Idealized Models in the New Classical Macroeconomics The new classical macroeconomics responds to the “Lucas critique” by providing microfoundations for macroeconomic relationships. There is disagreement, however, about how best to conduct quantitative policy evaluation, given the idealized nature of the models in question. An examination of whether a coherent methodological basis can be found for one method – the calibration strategy – and an assessment of the limits of quantitative analysis within idealized macroeconomic models, with special attention to the work of Herbert Simon.
John Pemberton, Why Idealized Models in Economics Have Limited Use A distinction is drawn between causal and non-causal idealized models. Models of the second sort, which include those specified by means of restrictive antecedent clauses, are widely used in economics, but do not provide generally reliable predictions; moreover, they contain no clues as to when their reliability will fail.This is illustrated by a critique of the Black and Scholes option pricing model. The high degree of causal complexity of economic systems means that causal idealized models are also of limited value.
Amos Funkenstein, The Revival of Aristotle’s Nature A response to Nancy Cartwright’s arguments for the primacy of the notion of capacity. Explores the question of why, historically, capacities were seemingly set aside in the physics of the Scientific Revolution, and suggests some doubts about the fit between the notion of capacity and post-Galilean natural science. Connections between the putative abandonment of capacities and the move to seeing nature as uniform, and implications for the correct understanding of idealization as a method.
12
Analytical Table of Contents
James R. Griesemer, The Informational Gene and the Substantial Body: On the Generalization of Evolutionary Theory by Abstraction Darwin’s theory of evolution dealt only with organisms in populations. Problems are raised for two central strategies for generalization by abstraction, including, in one case, contamination of the abstract theory by assumptions specific to particular mechanisms operating in particular sorts of material. A different foundation is needed for the generalization of evolutionary theory; and there are ramifications for the “units of selection” debate. A modification of Cartwright and Mendell’s account of abstraction is also proposed.
Nancy J. Nersessian, Abstraction via Generic Modeling in Concept Formation in Science Conceptual innovation in science often involves analogical reasoning. The notion of abstraction via generic modeling casts light on this process; generic modeling is the process of constructing a model which represents features common to a class of phenomena. Maxwell’s development of a model of the electromagnetic field as a state of a mechanical aether involved just this strategy. The study of generic modeling also yields insights into abstraction and idealization, and their roles in mathematization.
Margaret Morrison, Approximating the Real: The Role of Idealizations in Physical Theory An examination of the role and nature of idealization and abstraction in model construction and theory development. Parallels between current debates about the role of models and nineteenth-century debates surrounding Maxwellian electrodynamics. Response to realist claims that abstract models can be made representationally accurate by “adding back” parameters. Problems with the standard philosophical distinction between realistic and heuristic models. Two levels at which idealization operates, and the differing implications for the attempt to connect models to real systems.
Martin R. Jones, Idealization and Abstraction: A Framework A fundamental distinction is between idealizations as misrepresentations, and abstractions as mere omissions; other characteristic features of idealizations and abstractions are considered. A systematic proposal for clarifying talk of
Analytical Table of Contents
13
idealization and abstraction in both models and laws, degrees of idealization and abstraction, and idealization and abstraction as processes. Relations to the work of Cartwright and McMullin. Three ways in which idealization can occur in laws and our employment of them – there are quasi-laws, idealized laws, and ideal laws.
David S. Nivison, Standard Time Ancient Chinese astronomy and calendary divided time in simple, idealized ways with respect to, for example, lunar cycles, the seasons, and the period of Jupiter. The resulting schemes do not fit the data exactly, and astronomers employed complex rules of adjustment to correct for the mismatch. Nonetheless, the simpler picture was often treated as ideally true, and as capturing an order which was supposed to underlie the untidiness of observed reality.
James Bogen and James Woodward, Evading the IRS Critique of, and an alternative to, the view that the epistemic bearing of observational evidence on theory is best understood by examining Inferential Relations between Sentences (‘IRS’). Instead, we should attend to empirical facts about particular causal connections and about the error characteristics of detection processes. Both the general and the local reliability of detection procedures must be evaluated. Case studies, including attempts to confirm General Relativity by observing the bending of light, and by detecting gravity waves.
M. Norton Wise, Realism Is Dead An evaluation of Giere’s views of theory structure and representation in Explaining Science, and objections to his “constructive realism”: focus on the Hamiltonian version of classical mechanics shows that the theory should not be seen as a collection of abstract and idealized models; many nineteenth-century figures in classical physics did not treat mechanics as realistic, seeing its models as merely structurally analogous to real systems; a naturalistic theory of science ought to reflect that historical fact; and Giere’s realism does not take sufficient account of the social. Constrained social constructionism is the better view.
14
Analytical Table of Contents
Ronald N. Giere, Is Realism Dead? Response to Wise, defending an “enlightened post-modernism”: the Hamiltonian version of classical mechanics simply provides more general recipes for model building; structural analogy is a kind of similarity, and so fits with constructive realism; scientists’ own theories of science are not privileged; and although constructive realism makes room for the social, strategies of experimentation can overwhelm social factors, and individuals are basic. We should start with a minimal realism, and focus on the task of explaining how science has led to an increase in our knowledge of the world.
Kevin D. Hoover QUANTITATIVE EVALUATION OF IDEALIZED MODELS IN THE NEW CLASSICAL MACROECONOMICS
The new classical macroeconomics is today certainly the most coherent, if not the dominant, school of macroeconomic thought. The pivotal document in its two decades of development is Robert Lucas’s 1976 paper, “Econometric Policy Evaluation: A Critique.”1 Lucas argued against the then reigning methods of evaluating the quantitative effects of economic policies on the grounds that the models used to conduct policy evaluation were not themselves invariant with respect to changes in policy.2 In the face of the Lucas critique, the new classical economics is divided in its view of how to conduct quantitative policy analysis between those who take the critique as a call for better methods of employing theoretical knowledge in the direct empirical estimation of macroeconomic models, and those who believe that it shows that estimation is hopeless and that quantitative assessment must be conducted using idealized models. Assessing the soundness of the views of the latter camp is the main focus of this essay.
1. The Lucas Critique The Lucas critique is usually seen as a pressing problem for macroeconomics, the study of aggregate economic activity (GNP, inflation, unemployment, interest rates, etc.). Economists typically envisage individual consumers, workers, or firms as choosing the best arrangement of their affairs, given their preferences, subject to constraints imposed by limited resources. Moving up from an account of individual behavior to an account of economy-wide aggregates is problematic. Typical macroeconometric models before Lucas’s ⎯⎯⎯⎯⎯⎯⎯
1
See Hoover (1988), (1991) and (1992) for accounts of the development of new classical thinking. The “Lucas critique,” as it is always known, is older than Lucas’s paper, going back at least to the work of Frisch and Haavelmo in the 1930s; see Morgan (1990), Hoover (1994).
2
In: Martin R. Jones and Nancy Cartwright (eds.), Idealization XII: Correcting the Model. Idealization and Abstraction in the Sciences (Poznań Studies in the Philosophy of the Sciences and the Humanities, vol. 86), pp. 15-33. Amsterdam/New York, NY: Rodopi, 2005.
16
Kevin D. Hoover
paper consisted of a set of aggregate variables to be explained (endogenous variables). Each was expressed as a (typically, but not necessarily, linear) function of other endogenous variables and of variables taken to be given from outside the model (exogenous variables). The functional forms were generally suggested by economic theory. The functions usually contained free parameters; thus, in practice, no one believed that these functions held exactly. Consequently, the parameters were usually estimated using statistical techniques such as multivariate regression. The consumption function provides a typical example of a function of a macromodel. The permanent income hypothesis (Friedman 1957) suggests that consumption at time t (Ct) should be related to income at t (Yt) and consumption at t-1 as follows: C t = k(1 – µ)Y t + µC t–1 + e
(1)
where k is a parameter based on people’s tastes for saving, µ is a parameter based on the process by which they form their expectations of future income, and e is an unobservable random error indicating that the relationship is not exact. One issue with a long pedigree in econometrics is whether the underlying parameters can be recovered from an econometric estimation.3 This is called the identification problem. For example, suppose that what we want to know is the true relationship between savings and permanent income, k. If the theory that generated equation (1) is correct, then we would estimate a linear regression of the form C t = π1Y t + π2C t–1 + w
(2)
where w is a random error and π1 and π2 are free parameters to be estimated. We may then calculate k = π1 / (1– π2). The Lucas critique suggested that it was wrong to take an equation such as (1) in isolation. Rather, it had to be considered in the context of the whole system of related equations. What is more, given that economic agents were optimizers, it was wrong to treat µ as a fixed parameter, because people’s estimates of future income would depend in part on government policies that alter the course of the economy and, therefore, on GNP, unemployment, and other influences on personal income. The parameter itself would have to be modeled something like this: µ = f (Yt, Y t–1, C t–1, G)
⎯⎯⎯⎯⎯⎯⎯ 3
See Morgan (1990), ch. 6, for a history of this problem.
(3)
Quantitative Evaluation of Idealized Models in the New Classical Macroeconomics
17
where G is a variable indicating the stance of government policy. Every time government policy changes, µ changes. If equation (1) were a true representation of economic reality, the estimates of π1 and π2 in equation (2) would alter with every change in µ: they would not be autonomous.4 The absence of autonomy, in turn, plays havoc with identification: the true k can be recovered from π1 and π2 only if they are stable. Identification and autonomy are distinct streams in the history of econometrics; the Lucas critique stands at their confluence. Autonomy had been a neglected issue in econometrics for over forty years. Identification, in contrast, is a central part of every econometrics course.
2. Calibration versus Estimation New classical macroeconomists are radical advocates of microfoundations for macroeconomics. Explanations of all aggregate economic phenomena must be based upon, or at least consistent with, the microeconomics of rational economic agents. The new classicals stress constrained optimization and general equilibrium. A characteristic feature of new classical models is the rational expectations hypothesis. Expectations are said to be formed rationally if they are consistent with the forecasts of the model itself. This is usually seen to be a requirement of rationality, since, if the model is correct, expectations that are not rational would be systematically incorrect, and rational agents would not persist in forming them in such an inferior manner.5 All new classicals share these theoretical commitments, yet they do not all draw the same implications for quantitative policy evaluation. One approach stresses the identification aspects of the Lucas critique. Hansen and Sargent (1980), for instance, recommend starting with a model of individual preferences and constraints (i.e., “taking only tastes and technology as given,” to use the jargon). From this they derive the analogs of equations (1)-(3). In particular, using the rational expectations hypothesis, they derive the correctly parameterized analog to (3). They then estimate the reduced forms, the analogs to equation (2) (which may, in fact, be a system of equations). Statistical tests are used to determine whether the restrictions implied by the theory and represented in the analogs to equations (1) and (3) hold. In the current example, since we have not specified (3) more definitely, there is nothing to test in equation (1). This is because equation (1) is just identified; i.e., there is only one way to calculate its parameters from the estimated coefficients of equation (2). Nonetheless, in many cases, there is more than one way to ⎯⎯⎯⎯⎯⎯⎯
4
For the history of autonomy, see Morgan (1990), chs. 4 and 8. See Sheffrin (1983) for a general account of the rational expectations hypothesis, and Hoover (1988), ch. 1, for a discussion of the weaknesses of the hypothesis.
5
18
Kevin D. Hoover
calculate these parameters; the equation is then said to be overidentified. A statistical test of whether, within some acceptable margin, all of these ways yield the same estimated parameters is a test of overidentifying restrictions and is the standard method of judging the success of the theoretical model in matching the actual data. Hansen and Sargent’s approach assumes that the underlying theory is (or could be or should be) adequate, and that the important problem is to obtain accurate estimates of the underlying free parameters. Once these are known, they can be used to conduct policy analysis. Lucas (1987, p. 45) and Kydland and Prescott (1991) argue that Hansen and Sargent’s approach is inappropriate for macroeconomics. They do not dissent from the underlying theory used to derive the overidentifying restrictions. Instead, they argue that reality is sufficiently complex that no tractable theory can preserve a close mapping with all of the nuances of the data that might be reflected in the reduced forms. Sufficiently general reduced forms will pick up many aspects of the data that are not captured in the theory, and the overidentifying restrictions will almost surely be rejected. Lucas and Prescott counsel not attempting to estimate or test macroeconomic theories directly. Instead, they argue that macroeconomists should build idealized models that are consistent with microeconomic theory and that mimic certain key aspects of aggregate economic behavior. These may then be used to simulate the effects of policy. Of course, theory leaves free parameters undetermined for Lucas and Prescott, just as it does for Hansen and Sargent. Lucas and Prescott suggest that these may be supplied either from independently conducted microeconomic econometric studies, which do not suffer from the aggregation problems of macroeconomic estimation, or from searches over the range of possible parameter values for a combination of values that well matches the features of the economy most important for the problem at hand. Their procedure is known as calibration.6 In a sense, advocacy of calibration downplays the identification problem in the Lucas critique and emphasizes autonomy. Estimation in the manner of Hansen and Sargent is an extension of standard and well-established practices in econometrics. When the type of models routinely advocated by Prescott (e.g., Kydland and Prescott 1982) are estimated, they are rejected statistically (e.g., Altug 1989). Prescott simply dismisses such rejections as applying an inappropriate standard. Lucas and Prescott argue that models that are idealized to the point of being incapable of passing such statistical tests are nonetheless the preferred method for generating quantitative ⎯⎯⎯⎯⎯⎯⎯ 6
Calibration is not unique to the new classical macroeconomics, but is well established in the context of “computable general equilibrium” models common in the analysis of taxation and international trade; see Shoven and Whalley (1984). All the methodological issues that arise over calibration of new classical macromodels must arise with respect to computable general equilibrium models as well.
Quantitative Evaluation of Idealized Models in the New Classical Macroeconomics
19
evaluations of policies. The central issue now before us is: can models which clearly do not fit the data be useful as quantitative guides to policy? Who is right – Lucas and Prescott or Hansen and Sargent?
3. Models and Artifacts “Model” is a ubiquitous term in economics, and a term with a variety of meanings. One commonly speaks of an econometric model. Here one means the concrete specification of functional forms for estimation: equation (2) is a typical example. I call these observational models. The second main class of models are evaluative or interpretive models. An obvious subclass of interpretive/evaluative models are toy models. A toy model exists merely to illustrate or to check the coherence of principles or their interaction. An example of such a model is the simple exchange model with two goods and two agents (people or countries). Adam Smith’s famous “invisible hand” suggests that a price system can coordinate trade and improve welfare. In this simple model, agents are characterized by functions that rank their preferences over different combinations of goods and by initial endowments of the goods. One can check, first, whether there is a relative price between the two goods and a set of trades at that price that makes both agents satisfied with the arrangement; and, second, given such a price, whether agents are better off in their own estimation than they would have been in the absence of trade. No one would think of drawing quantitative conclusions about the working of the economy from this model. Instead, one uses it to verify in a tractable and transparent case that certain qualitative results obtain. Such models may also suggest other qualitative features that may not have been known or sufficiently appreciated.7 The utter lack of descriptive realism of such models is no reason to abandon them as test beds for general principles. Is there another subclass of interpretive/evaluative models – one that involves quantification? Lucas seems to think so: One of the functions of theoretical economics is to provide fully articulated, artificial economic systems that can serve as laboratories in which policies that would be prohibitively expensive to experiment with in actual economies can be tested out at much lower cost (Lucas 1980, p. 271).
Let us call such models benchmark models. Benchmark models must be abstract enough and precise enough to permit incontrovertible answers to the questions put to them. Therefore, ⎯⎯⎯⎯⎯⎯⎯ 7
Cf. Diamond (1984), p. 47.
20
Kevin D. Hoover
. . . insistence on the “realism” of an economic model subverts its potential usefulness in thinking about reality. Any model that is well enough articulated to give clear answers to the questions we put to it will necessarily be artificial, abstract, patently unreal (Lucas 1980, p. 271).
On the other hand, only models that mimic reality in important respects will be useful in analyzing actual economies. The more dimensions in which the model mimics the answers actual economies give to simple questions, the more we trust its answers to harder questions. This is the sense in which more “realism” in a model is clearly preferred to less (Lucas 1980, p. 272).
Later in the same essay, Lucas emphasizes the quantitative nature of such model building: Our task . . . is to write a FORTRAN program that will accept specific economic policy rules as “input” and will generate as “output” statistics describing the operating characteristics of time series we care about, which are predicted to result from these policies (Lucas 1980, p. 288).
For Lucas, Kydland and Prescott’s model is precisely such a program.8 One might interpret Lucas’s remarks as making a superficial contribution to the debate over Milton Friedman’s “Methodology of Positive Economics” (1953): must the assumptions on which a theory is constructed be true or realistic or is it enough that the theory predicts “as if” they were true? But this would be a mistake. Lucas is making a point about the architecture of models and not about the foundations of secure prediction. To make this clear, consider Lucas’s (1987, pp. 20-31) cost-benefit analysis of the policies to raise GNP growth and to damp the business cycle. Lucas’s model considers a single representative consumer with a constant-relative-riskaversion utility function facing an exogenous consumption stream. The model is calibrated by picking reasonable values for the mean and variance of consumption, the subjective rate of discount, and the constant coefficient of relative risk aversion. Lucas then calculates how much extra consumption consumers would require to compensate them in terms of utility for a cut in the growth rate of consumption and how much consumption they would be willing to give up to secure smoother consumption streams. Although the answers that Lucas seeks are quantitative, the model is not used to make predictions that might be subjected to statistical tests. Rather, it is used to set upper bounds to the benefits that might conceivably be gained in the real world. Its parameters must reflect some truth about the world if it is to be useful, but they could not be easily directly estimated. In that sense, the model is unrealistic. ⎯⎯⎯⎯⎯⎯⎯ 8
They do not say, however, whether it is actually written in FORTRAN.
Quantitative Evaluation of Idealized Models in the New Classical Macroeconomics
21
In a footnote, Lucas (1980, p. 272, fn. 1) cites Herbert Simon’s Sciences of the Artificial (1969) as an “immediate ancestor” of his “condensed” account. To uncover a more fully articulated argument for Lucas’s approach to modeling, it is worth following up the reference. For Simon, human artifacts, among which he must count economic models, can be thought of as a meeting point – an “interface” . . . – between an “inner” environment, the substance and organization of the artifact itself, and an “outer” environment, the surroundings in which it operates (Simon 1969, pp. 6, 7).
An artifact is useful, it achieves its goals, if its inner environment is appropriate to its outer environment. Simon’s distinction can be illustrated in the exceedingly simplified macroeconomic model represented by equations (1) and (3). The model converts inputs (Y t and G) into an output C t. Both the inputs and outputs, as well as the entire context in which such a model might be useful, can be considered parts of the outer environment. The inner environment includes the structure of the model – i.e., the particular functional forms – and the parameters, µ and k.9 The inner environment controls the manner in which the inputs are processed into outputs. The distinction between the outer and inner environments is important because there is some degree of independence between them. Clocks tell time for the outer environment. Although they may indicate the time in precisely the same way, say with identical hands on identical faces, the mechanisms of different clocks, their inner environments, may be constructed very differently. For determining when to leave to catch a plane, such differences are irrelevant. Equally, the inner environments may be isolated from all but a few key features of the outer environment. Only light entering through the lens for the short time that its shutter is open impinges on the inner environment of the camera. The remaining light is screened out by the opaque body of the camera, which is an essential part of its design. Our simple economic model demonstrates the same properties. If the goal is to predict the level of consumption within some statistical margin of error, then like the clock, other models with quite different functional forms may be (approximately) observationally equivalent. Equally, part of the point of the functional forms of the model is to isolate features which are relevant to achieving the goal of predicting C t – namely, Y t and G – and screening out irrelevant features of the outer environment, impounding any relevant but ignorable influences in the error term. Just as the camera is an opaque barrier to ambient light, the functional form of the consumption model is an opaque barrier to economic influences other than income and government policy. ⎯⎯⎯⎯⎯⎯⎯ 9
One might also include Ct-l, because, even though it is the lagged value of Ct (the output), it may be thought of as being stored “within” the model as time progresses.
22
Kevin D. Hoover
Simon factors adaptive systems into goals, outer environments, and inner environments. The relative independence of the outer and inner environments means that [w]e might hope to characterize the main properties of the system and its behavior without elaborating the detail of either the outer or the inner environments. We might look toward a science of the artificial that would depend on the relative simplicity of the interface as its primary source of abstraction and generality (Simon 1969, p. 9).
Simon’s views reinforce Lucas’s discussion of models. A model is useful only if it foregoes descriptive realism and selects limited features of reality to reproduce. The assumptions upon which the model is based do not matter, so long as the model succeeds in reproducing the selected features. Friedman’s “as if” methodology appears vindicated. But this is to move too fast. The inner environment is only relatively independent of the outer environment. Adaptation has its limits. In a benign environment we would learn from the motor only what it had been called upon to do; in a taxing environment we would learn something about its internal structure – specifically, about those aspects of the internal structure that were chiefly instrumental in limiting performance (Simon 1969, p. 13).
This is a more general statement of principles underlying Lucas’s (1976) critique of macroeconometric models. A benign outer environment for econometric models is one in which policy does not change. Changes of policy produce structural breaks in estimated equations: disintegration of the inner environment of the models. Economic models must be constructed like a ship’s chronometer, insulated from the outer environment so that . . . it reacts to the pitching of the ship only in the negative sense of maintaining an invariant relation of the hands on its dial to real time, independently of the ship’s motions (Simon 1969, p. 9).
Insulation in economic models is achieved by specifying functions whose parameters are invariant to policy. Again, this is easily clarified with the simple consumption model. If µ were a fixed parameter, as it might be in a stable environment in which government policy never changed, then equation (1) might yield an acceptable model of consumption. But in a world in which government policy changes, µ will also change constantly (the ship pitches, tilting the compass that is fastened securely to the deck). The role of equation (3) is precisely to isolate the model from changes in the outer environment by rendering µ a stable function of changing policy; µ changes, but in predictable and accountable ways (the compass mounted in a gimbal continually turns relative to the deck in just such a way as to maintain its orientation with the earth’s magnetic field).
Quantitative Evaluation of Idealized Models in the New Classical Macroeconomics
23
The independence of the inner and outer environments is not something that is true of arbitrary models; rather it must be built into models. While it may be enough in hostile environments for models to reproduce key features of the outer environment “as if” reality were described by their inner environments, it is not enough if they can do this only in benign environments. Thus, for Lucas, the “as if” methodology interpreted as an excuse for complacency with respect to modeling assumptions must be rejected. New classical economists argue that only through carefully constructing the model from invariants – tastes and technology, in Lucas’s usual phrase – can the model secure the benefits of a useful abstraction and generality. This is again an appeal to found macroeconomics in standard microeconomics. Here preferences and the production possibilities (tastes and technology) are presumed to be fixed, and the economic agent’s problem is to select the optimal combination of inputs and outputs. Tastes and technology are regarded as invariant partly because economists regard their formation as largely outside the domain of economics: de gustibus non est disputandum. Not all economists, however, would rule out modeling the formation of tastes or technological change. But for such models to be useful, they would themselves have to have parameters to govern the selection among possible preference orderings or the evolution of technology. These parameters would be the ultimate invariants from which a model immune to the Lucas critique would have to be constructed.
4. Quantification and Idealization Economic models are idealizations of the economy. The issue at hand, is whether, as idealizations, models can be held to a quantitative standard. Can idealized models convey useful quantitative information? The reason why one is inclined to answer these questions negatively is that models patently leave things out. Simon’s analysis, however, suggests that even on a quantitative standard that may be their principal advantage. To see this better, consider the analogy of physical laws. Nancy Cartwright (1983, esp. essays 3 and 6) argues that physical laws are instruments, human artifacts in precisely the sense of Simon and Lucas, the principal use of which is to permit calculations that would otherwise be impossible. The power of laws to rationalize the world and to permit complex calculation comes from their abstractness, definitiveness, and tractability. Lucas (1980, pp. 271, 272) says essentially the same thing about the power of economic models. The problem with laws for Cartwright (1989), however, is that they work only in ideal circumstances. Scientists must experiment to identify the actual working of a physical law: the book of nature is written in code or, more aptly
24
Kevin D. Hoover
perhaps, covered in rubbish. To break the code or clear the rubbish, the experimenter must either insulate the experiment from countervailing effects or must account for and, in essence, subtract away the influence of countervailing effects.10 What is left at the end of the well-run experiment is a measurement supporting the law – a quantification of the law. Despite its tenuousness in laboratory practice, the quantified law remains of the utmost importance. To illustrate, consider an experiment in introductory physics. To investigate the behavior of falling objects, a metal weight is dropped down a vertical track lined with a paper tape. An electric current is periodically sent through the track. The arcing of the electricity from the track to the weight burns a sequence of holes in the tape. These mark out equal times, and the experimenter measures the distances between the holes to determine the relation between time and distance. This experiment is crude. When, as a college freshman, I performed it, I proceeded as a purely empirically – minded economist might in what I thought was a true scientific spirit: I fitted the best line to the data. I tried linear, loglinear, exponential and quadratic forms. Linear fit best, and I got a C – on the experiment. The problem was not just that I did not get the right answer, although I felt it unjust at the time that scientific honesty was not rewarded. The problem was that many factors unrelated to the law of gravity combined to mask its operation – conditions were far from ideal. A truer scientific method would attempt to minimize or take account of those factors. Thus, had I not regarded the experiment as a naive attempt to infer the law of gravity empirically, but as an attempt to quantify a parameter in a model, I would have gotten a better grade. The model says that distance under the uniform acceleration of gravity is gt2 / 2. I suspect that given calculable margins of error, not only would my experiment have assigned a value to g, but that textbook values of g would have fallen within the range of experimental error – despite the apparent better fit of the linear curve.11 Even though the data had to be fudged and discounted to force it into the mold of the gravitational model, it would have been sensible to do so, because we know from unrelated experiments – the data from which also had to be fudged and discounted – that the quadratic law is more general. The law is right, and must be quantified, even though it is an idealization. Confirmation by practical application is important, although sometimes the confirmation is exceedingly indirect. Engineers would often not know where to begin if they did not have quantified physical laws to work with. But laws as ⎯⎯⎯⎯⎯⎯⎯
10
Cartwright (1989, secs. 2.3, 2.4), discusses the logic and methods of accounting for such countervailing effects. 11 An explicit analog to this problem is found in Sargent (1989), in which he shows that the presence of measurement error can make an investment-accelerator model of investment, which is incompatible with new classical theory, fit the data better, even when the data were in fact generated according to Tobin’s q-theory, which is in perfect harmony with new classical theory.
Quantitative Evaluation of Idealized Models in the New Classical Macroeconomics
25
idealizations leave important factors out. Thus, an engineer uses elementary physics to calculate loadings for an ideal bridge. Knowing that these laws fail to account for many critical factors, the engineer then specifies material strengths two, four, or ten times what the calculations showed was needed. Error is inevitable, so err on the side of caution. Better models reduce the margins of error. It is said that the Golden Gate Bridge could easily take a second deck because it was designed with pencil, paper, and slide rule, with simple models, and consequently greatly overbuilt. England’s Humber River Bridge, designed some forty years later with digital computers and more complex models is built to much closer tolerances and could not take a similar addition. The only confirmation that the bridges give of the models underlying their design is that they do not fall down. Overbuilding a bridge is one thing, overbuilding a fine watch is quite another: here close tolerances are essential. An engineer still begins with an idealized model, but the effects ignored by the model now have to be accounted for. Specific, unidealized knowledge of materials and their behavior is needed to correct the idealized model for departures from ideal conditions. Such knowledge is often based on the same sort of extrapolations that I rejected in the gravity experiment. Cartwright (1983, essay 6) refers to these extrapolations as “phenomenal laws.” So long as these extrapolations are restricted to the range of actual experience, they often prove useful in refining the fudge factors. Even though the phenomenal laws prove essential in design, the generality of idealized laws is a great source of efficiency: it is easier to use phenomenal laws to calculate departures from the ideal, than to attempt to work up from highly specific phenomenal laws. Again, the ideal laws find their only confirmation in the watch working as designed. Laws do not constitute all of physics: It is hard to find them in nature and we are always having to make excuses for them: why they have exceptions – big or little; why they only work for models in the head; why it takes an engineer with a special knowledge of materials and a not too literal mind to apply physics to reality (Cartwright 1989, p. 8).
Neither do formal models constitute all of economics. Yet despite the shortcomings of idealized laws, we know from practical applications, such as shooting artillery or sending rockets to the moon, that calculations based on the law of gravity get it nearly right and calculations based on linear extrapolation go hopelessly wrong. Cartwright (1989, ch. 4) argues that “capacities” are more fundamental than laws. Capacities are the invariant dispositions of the components of reality. Something like the notion of capacities must lie behind the proposal to set up “elasticity banks” to which researchers could turn when calibrating computable general equilibrium models (Shoven and Whalley 1984, p. 1047). An elasticity
26
Kevin D. Hoover
is the proportionate change of one variable with respect to another.12 Estimated elasticities vary according to the methods of estimation employed (e.g., functional forms, other controlling variables, and estimation procedures such as ordinary least squares regression or maximum likelihood estimation) and the set of data used. To “bank” disparate estimates is to assume that such different measures of elasticities somehow bracket or concentrate on a “true” value that is independent of the context of estimation. Laws, at best, describe how capacities compose under ideal circumstances. That models should represent the ways in which invariant capacities compose is, of course, the essence of the Lucas critique. Recognizing that models must be constructed from invariants does not itself tell us how to measure the strengths of the component capacities.
5. Aggregation and General Equilibrium Whether calibrated or estimated, real-business-cycle models are idealizations along many dimensions. The most important dimension of idealization is that the models deal in aggregates while the economy is composed of individuals. After all, the distinction between microeconomics and macroeconomics is the distinction between the individual actor and the economy as a whole. All new classical economists believe that one understands macroeconomic behavior only as an outcome of individual rationality. Lucas (1987, p. 57) comes close to adopting the Verstehen approach of the Austrians.13 The difficulty with this approach is that there are millions of people in an economy and it is not practical – nor is it ever likely to become practical – to model the behavior of each of them.14 Universally, new classical economists adopt representativeagent models, in which one agent or a few types of agents, stand in for the behavior of all agents.15 The conditions under which a single agent’s behavior can accurately represent the behavior of an entire class are onerous. Strict ⎯⎯⎯⎯⎯⎯⎯ 12
In a regression of the logarithm of one variable on the logarithms of others, the elasticities can be read directly as the value of the estimated coefficients. 13 For a full discussion of the relationship between new classical and Austrian economics see Hoover (1988), ch. 10. 14 In (Hoover 1984, pp. 64-66), and (Hoover 1988, pp. 218-220), I refer to this as the “Cournot problem,” since it was first articulated by Augustine Cournot (1927, p. 127). 15 Some economists reserve the term “representative-agent models” for models with a single, infinitely-lived agent. In a typical overlapping-generations model the new young are born at the start of every period, and the old die at the end of every period, and the model has infinitely many periods; so there are infinitely many agents. On this view, the overlapping-generations model is not a representative-agent model. I, however, regard it as one, because within any period one type of young agent and one type of old agent stand in for the enormous variety of people, and the same types are repeated period after period.
Quantitative Evaluation of Idealized Models in the New Classical Macroeconomics
27
aggregation requires not only that every economic agent have identical preferences but that these preferences are such that any individual agent would like to consume goods in the same ratios whatever their levels of wealth. The reason is straightforward: if agents with the same wealth have different preferences, then a transfer from one to the other will leave aggregate wealth unchanged but will change the pattern of consumption and possibly aggregate consumption as well; if all agents have identical preferences but prefer different combinations of goods when rich than when poor, transfers that make some richer and some poorer will again change the pattern of consumption and possibly aggregate consumption as well (Gorman 1953). The slightest reflection confirms that such conditions are never fulfilled in an actual economy. New classical macroeconomists insist on general equilibrium models. A fully elaborated general equilibrium model would represent each producer and each consumer and the whole range of goods and financial assets available in the economy. Agents would be modeled as making their decisions jointly so that, in the final equilibrium, production and consumption plans are individually optimal and jointly feasible. Such a detailed model is completely intractable. The new classicals usually obtain tractability by repairing to representativeagent models, modeling a single worker/consumer, who supplies labor in exchange for wages, and a single firm, which uses this labor to produce a single good that may be used indifferently for consumption or as a capital input into the production process. Labor, consumption, and capital are associated empirically with their aggregate counterparts. Although these models omit most of the details of the fully elaborated general equilibrium model, they nonetheless model firms and worker/consumers as making individually optimally and jointly consistent decisions about the demands for and supplies of labor and goods. They remain stripped-down general equilibrium models. One interpretation of the use of calibration methods in macroeconomics is that the practitioners recognize that highly aggregated, theoretical representative-agent models must be descriptively false, so that estimates of them are bound to fit badly in comparison to atheoretical (phenomenal) econometric models. The theoretical models are nonetheless to be preferred because useful policy evaluation is possible only within tractable models. In this, they are exactly like Lucas’s benchmark consumption model (see section III above). Calibrators appeal to microeconomic estimates of key parameters because information about individual agents is lost in the aggregation process. In general, these microeconomic estimates are not obtained using methods that impose the discipline of individual optimality and joint feasibility implicit in the general equilibrium model. Lucas (1987, pp. 46, 47) and Prescott (1986, p. 15) argue that the strength of calibration is that it uses multiple sources of information, supporting the belief that it is structured around true invariants. Again this
28
Kevin D. Hoover
comes close to endorsing a view of capacities as invariant dispositions independent of context. In contrast, advocates of direct estimation could argue that the idealized representative-agent model permits better use of other information not employed in microeconomic studies. Hansen and Sargent (1980, pp. 91, 92), for example, argue that the strength of their estimation method is that it accounts consistently for the interrelationships between constituent parts of the model; i.e., it enforces the discipline of the general equilibrium method – individual optimality and especially joint feasibility. The tradeoff between these gains and losses is not clear cut. Since both approaches share the representative-agent model, they also share a common disability: using the representative-agent model in any form begs the question by assuming that aggregation does not fundamentally alter the structure of the aggregate model. Physics may again provide a useful analogy. The laws that relate the pressures, temperatures and volumes of gases are macrophysics. The ideal laws can be derived from a micromodel: gas molecules are assumed to be point masses, subject to conservation of momentum, with a distribution of velocities. An aggregation assumption is also needed: the probability of the gas molecules moving in any direction is taken to be equal. Direct estimation of the ideal gas laws shows that they tend to break down – and must be corrected with fudge factors – when pushed to extremes. For example, under high pressures or low temperatures the ideal laws must be corrected according to van der Waals’s equation. This phenomenal law, a law in macrophysics, is used to justify alterations of the micromodel: when pressures are high one must recognize that forces operate between individual molecules. Despite some examples of macro-to-micro inferences analogous to the gas laws, Lucas’s (1980, p. 291) more typical view is that we must build our models up from the microeconomic to the macroeconomic. Unlike gases, human society does not comprise homogeneous molecules, but rational people, each choosing continually. To understand (verstehen) their behavior, one must model the individual and his situation. This insight is clearly correct. It is not clear in the least that it is adequately captured in the heroic aggregation assumptions of the representative-agent model. The analog for physics would be to model the behavior of gases at the macrophysical level, not as derived from the aggregation of molecules of randomly distributed momenta, but as a single molecule scaled up to observable volume – a thing corresponding to nothing ever known to nature.16
⎯⎯⎯⎯⎯⎯⎯ 16
A notable, non-new classical attempt to derive macroeconomic behavior from microeconomic behavior with appropriate aggregation assumptions is ( Durlauf 1989).
Quantitative Evaluation of Idealized Models in the New Classical Macroeconomics
29
6. Lessons for Econometrics Does the calibration methodology amount to a repudiation of econometrics? Clearly not. At some level, econometrics still helps to supply the values of the parameters of the models. Beyond that, whatever I have said in favor of calibration methods notwithstanding, the misgivings of econometricians such as Sargent are genuine. The calibration methodology, to date, lacks any discipline as stern as that imposed by econometric methods. For Lucas (1980, p. 288) and Prescott (1983, p. 11), the discipline of the calibration method comes from the paucity of free parameters. But one should note that theory places only loose restrictions on the values of key parameters. In the practice of the new classical calibrators, they are actually pinned down from econometric estimation at the microeconomic level or accounting considerations. In some sense, then, the calibration method would appear to be a kind of indirect estimation. Thus, it would be a mistake to treat calibration as simply an alternative form of estimation, although it is easy to understand why some critics interpret it that way. Even were there less flexibility in the parameterizations, the properties ascribed to the underlying components of the idealized representative-agent models – the agents, their utility functions, production functions, and constraints – are not subject to as convincing cross-checking as the analogous components in physical models usually are. The fudge factors that account for the discrepancies between the ideal model and the data look less like van der Waals’s equation, less like phenomenal laws, than like special pleading. Above all, it is not clear on what standards competing but contradictory models are to be compared and adjudicated.17 Some such standards are essential if any objective progress is to be made in economics.18 The calibration methodology is not, then, a repudiation of econometrics; yet it does contain some lessons for econometrics. In (Hoover 1994), I distinguish between two types of econometrics. Econometrics as observation treats econometric procedures as filters that process raw data into statistics. On this view, econometric calculations are not valid or ⎯⎯⎯⎯⎯⎯⎯ 17
Prescott (1983, p. 12), seems, oddly, to claim that the inability of a model to account for some real events is a positive virtue – in particular, that the inability of real-business-cycle models to account for the Great Depression is a point in their favor. He writes: “If any observation can be rationalized with some approach, then that approach is not scientific.” This seems to be a confused rendition of the respectable Popperian notion that a theory is more powerful the more things it rules out. But one must not mistake the power of a theory with its truth. Aside from issues of tractability, a theory that rationalizes only and exactly those events that actually occur, while ruling out exactly those events that do not occur is the perfect theory. In contrast, Prescott seems inadvertently to support the view that the more exceptions, the better the rule. 18 Watson (1993) develops a goodness-of-fit measure for calibrated models. It takes into account that, since idealization implies differences between model and reality that may be systematic, the errors-in-variables and errors-in-equations statistical models are probably not appropriate.
30
Kevin D. Hoover
invalid, but useful if they reveal theoretically interpretable facts about the world and not useful if they do not. Econometrics as measurement treats econometric procedures as direct measurements of theoretically articulated structures. This view is the classic Cowles Commission approach to structural estimation that concentrates on testing identified models specified from a priori theory.19 Many new classicals, such as Cooley and LeRoy (1985) and Sargent (1989), advocate econometrics as measurement. From a fundamental new classical perspective, they seem to have drawn the wrong lesson from the Lucas critique. Recall that the Lucas critique links traditional econometric concerns about identification and autonomy. New classical advocates of economics as observation overemphasize identification. Identification is achieved through prior theoretical commitment. The only meaning they allow for “theory” is general equilibrium microeconomics. Because such theory is intractable, they repair to the representative-agent model. Unfortunately, because of the failure of the conditions for exact aggregation to obtain, the representative-agent model does not represent the actual choices of any individual agent. The representative-agent model applies the mathematics of microeconomics, but in the context of econometrics as measurement it is only a simulacrum of microeconomics. The representative-agent model does not solve the aggregation problem; it ignores it. There is no reason to think that direct estimation will capture an accurate measurement of even the average behavior of the individuals who make up the economy. In contrast, calibrators use the representative-agent model precisely to represent average or typical behavior, but quantify that behavior independently of the representative-agent model. Thus, while it is problematic at the aggregate level, calibration can use econometrics as measurement, when it is truly microeconometric – the estimation of fundamental parameters from cross-section or panel data sets. Calibrators want their models to mimic the behavior of the economy; but they do not expect economic data to parameterize those models directly. Instead, they are likely to use various atheoretical statistical techniques to establish facts about the economy that they hope their models will ultimately imitate. Kydland and Prescott (1990, pp. 3, 4) self-consciously advocate a modern version of Burns and Mitchell’s “measurement without theory” – i.e., econometrics as observation. Econometrics as observation does not attempt to quantify fundamental invariants. Instead it repackages the facts already present in the data in a manner that a well calibrated model may successfully explain.
⎯⎯⎯⎯⎯⎯⎯
19
For a general history of the Cowles Commission approach, see Epstein (1987), ch. 2.
Quantitative Evaluation of Idealized Models in the New Classical Macroeconomics
31
7. Conclusion The calibration methodology has both a wide following and a substantial opposition within the new classical school. I have attempted to give it a sympathetic reading. I have concentrated on Prescott and Lucas, as its most articulate advocates. Calibration is consistent with appealing accounts of the nature and role of models in science and economics, of quantification and idealization. The practical implementation of calibration methods typical of new classical representative-agent models is less convincing. The calibration methodology stresses that one might not wish to apply standard measures of goodness-of-fit (e.g., R2 or tests of overidentifying restrictions) such as are commonly applied with the usual econometric estimation techniques, because it is along only selected dimensions that one cares about the model’s performance at all. This is completely consistent with Simon’s account of artifacts. New classical economics has traditionally been skeptical about discretionary economic policies. They are therefore more concerned to evaluate the operating characteristics of policy rules. For this, the fit of the model to a particular historical realization is largely irrelevant, unless it assures that the model will also characterize the future distribution of outcomes. The implicit claim of most econometrics is that it does assure a good characterization. Probably most econometricians would reject calibration methods as coming nowhere close to providing such assurance. Substantial work remains to be done in establishing objective, comparative standards for judging competing models. Fortunately, even those converted to the method need not become Lucasians: methodology underdetermines substance. Simon, while providing Lucas with a foundation for his views on modeling, nonetheless prefers a notion of “bounded rationality” that is inconsistent with the rational expectations hypothesis or Lucas’s general view of humans as efficient optimizers.20 Favero and Hendry (1989) agree with Lucas over the importance of invariance, but seek to show that not only can invariance be found at the level of aggregate econometric relations (e.g., in the demand-for-money function), but that this rules out rational expectations as a source of noninvariance.21 Finally, to return to a physical analog, economic modeling is like the study of cosmology. Substantial empirical work helps to determine the values of key constants; their true values nonetheless remain doubtful. Different values within the margins of error, even given similarly structured models, may result in very ⎯⎯⎯⎯⎯⎯⎯ 20
E.g., Simon (1969, p. 33) writes: “What do these experiments tell us? First, they tell us that human beings do not always discover for themselves clever strategies that they could readily be taught (watching a chess master play a duffer should also convince us of that).” 21 Favero and Hendry (1989) reject the practical applicability of the Lucas critique for the demand for money in the U.K.; Campos and Ericsson (1988) reject it for the consumption function in Venezuela.
32
Kevin D. Hoover
different conclusions (e.g., that the universe expands forever or that it expands and then collapses). Equally, the same values, given the range of competing models, may result in very different conclusions. Nevertheless, we may all agree on the form that answers to cosmological or economic questions must take, without agreeing on the answers themselves.* Kevin D. Hoover Department of Economics University of California, Davis
[email protected] REFERENCES Altug, S. (1989). Time-to-Build and Aggregate Fluctuations: Some New Evidence. International Economic Review 30, 889-920. Campos, J. and Ericsson, N. R. (1988). Econometric Modeling of Consumers’ Expenditure in Venezuela. Board of Governors of the Federal Reserve System International Finance Discussion Paper, no. 325. Cartwright, N. (1983). How the Laws of Physics Lie. Oxford: Clarendon Press. Cartwright, N. (1989). Nature’s Capacities and their Measurement. Oxford: Clarendon Press. Cooley, T. F. and LeRoy, S. F. (1985). Atheoretical Macroeconometrics: A Critique. Journal of Monetary Economics 16, 283-308. Cournot, A. ([1838] 1927). Researches into the Mathematical Principles of the Theory of Wealth. Translated by Nathaniel T. Bacon. New York: Macmillan. Diamond, P. A. (1984). A Search-Equilibrium Approach to the Micro Foundations of Macroeconomics: The Wicksell Lectures, 1982. Cambridge, Mass.: MIT Press. Durlauf, S. N. (1989). Locally Interacting Systems, Coordination Failure, and the Behavior of Aggregate Activity. Unpublished typescript, November 5th. Epstein, R. J. (1987). A History of Econometrics. Amsterdam: North-Holland. Favero, C. and Hendry, D. F. (1989). Testing the Lucas Critique: A Review. Unpublished typescript. Friedman, M. (1953). The Methodology of Positive Economics. In: Essays in Positive Economics. Chicago: Chicago University Press. Friedman, M. (1957). A Theory of the Consumption Function. Princeton: Princeton University Press. Gorman, W. M. (1953). Community Preference Fields. Econometrica 21, 63-80. Hansen, L. P. and Sargent, T. J. (1980). Formulating and Estimating Dynamic Linear Rational Expectations Models. In R. E. Lucas, Jr. and T. J. Sargent (eds.), Rational Expectations and Econometric Practice. London: George Allen & Unwin. Hoover, K. D. (1984). Two Types of Monetarism. Journal of Economic Literature 22, 58-76. Hoover, K. D. (1988). The New Classical Macroeconomics: A Skeptical Inquiry. Oxford: Blackwell.
⎯⎯⎯⎯⎯⎯⎯ *
I thank Thomas Mayer, Kevin Salyer, Judy Klein, Roy Epstein, Nancy Cartwright, and Steven Sheffrin for many helpful comments on an earlier version of this paper.
Quantitative Evaluation of Idealized Models in the New Classical Macroeconomics
33
Hoover, K. D. (1991). Scientific Research Program or Tribe? A Joint Appraisal of Lakatos and the New Classical Macroeconomics. In M. Blaug and N. de Marchi (eds.), Appraising Economic Theories: Studies in the Application of the Methodology of Research Programs. Aldershot: Edward Elgar. Hoover, K. D. (1992). Reflections on the Rational Expectations Revolution in Macroeconomics. Cato Journal 12, 81-96. Hoover, K. D. (1994). Econometrics as Observation: The Lucas Critique and the Nature of Econometric Inference. Journal of Economic Methodology 1, 65-80. Kydland, F. E. and Prescott, E. C. (1982). Time to Build and Aggregate Fluctuations. Econometrica 50, 1345-1370. Kydland, F. E. and Prescott, E. C. (1990). Business Cycles: Real Facts and a Monetary Myth. Federal Reserve Bank of Minneapolis Quarterly Review 14, 3-18. Kydland, F. E. and Prescott, E. C. (1991). The Econometrics of the General Equilibrium Approach to Business Cycles. Scandinavian Journal of Economics 93, 161-78. Lucas, R. E., Jr. (1976). Econometric Policy Evaluation: A Critique. Reprinted in Lucas (1981). Lucas, R. E., Jr. (1980). Methods and Problems in Business Cycle Theory. Reprinted in Lucas (1981). Lucas, R. E., Jr. (1981). Studies in Business-Cycle Theory. Oxford: Blackwell. Lucas, R. E., Jr. (1987). Models of Business Cycles. Oxford: Blackwell. Morgan, M. S. (1990). The History of Econometric Ideas. Cambridge: Cambridge University Press. Prescott, E. C. (1983). ‘Can the Cycle be Reconciled with a Consistent Theory of Expectations?’ or A Progress Report on Business Cycle Theory. Federal Reserve Bank of Minneapolis Research Department Working Paper, No. 239. Prescott, E. C Prescott, E. C. (1986). Theory Ahead of Business Cycle Measurement. Federal Reserve Bank of Minneapolis Quarterly Review 10, 9-22. Sargent, T. J. (1989). Two Models of Measurements and the Investment Accelerator. Journal of Political Economy 97, 251-287. Sheffrin, S. M. (1983). Rational Expectations. Cambridge: Cambridge University Press. Shoven, J. B. and Whalley, J. (1984). Applied General-equilibrium Models of Taxation and International Trade. Journal of Economic Literature 22, 1007-1051. Simon, H. A. (1969). The Sciences of the Artificial. Cambridge, Mass.: The MIT Press. Watson, M. W. (1993). Measures of Fit for Calibrated Models. Journal of Political Economy 101, 1011-41.
This page intentionally left blank
John Pemberton WHY IDEALIZED MODELS IN ECONOMICS HAVE LIMITED USE
1. Introduction This paper divides idealized models into two classes – causal and non-causal – according to whether the idealized model represents causes or not. Although the characterization of a causal idealized model may be incomplete, it is sufficiently well-defined to ensure that idealized models specified using restrictive antecedent clauses are non-causal. The contention of this paper is that whilst such idealized models are commonly used in economics, they are unsatisfactory; they do not predict reliably. Moreover, notions of causation that cut across such models are required to suggest when the idealized model will provide a sufficiently good approximation and when it will not. Doubt is cast on the ability of simple causal idealized models to capture sufficiently the causal complexity of economics in such a way as to provide useful predictions. The causalist philosophical standpoint of this paper is close to that of Nancy Cartwright.
2. Causal and Non-Causal Idealized Models For the purposes of this paper a “causal idealized model” is an idealized model that rests on simple idealized causes. The idealized models represent causes, or the effects of causes, that operate in reality. The inverse square law of gravitational attraction has an associated idealized model consisting of forces operating on point masses, and this is a causal idealized model by virtue of the fact that the forces of the model are causes. The standard idealized models of the operation of a spring, a pendulum or an ideal gas are causal by virtue of the fact that there are immediately identifiable causes that underpin these models. Cartwright shows, in Nature’s Capacities and their In: Martin R. Jones and Nancy Cartwright (eds.), Idealization XII: Correcting the Model. Idealization and Abstraction in the Sciences (Poznań Studies in the Philosophy of the Sciences and the Humanities, vol. 86), pp. 35-46. Amsterdam/New York, NY: Rodopi, 2005.
36
John Pemberton
Measurement, how certain idealized models of the operation of lasers were shown to be causal (Cartwright 1989, pp. 41-55). Causal idealized models can on occasion tell us what happens in real, nonideal situations when the causes identified in the idealized model are modeled sufficiently accurately and operate in a sufficiently undisturbed way in the real situation under consideration. It is the presence and operation of the identified causes in reality that, under the right circumstances, allow the idealized model to approximate the behavior of that reality. Empirical evidence may allow us to calibrate the degree of approximation. A “non-causal idealized model” is an idealized model that is not causal – it does not attempt to capture causes or the effects of causes that operate in reality. It is not apparent how a non-causal idealized model approximates the behavior of real, non-ideal situations. On occasion a model may regularly and accurately predict the behavior of a real system, although the causes that underpin such behavior have not been identified. Such was the case with the early models of lasers – Cartwright’s account tells how much importance the early laser scientists attached to finding the causes. Such an idealized model may be causal even though the causes are unknown.
3. Non-Causal Idealized Models in Economics Many idealized models used in economics are non-causal. Such idealized models are generally identified by restrictive antecedent clauses which are true of nothing actual. An example of such a restrictive antecedent clause occurs in the assumption of perfect knowledge that is used most commonly in relation to consumers or producers to underpin the achievement of certain maximizing behavior. The assumption is of course false of all actual consumers or producers; it is indeed true of nothing actual. It partially defines an idealized model – a model in which behavior is simpler and more predictable than in reality. Friedman’s (1953) essay “The Methodology of Positive Economics” still dominates the methodology of economics despite its apparent shortcomings (Friedman 1953). Nagel suggests that the important sense in which Friedman defends “unrealistic assumptions” is that they are restrictive antecedent clauses (Nagel 1963, pp. 217-18). Nagel’s analysis highlights an ambiguity underlying Friedman’s paper concerning the role of assumptions in idealized models: 1. The assumptions are merely a device for stating the conclusion so that it is only the conclusion which is descriptive of reality. In this case the assumptions and the idealized model are of little interest because they cannot support in any
Why Idealized Models in Economics Have Limited Use
37
way the validity of the conclusion which must stand or fall in accordance with available empirical evidence. Nagel’s example of such an assumption is “firms behave as if they were seeking rationally to maximize returns” (1963, p. 215). The assumption concerns the behavior of firms but does not require that they do actually rationally seek to maximize returns – merely that they behave as if this were the case. 2. The assumptions are understood as dealing with “pure cases” and the model is then descriptive of the pure case. The assumptions are then restrictive antecedent clauses. Nagel comments: “laws of nature formulated with reference to pure cases are not useless. On the contrary, a law so formulated states how phenomena are related when they are unaffected by numerous factors whose influence may never be completely eliminable but whose effects generally vary in magnitude with differences in attendant circumstances under which the phenomena recur. Accordingly, discrepancies between what is asserted for pure cases and what actually happens can be attributed to factors not mentioned in the law” (1963, pp. 215-16). Nagel labels idealized models “pure cases,” and suggests that the restrictive antecedent clauses that define them succeed by eliminating extraneous causes not dealt with by the law (1963, p. 216). The simple idealized world is tractable to analysis. The antecedent clauses define idealized models that are clearly noncausal. The key question is how behavior in such an idealized model relates to the behavior of systems in the real world. Behavior in the idealized model is derived using deductive analysis that appeals to the simplifying assumptions – the restrictive antecedent clauses. In practice, an implicit assumption is made but never stated in economics that the derived behavior pictured in the idealized model does approximate the behavior of systems in reality – this is here termed the “Approximate Inference Assumption” and is discussed below. Consider, for instance, the idealized model of perfect competition. The assumptions are normally stated broadly as follows (McCormick et al. 1983, p. 348): 1. Producers aim to maximize their profits and consumers are interested in maximizing their utility. 2. There are a large number of actual and potential buyers and sellers. 3. All actual and potential buyers and sellers have perfect knowledge of all existing opportunities to buy and sell. 4. Although tastes differ, buyers are generally indifferent among all units of the commodity offered for sale. 5. Factors of production are perfectly mobile.
38
John Pemberton
6. Productive processes are perfectly divisible; that is, constant returns to scale prevail. 7. Only pure private goods are bought and sold. Conclusions are typically derived by economists concerning behavior within such conditions over both the short run and the long run. The structure is causally complex. The short run/long run distinctions are not easy to explicate, so that it is far from clear, even in principle, that the behavior pictured in the idealized model could be produced in reality. How close is the behavior pictured in the model to the behavior of a given real economy? It is not clear how a user of this non-causal idealized model could begin to answer this question. The idealized model of perfect competition is a causally complex structure which, if the economists are right, allows certain functional relationships (in Russell’s (1913) sense) that are the effects of the behavior of firms, to be exhibited within the ideal circumstances prescribed by the assumptions. It is the contention of this paper that such an idealized model has very limited use in predicting the behavior of real economies.
4. The “Approximate Inference Assumption” The Approximate Inference Assumption (AIA) states that if a situation is “close” in some sense to the one captured by an idealized model then it may be inferred that the behavior in that situation is approximately that pictured in the idealized model. For a causal idealized model, the presence of the causes modeled in the real situation provides some basis for supposing that under the right circumstances such an approximation may hold. In the case of non-causal idealized models, the AIA is unsound. Chaos theory leads us to expect that in many situations the AIA is incorrect. Small changes to the boundary conditions of a model can lead to major changes in the behavior of the system. There are thus serious theoretic problems concerning the applicability of the AIA. Nevertheless Gibbard and Varian comment: It is almost always preposterous to suppose that the assumptions of the applied model are exactly true . . . The prevailing view is that when an investigator applies a model to a situation he hypothesizes that the assumptions of the applied model are close enough to the truth for his purposes (1978, pp. 668-9).
On this view of the behavior of economists, which seems to me to be correct, economists do indeed rely upon the AIA even in the case of non-causal idealized models.
Why Idealized Models in Economics Have Limited Use
39
5. Mathematical Economics Mathematical economics requires highly idealized models of reality. Within the mathematical model, results are derived and these are deemed to be descriptive of reality. In order to apply the results of the mathematical model to reality, the AIA is usually required. A particularly good example is Arrow and Debreu’s proof of the existence of a general equilibrium within an economy (Debreu 1959). The proof rests on modeling the economy as a multi-dimensional vector space in which certain sets of points are assumed to be compact – a quality which only in the loosest sense could be deemed to be descriptive of reality. This idealization of an economy is non-causal. Once the nature of the relationship between the mathematical model and reality is brought into question, and the problems with the AIA accepted, it becomes clear that the proof of the existence of an equilibrium within an economy is simply an artifact of the mathematical model, and does not demonstrate the existence of a corresponding counterpart in reality.
6. The Black and Scholes Argument The Black and Scholes (hereafter ‘B&S’) paper on option pricing is a first-class example in its economic field – mathematical in style and almost universally accepted (Black and Scholes 1973). For this reason it is also a good example of economics that, for its practical application and in order to claim empirical content, rests wholly upon the AIA to make inferences from a non-causal idealized model to reality. This paper uses the B&S argument to illustrate the use of the AIA and to give an example of its failure. B&S make the following assumptions: (a) The short-term interest rate is known and is constant through time. (b) The stock price follows a random walk in continuous time with a variance rate proportional to the square root of the stock price. Thus the distribution of the stock price at the end of any finite interval is lognormal. The variance rate of return on the stock is constant. (c) The stock pays no dividends or other distributions. (d) The option is “European,” that is, it can only be exercised at maturity. (e) There are no transaction costs in buying or selling the stock or the option. (f) It is possible to borrow any fraction of the price of a security to buy it or to hold it at the short-term interest rate. (g) There are no penalties for short selling. A seller who does not own a security will simply accept the price of the security from a buyer, and
40
John Pemberton
will agree to settle with the buyer on some future date by paying him an amount equal to the price of the security on that date. (a), (e), (f) and (g) are false. (c) is generally false. (d) is false if the option is American. In the case of (b) a far weaker assumption, namely that the price of the stock is a continuous function of time, is false. These assumptions function as restrictive antecedent clauses that define a non-causal idealized model. Using these assumptions B&S derive a solution to the option-pricing problem. By the end of their paper it is clear that the authors believe the solution in the idealized model situation is applicable to real options. It is rather surprising and worthy of note that many followers of B&S have not only employed, implicitly, the AIA but appear to have used a stronger assumption – a “precise inference assumption” – to the effect that the idealized solution is precise in real situations. Cox and Rubenstein for instance write a section of their leading textbook on option pricing under the title “An Exact Option Pricing Formula” (1985, pp. 165-252). B&S themselves conclude that: the expected return on the stock does not appear in the equation. The option value as a function of the stock price is independent of the expected return on the stock (1973, p. 644).
The failure of the expected return on the stock to appear as a parameter of the value in the option is a direct result of the powerful assumptions that B&S employ for defining their idealized model. They have not succeeded in showing that in real situations the expected return on the stock is not a parameter in the value of an option. This logically incorrect conclusion demonstrates their use of the AIA. (This is the equivalent of the conclusion that walking uphill is as quick as walking downhill in the Narrow Island analogy below.) Many economists have sought to demonstrate the relevance of the B&S solution to real options by showing that similar results hold under different (usually claimed to be weaker) assumptions than those used for the B&S idealized model. Beenstock’s attempt in “The Robustness of the Black-Scholes Option Pricing Model” (1982) is typical. Unfortunately, these “relaxations” of the assumptions merely tell us what would happen in ideal circumstances and do little to bridge the gap to the real world. Beenstock’s first conclusion is that “[o]ption prices are sensitive to the stochastic processes that determine underlying stock prices . . . Relaxation of these assumptions can produce large percentage changes in option prices” (1982, p. 40). The B&S solution is not necessarily a good approximation even in the carefully controlled idealized situations where all except one of their assumptions are held constant. The AIA simply does not work. A more tangible practical example of the failure of the AIA arises when the stock price movement is discontinuous. Discontinuities arise in a wide range of
Why Idealized Models in Economics Have Limited Use
41
practical situations, but perhaps most markedly in relation to bids. A situation arose recently where one morning a major stake in a top 100 company changed hands. The purchaser made known its intention to make a statement concerning its intentions at 1pm. Expert analysts considered there to be a significant possibility of a bid – probabilities in the 10% to 50% range were typical assessments. In the event of a bid, the stock price would rise by some 20% or more. In the event of no bid, the stock price would drop some 10%. For a slightly in-the-money, short-dated call option, cash receipt at expiry would almost certainly be zero if no bid were received, and close to 20%-odd of the stock price per underlying share if a bid were forthcoming. Under any practical method of valuation, the value of the option must be close to the probability of a bid times 20%-odd of the stock price. The B&S solution simply does not work in this situation. (This breakdown of the B&S solution is the equivalent of the breakdown of Professor White’s solution for villages on opposite coasts at broader parts of Narrow Island (see below). The assumption that stock price moves are continuous is a good one most of the time, but is totally wrong here, just as Professor White’s assumption that the island is a line breaks down for the corresponding case.) Practicing options experts recognize the shortcomings of B&S and will commonly adjust their results for the non-lognormality of the stock price distribution since by an ad hoc increase to the stock price volatility estimate, empirical evidence shows that most real stock price distributions have higher kurtosis than the lognormal distribution. But B&S provides no basis for such adjustments. The key question is “If a real situation is close to the B&S assumptions, how close is it to the B&S conclusion?” The major complaint about the B&S derivation is that it does not allow an answer to this question. Their precise solution to the idealized case tells us nothing about real situations. In the case of the imminent bid announcement, B&S breaks down entirely. What characteristics of the real situation will ensure the B&S solution “works” sufficiently well? The B&S approach provides no answer.
7. The Narrow Island Analogy In the Southern Seas, some way to the east and south of Zanzibar is a thin strip of land that modern visitors call “Narrow Island.” An indigenous people inhabit the island whose customs derive from beyond the mists of time. At the northern end of the island is a hill cut at its midpoint by a high cliff which crashes vertically into the sea. On the headland above the cliff, which to local people is a sacred site, a shrine has been erected. It is the custom of the island that all able-bodied adults walk to the shrine to pay their respects every seventh day.
42
John Pemberton
Despite its primitive condition the island possesses the secret of accurate time-keeping using instruments that visitors recognize as simple clocks. In addition to a traditional name, each village has a “numeric name,” which is the length of time it takes to walk to the shrine, pay due respects and return to the village. The island being of a fair length, the numeric names stretch into the thousands. Many years ago one of the first visitors to the island from Europe was a traveler the islanders called Professor White. Modern opinion has it that White was an economist. What is known is that at the time of his visit the local people were wrestling with the problem of establishing how long it took to walk between villages on the island. The argument of Professor White is recorded as follows: The problem as it stands is a little intractable. Let us make some assumptions. Suppose that the island is a straight line. Suppose the speed of walking is uniform. Then the time taken to walk between two villages is half the absolute difference between their numeric names.
The islanders were delighted with this solution and more delighted still when they found how well it worked in practice. It was noted with considerable interest that the time taken to walk between two villages is independent of the height difference between them. As the island is quite hilly this astonishing result was considered an important discovery. Stories tell that so great was the islanders’ enthusiasm for their new solution that they attempted to share it with some of their neighboring islands. To this day there is no confirmation of the dreadful fate which is said to have befallen an envoy to Broad Island, inhabited by an altogether more fearsome people, who were apparently disappointed with the Narrow’s solution. Although the Narrow Islanders have used their solution happily for many years, more recently doubts have begun to emerge. In some parts of Narrow, where the island is slightly broader, reports suggest that the time between villages on opposite coasts is greater than Professor White’s solution would suggest. Others who live near the hills quite frankly doubt that the time taken to walk from the bottom of the hill to the top is the same as the time taken to walk from the top to the bottom. It is a tradition amongst the islanders that land is passed to sons upon their marriage. Upon the marriage of a grandson the grandfather moves to a sacred village near the middle of the island which is situated in the lee of a large stone known as the McCarthy. The McCarthy Stoners report that Professor White’s solution suggests it is quicker to walk to villages further south than it is to walk to villages that appear to be their immediate neighbors. The island’s Establishment continues to point out that Professor White’s solution gives “the time to walk between two villages,” so that the time taken on
Why Idealized Models in Economics Have Limited Use
43
any particular journey – resting as it does on a particular person in particular climactic conditions – is hardly adequate contrary evidence. Those who would talk of average times are considered to be wide of the mark. Whilst few islanders seriously doubt the veracity of the theory, it is reported that many have taken to making ad hoc adjustments to the official solution, a practice that is difficult to condone. Last year a visitor to the island was impolite enough to question the logic of Professor White’s solution itself. His main points were as follows: 1. The “time taken to walk between two villages” is not well defined. A workable definition might be “the average time taken to walk between the two villages in daylight hours, by able-bodied men between the ages of eighteen and forty, traveling at a comfortable pace, without rest stops, during normal climatic conditions, e.g., excluding unduly wet or windy weather.” 2. Both of the assumptions are incorrect. The conclusion is not always approximately correct. 3. A more robust solution which should provide answers that are approximately correct for all pairs of villages is to regress the average times as proposed in (1) against the two principal causal factors underlying journey times – horizontal distance and height of climb, positive or negative. If a linear equation does not provide a sufficiently good fit, a higher order polynomial might be used. To the Narrow Islanders such talk was incomprehensible.
8. A Robust Solution to the Option Pricing Problem The Narrow Island analogy shows how the use of a non-causal idealized model breaks down. It may represent a good approximation most of the time, but on occasion it is no approximation at all. Moreover, the model itself provides no clues as to when it is, and when it is not, applicable. We have this knowledge (if at all) from a consideration of broader factors; it would seem these must include causal factors. A causal idealized model always provides a good approximation whenever it captures enough of the causes sufficiently accurately, and the causes modeled operate in reality in a sufficiently undisturbed way. Often our knowledge of the situation will allow us to judge whether this is likely to be the case. The visitor’s solution to the Narrow Island problem is correct; it is a robust causal solution. A similar solution exists for the option pricing problem.
44
John Pemberton
In the absence of any generally accepted definition of investment value, we may choose as a robust measure the expected value of receipts using a discount rate appropriate to the net liabilities. The causal parameters of value are those which affect the stock price at expiry. These may be categorized as known and unknown causes. Known causes: 1. Expected growth of stock price. Unless the dividend yield is unusually high, growth in stock price is part of the anticipated investment return. 2. Bifurcating event. To the extent that the market is efficient (all causes known to any market participant are reflected in the stock price, say), the probability density function (pdf) that describes the stock price at expiry should be evenly distributed about the central expectation. A bifurcating event induces a double-humped pdf. 3. Market inefficiencies. If we have insight into a stock which differs from that of other market participants then we may on occasion anticipate a movement in the market price of the stock during the period to expiry. Unknown causes: John Stuart Mill referred to “disturbing causes,” causes that do not appear explicitly within the model (1967, pp. 330-31). Such unknown causes may be allowed for by choosing a pdf that allows for a range of outcomes. It is usual to use a lognormal distribution for stock prices, but empirical evidence suggests that the chosen distribution should have higher kurtosis than this. A pdf may be chosen that makes allowance for all the causes, both known and unknown, and the discounted expected value calculated to provide a robust approximate measure of value. Sensitivity to changing assumptions may be checked.
9. Causal Idealized Models in Economics In the physical sciences real working systems can sometimes be constructed that arrange for causes to operate in just the right way to sustain the prescribed behavior of the system. Pendula, lasers and car engines are examples. Whilst the systems are working, aspects of their behavior may be described by functional relationships (again, in Russell’s (1913) sense) between variables. The design of the system ensures consistency, repeatability and reversibility. Real economies are not so neat; they have a complex causal structure, and are more akin to machines with many handles, each of which is turned continually and independently. The state of the machine is continually changing and evolving as the handles are turned. The effect of turning a single handle is
Why Idealized Models in Economics Have Limited Use
45
not in general consistent, repeatable or reversible. Simple functional relationships between variables are at best approximate. Simple mathematical models do not capture such a complicated process, but may on occasion approximate part of the process for some duration. Economists do on occasion use causal idealized models. Examples are equilibrium models where causes are identified that tend to push the economic system back towards equilibrium. One such model equates money supply with money demand. Economics textbooks argue that behavioral patterns create causes that ensure equality holds “in equilibrium.” Such causal idealized models leave out the vast bulk of causes and simplify to an extreme extent. It is far from clear that such models can demonstrate any predictive success – their usefulness is at best very limited.
10. Conclusion Economists’ use of non-causal idealized models is problematic. No account can be provided in terms of the idealized model as to when the model will provide a sufficient approximation to reality. Such models do not predict reliably. Moreover, the causal complexity of economies suggests that they may be intractable to simple models, so that causal idealized models may also have very limited predictive success. John Pemberton Institute of Actuaries, London
[email protected] REFERENCES Beenstock, M. (1982). The Robustness of the Black-Scholes Option Pricing Model. The Investment Analyst, October 1982, 30-40. Black, F. and Scholes, M. (1973). The Pricing of Options and Corporate Liabilities. Journal of Political Economy 82, 637-54. Cartwright, N. (1989). Nature’s Capacities and their Measurement. Oxford: Clarendon Press. Cox, J. C. and Rubinstein, M. (1985). Options Markets. Englewood Cliffs, NJ: Prentice-Hall. Debreu, G. (1959). Theory of Value: An Axiomatic Analysis of Economic Equilibrium. New York: Wiley. Friedman, M. (1953). The Methodology of Positive Economics. In: Essays in Positive Economics, pp. 637-54. Chicago: University of Chicago Press. Gibbard, A. and Varian, H. (1978). Economic Models. Journal of Philosophy 75, 664-77. McCormick, B. J., Kitchin, P. D., Marshall, G. P., Sampson, A. A., and Sedgwick, R. (1983). Introducing Economics. 3rd edition. Harmondsworth: Penguin.
46
John Pemberton
Mill, J. S. (1967). On the Definition of Political Economy. In: J. M. Robson (ed.), Collected Works, Vols. 4-5: Essays on Economics and Society, pp. 324-42. Toronto: Toronto University Press. Nagel, E. (1963). Assumptions in Economic Theory. American Economic Review: Papers and Proceedings 53, 211-19. Russell, B. (1913). On the Notion of Cause. Proceedings of the Aristotelian Society 13, 1-26.
Amos Funkenstein THE REVIVAL OF ARISTOTLE’S NATURE
1. The Problem In her recent book on Nature’s Capacities and their Measurement (1989), Nancy Cartwright argued forcefully for the recognition of “capacities” as an indispensable ingredient of causal scientific explanations. Idealizations in science assume them anyhow; abstract laws explain their causal impact; symbolic representation secures their proper formulation and formalization. So very satisfactory is the picture of the language and operation of science thus emerging that one wonders why the language of capacities, which indeed once dominated the language of science, was ever abandoned. Aristotle had based his systematic understanding of nature on potentialities (δυνάµεις) and their actualization; his terminology and perspective ruled throughout the Middle Ages, and was abandoned – at least explicitly – only in the seventeenth century. But why? The historical retrospection may also lead us towards some more systematic insights into the difficulties involved in the notion of nature’s capacities. These difficulties may not be insurmountable: and it may or may not turn out to be the case that we stand more to gain than to lose by readopting the lost idiom with due modification. With this hope in mind I shall begin a historical account which, needless to say, is highly schematized and therefore not above suspicion of bias or error.
2. Aristotle and the Aristotelian Tradition Aristotle’s language of “potentialities” and their “actualization” reflects his inability or unwillingness to view nature as uniform, homogeneous, always and everywhere “true to itself” – uniform in the sense that it is always and everywhere, “in Europe and in America,” ruled by the same laws (Newton 1687,
In: Martin R. Jones and Nancy Cartwright (eds.), Idealization XII: Correcting the Model. Idealization and Abstraction in the Sciences (Poznań Studies in the Philosophy of the Sciences and the Humanities, vol. 86), pp. 47-58. Amsterdam/New York, NY: Rodopi, 2005.
48
Amos Funkenstein
p. 402).1 This latter view, I shall argue, eventually drove out the Aristotelian language of capacities, their realization or the lack thereof (“privation”). Indeed, Aristotle’s “nature” is rather a ladder of many “natures,” classified in orders down to the most specific nature. The “nature” of sublunar bodies was to him unlike the “nature” of celestial bodies. The former are made of a stuff which, by nature (φσύει), comes-to-be and decays; and its natural motion is “upwards” or “downwards” in a straight line – sublunar bodies are, of necessity, either light or heavy (1922, ã 2.301 a 20ff.). The latter are imperishable, and move by their nature in a perfect circle.2 Further down the ladder, we come across more particularized natures (forms) – until we reach the most specialized nature, the species specialissima. In order to secure the objectivity of his classification of nature, Aristotle commits himself to a heavy ontological, or at least a priori, assumption: namely, that a specific difference within a genus (say, rationality within mammals) can never appear in another genus (say, metals);3 at best he admits analogous formations (1912, A 4.644 a 15ff.). To this view of “natures” of diverse groups of things Aristotle’s predicatelogic was a most suitable “instrument” (organon). All scientific propositions, inasmuch as they gave account of the nature of things, should be able to be cast into the form S ε P. But what if S, a subject, actually lacks a predicate which “by nature” belongs to it? In the case of such an unrealized “capacity” Aristotle spoke of “privation” (στερήσις). So important to his scientific enterprise was this concept that he did not shy away from ranking it, together with form (ει’δός) and matter (υ‛πουείµενον), as chief “causes” or “principles” of all there is (1957, Λ 2, 1069 b 32-4; ∆ 22, 1022 b 22-1023 a 7; and cf. Wolfson 1947). But it is an ambiguous notion. Already the logicians of Megara recognized that it commits Aristotle to the strange view that a log of wood, even if it stays under water for the duration of its existence, is nonetheless “burnable” (Kneale and Kneale 1962, pp. 117-28). Worse still, while the negation of a negation, assuming the principle of excluded middle, is perfectly unequivocal, the negation of a privation is not; it either negates the proposition that S ε P or says that the proposition S ε P is a category mistake, but not both.4 ⎯⎯⎯⎯⎯⎯⎯ 1
I have elaborated this demand for homogeneity in my (1986), pp. 29, 37-9, and 63-97. The Greek “obsession with circularity” (Koyré) can be said to have ruled, unchallenged, until Kepler – even while any other astronomical presupposition, e.g., geocentricity, was debated. An exception of sorts – in very vague terms – were the atomists: cf. Cicero (1933), 1.10.24. 3 Aristotle (1958), Z 6.144 b 13ff., and (1957), Z 12, 1038 a 5-35. That Aristotle actually dropped this requirement in his biological research was argued by Furth (1988, pp. 97-8). He also joins those who maintain that Aristotle’s main objective was the complete description of species, not their hierarchical classification. On the methodological level, this principle corresponds to the demand not to apply principles of one discipline to another: Aristotle (1960), A 7.75 a 38-75 b 6. Cf. Livesey (1982), pp. 1-50, and Lloyd (1987), p. 184ff. 4 Because of this ambiguity, it could serve Maimonides’s negative theology as a way to generate divine attributes; cf. my (1986), p. 53, nn. 41 and 44. 2
The Revival of Aristotle’s Nature
49
The coalescence of properties that makes a singular thing into that specific thing is its “form” – medieval schoolmen will speak of the “substantial form.” The form determines the capacities of a thing: if an essential property is missing “by accident” from that thing, we speak of a privation, as when we say that Homer was blind or Himmler inhuman. For, below the level of the infima species, what particularizes an object (down to its singular features) is matter, not form: wherefore Aristotle could never endorse Leibniz’s principle of the identity of indiscernibles. Only some unevenness in their matter explains why one cow has a birthmark on her left shoulder while her twin has none. Matter, however, can by definition not be cognited – if cognition is, as Aristotle thought, an assimilatory process of “knowing the same by the same,” an identity of the form of the mind with that of the object cognited.5 There is no direct knowledge of singulars qua singulars. Nor is it needed for the purposes of science, which always consists of knowledge of common features, of universals. Such, in broad outlines, was the meaning of “nature” in the most successful body of scientific explanations during antiquity and in the Middle Ages. Indeed, it articulated a deep-seated linguistic intuition. The Greek word φύσις, much as the Latin natura, is derived from the verb “to be born, to become” (φύω, nasci). The “nature” of a thing is the set of properties with which this thing came into the world – in contrast to acquired, “mechanical” or “accidental” properties. Milk teeth are called, in ancient Greek, “natural teeth” (φύσικοι o’δόντες). A “slave by nature” is a born slave, as against an “accidental” slave captured and sold by pirates or conquerors (Aristotle 1957, A5, 1254 b 25-32, and A6, 1255 a 3ff.). At times, custom will be called “second nature.” In short: the term “nature” was attached, in common discourse, to concrete entities. The set of properties with which something is born was its “nature.”6 Aristotle accommodated this intuition – with the important condition that such essential properties be thought of as capacities, as discernible potentialities to be actualized. Were there, in antiquity, other conceptions than Aristotle’s hierarchy of “natures,” anticipations of a more modern sense of the uniformity of nature? Indeed there were, but they remained, through the Middle Ages, vague and marginal. From its very beginning in the sixth century, Greek science sought the “causes” (αι’τία) of all things. This demand of the φυσιολόγοι reached its culmination with Parmenides, who postulated the existence of but one indivisible being (τò ‛o′ν), without differentiation or internal qualification, ⎯⎯⎯⎯⎯⎯⎯ 5
‛′µοιον ωˆ o‛µοίω see Schneider (1923). It is the basis for On the Aristotelian assumption of o Aristotle’s distinction between a “passive” intellect, which “becomes everything,” and an active intellect, which “makes everything” (πάντα γίγνεται – πάντα ποιειˆν) (1961, Γ5, 430 a 14-15). This distinction leads to later, far-reaching theories of the “active intellect.” 6 Hence also the conviction that, explaining the origin of a nation, one uncovers its nature or character: cf. Funkenstein (1981), especially pp. 57-9.
50
Amos Funkenstein
without motion or change, negation or privation: it has, indeed, only a general “nature.” Every subsequent Greek cosmological theory had to take up the Parmenidean challenge; each of them tried to gain back something from the domain of “illusion” or “opinion” for the domain of truth. Democritus, Plato, and Aristotle, each of them in his own way, divided Parmenides’s one “being” into many – be they atoms, forms, or natures. Only the Stoics renewed the idea of the homogeneity of the cosmos, a cosmos filled with a homogeneous, continuous matter suspended in an infinite space and penetrated everywhere by one force – the πνευˆµα – which holds it together by giving it tension (τόνος) (Sambursky 1965; Lapidge 1978). The Stoic cosmos, like their ideal world-state (κοσµόλογις), was subject everywhere to the same rules (ι’σονοµία). Instead of Aristotle’s “natures” or “essential forms” they looked for forces: indeed, the very term δύναµις – Aristotle’s “potentiality” – came to mean an active, internal, actual force (Sambursky 1965, pp. 217-9; Lapidge 1978, pp. 163-5 (ποˆιον); Frede 1987, pp. 125-50). Inasmuch as an entity, a body, is saturated with πνευˆµα, it desires to preserve its existence: “omnis natura vult esse conservatrix sua.” All forces seemed to them as so many expressions of one force, which is God and nature at once – and of course also “matter”: “Cleanthes totius naturae mentis atque animo hoc nomen [dei] tribuit” (Cicero 1933, 1, 14, 37 and 4, 17, 16). Here nature is a totality. But its “capacities” are always actualized. In fact, Stoic logic, like the Megarian, did not recognize unactualized possibilities: the possible is that which happens at some time (Mates 1953, pp. 6, 36-41). Still another source for a more uniform “nature” was, in later antiquity as in the Middle Ages, the Neoplatonic conception of light as prime matter and all pervasive: “natura lucis est in omnibus” (Witelo 1908, 7.1, 8.1-4, 9.1-2; cf. also Crombie 1962, pp. 128-34). Close attention to Witelo’s wording reveals the two almost incompatible notions of “nature” at work: light is a definite body with a specific “nature” – this is the Aristotelian heritage; but it is in everything and everywhere – this is the Stoic, and sometimes Neoplatonic, sense of a homogeneous universe. A minority tradition in the Middle Ages, this view became popular, in many varieties, during the Renaissance. In their campaign against Aristotelianism, Renaissance philosophers of nature tried to revive alternative traditions – Presocratic, Stoic and Epicurean (Barker 1985; Barker and Goldstein 1984).
3. The Decline of the Aristotelian-Thomistic “Forms” While these deviant traditions were not without importance in the preparation for an alternative scheme of nature, a more serious and enduring challenge to Aristotle arose within the scholastic discourse itself in the turn from the
The Revival of Aristotle’s Nature
51
thirteenth to the fourteenth centuries. Aristotle relegated the differences between singulars of the same species specialissima to matter; matter is his (and Thomas’s) “principle of individuation.” Consequently, since knowledge is always knowledge of forms (whereby the knower and the known become identical),7 and since matter is that substrate which remains when forms are abstracted, individuals (singulars) can be known only by inference. As mentioned above, Aristotle could never have endorsed Leibniz’s principle of the identity of indiscernibles (and nor could Thomas) (Funkenstein 1986, pp. 135-40, 309-10). The theological difficulties of this position were numerous. It prevented God from creating more than one separate intelligence (angel) of the same species, since separate intelligences lack matter. It made individual post-existence almost impossible to account for. It posed difficulties for the beatific vision – knowledge of individuals qua individuals without the mediation of the senses (Day 1947). For these and other reasons, the whole scholastic discourse was changed between Thomas Aquinas and William of Ockham; and much of the change must be credited to Duns Scotus, who gave up “substantial forms” for the sake of a coalescence of many common forms in every singular being down to an individual form of “haecceity.” He also postulated an immediated, “intuitive” cognition of singulars qua existents: form, not matter, particularizes every individual down to its very singularity (Tachau 1988; Hochstetter 1927; Funkenstein 1986, p. 139, n. 42). Ockham and his followers even denied the need for a “principle of individuation”: individual things exist because they do, and their intuitive cognition is the ultimate basis for simple terms (incomplexa) which build propositions (complexa), scientific or other. “Forms,” “natures,” “universals” are no more than inductive generalizations of properties. They ceased to be the backbone of the understanding of “nature” (cf. also Eco 1988, pp. 205-10, and Pererius 1592, pp. 87-8). Whatever Aristotle’s physics declared to be conceptually impossible and contrary to nature, now became a possibility from the vantage point of God’s absolute power (de potentia Dei absoluta) – as long as it did not violate the principle of non-contradiction (see Funkenstein 1986, pp. 115-52, and the literature mentioned there). If he so wished, God could make stones move habitually upwards, or make the whole universe move rectilinearly in empty space, or save humankind not by sending his son, but rather with a stone, a donkey, or a piece of wood. With their penchant for devising counterfactual worlds, the schoolmen of the fourteenth century sought to enlarge the domain of God’s omnipotence without violating any purely logical principle; they virtually introduced Leibniz’s distinction between logical and less-than-logical necessity, ⎯⎯⎯⎯⎯⎯⎯ 7
Above, n. 5.
52
Amos Funkenstein
a distinction the ancients never made. Every order or nature was to them at best an empirical, contingent fact. How far did the new discourse really permeate the explanation of natural phenomena? A telling example of the sixteenth century shows how indeed Aristotle’s “natures,” “forms,” or Thomas’s “substantial forms” (with their natural capacities) were in fact dethroned, and turned into mere properties or forces. In his treatise on “Fallacious and Superstitious Sciences,” Benedict Pereira – whose works were read by Galileo – summed up, among other matters, the history and the état de question of discussions concerning the value of alchemy, and concluded: No philosophical reason exists, either necessary or very probable, by which it can be demonstrated that chemical accounts of making true gold are impossible (Pererius 1592, pp. 87-8).
The arguments against turning baser metals into gold that he found in the Aristotelian-Thomistic tradition (Avicenna, Averroës, Thomas, Aegidius Romanus) were all variations on the assumption that a substantial form induced by a natural agent cannot be induced artificially. Artificial gold may, they said, resemble natural gold, but it will always differ from it in some essential properties (amongst which is the therapeutic value of gold for cardiac diseases): just as a louse generated inorganically from dirt will differ in specie from a louse with respectable organic parents (Pererius 1592, pp. 75-82, esp. p. 78). Pereira discards this argument as unwarranted. The natural agent that molds other metals into gold is the sun’s heat, a property shared by fire, which can certainly be induced artificially – at least in principle. Yet Pereira, too, believed that it is, with our technical means, impossible to do so and that claims of success are fraudulent. While possible in principle, given our tools, knowledge and the fact that no case has been proven, it seems to him that it will at best be much more expensive to generate gold artificially, even should we know one day how to do so, than it is to mine it. The “forms” have visibly turned into qualities and forces – here as elsewhere in the natural philosophy of the Renaissance. Empirical evidence rather than a priori argument governs even the scholastic dispute. And the distance between the natural and the artificial seems considerably undermined. A new language of science was about to be forged. The role of fourteenth century scholastic debates in this process was largely a critical one: to purge the scientific discourse of excess ontological baggage.
The Revival of Aristotle’s Nature
53
4. Galileo’s Idealizations It became, then, increasingly unfashionable to ascribe natural “capacities” or “tendencies” to specific substances in an absolute sense. Now it is worthwhile to look at the process by which they were actually dethroned, their place to be occupied by “forces” or “tendencies” common to all bodies. Seldom can the actual point of transition from medieval to classical physics be identified as clearly as in the case of Galileo. As a young professor of mathematics in Pisa he tried to account for the free fall of bodies in the following way. When a body is heaved or thrown upwards, it accumulates an “impetus.”8 It is also endowed with more or less “gravity,” depending on its specific weight. Now, as long as g) represent the inferential relation used to characterize confirmation. A naive HD answer to our question would be that if o ---> h, evidence which makes o true provides epistemic support for the claim represented by h or the theory to which that claim belongs and evidence which makes ~o true provides epistemic support for the rejection of the claim represented by h. A simplified positive instance answer would be that the evidence represented by o provides epistemic support for the claim represented by h if o is true and o ---> h, while evidence counts against the claim if o is true and o ---> ~h.11
IV We have emphasized that the IRS depicts confirmation as depending upon formal relations between sentences in a first order language, even though many data are photographs, drawings, etc., which are not sentences in any language, let alone a first order one. This is enough to establish that the claim that confirmation captures what is essential to evidential relevance is not trivial. In fact that claim is problematic. Hempel’s raven paradox illustrates one of its ⎯⎯⎯⎯⎯⎯⎯ 10
Different versions of HD and positive instance theories add different conditions on confirmation to meet counterexamples which concern them. For example, o may be required to have a chance of being false, to be consistent with the theory whose claims are to be tested, to be such that its denial would count against the claim it would support, etc. The details of such conditions do not affect our arguments. Thus our discussion frequently assumes these additional conditions are met so that its being the case that o ---> h is sufficient for confirmation of the claims represented by h by the evidence represented by o. 11 See the previous note. For examples of this view, see Braithwaite (1953) and Glymour (1980), ch. V.
240
James Bogen and James Woodward
problems. Replacing the natural language predicate ‘is a raven’ with F, and ‘is black’ with G, let a hypothesis sentence (h1), (x) (Fx ⊃ Gx), represent the general claim (C) “All ravens are black.”12 Where a is a name, Fa & Ga is an instance of (h1). But (h1) is logically equivalent to (h2), (x)(~Gx ⊃ ~Fx). Now ~Fa & ~Ga entails ~Ga ⊃ ~Fa, an instance of (h2). Thus, ~Fa & ~Ga ---> (h2). But as Hempel observes, ~Fa & ~Ga is true when the referent of a is a red pencil (Hempel 1965, p. 15f. Cf. Glymour 1980, p. 15ff.). Therefore, assuming that evidence which confirms a claim also confirms claims which are logically equivalent to it, why shouldn’t the observation of a red pencil confirm (C)? If it does, this version of IRS allows evidence (e.g., red pencil observations) to confirm theoretical claims (like “All ravens are black”) to which it is epistemically quite irrelevant. Since the premises of a deductively valid argument cannot fail to be relevant to its conclusion, this (along with such related puzzles as Goodman’s grue riddle and Glymour’s problem of irrelevant conjunction (Glymour 1980, p. 31), points to a serious disanalogy between deductive validity and confirmation. While the deductive validity of an argument guarantees in every case that if its premises are true, then one has a compelling reason to believe its conclusion, the evidence represented by o can be epistemically irrelevant to the hypothesis represented by h even though o ---> h. The most popular response to such difficulties is to tinker with IRS by adding or modifying the formal requirements for confirmation. But close variants of the above puzzles tend to reappear in more complicated IRS models.13 We think this is symptomatic of the fact that evidential relevance depends upon features of the causal processes by which the evidence is produced and that the formal resources IRS has at its disposal are not very good at capturing or tracking these factors or the reasoning which depends upon them. This is why the tinkering doesn’t work. An equally serious problem emerges if we consider the following analogy: Just as we can’t tell whether we must accept the conclusion of a deductively valid argument unless we can decide whether its premises are true, the fact that o ---> h doesn’t give us any reason to believe a theoretical claim unless we can decide whether o is true. To see why this is a problem for the IRS consider the test Priestley and Lavoisier used to show that the gas produced by heating oxides of mercury, iron, and some other metals differ from atmospheric air.14 Anachronistically described, it depends on the fact that the gas in question was oxygen and that oxygen combines with what Priestley called “nitrous air” (nitric ⎯⎯⎯⎯⎯⎯⎯ 12
Examples featuring items which sound more theoretical than birds and colors are easily produced. For an illustration of this point in connection with Glymour’s treatment of the problem of irrelevant conjunction, see Woodward (1983). 14 This example is also discussed in Bogen and Woodward (1992). 13
Evading the IRS
241
oxide) to produce water-soluble nitrous oxide. To perform the test, one combines measured amounts of nitric oxide and the gas to be tested over water in an inverted graduated tube sealed at the top. As the nitrous oxide thus produced dissolves, the total volume of gas decreases, allowing the water to rise in the tube. At fixed volumes, the more uncompounded oxygen a gas contains, the greater will be the decrease in volume of gas. The decrease is measured by watching how far the water rises. In their first experiments with this test, Priestley and Lavoisier both reported that the addition of “nitrous air” to the unknown gas released from heated red oxide of mercury decreased the volume of the latter by roughly the amount previously observed for atmospheric air. This datum could not be used to distinguish oxygen from atmospheric air. In later experiments Priestley obtained data which could be used to make the distinction. When he added three measures of “nitrous air” to two measures of the unknown gas, the volume of gas in the tube dropped to one measure. Lavoisier eventually “obtained roughly similar results” (see Conant 1957; Lavoisier 1965, pt. I, chs. 1-4). The available equipment and techniques for measuring gases, for introducing them into the graduated tube, and for measuring volumes were such as to make it impossible for either investigator to obtain accurate measures of the true decreases in volume (Priestley 1970, pp. 23-41). Therefore an IRS account which thinks of the data as putative measures of real decreases should treat observation sentences representing the data from the later as well as from the earlier experiments as false. But while unsound deductive arguments provide no epistemic support for their conclusions, the inaccurate data from Priestley’s and Lavoisier’s later experiments provide good reason to believe the phenomena claim for which they were used to argue. Alternatively, suppose the data were meant to report how things looked to Priestley and Lavoisier instead of reporting the true magnitudes of the volume decreases. If Priestley and Lavoisier were good at writing down what they saw, observation sentences representing the worthless data from the first experiments should be counted as true along with observation sentences representing the epistemically valuable data from the later experiments. But while all deductively sound arguments support their conclusions, only data from the later experiments supported the claim that the gas released from red oxide or mercury differs from atmospheric air. Here the analogy between deductive soundness and confirmation by good evidence goes lame unless the IRS has a principled way to assign true observation sentences to the inaccurate but epistemically valuable data from the later experiments, and false observation sentences to the inaccurate but epistemically worthless data from the earlier experiments. If truth values must be allocated on the basis of something other than the accuracy of the data they represent, it is far from clear how the IRS is to allocate them.
242
James Bogen and James Woodward
To avoid the problem posed by Priestley’s and Lavoisier’s data the IRS must assign true observation sentences to epistemically good evidence and false ones to epistemically bad evidence. What determines whether evidence is good or bad? The following example illustrates our view that the relevance of evidence to theory and the epistemic value of the evidence depends in large part upon causal factors. If we are right about this, it is to be expected – as we will argue in sections VII and VIII – that decisions about the value of evidence depend in large part upon a sort of causal reasoning concerned with what we are calling reliability.
V Curtis and Campbell, Eddington and Cottingham (among others) produced astronomical data to test Einstein’s theory of general relativity. One of the phenomena general relativity can be used to make predictions about is the deflection of starlight due to the gravitational influence of the sun. Eddington and the others tried to produce data which would enable investigators to decide between three competing claims about this phenomenon: (N) no deflection at all, (E) deflection of the magnitude predicted by general relativity, and (NS) deflection of a different magnitude predicted by Soldner from Newtonian physics augmented by assumptions needed to apply it to the motion of light.15 (N) and (NS) would count against general relativity while (E) would count in favor of it. The data used to decide between these alternatives included photographs of stars taken in daytime during a solar eclipse, comparison photographs taken at night later in the year when the starlight which reached the photographic equipment would not pass as near to the sun, and check photographs of stars used to establish scale. To interpret the photographs, the investigators would have to establish their scale, i.e., the correspondence of radial distances between stars shown on an accurate star map to linear distances between star images on the photographs. They would have to measure differences between the positions of star images on the eclipse and the comparison photographs. They would have to calculate the deflection of starlight in seconds of arc from displacements of the star images together with the scale. At each step of the way they would have to correct for errors of different kinds from different sources (Earman and Glymour 1980, p. 59). The evidential bearing of the photographic data on Einstein’s theory is an instance of what IRS accounts of confirmation are supposed to explain. This evidential bearing depended upon two considerations: (1) the usefulness of the data in discriminating between phenomena claims (N), (NS), (E), and (2) the ⎯⎯⎯⎯⎯⎯⎯ 15 Soldner’s is roughly the same as a value predicted from an earlier theory of Einstein’s. See Pais (1982), p. 304.
Evading the IRS
243
degree to which (E), the value predicted by general relativity, disagrees with predictions based on the competitor theories under consideration. (1) belongs to the first of the two stages of theory testing we mentioned in section I: the production and interpretation of data to answer a question about phenomena. (2) belongs to the second of these stages – the use of a phenomena claim to argue for or against part of a theory. With regard to the first of these considerations, evidential relevance depends upon the extent (if any) to which differences between the positions of star images on eclipse and comparison pictures are due to differences between paths of starlight due to the gravitational influence of the sun. Even if the IRS has the resources to analyze the prediction of (E) from Einstein’s theory, the relevance of the data to (E) would be another matter.16 Assuming that the sun’s gravitational field is causally relevant to differences between positions of eclipse and comparison images, the evidential value of the data depended upon a great many other factors as well. Some of these had to do with the instruments and techniques used to measure distances on the photograph. Some had to do with the resources available to the investigator for deciding whether and to what extent measured displacements of star images were due to the deflection of starlight rather than extraneous influences. As Fig. 1 indicates, one such factor was change in camera angle due to the motion of the earth. Another was parallax resulting from the distance between the geographical locations from which Eddington’s eclipse and comparison pictures were taken.17
⎯⎯⎯⎯⎯⎯⎯ 16
In the discussion which follows, we ignore the fact that the deflection values calculated from the best photographic data differed not only from (N) and (NS), but also (albeit to a lesser extent) from (E). Assuming (as we do) that the data supported general relativity, this might mean that although (E) is correct, its discrimination does not require it to be identical to the value calculated from the photographs. Alternatively, it might mean that (E) is false, but that just as inaccurate data can make it reasonable to believe a phenomenon-claim, some false phenomena claims provide epistemically good support for theoretical claims in whose testing they are employed. Deciding which if either of these alternatives is correct falls beyond the scope of this paper. But since epistemic support by inaccurate data and confirmation by false claims are major difficulties for IRS, the disparities between (E) and magnitudes calculated from the best data offer no aid and comfort to the IRS analysis. Important as they are in connection with other epistemological issues, these disparities will not affect the arguments of this paper. 17 Eddington and Cottingham took eclipse photographs from Principe, but logistical complications made it necessary for them to have comparison pictures taken from Oxford. In addition to correcting for parallax, they had to establish scale for photographs taken from two very different locations with very different equipment (Earman and Glymour 1980, pp. 73-4).
244
James Bogen and James Woodward
Fig. 1. As the earth moves from its position at one time, t1, to its position at a later time, t2, the positions of the eclipse and comparison cameras change relative to the stars.
Apart from these influences, a number of factors including changes in temperature could produce mechanical effects in the photographic equipment sufficient to cause significant differences in scale (Earman and Glymour 1980). Additional complications arose from causes involved in the process of interpretation. One procedure for measuring distances between star images utilizes a low power microscope equipped with a cross hair. Having locked a photograph onto an illuminated frame, the investigator locates a star image (or part of one) against the cross hair and slowly turns a crank until the image whose distance from the first is to be measured appears against the cross hair. At each turn of the crank a gadget registers the distance traversed by the microscope in microns and fractions of microns. The distance is recorded, the photograph is removed from the frame, and the procedure is repeated with the next photograph.18 If the photographs are not oriented in the same way on the frame, image displacements will be measured incorrectly (Earman and Glymour 1980, p. 59). ⎯⎯⎯⎯⎯⎯⎯ 18
We are indebted to Alma Zook of the Pomona College physics department for showing and explaining the use of such measuring equipment to us.
Evading the IRS
245
The following drawing of a star image from one of Curtis’s photographs19 illustrates effects (produced by the focus and by motions of the camera) which make this bit of data epistemically irrelevant to the testing of general relativity theory by rendering it useless in deciding between (E), (N), and (NS).
The epistemic defects of Curtis’s star image are not due to the failure of inferential connections between an observation sentence and a hypothesis sentence. Nor are they due to the falsity of an observation sentence. By the same token, the epistemic value of the best photographs was not due to the truth of observation sentences or to the obtaining of inferential connections between them and hypothesis sentences. The evidential value of the starlight data depended upon non-logical, extra-linguistic relations between non-sentential features of photographs and causes which are not sentential structures. At this point we need to say a little more about a difficulty we mentioned in section I. Observation sentences are supposed to represent evidence. But the IRS tends to associate evidence with sentences reporting observations, and even though some investigations use data of this sort, the data which supported (E) were not linguistic items of any sort, let alone sentences. They were photographs. This is not an unusual case. So many investigations depend upon nonsentential data that it would be fatal for the IRS to maintain that all scientific evidence consists of observation reports (let alone the expressions in first order logic we are calling observation sentences). What then do observation sentences represent? The most charitable answer would be that they represent whatever data are actually used as evidence, even where the data are not observation reports. But this does not tell us which observation sentences to use to represent the photographs. Thus a serious difficulty is that for theory testing which involves non-sentential evidence, the IRS provides no guidance for the construction of the required observation sentences. Lacking an account of what observation sentences the IRS would use to represent the photographs, it is hard to talk about what would decide their truth values. But we can say this much: whatever the observation sentences may be, their truth had better not depend upon how well the photographs depicted the true positions of the stars. The photographs did not purport to show (and were not used to calculate) their actual positions or the true magnitudes of distances between them. They could represent true positions of (or distances between) stars with equal accuracy only if there were no significant20 discrepancies ⎯⎯⎯⎯⎯⎯⎯ 19
From a letter from Campbell to Curtis, reproduced in Earman and Glymour (1980), p. 67. We mean measurable discrepancies not accounted for by changes in the position of the earth, differences in the location of the eclipse and comparison equipment, etc.
20
246
James Bogen and James Woodward
between the positions of star images on the eclipse photographs and star images on the comparison photographs. But had there been no such discrepancies the photographs would have argued against (E). Thus to require both the eclipse and the comparison photographs to meet the same standard of representational accuracy would be to rule out evidence needed to support (E). Furthermore, the truth values of the observation sentences had better not be decided solely on the basis of whether the measurements of distances between their star images meet some general standard of accuracy specified independently of the particular investigation in question. In his textbook on error analysis, John Taylor points out that even though measurements can be too inaccurate to serve their purposes . . . it is not necessary that the uncertainties [i.e., levels of error] be extremely small . . . This . . . is typical of many scientific measurements, where uncertainties have to be reasonably small (perhaps a few percent of the measured value), but where extreme precision is often quite unnecessary (Taylor 1982, p. 6).
We maintain that what counts as a “reasonably small” level of error depends upon the nature of the phenomenon under investigation, the methods used to investigate it, and the alternative phenomena claims under consideration. Since these vary from case to case no single level of accuracy can distinguish between acceptable and unacceptable measurements for every case. Thus Priestley’s nitric oxide test tolerates considerably more measurement error than did the starlight bending investigations. This means that in order to decide whether or not to treat observation sentences representing Eddington’s photographs and measurements as true, the IRS epistemologist would have to know enough about local details peculiar to their production and interpretation to find out what levels of error would be acceptable. Suppose that one responds to this difficulty by stipulating that whatever observation sentences are used to represent photographs are to be called true if the photographs constitute good evidence and false if they do not. This means that the truth values of the observation sentences will depend, for example, upon whether the investigators could rule out or correct for the influences of such factors as mechanical changes in the equipment, parallax, sources of measurement error, etc., as far as necessary to allow them to discriminate correctly between (E), (N), and (NS). We submit that this stipulation is completely unilluminating. The notion of truth as applied to an observation sentence is now unconnected with any notion of representational correctness or accuracy (i.e., it is unclear what such sentences are supposed to represent or correspond to when they are true). Marking an observation sentence as true is now just a way of saying that the data associated with the sentence possess various other features that allow them to play a role in reliable discrimination. It is better to focus directly on the data and the processes that generate them and to drop the role of an observation sentence as an unnecessary intermediary.
Evading the IRS
247
VI Recall that an important part of the motivation for the development of IRS was the question of what objective factors do or should determine a scientist’s decision about whether a given body of evidence warrants the acceptance of a hypothesis. We have suggested that the evidential value of data depends upon complex and multifarious causal connections between the data, the phenomenon of interest, and a host of other factors. But it does not follow from this that scientists typically do (or even can) know much about the fine details of the relevant causal mechanisms. Quite the contrary, as we have argued elsewhere, scientists can seldom if ever give, and are seldom if ever required to give, detailed, systematic causal accounts of the production of a particular bit of data or its interaction with the human perceptual system and with devices (like the measuring equipment used by the starlight investigators) involved in its interpretation.21 But even though it does not involve systematic causal explanation, we believe that a kind of causal reasoning is essential to the use of data to investigate phenomena. This reasoning focuses upon what we have called general and local reliability. The remainder of this paper discusses some features of this sort of reasoning, and argues that its objectivity does not depend upon, and is not well explained in terms of the highly general, content independent, formal criteria sought by the IRS.
VII We turn first to a more detailed characterization of what we mean by general reliability. As we have already suggested, general reliability is a matter of longrun error characteristics. A detection process is generally reliable, when used in connection with a body of data, if it has a satisfactorily high probability of outputting, under repeated use, correct discriminations among a set of competing phenomenon-claims and a satisfactorily low probability of outputting incorrect discriminations. What matters is thus that the process discriminates correctly among a set of relevant alternatives, not that it discriminates correctly among all logically possible alternatives. Whether or not a detection process is generally reliable is always an empirical matter, having to do with the causal characteristics of the detection process and its typical circumstances of use, rather than with any formal relationship between the data that figure in such a process and the phenomenon-claims for which they are evidence. The notion of general reliability thus has application in those contexts in which we can provide a non-trivial characterization of what it is to repeat a process of data ⎯⎯⎯⎯⎯⎯⎯ 21
Bogen and Woodward (1988). For an excellent illustration of this, see Hacking (1983), p. 209.
248
James Bogen and James Woodward
production and interpretation (we shall call this a detection process, for brevity) and where this process possesses fairly stable, determinate error characteristics under repetition that are susceptible of empirical investigation. As we shall see in section VIII, these conditions are met in many, but by no means all the contexts in which data are used to assess claims about phenomena. Where these conditions are not met, we must assess evidential support in terms of a distinct notion of reliability, which we call local reliability. Here is an example illustrating what we have in mind by general reliability.22 Traditionally paleoanthropologists have relied on fossil evidence to infer relationships among human beings and other primates. The 1960s witnessed the emergence of an entirely distinct biochemical method for making such inferences, which involved comparing proteins and nucleic acids from living species. This method rests on the assumption that the rate of mutation in proteins is regular or clocklike; with this assumption one can infer that the greater the difference in protein structure among species, the longer the time they have been separated into distinct species. Molecular phylogeny (as such techniques came to be called) initially suggested conclusions strikingly at variance with the more traditional, generally accepted conclusions based on fossil evidence. For example, while fossil evidence suggested an early divergence between hominids and other primates, molecular techniques suggested a much later date of divergence – that hominids appeared much later than previously thought. Thus while paleoanthropologists classified the important prehistoric primate Ramapithicus as an early hominid on the basis of its fossil remains, the molecular evidence seemed to suggest that Ramapithicus could not be a hominid. Similarly, fossil and morphological data seemed to suggest that chimpanzees and gorillas were more closely related to each other than to humans, while molecular data suggested that humans and chimpanzees were more closely related. The initial reaction of most paleoanthropologists to these new claims was that the biochemical methods were unreliable, because they produced results at variance with what the fossils suggested. It was suggested that because the apparent rates of separation derived from molecular evidence were more recent than those derived from the fossil record, this showed that the molecular clock was not steady and that there had been a slow-down in the rate of change in protein structure among hominids. This debate was largely resolved in favor of the superior reliability of molecular methods. The invention of more powerful molecular techniques based on DNA hybridization, supported by convincing statistical arguments that the rate of mutation was indeed clocklike, largely corroborated the results of earlier molecular methods. The discovery of ⎯⎯⎯⎯⎯⎯⎯ 22
Details of this example are largely taken from Lewin (1987).
Evading the IRS
249
additional fossil evidence undermined the hominid status of Ramapithicus and supported the claim of a late divergence between hominids and other primates. This example illustrates what we have in mind when we ask whether a measurement or detection technique is generally reliable. We can think of various methods for inferring family trees from differences in protein structure and methods for inferring such relationships from fossil evidence as distinct measurement or detection techniques. Any particular molecular method is assumed to have fairly stable, determinate error characteristics which depend upon empirical features of the method: if the method is reliable it will generally yield roughly correct conclusions about family relationship and dates of divergence; if it is unreliable it will not. Clearly the general reliability of the molecular method will depend crucially on whether it is really true that the molecular clock is regular. Similarly, the reliability of the method associated with the use of fossil evidence also depends upon a number of empirical considerations – among them the ability of human beings to detect overall patterns of similarity based on visual appearance that correlate with genetic relationships. What the partisans of fossils and of molecular methods disagree about is the reliability of the methods they favor, in just the sense of reliability as good error characteristics described above. Part of what paleoanthropologists learned as they became convinced of the superior reliability of molecular methods, was that methods based on similarity of appearance were often less reliable than they had previously thought, in part because judgements of similarity can be heavily influenced by prior expectations and can lead the investigator to think that she sees features in the fossil evidence that are simply not there.23 Issues of this sort about general reliability – about the long-run error characteristics of a technique or method under repeated applications – play a central role in many areas of scientific investigation. Whenever a new instrument or detection device is introduced, investigators will wish to know about its general reliability – whether it works in such a way as to yield correct discriminations with some reasonable probability of success, whether it can be relied upon as a source of information in some particular area of application. Thus Galileo’s contemporaries were interested not just in whether his telescopic observations of the rough and irregular surface of the moon were correct, but with the general reliability of his telescope – with whether its causal characteristics were such that it could be used to make certain kinds of discrimination in astronomical applications with some reasonable probability of correctness or with whether instead what observers seemed to see through the telescope were artifacts, produced by imperfections in the lenses or some such source. ⎯⎯⎯⎯⎯⎯⎯ 23
See especially Lewin (1987), p. 122ff.
250
James Bogen and James Woodward
Similarly in many contexts in which human perceivers play an important role in science one can ask about their general reliability at various perceptual detection tasks, where this has to do with the probability or frequency with which perceivers make the relevant perceptual discriminations correctly, under repeated trials. Determinations of personal error rates in observational sciences like astronomy make use of this understanding of reliability.24 Similarly one can ask whether an automated data reduction procedure which sorts through batches of photographs selecting those which satisfy some preselected criterion is operating reliably, where this has to do with whether or not it is in fact classifying the photographs according to the indicated criterion with a low error rate. There are several general features of the above examples which are worth underscoring. Let us note to begin with that the question of whether a method, technique, or detection device and the data it produces are reliable always depends very much on the specific features of the method, technique, or instrument in question. It is these highly specific empirical facts about the general reliability of particular methods of data production and interpretation, and not the formal relationships emphasized by IRS, that are relevant to determining whether or not data are good evidence for various claims about phenomena. For example, it is the reliability characteristics of Galileo’s telescope that insure the evidential relevance of the images that it produces to the astronomical objects he wishes to detect, and it is the reliability characteristics of DNA hybridization that insure the evidential relevance of the biochemical data it produces to the reconstruction of relationships between species. How is the general reliability of an instrument or detection technique ascertained? We (and others) have discussed this issue at some length elsewhere and readers are referred to this discussion for a more detailed treatment.25 A wide variety of different kinds of considerations having to do, for example, with the observed effects of various manipulations and interventions into the detection process, with replicability, and with the use of various calibration techniques play an important role. One point that we especially wish to emphasize, and which we will make use of below, is that assessing the general reliability of an instrument or detection technique does not require that one ⎯⎯⎯⎯⎯⎯⎯ 24
For additional discussion, see Bogen and Woodward (1992). See Bogen and Woodward (1988) and Woodward (1989). Although, on our view, it is always a matter of empirical fact whether or not a detection process is generally reliable, we want to emphasize that there is rarely if ever an algorithm or mechanical procedure for deciding this. Instead it is typically the case that a variety of heterogeneous considerations are relevant, and building a case for general reliability or unreliability is a matter of building a consensus that most of these considerations, or the most compelling among them, support one conclusion rather than another. As writers like Peter Galison (1987) have emphasized, reaching such a conclusion may involve an irreducible element of judgement on the part of experimental investigators about which sources of error need to be taken seriously, about which possibilities are physically realistic, or plausible and so forth. Similar remarks apply to conclusions about local reliability. (Cf. n. 42.)
25
Evading the IRS
251
possess a general theory that systematically explains the operation of the instrument or technique or why it is generally reliable. There are many cases in the history of science involving instruments and detection techniques that investigators reasonably believed to be generally reliable in various standard uses even though those investigators did not possess a general explanatory theory of the operation of these instruments and techniques. Thus it was reasonable of Galileo and his contemporaries to believe that his telescope was generally reliable in many of its applications, even though Galileo lacked an optical theory that explained its workings; it is reasonable to believe that the human visual system can reliably make various perceptual discriminations in specified circumstances even though our understanding of the operation of the visual system is still rudimentary; it may be reasonable to believe that a certain staining technique reliably stains certain cells and doesn’t produce artifacts even though one doesn’t understand the chemistry of the staining process, and so on. We may contrast the picture we have been advocating, according to which evidential relevance is carried by the reliability characteristics of highly specific processes of data production and interpretation, with the conception of evidential relevance which is implicit in IRS. According to that conception, the relevance of evidence to hypotheses is a matter of observation sentences standing in various highly general, structural or inferential relations to those hypotheses, relationships which, according to IRS, are exemplified in many different areas of scientific investigation. Thus the idea is that the evidential relevance of biochemical data to species relationships or the evidential relevance of the images produced by Galileo’s telescope to various astronomical hypotheses is a matter of the obtaining of some appropriate formal relationship between sentences representing these data, the hypotheses in question, and perhaps appropriate background or auxiliary assumptions. On the contrasting picture we have defended, evidential relevance is not a matter of any such formal relationship, but is instead a matter of empirical fact – a matter of there existing empirical relationships or correlations between data and phenomena which permit us to use the data to discriminate among competing claims about phenomena according to procedures that have good general error characteristics. Evidential relevance thus derives from an enormous variety of highly domainspecific facts about the error characteristics of various quite heterogeneous detection and measurement processes, rather than from the highly general, domain-independent formal relationships emphasized in IRS accounts. Our alternative conception seems to us to have several advantages that are not shared by IRS accounts. First, we have already noted that a great deal of data does not have an obvious sentential representation and that, even when such representations are available, they need not be true or exactly representationally accurate for data to play an evidential role. Our account helps to make sense of these facts. There is nothing in the notion of general reliability
252
James Bogen and James Woodward
that requires that data be sentential in structure, or have a natural sentential representation, or have semantic characteristics like truth or exact representational accuracy. Data can figure in a generally reliable detection process, and features of data can be systematically correlated with the correctness or incorrectness of different claims about phenomena without the data being true or even sententially representable. For example, when a pathologist looks at an x-ray photograph and produces a diagnosis, or when a geologist looks at a rock and provides an identification of its type, all that we require, in order for these claims to be credible or evidentially well-supported, is that the relevant processes of perceptual detection and identification be generally reliable in the sense of having good error characteristics, and that we have some evidence that this is the case. It isn’t necessary that we be able to provide sentential representations of what these investigators perceive or to exhibit their conclusions as the result of the operation of general IRS-style inductive rules on sentential representations of what they see. Similarly, in the case of the Priestley/Lavoisier example, the characteristics of Priestley’s detection procedure may very well be such that it can be used to reliably discriminate between ordinary air and oxygen on the basis of volume measurements, in the sense that repeated uses of the procedure will result in correct discriminations with high probability, even though the volume measurements on which the discrimination is based are inaccurate, noisy and in fact false if taken as reports of the actual volume decrease. There is a second reason to focus on reliability in preference to IRS-style confirmation relations. According to the IRS, evidence e provides epistemic support for a theoretical claim when the observation sentence, o, which corresponds to the evidence stands in the right sort of formal relationship to the hypothesis sentence, h, which represents the theoretical claim. Our worries so far have centered around the difficulties of finding a true observation sentence o which faithfully represents the evidential significance of e, and a hypothesis sentence h which faithfully represents the content of the theoretical claim. But quite apart from these difficulties there is a perennial internal puzzle for IRS accounts. Given that within these accounts o does not, even in conjunction with background information, entail h, why should we suppose that there is any connection between o’s being true and o and h instantiating the formal relationships specified in these accounts, and h’s being true or having a high probability of truth or possessing some other feature associated with grounds for belief? For example, even if a true observation sentence representing Priestley’s data actually did entail a positive instance of a hypothesis sentence representing the claim that a certain sort of gas is not ordinary air, why would that make the latter claim belief-worthy? We think that it is very hard to see what the justification of a non-deductive IRS-style method or criterion of evidential support could possibly consist in except the provision of grounds that the
Evading the IRS
253
method or criterion has good (general) reliability or error characteristics under repeated use. That is, it is hard to see why we should believe that the truth of the observation sentence o together with the fact that the relationship between o and hypothesis h satisfies the pattern recommended by, for example, hypotheticodeductivism or bootstrapping provides a reason for belief in h if it were not true that cases in which such patterns are instantiated turn out, with some reasonable probability, to be cases in which h is true, or were it not at least true that cases in which such patterns are instantiated turn out more frequently to be cases in which h is true than cases in which such patterns are not instantiated.26 However, it seems very unlikely that any of the IRS-style accounts we have considered can be given such a reliabilist justification. IRS accounts are, as we have seen, subject matter and context-independent; they are meant to supply universal criteria of evidential support. But it is all too easy to find, for any IRS account, not just hypothetical, but actual cases in which true observation sentences stand in the recommended relationship to hypothesis h and yet in which h is false: cases in which positive instances instantiate a hypothesis and yet the hypothesis is false, cases in which true observation sentences are deduced from a hypothesis and yet it is false, and so forth. Whether accepting h when it stands in the relationship to o described in one’s favorite IRS schema and o is true will lead one to accept true hypotheses some significant fraction of the time will depend entirely on the empirical details of the particular cases to which the schema in question is applied. But this is to say that the various IRS schemas we have been considering when taken as methods for forming beliefs or accepting hypotheses either have no determinate error characteristics at all when considered in the abstract (their error characteristics vary wildly, depending on the details of the particular cases to which they are applied) or at least no error characteristics that are knowable by us. Indeed, the fact that the various IRS accounts we have been considering cannot be given a satisfying reliabilist justification is tacitly conceded by their proponents, who usually do not even try to provide such a justification.27 ⎯⎯⎯⎯⎯⎯⎯ 26
For a general argument in support of this conclusion see Friedman (1979). One can think of Larry Laudan’s recent naturalizing program in philosophy of science which advocates the testing of various philosophical theses about scientific change and theory confirmation against empirical evidence provided by the history of science as (among other things) an attempt to carry out an empirical investigation of the error or reliability characteristics of the various IRS confirmation schemas (Donovan et al. 1988). We agree with Laudan that vindicating the various IRS models would require information about long-run error characteristics of the sort for which he is looking. But for reasons described in the next paragraph in the text, we are much more pessimistic than Laudan and his collaborators about the possibility of obtaining such information. 27 Typical attempts to argue for particular IRS models appeal instead to (a) alleged paradoxes, and inadequacies associated with alternative IRS approaches, (b) various supposed intuitions about evidential support, and (c) famous examples of successful science that are alleged to conform to the model in question. (Cf. Glymour 1980.) But (a) is compatible with and perhaps even supports
254
James Bogen and James Woodward
By contrast, there is no corresponding problem with the notion of general reliability as applied to particular instruments or detection processes. Such instruments and processes often do have determinate error characteristics, about which we can obtain empirical evidence. Unlike the H-D method or the method associated with bootstrapping, the reliability of a telescope or a radioactive dating technique is exactly the sort of thing we know how to investigate empirically and regarding which we can obtain convincing evidence. There is no puzzle corresponding to that raised above in connection with IRS accounts about what it means to say that a dating technique has a high probability of yielding correct conclusions about the ages of certain fossils or about why, given that we have applied a reliable dating technique and have obtained a certain result, we have good prima facie grounds for believing that result. In short, it is the use of specific instruments, detection devices, measurement and observational techniques, rather than IRS-style inductive patterns, that are appropriate candidates for justification in terms of the idea of general reliability. Reflection on a reliabilist conception of justification thus reinforces our conclusion that the relevance of evidence to hypothesis is not a matter of formal, IRS-style inferential relations, but rather derives from highly specific facts about the error characteristics of various detection processes and instruments.
VIII In addition to the question of whether some type of detection process or instrument is generally reliable in the repeatable error characteristics sense described above, scientists also are interested in whether the use of the process on some particular occasion, in a particular detection task, is reliable – with whether the data produced on that particular occasion are good evidence for some phenomenon of interest. This is a matter of local reliability. While in those cases in which a detection process has repeatable error characteristics, information about its general reliability is always evidentially relevant, there are many cases in which the evidential import of data cannot be assessed just in skepticism about all IRS accounts of evidence, and with respect to (b), it is uncontroversial that intuitions about inductive support frequently lead one astray. Finally, from a reliabilist perspective (c) is quite unconvincing. Instead, what needs to be shown is that scientists systematically succeed in a variety of cases because they accept hypotheses in accord with the recommendations of the IRS account one favors. That is, what we need to know is not just that there are episodes in the history of science in which hypotheses stand in the relationship to true observation sentences described by, say, a bootstrap methodology and that these hypotheses turn out to be true or nearly so, but what the performance of a bootstrap methodology would be, on a wide variety of different kinds of evidence, in discriminating true hypotheses from false hypotheses – both what this performance is absolutely and how it compares with alternative methods one might adopt. (As we understand it, this is Glymour’s present view as well.)
Evading the IRS
255
terms of general reliability. For example, even if I know that some radioactive dating technique is generally reliable when applied to fossils, this still leaves open the question of whether the date assigned to some particular fossil by the use of the technique is correct: it might be that this particular fossil is contaminated in a way that gives us mistaken data, or that the equipment I am using has misfunctioned on this particular occasion of use. That the dating process is generally reliable doesn’t preclude these possibilities. Some philosophers with a generalist turn of mind will find it tempting to try to reduce local reliability to general reliability: it will be said that if the data obtained from a particular fossil are mistaken because of the presence of a contaminant, then if that very detection process is repeated (with the contaminant present and so forth) on other occasions, it will have unfavorable error characteristics, and this is what grounds our judgement of reliability or evidential import in the particular case. As long as we take care to specify the relevant detection processes finely enough, all judgements about reliability in particular cases can be explicated in terms of the idea of repeated error characteristics. Our response is not that this is necessarily wrong, but that it is thoroughly unilluminating at least when understood as an account of how judgements of local reliability are arrived at and justified. As we shall see below, many judgements of local reliability turn on considerations that are particular or idiosyncratic to the individual case at hand. Often scientists are either unable to describe in a non-trivial way what it is to repeat the measurement or detection process that results in some particular body of data or lack (and cannot get) information about its long-run error characteristics. It is not at all clear to us that whenever a detection process is used on some particular occasion, and a judgement about its local reliability is reached on the basis of various considerations, there must be some description of the process, considerations, and judgements involved that exhibits them as repeatable. But even if this is the case, this description and the relevant error characteristics of the process when repeated often will be unknown to the individual investigator – this information is not what the investigator appeals to in reaching his judgement about local reliability or in defending his judgement. What then are the considerations which ground judgements of local reliability and how should we understand what it is that we are trying to do when we make such judgements? While the relevant considerations are, as we shall see, highly heterogeneous, we think that they very often have a common point or pattern, which we will now try to describe. Put baldly, our idea is that judgements of local reliability are a species of singular causal inference in which one tries to show that the phenomenon of interest causes the data by means of an eliminativist strategy – by ruling out other possible causes of the
256
James Bogen and James Woodward
data.28 When one makes a judgement of local reliability one wants to ascertain on the basis of some body of data whether some phenomenon of interest is present or has certain features. One tries to do this by showing that the detection process and data are such that the data must have been caused by the phenomenon in question (or by a phenomenon with the features in question) – that all other relevant candidates for causes of the data can be ruled out. Since something must have caused the data, we settle on the phenomenon of interest as the only remaining possibility. For example, in the fossil dating example above, one wants to exclude (among other things) the possibility that one’s data – presumably some measure of radioactive decay rate, such as counts with a Geiger counter – were caused by (or result in part from a causal contribution due to) the presence of the contaminant. Similarly, as we have already noted, showing that some particular bubble chamber photograph was evidence for the existence of neutral currents in the CERN experiments of 1973 requires ruling out the possibility that the particular photograph might have been due instead to some alternative cause, such as a high energy neutron, that can mimic many of the effects of neutral currents. The underlying idea of this strategy is nicely described by Allan Franklin in his recent book Experiments, Right or Wrong (1990). Franklin approvingly quotes Sherlock Holmes’s remark to Watson, “How often have I said to you that when you have eliminated the impossible, whatever remains, however improbable, must be the truth?” and then adds, “If we can eliminate all possible sources of error and alternative explanations, then we are left with a valid experimental result” (1990, p. 109). Here is a more extended example designed to illustrate what is involved in local reliability and the role of the eliminative strategy described above.29 In experiments conducted in the late 1960s, Joseph Weber, an experimentalist at the University of Maryland, claimed to have successfully detected the phenomenon of gravitational radiation. The production of gravity waves by massive moving bodies is predicted (and explained) by general relativity. However, gravitational radiation is so weakly coupled to matter that detection of such radiation by us is extremely difficult. Weber’s apparatus initially consisted of a large metal bar which was designed to vibrate at the characteristic frequency of gravitational radiation emitted by relatively large scale cosmological events. The central problem of ⎯⎯⎯⎯⎯⎯⎯ 28
As with judgements about general reliability, we do not mean to suggest that there is some single method or algorithm to be employed in this ruling out of alternatives. For example, ruling out an alternative may involve establishing an observational claim that is logically inconsistent with the alternative (Popperian falsification), but might take other forms as well; for example, it may be a matter of finding evidence that renders the alternative unlikely or implausible or of finding evidence that the alternative should but is not able to explain. 29 The account that follows draws heavily on Collins (1975) and Collins (1981). Other accessible discussions of Weber’s experiment on which we have relied include Davis (1980), esp. pp. 102-117, and Will (1986).
Evading the IRS
257
experimental design was that to detect gravitational radiation one had to be able to control or correct for other potential disturbances due to electromagnetic, thermal, and acoustic sources. In part, this was attempted by physical insulation of the bar, but this could not eliminate all possible sources of disturbance; for example, as long as the bar is above absolute zero, thermal motion of the atoms in the bar will induce random vibrations in it. One of the ways Weber attempted to deal with this difficulty was through the use of a second detector which was separated from his original detector by a large spatial distance – the idea being that genuine gravitational radiation, which would be cosmological in origin, should register simultaneously on both detectors while other sorts of background events which were local in origin would be less likely to do this. Nonetheless, it was recognized that some coincident disturbances will occur in the two detectors just by chance. To deal with this possibility, various complex statistical arguments and other kinds of checks were used to attempt to show that it was unlikely that all of the coincident disturbances could arise in this way. Weber also relied on facts about the causal characteristics of the signal – the gravitational radiation he was trying to detect. The detectors used by Weber were most sensitive to gravitational radiation when the direction of propagation of given radiation was perpendicular to the axes of the detectors. Thus if the waves were coming from a fixed direction in space (as would be plausible if they were due to some astronomical event), they should vary regularly in intensity with the period of revolution of the earth. Moreover, any periodic variations due to human activity should exhibit the regular twenty-four hour variation of the solar day. By contrast, the pattern of change due to an astronomical source would be expected to be in accordance with the sidereal day which reflects the revolution of the earth around the sun, as well as its rotation about its axis, and is slightly shorter than the solar day. When Weber initially appeared to find a significant correlation with sidereal, but not solar, time in the vibrations he was detecting, this was taken by many other scientists to be important evidence that the source of the vibrations was not local or terrestrial, but instead due to some astronomical event. Weber claimed to have detected the existence of gravitational radiation from 1969 on, but for a variety of reasons his claims are now almost universally doubted. In what follows, we concentrate on what is involved in Weber’s claim that his detection procedure was locally reliable and how he attempted to establish that claim. As we see it, what Weber was interested in establishing was a singular causal claim: he wanted to show that at least some of the vibrations and disturbances his data recorded were due to gravitational radiation (the phenomenon he was trying to detect) and (hence) that such radiation existed. The problem he faced was that a number of other possible causes or factors besides gravitational radiation might in principle have caused his data. Unless
258
James Bogen and James Woodward
Weber could rule out, or render implausible or unlikely, the possibility that these other factors might have caused the disturbances, he would not be justified in concluding that the disturbances are due to the presence of gravitational radiation. The various experimental strategies and arguments described above (physical isolation of the bar, use of a second detector, and so forth) are an attempt to do just this – to make it implausible that the vibrations in his detector could have been caused by anything but gravitational radiation. For example, in the case of the sidereal correlation the underlying argument is that the presence of this pattern or signature in the data is so distinctive that it could only have been produced by gravitational radiation rather than by some other source. We will not attempt to describe in detail the process by which Weber’s claims of successful detection came to be criticized and eventually disbelieved. Nonetheless it is worth noting that we can see the underlying point of these criticisms as showing that Weber’s experiment fails to conform to the eliminative pattern under discussion – what the critics show is that Weber has not convincingly ruled out the possibility that his data were due to other causes besides gravitational radiation. Thus, for example, the statistical techniques that Weber used turned out to be problematic – indeed, an inadvertent natural experiment appeared to show that the techniques lacked general reliability in the sense described above. (Weber’s statistical techniques detected evidence for gravitational radiation in data provided by another group which, because of a misunderstanding on Weber’s part about synchronization, should have been reported as containing pure noise.) Because of this, Weber could no longer claim to have convincingly eliminated the possibility that all of the disturbances he was seeing in both detections were due to the chance coincidence of local causes. Secondly, as Weber continued his experiment and did further analysis of his data, he was forced to retract his claim of sidereal correlation. Finally, and perhaps most fundamentally, a number of other experiments, using similar and more sensitive apparatus, failed to replicate Weber’s results. Here the argument is that if in fact gravitational radiation was playing a causal role in the production of Weber’s data such radiation ought to interact causally with other similar devices; conversely, failure to detect such radiation with a similar apparatus, while it does not tell us which alternative cause produced Weber’s data, does undermine the claim that it was due to gravitational radiation. Much of what we have said about the advantages of the notion of general reliability vis-à-vis IRS-style accounts holds as well for local reliability. When we make a judgement of local reliability about certain data – when we conclude, for example, that some set of vibrations in Weber’s apparatus were or were not evidence for the existence of gravitational radiation – what needs to be established is not whether there obtains some appropriate formal or logical relationship of the sort IRS models attempt to capture, but rather whether there
Evading the IRS
259
is an appropriate causal relationship leading from the phenomenon to the data. Just as with general reliability, the causal relationships needed for data to count as locally reliable evidence for some phenomenon can hold even if data lack a natural sentential representation that stands in the right formal relationship to the phenomenon-claim in question. Conversely, a sentential representation of the data can stand in what (according to some IRS accounts of confirmation) is the right formal relationship to a hypothesis and yet nonetheless fail to evidentially support it. Weber’s experiment also illustrates this point: Weber obtained data which (or so he was prepared to argue) were just what would be expected if general relativity were true (and gravitational radiation existed). On at least some natural ways of representing data by means of observation sentences, these sentences stand in just the formal relationships to general relativity which according to H-D and positive instance accounts, are necessary for confirmation. Nonetheless this consideration does not show that Weber’s data were reliable evidence for the existence of gravitational radiation. To show this Weber must show that his data were produced by a causal process in which gravitational radiation figures. This is exactly what he tries, and fails, to do. The causally driven strategies and arguments described above would make little sense if all Weber needed to show was the existence of some appropriate IRS-style formal relationship between a true sentential representation of his data and the claim that gravitational radiation exists. Similarly, as we have already had occasion to note, merely producing bubble chamber photographs that have just the characteristic patterns that would be expected if neutral currents were present – producing data which conform to this hypothesis or which have some description which is derivable from the hypothesis – is not by itself good evidence that neutral currents are present. To do this one must rule out the possibility that this data was caused by anything but neutral currents. And as we have noted, this involves talking about the causal process that has produced the data – a consideration which is omitted in most IRS accounts. As we have also argued, a similar point holds in connection with the Eddington solar eclipse expedition. What Eddington needs to show is that the apparent deflection of starlight indicated by the photographic plates is due to the causal influence of the sun’s gravitational field, as described by general relativity, rather than to more local sources, such as changes in the plates due to variations in temperature. Once we understand Eddington’s reasoning as reasoning to the existence of a cause in accordance with an eliminative strategy, various features of that reasoning that seem puzzling on IRS treatments – that it is not obvious how to represent all of the evidentially relevant features of the photographs in terms of true observation sentences and auxiliaries and that the values calculated from the photographs don’t exactly coincide with (E) but are nonetheless taken to support (E) – fall naturally into place.
260
James Bogen and James Woodward
IX There is a common element to a number of the difficulties with IRS models that we have discussed that deserves explicit emphasis. It is an immediate consequence of our notions of general and local reliability that the processes that produce or generate data are crucial to its evidential status. Moreover, it is often hard to see how to represent the evidential relevance of such processes in an illuminating way within IRS-style accounts. And in fact the most prominent IRS models simply neglect this element of evidential assessment. The tendency within IRS models is to assume, as a point of departure, that one has a body of evidence, that it is unproblematic how to represent it sententially, and to then try to capture its evidential relevance to some hypothesis by focusing on the formal or structural relationship of its sentential representation to that hypothesis. But if the processes that generated this evidence make a crucial difference to its evidential significance, we can’t as IRS approaches assume, simply detach the evidence from the processes which generated it, and use a sentential representation of it as a premise in an IRS-style inductive inference. To make this point vivid, consider (P) a collection of photographs which qua photographs are indistinguishable from those that in fact constituted evidence for the existence of neutral current interactions in the CERN experiments of 1973. Are the photographs in (P) also evidence for the existence of neutral currents? Although many philosophers (influenced by IRS models of confirmation) will hold that the answer to this question is obviously yes, our claim is that on the basis of the above information one simply doesn’t know – one doesn’t know whether the photographs are evidence for neutral currents until one knows something about the processes by which they are generated. Suppose that the process by which the photographs were produced failed to adequately control for high energy neutrons. Then our claim is that such photographs are not reliable evidence for the existence of neutral currents, even if the photographs themselves look no different from those that were produced by experiments (like the CERN experiment) in which there was adequate control for the neutron background. It is thus a consequence of our discussion of general and local reliability that the evidential significance of the same body of data will vary, depending upon what it is reasonable to believe about how it was produced. We think that the tendency to neglect the relevance of the data-generating processes explains, at least in large measure, the familiar paradoxes which face IRS accounts. Consider the raven paradox, briefly introduced in section IV above. Given our discussion so far it will come as no surprise to learn that we think the culprit in this case is the positive instance criterion itself. Our view is that one just can’t say whether a positive instance of a hypothesis constitutes evidence for it, without knowing about the procedure by which the positive
Evading the IRS
261
instance was produced or generated. A possibility originally introduced by Paul Horwich (1982) makes this point in a very striking way: suppose that you are told that a large number of ravens have been collected, and that they have all turned out to be black. You may be tempted to suppose that such observations support the hypothesis that (h1) all ravens are black. Suppose, however, you then learn how this evidence has been produced: a machine of special design which seizes all and only black objects and stores them in a vast bin has been employed, and all of our observed ravens have come from this bin. In the bin, we find, unsurprisingly, in addition to black shoes, old tires and pieces of coal, a number of black ravens and no non-black ravens. Recall that our interest in data is in using it to discriminate among competing phenomenon-claims. Similarly, when we investigate the hypothesis that all ravens are black, our interest is in obtaining evidence that differentially supports this hypothesis against other natural competitors. That is, our interest is in whether there is evidence that provides some basis for preferring or accepting this hypothesis in contrast to such natural competitors as the hypothesis that ravens come in many different colors, including black. It is clear that the black ravens produced by Horwich’s machine do not differentially support the hypothesis that all ravens are black or provide grounds for accepting it rather than such competitors. The reason is obvious: the character of the evidencegathering or data-generating procedure is such that it could not possibly have discovered any evidence which is contrary to the hypothesis that all ravens are black, or which discriminates in favor of a competitor to this hypothesis, even if such evidence exists. The observed black ravens are positive instances of the hypothesis that all ravens are black, but they do not support the hypothesis in the sense of discriminating in favor of it against natural competitors because of the way in which those observations have been produced or generated. If observations of a very large number of black ravens had been produced in some other way – e.g., by a random sampling process, which had an equal probability of selecting any raven (black or non-black) or by some other process which was such that there was some reason to think that the evidence it generated was representative of the entire population of ravens – then we would be entitled to regard such observations as providing evidence that favors the hypothesis under discussion. But in the absence of a reason to think that the observations have been generated by some such process that makes for reliability, the mere accumulation of observations of black ravens provides no reason for accepting the hypothesis that all ravens are black in contrast to its natural competitors. Similar considerations apply to the question of whether the observation of non-black, non-ravens supports the hypothesis that (h2), “All non-black things are non-ravens.” As a point of departure, let us note that it is less clear than it is in the case of (h1) what the “natural” serious alternatives to (h2) are. The hypothesis (h3) that “All non-black things are ravens” is a competitor to (h2) – it
262
James Bogen and James Woodward
is inconsistent with (h2) on the supposition that there is at least one non-black thing – but not a serious competitor since every investigator will have great confidence that it is false prior to beginning an investigation of (h2). Someone who is uncertain whether (h2) is true will not take seriously the possibility that (h3) is true instead and for this reason evidence that merely discriminates between (h2) and (h3) but not between (h2) and its more plausible alternatives will not be regarded as supporting (h2). Thus while the observation of a white shoe does indeed discriminate between (h2) and (h3) this fact by itself does not show that the observation supports (h2). Presumably the best candidates for serious specific alternatives to (h2) are various hypotheses specifying the conditions (e.g., snowy regions) under which non-black ravens will occur. But given any plausible alternative hypothesis about the conditions under which a non-black raven will occur, the observation of a white shoe or a red pencil does nothing to effectively discriminate between (h2) and this alternative. For example, these observations do nothing to discriminate between (h2) and the alternative hypotheses that there are white ravens in snowy regions. As far as these alternatives go, then, there is no good reason to think of an observation of a white shoe as confirming (h2). There are other possible alternatives to (h2) that one might consider. For example, there are various hypotheses, (hp), specifying that the proportion of ravens among non-black things is some (presumably very small) positive number p for various values of p. There is also the generic, non-specific alternative to (h2) which is simply its denial (h4), “Some non-black things are ravens.” For a variety of reasons these alternatives are less likely to be of scientific interest than the alternatives considered in the previous paragraph. But even if we put this consideration aside, there is an additional problem with the suggestion that the observation of a white shoe confirms (h2) because it discriminates between (h2) and one or more of these alternatives. This has to do with the characteristics of the processes involved in the production of such observations. In the case of (h1), “All ravens are black,” we have some sense of what it would mean to sample randomly from the class of ravens or at least to sample a “representative” range of ravens (e.g., from different geographical locations or ecological niches) from this class. That is, we have in this case some sense of what is required for the process that generates relevant observations to be unbiased or to have good reliability characteristics. If we observe enough ravens that are produced by such a process and all turn out to be black, we may regard this evidence as undercutting not just those competitors to (h1) that claim that all ravens are some uniform non-black color but also those alternative hypotheses that claim that various proportions of ravens are non-black, or the generic alternative hypothesis that some ravens are non-black. Relatedly, observations of non-black ravens produced by such a process might confirm some alternative hypothesis to (h1) about the proportion
Evading the IRS
263
of ravens that are non-black or the conditions under which we may expect to find them. By contrast, nothing like this is true of (h2). It is hard to understand even what it might mean to sample in a random or representative way from the class of non-black things and harder still to envision a physical process that would implement such a sampling procedure. It is also hard to see on what basis one might argue that a particular sample of non-black things was representative of the entire range of such things. As a result, when we are presented with even a very generous collection of objects consisting of white shoes, red pencils and so on, it is hard to see on what sort of basis one might determine whether the procedure by which this evidence was produced had the right sort of characteristics to enable us to reliably discriminate between (h2) and either the alternatives (hp) or (h4), and hence hard to assess what its evidential significance is for (h2). It is thus unsurprising that we intuitively judge the import of such evidence for (h2) to be at best unclear and equivocal. On our analysis, then, an important part of what generates the paradox is the mistaken assumption, characteristic of IRS approaches, that evidential support for a claim is just a matter of observation sentences standing in some appropriate structural or formal relationship to a hypothesis sentence (in this case the relationship captured by the positive instance criterion) independently of the processes which generate the evidence and independently of whether the evidence can be used to discriminate between the hypothesis and alternatives to it. It might be thought that while extant IRS accounts have in fact neglected the relevance of those features of data-generating processes that we have sought to capture with our notions of general and local reliability, there is nothing in the logic of such accounts that requires this omission. Many IRS accounts assign an important role to auxiliary or background assumptions. Why can’t partisans of IRS represent the evidential significance of processes of data generation by means of these assumptions? We don’t see how to do this in a way that respects the underlying aspirations of the IRS approach and avoids trivialization. The neglect of data generating processes in standard IRS accounts is not an accidental or easily correctable feature of such accounts. Consider those features of data generation captured by our notion of general reliability. What would the background assumptions designed to capture this notion within an IRS account look like? We have already argued that in order to know that an instrument or detection process is generally reliable, it is not necessary to possess a general theory that explains the operation of the instrument or the detection process. The background assumptions that are designed to capture the role of general reliability in inferences from data to phenomena thus cannot be provided by general theories that explain the operation of instruments or detection processes. The information that grounds judgements of general reliability is, as we have seen, typically
264
James Bogen and James Woodward
information from a variety of different sources – about the performance of the detection process in other situations in which it is known what results to expect, about the results of manipulating or interfering with the detection process in various ways, and so forth. While all of this information is relevant to reliability, no single piece of information of this sort is sufficient to guarantee reliability. Because this is the case and because the considerations which are relevant to reliability are so heterogeneous and so specific to the particular detection process we want to assess, it is not clear how to represent such information as a conventional background or auxiliary assumption or as a premise in an inductive inference conforming to some IRS pattern. Of course we can represent the relevant background assumptions by means of the brute assertion that the instruments and detection processes with which we are working are generally reliable. Then we might represent the decision to accept phenomenon-claim P, on the basis of data D produced by detection process R as having something like the following structure: (1) If detection process R is generally reliable and produces data having features D, it follows that phenomenon-claim P will be true with high probability. (2) Detection process R is generally reliable and has produced data having features D; therefore (3) phenomenon-claim P is true with high probability (or alternatively (4) phenomenon-claim P is true). The problem with this, of course, is that the inference from data to phenomenon now no longer looks like an IRS-style inductive inference at all. The resulting argument is deductive if (3) is the conclusion. If (4) is the conclusion, the explicitly inductive step is trivial – a matter of adopting some rule of acceptance that allows one to accept highly probable claims as true. All of the real work is done by the highly specific subject-matter dependent background claim (2) in which general reliability is asserted. The original aspiration of the IRS approach, which was to represent the goodness of the inference as a matter of its conforming to some highly general, subject-matter independent pattern of argument – with the subject-matter independent pattern supplying, so to speak, the inductive component to the argument – has not been met.30 ⎯⎯⎯⎯⎯⎯⎯ 30
Although we lack the space for a detailed discussion, we think that a similar conclusion holds in connection with judgements of local reliability. If one wished to represent formally the eliminative reasoning involved in establishing local reliability, then it is often most natural to represent it by means of the deductively valid argument pattern known as disjunctive syllogism: one begins with the premise that some disjunction is true, shows that all of the disjuncts save one are false, and concludes that the remaining disjunct is true. But, as in the case of the representation of the argument appealing to general reliability considered above, this formal representation of eliminative reasoning is obvious and trivial; the really interesting and difficult work that must be done in connection with assessing such arguments has to do with writing down and establishing the truth of their premises: has one really considered all the alternatives, does one really have good grounds for considering all but one to be false? Answering such questions typically requires a great deal of subject-matter specific causal knowledge. Just as in the case of general reliability, the original IRS
Evading the IRS
265
Here is another way of putting this matter: someone who accepts (1) and (2) will find his beliefs about the truth of P significantly constrained, and constrained by empirical facts about evidence. Nonetheless the kind of constraint provided by (1) and (2) is very different from the kinds of non-deductive constraints on hypothesis choice sought by proponents of IRS models. Consider again the passage quoted from Hempel in section II. As that passage suggests, the aim of the IRS approach is to exhibit the grounds for belief in hypotheses like (3) or (4) in a way that avoids reference to “personal” or “subjective” factors and to subject-matter specific considerations. Instead the aim of the IRS approach is to exhibit the grounds for belief in (3) or (4) as resulting from the operation of some small number of general patterns of non-deductive argument or evidential support which recur across many different areas of inquiry. If (2) is a highly subject-matter specific claim about, say, the reliability of a carbon-14 dating procedure when applied to a certain kind of fossil or (even worse) a claim that asserts the reliability of a particular pathologist in correctly discriminating benign from malignant lung tumors when she looks at x-ray photographs, reference to “subject-matter specific” or “personal” considerations will not have been avoided. A satisfactory IRS analysis would begin instead with some sentential characterization of the data produced by the radioactive dating procedure or the data looked at by the pathologist, and then show us how this data characterization supports (3) or (4) by standing in some formally characterizable relationship to it that can be instantiated in many different areas of inquiry. That is, the evidential relevance of the data to (3) or (4) should be established or represented by the instantiation of some appropriate IRS pattern, not by a highly subject-matter specific hypothesis like (2). If our critique of IRS is correct, this is just what cannot be done. As the passage quoted from Hempel makes clear, IRS accounts are driven in large measure by a desire to exhibit science as an objective, evidentially constrained enterprise. We fully agree with this picture of science. We think that in many scientific contexts, evidence has accumulated in such a way that only one hypothesis from some large class of competitors is a plausible candidate for belief or acceptance. Our disagreement with IRS accounts has to do with the nature or character of the evidential constraints that are operative in science, not with whether such constraints exist. According to IRS accounts these constraints derive from highly general, domain-independent, formally characterizable patterns of evidential support that appear in many different areas of scientific investigation. We reject this claim, as well as Hempel’s implied suggestion that either the way in which evidence constrains belief must be capturable within an
aspiration of finding a subject-matter independent pattern of inductive argument in which the formal features of the pattern do interesting, non-trivial work of a sort that might be studied by philosophers has not been met.
266
James Bogen and James Woodward
IRS-style framework or else we must agree that there are no such constraints at all. On the contrasting picture we have sought to provide, the way in which evidence constrains belief should be understood instead in terms of non-formal subject-matter specific kinds of empirical considerations that we have sought to capture with our notions of general and local reliability. On our account, many well-known difficulties for IRS approaches – the various paradoxes of confirmation, and the problem of explaining the connection between a hypothesis’s standing in the formal relationships to an observation sentence emphasized in IRS accounts and its being true – are avoided. And many features of actual scientific practice that look opaque on IRS approaches – the evidential significance of data generating processes or the use of data that lacks a natural sentential representation, or that is noisy, inaccurate or subject to error – fall naturally into place.31 James Bogen Department of Philosophy Pitzer College (Emeritus) and University of Pittsburgh
[email protected] James Woodward Division of Humanities and Social Sciences California Institute of Technology
[email protected] ⎯⎯⎯⎯⎯⎯⎯ 31
We have ignored Bayesian accounts of confirmation. We believe that in principle such accounts have the resources to deal with some although perhaps not all of the difficulties for IRS approaches described above. However, in practice the Bayesian treatments provided by philosophers often fall prey to these difficulties, perhaps because those who construct them commonly retain the sorts of expectations about evidence that characterize IRS-style approaches. Thus while there seems no barrier in principle to incorporating information about the process by which data has been generated into a Bayesian analysis, in practice many Bayesians neglect or overlook the evidential relevance of such information – Bayesian criticisms of randomization in experimental design are one conspicuous expression of this neglect. For a recent illustration of how Bayesians can capture the evidential relevance of data generating processes in connection with the ravens paradox, see Earman (1992); for a rather more typical illustration of a recent Bayesian analysis that fails to recognize the relevance of such considerations, see the discussion of this paradox in Howson and Urbach (1989). As another illustration of the relevance of the discussion in this paper to Bayesian approaches, consider that most Bayesian accounts require that all evidence have a natural representation by means of true sentences. These accounts thus must be modified or extended to deal with the fact that such a representation will not always exist. For a very interesting attempt to do just this, see Jeffrey (1989).
Evading the IRS
267
REFERENCES Bogen, J. and Woodward, J. (1988). Saving the Phenomena. The Philosophical Review 97, 303-52. Bogen, J. and Woodward, J. (1992). Observations, Theories, and the Evolution of the Human Spirit. Philosophy of Science 59, 590-611. Braithwaite, R. (1953). Scientific Explanation. Cambridge: Cambridge University Press. Collins, H. M. (1975). The Seven Sexes: A Study in the Sociology of a Phenomenon, or the Replication of Experiments in Physics. Sociology 9, 205-24. Collins, H. M. (1981). Son of Seven Sexes: The Social Deconstruction of a Physical Phenomenon. Social Studies of Science 11, 33-62. Conant, J. B. (1957). The Overthrow of the Phlogiston Theory: The Chemical Revolution of 17751789. In: J. B. Conant and L. K. Nash (eds.), Harvard Case Histories in Experimental Science, vol. 1. Cambridge, Mass.: Harvard University Press. Davis, P. (1980). The Search for Gravity Waves. Cambridge: Cambridge University Press. Donovan, A., Laudan, L., and Laudan, R. (1988). Scrutinizing Science. Dordrecht: Reidel. Earman, J. (1992). Bayes or Bust? A Critical Examination of Bayesian Confirmation Theory. Cambridge, Mass.: The MIT Press. Earman, J. and Glymour, C. (1980). Relativity and Eclipses. In: J.L. Heilbron (ed.), Historical Studies in the Physical Sciences, vol. 11, Part I. Feyerabend, P. K. (1985). Problems of Empiricism. Cambridge: Cambridge University Press. Franklin, A. (1990). Experiment, Right or Wrong. Cambridge: Cambridge University Press. Friedman, M. (1979). Truth and Confirmation. The Journal of Philosophy 76, 361-382. Galison, P. (1987). How Experiments End. Chicago: University of Chicago Press. Glymour, C. (1980). Theory and Evidence. Princeton: Princeton University Press. Goldman, A. (1986). Epistemology and Cognition. Cambridge, Mass: Harvard University Press. Hacking, I. (1983). Representing and Intervening. Cambridge: Cambridge University Press. Hempel, C. G. (1965) Aspects of Scientific Explanation. New York: The Free Press. Howson, C. and Urbach, P. (1989). Scientific Reasoning: The Bayesian Approach. La Salle, Ill.: Open Court. Humphreys, P. (1989). The Chances of Explanation. Princeton: Princeton University Press. Jeffrey, R. (1989). Probabilizing Pathology. Proceedings of the Aristotelian Society 89, 211-226. Lavoisier, A. (1965). Elements of Chemistry. Translated by W. Creech. New York: Dover. Lewin, R. (1987). Bones of Contention. New York: Simon and Schuster. Mackie, J. L. (1963). The Paradox of Confirmation. The British Journal for the Philosophy of Science 13, 265-277. Merrill, G. H. (1979). Confirmation and Prediction. Philosophy of Science 46, 98-117. Miller, R. (1987). Fact and Method. Princeton: Princeton University Press. Pais, A. (1982). ‘Subtle is the Lord. . .’: The Science and Life of Albert Einstein. Oxford: Oxford University Press. Popper, K. R. (1959). The Logic of Scientific Discovery. New York: Harper & Row. Priestley, J. (1970). Experiments and Observations on Different Kinds of Air, and Other Branches of Natural Philosophy Connected with the Subject. Vol. 1. Reprinted from the edition of 1790 (Birmingham: Thomas Pearson). New York: Kraus Reprint Co. Reichenbach, H. (1938). Experience and Prediction: An Analysis of the Foundations and the Structure of Knowledge. Chicago: University of Chicago Press. Schlesinger, G. (1976). Confirmation and Confirmability. Oxford: Clarendon Press. Taylor, J. (1982). An Introduction to Error Analysis. Oxford: Oxford University Press. Will, C. (1986). Was Einstein Right? New York: Basic Books. Woodward, J. (1983). Glymour on Theory Confirmation. Philosophical Studies 43, 147-157. Woodward, J. (1989). Data and Phenomena. Synthese 79, 393-472.
This page intentionally left blank
M. Norton Wise REALISM IS DEAD
1. Introduction In Explaining Science: A Cognitive Approach (1988), Ron Giere is attempting to deal with three opponents at once: the sterility of philosophy of science over the last thirty years, so far as its having any relevance to scientific practice; the pragmatic empiricism of Bas van Fraassen with its associated anti-realism (van Fraassen, 1980); and the “strong program” of social construction, with its radical relativism. Thus he is playing three roles. With respect to irrelevance he is playing reformer to the profession, if not heretic, while vis-à-vis van Fraassen he is one of the knights of philosophy engaged in a friendly joust; but relativism is the dragon that must be slain, or at least caged if it cannot be tamed. Cognitive science is Giere’s steed and realism his lance, suitable both for the joust and the battle with the dragon. While the mount is sturdy, I shall suggest, the lance of realism is pure rubber. Since van Fraassen is well able to defend himself, and since I have no pretensions to being one of the knights of philosophy, I will play the dragon of social construction. Let me begin, however, by generating a historical context for this new approach to philosophy of science.
2. A New Enlightenment Anyone familiar with the naturalistic philosophy of science propagated during the French enlightenment by d’Alembert, Condillac, Lavoisier, Condorcet and many others will recognize that Explaining Science announces a new enlightenment. Like the old one it depends on a psychological model. While the old one was based on the sensationalist psychology of Locke and Condillac, the new one is to be based on cognitive science. Thereby the faculties of memory, reason, and imagination are to be replaced by those of representation and judgement. The old “experimental physics of the mind,” to borrow Jean
In: Martin R. Jones and Nancy Cartwright (eds.), Idealization XII: Correcting the Model. Idealization and Abstraction in the Sciences (Poznań Studies in the Philosophy of the Sciences and the Humanities, vol. 86), pp. 269-285. Amsterdam/New York, NY: Rodopi, 2005.
M. Norton Wise
270
d’Alembert’s phrase for Lockean psychology, is to be replaced by a new science, more nearly a “theoretical physics” of the mind, but in both cases the faculties of the mind ground a naturalistic philosophy (Giere 1988). The two philosophies are most strikingly parallel in that both rely on a triadic relation between a linguistic element, a perceptual element, and the real world. The old philosophy presented the three aspects of knowledge as words, ideas, and facts. Lavoisier, in his Elements of Chemistry, essentially quoting Condillac, put it as follows: “every branch of physical science must consist of three things; the series of facts which are the objects of the science, the ideas which represent these facts, and the words by which these ideas are expressed. Like three impressions of the same seal, the word ought to produce the idea, and the idea to be a picture of the fact” (Lavoisier 1965). In the new philosophy, in place of word, idea, and fact, we have relational correlates: proposition, model, and real system. model (idea)
proposition (word)
real system (fact)
The heart of the new scheme is “representation,” which refers to our everyday ordinary ability to make mental maps or mental models of real systems and to use those models to negotiate the world. With models replacing ideas, representation replaces reason, especially deductive reason, which seems to play a very limited role in the problem solving activity of practical life, whether of cooks, carpenters, chess players, or physicists. Given the stress on everyday life, practice is the focus of analysis. Explaining Science: A Cognitive Approach explains scientific practice in terms of mental practice. It is therefore critical for Giere to change our conception of what scientific practice is, to show us, in particular, that the understanding and use of theories is based on representation, not reason, or not deductive reason. Similarly, there can be no question of universal rational criteria for accepting or rejecting a theory. There is only judgement, and judgement is practical. It cannot, however, according to Giere, be understood in terms of the rational choice models so prominent today in the social sciences. Instead, he develops a “satisficing” model. I am going to leave the critique of this model of judgement to others and content myself with representation. I would observe, however, that in Giere’s model of science the traditional categories of theory and theory choice get translated into
Realism Is Dead
271
representation and judgement, respectively. Thus I am going to be discussing only the nature of theory, which he treats in chapters 3-5. Theory choices are made in chapters 6-8. I would also observe that science in Giere’s model is highly theorydominated. Experiment is regularly described as providing data for theory testing; it has no life of its own. To appreciate his project, therefore, we must extract the term “practice” from the practical activity of the laboratory, where knowledge of materials, instruments, apparatus, and how to make things work is at issue, and transfer it to the use of theories, to theoretical practice. Here the practical activity is that of producing and using theoretical models, or representations. The basic idea is that theoretical models function much like apparatus. They are schemata of the real systems in nature, such as springs, planets, and nuclei, which allow us to manipulate those systems and generally to interact with them. Now I want to applaud this attention to practice, which is so prominent in recent history of science. And I want especially to applaud its extension to theoretical practice, which has not been so prominent. But Giere’s treatment has a consequence which I think unfortunate. It completely reduces theory to practice. This reduction will provide the first focus of my critical remarks. The second will be his realism and the third his use of this realism to defeat social constructivists. The relation of these three issues can be understood from the diagram above. First, a theory is to consist, as it does for Nancy Cartwright in How the Laws of Physics Lie, not merely of linguistic propositions – such as Newton’s laws – but also of a family of abstract models – such as the harmonic oscillator – which make sense of the propositions and which relate them to the world (Cartwright 1983, esp. chs. 7, 8). These models are constructs, or ideal types. They embody the propositions and thereby realize them. In Giere’s view, however, the propositions as such have no independent status. They are to be thought of as implicit in the models and secondary to them. And in learning to understand and use the theory one learns the models, not so much the propositions. This is the reduction of theory to practice. Secondly, the relation of the models to real systems is one of similarity. A model never captures all aspects of a real system, nor is the relation a one-to-one correspondence between all aspects of the model and some aspects of the real system. But in some respects and to some degree the model is similar to the real system. This simulacrum model, so far as I can tell, is also the same as Cartwright’s, but while she calls it anti-realism, Giere calls it realism (Cartwright 1983, ch. 8; Giere 1988, chs. 3, 4). I will call the difference a word game and conclude that realism is dead. Thirdly, the similarity version of realism is supposed to defeat the relativism of social constructivists. Perhaps it does, but it is far too heavy a weapon. Just before getting into these issues it may be useful to give one more contextual reflection. It is difficult today to discuss notions of representation
272
M. Norton Wise
without simultaneously discussing postmodernism, which has made representation into the shibboleth of the present. But the ideals of the Enlightenment – unity, simplicity, communicability, rationality, certainty – are precisely what postmodernism has been ridiculing for the last decade. Thus we ought to suspect that Giere in some sense shares the postmodern condition. Simplicity is no longer the name of nature, complexity is. Indeed I believe he does share this condition and that it is basically a healthy one. He thoroughly rejects the Enlightenment dream of the one great fact, or the one law which would subsume all knowledge under a unified deductive scheme. He recognizes that theoretical practice is not unified in this way; it is not a set of abstract propositions and deductions but a collection of models and strategies of explanation which exhibit the unity of a species, an interbreeding population containing a great deal of variation and evolving over time. To extend the metaphor, the relations within and between theories, between subdisciplines, and between disciplines will take us from varieties and species to entire ecological systems evolving over time. Cognitive science provides the right sort of model for this practical scheme because it is not a unified discipline in the usual sense. It is a loose cluster of disciplines, or better, of parts of disciplines, a veritable smorgasbord with a bit of logic here, of artificial intelligence there, and a smattering of anthropology on the side. It is a collection of techniques of investigation and analysis, developed originally in a variety of disciplines, which are now applied to a common object, the mind. The cluster of practices is held together only by the desire to understand mental function. Cognitive science is an excellent example of the patchwork systems of postmodern theory. It also represents the mainstream direction in contemporary universities of much research organization and funding, the cross-disciplinary approach to attacking particular issues. Thus Giere’s introductory chapter, “Toward a Unified Cognitive Theory of Science” (1988, ch. 1) is appropriately titled if we remember that the unity involved is not unity in the Enlightenment sense. He is hostile to disciplinary identification of the object of study, as in philosophy’s hang-up on rationality and sociology’s on the social. Entities like cognitive science will hold together only so long as their components fertilize each other intellectually or support each other institutionally, that is, so long as fruitful interaction occurs at the boundaries between its component practices. Here, if the evolutionary metaphor pertains, we should expect to find competition, cooperation, and exchange. That is, we should expect to find that social processes are the very heart of the unity of practice. But that is what we do not find. Seen in the light of postmodern theories of practice, my problems with Giere’s theory come at its two ends: in the complete reduction of the old Enlightenment unity of ideas to a new enlightenment unity of practices, and in the failure to make the unity of practices into a truly social interaction.
Realism Is Dead
273
3. Reduction of Theory to Practice in Mechanics Because Giere sets up his scheme with respect to classical mechanics, I will discuss in some detail his rendering of it. He has examined a number of textbooks at the intermediate and advanced level and contends that the presentation conforms to his scheme, that physicists understand and use classical mechanics in the way they ought to if representation via models is the real goal of the theory. I will attempt to show that these claims are seriously distorting, both historically and in the present. The problems begin at the beginning, when he throws into one bag intermediate and advanced textbooks, which are designed for undergraduate and graduate courses, respectively. The intermediate ones are intermediate rather than elementary only because they use the vector calculus to develop a wide range of standard problems, which I agree function as models in the above sense. But these texts base their treatment on Newton’s laws of motion and on the concept of force, that is, on the principles of elementary mechanics. What most physicists would call advanced texts, which are based on extremum conditions like the principle of least action and Hamilton’s principle, are said to differ from the sort based on Newton’s laws “primarily in the sophistication of the mathematical framework employed” (Giere 1988, p. 64). Nothing could be farther from the truth. The entire conceptual apparatus is different, including the causal structure. And it includes large areas of experience to which Newton’s laws do not apply. Thus Giere’s phrase “the Hamiltonian version of Newton’s laws,” rather than the usual “Hamiltonian formulation of mechanics,” betrays a serious distortion in the meaning that “advanced” mechanics has had for most physicists since around 1870 (Giere 1988, p. 99). This is significant in the first instance because Giere wants us to believe that the reason it doesn’t seem to matter much in mechanics textbooks, and in the learning and doing of physics, whether or which of Newton’s laws are definitions, postulates, or empirical generalizations is that the theory is to be located not so much in these laws themselves as in the model systems which realize them, like the harmonic oscillator and the particle subject to an inversesquare central force. But a more direct explanation would be that these laws are not actually considered foundational by the physicists who write the textbooks. These writers are teaching the practical side of a simplified theory which has widespread utility. Foundations are discussed in advanced texts, where extremum conditions, symmetry principles, and invariance properties are at issue. Debate within a given context, I assume, normally focuses on what is important in that context. We should not look to intermediate textbooks for a discussion of foundations. If we do we will be in danger of reducing theory to practice, and elementary practice at that.
274
M. Norton Wise
I would like to reiterate this first point through a brief glance at the content of mechanics texts in historical terms. If we look at the great French treatises and textbooks of Lagrange, Laplace, Poisson, Duhamel, Delaunay and others through the nineteenth century we will not find Newton’s laws at all. The foundations of French mechanics, including the standard textbooks of the École Polytechnique, were d’Alembert’s principle and the principle of virtual velocities, a generalized form of the balance principle which Lagrange used to reduce dynamics to statics. In Britain, of course, Newton’s laws were foundational, and judging from the amount of ink spilt, their status mattered considerably: from their meaning, to how many were necessary, to their justification. William Whewell’s Cambridge texts are instructive here (1824, 1832). After mid-century Newton’s laws did take on a standard form, at least in Britain, but only when they had been superseded. In the new physics, which for simplicity may be dated from Thomson and Tait’s Treatise on Natural Philosophy of 1867, energy functions replace force functions and extremum principles replace Newton’s laws (Kelvin and Tait 1879-83). Force is still a meaningful concept, but a secondary one, literally derivative, being defined as the derivative of an energy function (Smith and Wise 1989). Now this is all rather important because the new theory promised to penetrate thermodynamics and electromagnetic fields. In these areas Newton’s laws had little purchase. My point is that the value of the new theory lay not so much in supplying a more powerful way to solve old problems, as in suggesting a different conceptual base which might encompass entirely new realms of experience. The value of theory lay not so much in its power to solve problems as in its power to unify experience. The monolithic attempt to reduce theory to practice misses this central point. Cartwright and I may fully agree with Giere that theories consist in the general propositions and idealized model systems together, indeed we may agree that so far as use of the theory to solve problems is concerned, the models are what count. But one does not thereby agree that the general propositions ought to be thought of as what is implicit in the models. The propositions are what enable one to recognize the diversity of models as of one kind, or as having the same form. The power of theory lies in its unifying function. In this sense, theoretical strategy is not the same as practical strategy. Inventing a chess game is not the same as playing chess. Similarly, theoretical physicists are not the same sort of bird as mathematical physicists and neither is of the same sort as an experimentalist. Although they all interbreed, the evolution of physics has produced distinct varieties. I suspect that at the level of professional physicists Giere’s naturalism reduces the theoretical variety to the mathematical one. To capture the essential difference between theoretical and practical strategies the diagram below may be helpful. The top wedge represents the strategy of a theorist in attempting to encompass as many different natural systems as
Realism Is Dead
275
possible under one set of propositions. If the propositions are taken as the Hamiltonian formulation of mechanics, then the theorist hopes to include not only classical mechanics itself, but thermodynamics, geometrical optics, electromagnetic theory, and quantum mechanics. The ideal is unity under deduction, although the deduction must be constructed in each case by factoring in a great deal of specialized information not contained in the propositions.
Hamilton’s principle
Logic Psychology Anthropology Artificial intelligence Neurology
Classical mechanics Thermodynamics Electromagnetism Geometrical optics Quantum mechanics
Mental function
The lower wedge, directed oppositely, represents the practical strategy of solving a particular problem (at the point of the wedge) by bringing to bear on it whatever resources are available: bits of theory, knowledge of materials, phenomenological laws, standard apparatus, mathematical techniques, etc. One assembles a wide variety of types of knowledge and tries to make them cohere with respect to a single object of interest. If classical mechanics epitomizes the theoretical wedge, cognitive science epitomizes the practical one. Of course neither strategy is ever fully realized in a pure form. But as strategies they seem to be very different in kind and to serve very different purposes. Perhaps a naturalist would argue that that is why the theoretical and the practical are so ubiquitously differentiated in everyday life.
4. Similarity Realism The historical picture takes on a somewhat different tint with respect to realism. Giere argues that theoretical models represent real systems in some respects and to some degree and that in these respects and degrees they are realistic. Nearly
276
M. Norton Wise
all of the formulators of the new mechanics, however, rejected precisely this version of the realism of theories. Thomson and Tait labeled their theoretical development “abstract dynamics” to differentiate it from the realistic theory they lacked, namely “physical dynamics.” Abstract dynamics worked with rigid rods, frictionless surfaces, perfectly elastic collisions, point particles and the like, not with the properties of real materials. They were singularly unimpressed with the fact that abstract dynamics yielded approximately correct results in certain idealized situations, because the theory actually violated all experience. Most simply, its laws were reversible in time, which meant that it contradicted the second law of thermodynamics and therefore could not be anything like correct physically. They suspected that its fundamental flaw lay in the fact that it dealt with only a finite number of variables. It could be formulated, therefore, only for a finite system of discrete particles and would not apply to a continuum, which Thomson especially believed to be the underlying reality (Smith and Wise 1989, chs. 11, 13, 18). But the argument does not depend on this continuum belief. As I interpret Thomson, he would say that the realist position is vitiated not by the fact that the theory fails to reproduce natural phenomena in some respect or to some degree but by the fact that it straightforwardly contradicts all empirical processes in its most fundamental principles. He would not understand the point of Giere’s contention with respect to the ether that its non-existence is “not [a good basis] for denying all realistically understood claims about similarities between ether models and the world” (Giere 1988, p. 107; emphasis in original). Why not, pray tell? Why not simply call the similarities analogies and forget the realism? Thomson was a realist, most especially about the ether. But to get at reality he started at the far remove from abstract theory and abstract models, namely at the phenomenological end, with the directly observable properties of known materials. For example, he argued for the reality of his elastic-solid model of ether partly on the grounds that the ether behaved like Scotch shoemaker’s wax and caves-foot jelly. He attempted always to construct realistic models which relied on the practical reality of familiar mechanical systems rather than on the mathematical structure of idealized hypothetical models. From the perspective of this contrast between abstract dynamics and practical reality, the point of reformulating mechanics in terms of energy functions and extremum conditions was not to obtain a more realistic theory but to obtain a more general one, and one that was thus more powerful in the sense of organizing a greater range of experience. To make abstract mechanics subsume thermodynamics one could represent matter as composed of a finite number of hard atoms interacting via forces of attraction and repulsion, but then one would have to add on a randomizing assumption in order to get rid of the effects of finiteness and timereversibility in the equations of mechanics. The resulting theory, doubly false,
Realism Is Dead
277
certainly did not count as realistic among its British analysts: Thomson, Tait, Maxwell, and others. Maxwell put this point in its strongest form, to the effect that the goal of theoretical explanation in general, and especially of the new mechanics, was not to discover a particular concrete model which reproduced the observed behavior of the system in question but to discover the most general formulation possible consistent with this observed behavior. Thus one sought the most general energy function for the system, a Lagrangian or a Hamiltonian, which would yield empirically correct equations for its motion, this energy function being specified in terms of observable coordinates alone, like the input and output coordinates of a black box, or more famously like the bell ropes in a belfry, independent of any particular model of the interior workings of the belfry. For every such Lagrangian or Hamiltonian function, Maxwell observed, an infinite variety of concrete mechanical models might be imagined to realize it. He himself exhibited uncommon genius in inventing such models, but unlike his friend Thomson, he was much more sanguine about the value of unrealistic ones, regarding them as guarantors of the mechanical realizability of the Lagrangian in principle. He did not suppose that a similarity between the workings of a given model and observations on a real system indicated that the system was really like the model, but only analogous to it. Being analogous and being like were two different things. Similar remarks could be made for the perspective on generalized dynamics and on mechanical models of Kirchhoff, Mach, Hertz, and Planck. Even Boltzmann, the most infamous atomistic-mechanist of the late nineteenth century expressed himself as in agreement with Maxwell on the relation of models to real systems. Since Boltzmann, in his 1902 article on “Models” for the Encyclopaedia Britannica, cites the others to support his view, I will let him stand for them all. Boltzmann remarks that “On this view our thoughts stand to things in the same relation as models to the objects they represent . . . but without implying complete similarity between thing and thought; for naturally we can know but little of the resemblance of our thoughts to the things to which we attach them.” So far he does not diverge strikingly from Giere on either the nature or the limitations of similarity. But while Giere concludes realism from limited similarity, Boltzmann concludes that the “true nature and form” of the real system “must be regarded as absolutely unknown” and the workings of the model “looked upon as simply a process having more or less resemblance to the workings of nature, and representing more or less exactly certain aspects incidental to them.” Citing Maxwell on mechanical models, he observes that Maxwell “did not believe in the existence in nature of mechanical agents so constituted, and that he regarded them merely as means by which phenomena could be reproduced, bearing a certain similarity to those actually existing . . . The question no longer being one of ascertaining the actual internal structure of
278
M. Norton Wise
matter, many mechanical analogies or dynamical illustrations became available, possessing different advantages.” For Maxwell, “physical theory is merely a mental construction of mechanical models, the working of which we make plain to ourselves by the analogy of mechanisms we hold in our hands, and which have so much in common with natural phenomena as to help our comprehension of the latter” (Boltzmann 1974, p. 214, 218, emphasis added). Structural analogy, not realism, is the relation of similarity between models and natural phenomena. I do not see that Giere has shown more. His appeal to the category of representation in cognitive science does not help. To drive one last nail in this coffin of realism, let me tell a little story about G. F. Fitzgerald, a second generation Maxwellian who attempted in 1896 to convince his friend William Thomson, Lord Kelvin since 1892, of the wrong-headed nature of Thomson’s elastic-solid model of the ether. The debate between them had been going on for twenty years already, with Thomson insisting that the ether had to be like an elastic solid, given the nature of all known materials, and complaining that Maxwell’s equations did not give a sufficiently definite mechanical model based on force and inertia. To rely on Maxwell’s equations as the basis of electromagnetic theory Kelvin regarded as “nihilism,” the denial that reality could be known. Fitzgerald in turn argued that the elastic-solid supposition was unwarranted. In spite of a limited analogy, the ether might not be at all like any known matter. The matter of the ether, to him, was simply that which obeyed the mathematical laws invented by Maxwell and corroborated by experiment. “To work away upon the hypothesis” that the ether was an elastic solid, therefore, was “a pure waste of time” (Smith and Wise 1989, ch. 13). To this condemnation of his realistically interpreted analogy, Kelvin retorted: the analogy “is certainly not an allegory on the banks of the Nile. It is more like an alligator. It certainly will swallow up all ideas for the undulatory theory of light, and dynamical theory of E & M not founded on force and inertia. I shall write more when I hear how you like this.” The answer came, “I am not afraid of your alligator which swallows up theories not founded on force and inertia . . . I am quite open to conviction that the ether is like and not merely in some respects analogous to an elastic solid, but I will . . . wait till there is some experimental evidence thereof before I complicate my conceptions therewith.” Oliver Heaviside put it more succinctly, remarking to Fitzgerald that Lord Kelvin “has devoted so much attention to the elastic solid, that it has crystallized his brain” (Smith and Wise 1989, ch. 13). Now I am not suggesting that Ron Giere’s realism has crystallized his brain; but like Fitzgerald, I fail to see what the claim of realism about theories adds to that of analogy. I would suggest further, that a naturalistic theory of science, which is supposed to represent the actual behavior of theoretical physicists, ought to consider the history of that behavior. I have attempted to show with the above examples that a very significant group of theoreticians over a fifty-year
Realism Is Dead
279
period examined the similarity relations that Giere considers an argument for realism and drew the opposite conclusion. They opted for nihilism. Given the prominent role of these people in the evolution of physics, it seems that an evolutionary naturalism in particular, ought not to make the successful pursuit of physics depend on realism about theories. I therefore advocate following their nihilistic example with respect to the realism-antirealism game. But this conclusion does not follow only from history. A reading of many contemporary theorists supports it. Stephen Hawking, for example, in his popular little book, A Brief History of Time (1988), contends that theoreticians are merely engaged in making up more or less adequate stories about the world. The same view appears at length in a book called Inventing Reality: Physics as Language (1988), by Bruce Gregory, associate director of the Harvard-Smithsonian Center for Astrophysics. “A physicist is no more engaged in painting a ‘realistic’ picture of the world than a ‘realistic’ painter is,” Gregory opines, and again, “[p]hysical theories do not tell physicists how the world is; they tell physicists what they can predict reliably about the behavior of the world.” The preceding two sections suggest that Giere’s reduction of theory to practice and his similarity realism are linked. Actually I think that if we reject the former we automatically reject the latter. I will illustrate this linkage with a final example from mechanics, again emphasizing the more sophisticated versions which rely on extremum conditions and the variational calculus. The most important theoretical goal in using such formulations is to be able to encompass within a single formalism a wide variety of quite different fields of physics which employ different mathematical relations, such as Newton’s laws, Maxwell’s equations, and the Schrödinger equation. Goldstein, the author of one of Giere’s advanced textbooks and the one my entire generation of physicists was brought up on, remarks that “[c]onsequently, when a variational principle is used as the basis of the formulation, all such fields will exhibit, at least to some degree, a structural analogy” (Goldstein 1950, p. 47). This is the same view that Maxwell promoted about formal analogy. It means, to give Goldstein’s simple example, that an electrical circuit containing inductance, capacitance, and resistance can be represented by a mechanical system containing masses, springs, and friction. The similarity, however, has never induced physicists to call the representation realistic. The same argument applies to the relation between theoretical models and real systems.
5. Realism in the Laboratory Very briefly, I shall comment on a different kind of realism which appears in a chapter of Giere’s book entitled “Realism in the Laboratory.” The focus shifts from theoretical models to theoretical entities and the argument for realism
280
M. Norton Wise
shifts from similarity to manipulation and control. Here comparisons with Ian Hacking’s Representing and Intervening are unavoidable, as Giere acknowledges. One thinks particularly of Hacking’s phrase “If you can spray it, it’s real,” for which Giere’s alternative is, “Whatever can be physically manipulated and controlled is real” (Hacking 1983; Giere 1988). He has as little time for philosophers and sociologists who don’t believe in protons as he would have for soldiers who don’t believe in bullets, and for much the same reasons. Both can be produced and used at will with predictable effect. Protons, in fact, have the status in many experiments, not of hypothetical theoretical entities, but of research tools. The part of this discussion I like best is the three pages on technology in the laboratory. The subject has become a popular one in the last five years, at least in the history of science. In Giere’s naturalistic scheme, technology provides the main connector between our cognitive capacity for representation and knowledge of, for example, nuclear structure. He suggests that the sort of knowledge bound up in technology is an extension of our sensorimotor and preverbal representational systems. It is both different in kind and more reliable than knowledge requiring symbolic and verbal manipulation. It is knowledge of the everyday furniture of the laboratory, which allows one to act in the world of the nucleus. Here we seem finally to be talking about experimental practice and the life of an experimental laboratory. Theory testing is certainly involved, but is by no means the essence. I can only complain that three pages hardly count in a subject so large as this one. I do have two reservations, however. First, there seems to be a tendency to slide from the reality of a thing called a proton to the reality of the model of the proton, including its properties and characteristics. The model, even for the simplest purposes, involves a quantum mechanical description of elusive properties like spin, which is not spin at all in our everyday sense of a spinning ball. Does this assigned property have the same reality status as the proton itself? One need not get into the status of quantum mechanical models to raise this problem. Historically, many entities have been known to exist, and have been manipulated and controlled to an extraordinary degree, on the basis of false models. Electric current is an outstanding example. Our ability to manipulate and control a proton may well guarantee the reality of the proton without guaranteeing the reality of any model of it. This point goes back immediately to the previous remarks about similarity realism and theoretical models. My second reservation has to do with tool use. How are we to differentiate between material tools like magnets, detectors, and scattering chambers, on the one hand, and intellectual tools like mathematical techniques, on the other hand? Both are used as agents of manipulation and control of protons and both are normally used without conscious examination of principles or foundations.
Realism Is Dead
281
Both are the stuff of practice. Once again then, any arguments about laboratory practice are going to have to tie back into those about theoretical practice. A more thoroughgoing analysis will be required to show their similarities and differences.
6. Social Construction and Constrained Relativism If it is true, as I have argued, that Giere’s realism about theories is just a game with words, why does he pursue the game? The only answer I can see is that he is anxious to defeat the radical relativism which he associates with sociologists or social constructivists. That is, Giere’s realism serves primarily as a guard against relativism in the no-constraints version, which maintains that nature puts no constraints on the models we make up for explaining it. All models are in principle possible ones and which one is chosen is merely a matter of social negotiation. The problem of explaining the content of science, therefore, is exclusively that of explaining how one scheme rather than another comes to have currency among a particular social group. To move all the way to a realist position to defeat this form of relativism, however, seems to be bringing up cannons where peashooters would do. A simpler argument would be to reject no-constraints relativism on pragmatic grounds, on the grounds that there is no evidence for it and a great deal of evidence against it. It is so difficult to make up empirically adequate theories of even three or five different kinds, that we have no reason to believe we could make up ten, let alone an infinite number, and much less every conceivable kind. Constructing castles in the air has not proved a very successful enterprise. The construction of magnetic monopoles has fared little better. Arguing from all experience with attempts to construct adequate models, therefore, particularly from the empirical failure of so many of the ones that have been constructed, we must suppose that nature does put severe constraints on our capacity to invent adequate ones. Radical relativism, for a naturalist, ought to be rejected simply because it does not conform to experience. This argument leaves a social constructivist perfectly free to claim that, within the limits of empirical consistency, even though these limits are severe, all explanations are socially constructed. Among the reformed school, Andrew Pickering, Steven Shapin, Simon Schaffer, and Bruno Latour all hold this view. As Pickering puts it, nature exhibits a great deal of resistance to our constructions. Almost no social constructivists, to my knowledge, presently subscribe to the no-constraints relativism of the strong program except a few devotees
282
M. Norton Wise
remaining in Edinburgh and Bath.1 For those interested in the new constellation of social construction I would recommend a recent issue of Science in Context, with contributions from Tim Lenoir, Steve Shapin, Simon Schaffer, Peter Galison, and myself, among others. Bruno Latour’s Science in Action (1987) also presents an exceedingly interesting position. He extends the social negotiations of the old strong program to negotiations with nature. For all of these investigators, nature exhibits such strong resistance to manipulation that no-constraints relativism has become irrelevant. And the constrained relativism which they do adopt does not differ significantly from what Giere calls realism. There are, however, significant reasons for not calling it realism. Constraints act negatively. They tell us which of the options we have been able to invent are possibly valid ones, but they do not invent the options and they do not tell us which of the possible options we ought to pursue. This suggests immediately that in order to understand how knowledge gets generated we must analyze the social and cultural phenomena which, over and above the constraints which nature is able to exert, are productive of scientific knowledge. Actually this position seems to be Giere’s own. Noting that his constructive realism was invented to counter van Fraassen’s constructive empiricism, he adds, “The term emphasizes the fact that models are deliberately created, ‘socially constructed’ if one wishes, by scientists. Nature does not reveal to us directly how best to represent her. I see no reason why realists should not also enjoy this insight” (Giere 1988, p. 93). They should, but having savored its delights, they should give up realism. The position I am advocating has strong instrumentalist aspects. People use what works, or what has utility. But the instrumentalism of philosophers does not normally take into account the social relativity of utility. What works is relative to what purposes one has, and purposes are generally social. Thus it is no good explaining the growth of Maxwellian electromagnetic theory in Britain simply in terms of the fact that it gave a satisfactory account of the behavior of light. Satisfactory to whom? Certainly not to all mathematical physicists in Britain and certainly not to any mathematical physicists on the Continent between 1863, when Maxwell first published his theory, and the mid-1890s when it was superseded by Lorentz’s electron theory. Social historians contend that differences like these require a social interpretation of how scientific explanations get constituted. Philosophers of science, including instrumentalists and pragmatists, have generally had nothing to offer. They do not incorporate the social into the essence of science, but leave it on the borders, perpetuating the internal-external dichotomy. ⎯⎯⎯⎯⎯⎯⎯ 1
Shapin, one of the original Edinburgh group, never accepted the social determinist position. He has removed to San Diego. Pickering has removed to Illinois, but not before feuding with Bloor on the subject. Latour has been similarly sparring with Collins.
Realism Is Dead
283
Ron Giere is somewhat unique on this score, in that he is anxious not to exclude the social and historical contingency of science, but he too offers little that is explicitly social. Representation is for him an activity of individuals. They may achieve their individual representations in a social setting but the process of interaction is not a constitutive part of their representations. With this issue I would like to return to my starting point in the Enlightenment. The new enlightenment shares a standard flaw with the old. Based on individual psychology, it remains individualist in its philosophy of science. It allots no conceptual space, for example, to the political economy that permeates the practice of science and acts as a highly productive force in its advancement. It does not explicitly recognize that individuals are socially constructed, just as the objects are that individuals explain. Until the social process is brought in explicitly, philosophy of science will have little to offer to the social history of science.
7. Realism Is Dead In conclusion, I would like to return once more to the simulacrum model in order to suggest a somewhat different view of it. Previously I have attempted to show that if the behavior of physicists is to be our arbiter, then similarity between theoretical models and real systems cannot be used as an argument for the realism of theories. On the other hand, the similarity remains, and remains highly useful, whether we play the realist-antirealist game or not. It is just a game. But suppose we take the nihilist view a bit more seriously, and that we remodel Nietzsche’s famous line about God. The question is no longer whether God exists or not; God is dead. Similarly, we are beyond realism and antirealism. The question is no longer whether reality speaks to us through our models or not. Realism is dead. We have collapsed reality into our models and they have become reality for us. That seems to me the proper attitude to take to the simulacrum scheme. Here I am borrowing the argument of one of the more notorious postmodernists, Jean Baudrillard, who uses the term ‘simulacrum’ to describe the relation between image and reality in contemporary film and literature. I find his attitude particularly appropriate to Giere’s argument just because the argument identifies realism with similarity. It rescues reality by turning it into its opposite, the image. It thereby collapses reality into our models of it. Like film images, the models become more real than reality itself. They become “hyperreal,” to use Baudrillard’s term. One could easily take a critical view of this process. I would like instead to look at its positive side. It means that our models are reality-forming. They are so detailed, so technically perfect, that we lose the consciousness of their being representations. They function for us as total simulacra, or reality itself. Of
284
M. Norton Wise
course we know that models pick out certain aspects for emphasis, subordinate others, and ignore or even suppress the remainder, but if the model is our means for interacting with the world then the aspects that it picks out define the reality of the world for us. This attitude gains its strongest support from technology. Whenever we embody models in material systems and then use those systems to shape the world, we are shaping reality. Prior to the twentieth century, it might be argued, technology merely enhanced or modified already existing materials and energy sources. But that is certainly no longer the case. We regularly produce substances and whole systems that have no natural existence except as artifacts of our creation. This is most startling in the case of genetic engineering, where life itself is presently being shaped, but it applies equally to Teflon and television. I would stress that we typically accomplish these creations by manipulating models which we attempt to realize in the world. Often nature is recalcitrant and we have to return to our models. But far from representing a preexisting nature, the models create the reality that we then learn to recognize as natural. The process of creation is most obvious with respect to computer simulations. They have become so sophisticated, and so essential to basic research in all of the sciences, that they often substitute for experimental research. A computer-simulated wind tunnel can have great advantages over a “real” one as a result of increased control over relevant variables and elimination of irrelevant ones. More profoundly, the entire field of chaos theory has emerged as a result of computer simulations of the behavior of non-linear systems. The computer-generated pictures make visible distinct patterns of behavior where previously only chaotic motion appeared. They show how such patterns can be generated from simple codes iterated over and over again. Of course, it is possible to hold that the computer simulations merely discover a reality that was actually there all along. But this way of talking is precisely the target of the assertion that realism is dead. I prefer to say that the simulations create one of the constructions of reality possible under the constraints of nature. They show that it is possible coherently to represent chaotic systems in terms of iterated codes. But this representation is an artifact of the computer, or an artifact of an artifact. Herein lies the lesson of the simulacrum scheme. We are the creators. M. Norton Wise Department of History University of California, Los Angeles
[email protected] Realism Is Dead
285
REFERENCES Boltzmann, L. (1974). Theoretical Physics and Philosophical Problems: Selected Writings. Edited by B. McGuinness, with a foreword by S. R. de Groot; translated by P. Foulkes. Dordrecht: Reidel. Cartwright, N. (1983). How the Laws of Physics Lie. Oxford: Clarendon Press. Giere, R. N. (1988). Explaining Science: A Cognitive Approach. Chicago: University of Chicago Press. Goldstein, H. (1950). Classical Mechanics. Cambridge, Mass.: Addison-Wesley. Gregory, B. (1988). Inventing Reality: Physics as Language. New York: Wiley. Hacking, I. (1983). Representing and Intervening: Introductory Topics in the Philosophy of Natural Science. Cambridge: Cambridge University Press. Hawking, S. W. (1988). A Brief History of Time: From the Big Bang to Black Holes. Toronto, New York: Bantam. Latour, B. (1987). Science in Action: How to Follow Scientists and Engineers through Society. Cambridge, Mass.: Harvard University Press. Lavoisier, A.-L. (1965). Elements of Chemistry: In a New Systematic Order, Containing all the Modern Discoveries. Translated by R. Kerr. New York: Dover. Kelvin, W. Thomson and Tait, P. G. (1879-83). Treatise on Natural Philosophy. New edition. 2 vols. Cambridge: Cambridge University Press. Smith, C. and Wise, M. N. (1989). Energy and Empire: A Biographical Study of Lord Kelvin. Cambridge: Cambridge University Press. Van Fraassen, B. C. (1980). The Scientific Image. Oxford: Clarendon Press. Whewell, W. (1824). An Elementary Treatise on Mechanics: Designed for the Use of Students in the University. 2nd edition. Cambridge: Printed by J. Smith for J. Deighton. Whewell, W. (1832). On the Free Motion of Points, and on Universal Gravitation, including the Principal Propositions of Books I. and III. of the Principia; the First Part of a New Edition of a Treatise on Dynamics. Cambridge: For J. and J. J. Deighton.
286
M. Norton Wise
This page intentionally left blank
Ronald N. Giere IS REALISM DEAD?
1. A New Enlightenment? I appreciate Norton Wise’s comparison of my project in Explaining Science (1988) with that of Enlightenment scientists and philosophers (Wise, this volume). When rejecting one’s immediate philosophical predecessors, it is comforting to be able to portray oneself not as a heretic who has abandoned philosophy, but as a reformer who would return philosophy to the correct path from which his predecessors had strayed. But we cannot simply return to the ideals of the Enlightenment. Some doctrines that were fundamental to the Enlightenment picture of science must be rejected. In particular, I think we must reject the idea that the content of science is encapsulated in universal laws. And we must reject the notion that there are universal principles of rationality that justify our belief in the truth of universal laws. As Wise notes, these latter are typically “postmodern” themes, and, as such, are usually posed in explicit opposition to the modernism of the Enlightenment. It is my view that this opposition must be overcome. The overall project for the philosophy of science now is to develop an image of science that is appropriately postmodern while retaining something of the Enlightenment respect for the genuine accomplishments of modern science. To many the idea of an “enlightened postmodernism” may seem contradictory. I see it merely as an example of postmodern irony.
2. Mechanics Wise has most to say about my discussion of theories based on an analysis of classical mechanics seen through the eyes of contemporary textbooks. I will not argue with Wise about the history of mechanics, since he is an expert on that subject and I am not. I concede the point that the difference between intermediate and advanced mechanics texts is not just that the advanced texts
In: Martin R. Jones and Nancy Cartwright (eds.), Idealization XII: Correcting the Model. Idealization and Abstraction in the Sciences (Poznań Studies in the Philosophy of the Sciences and the Humanities, vol. 86), pp. 287-293. Amsterdam/New York, NY: Rodopi, 2005.
288
Ronald N. Giere
use a more sophisticated mathematical framework. I can safely admit, as Wise claims, that the Hamiltonian formulation of mechanics “includes large areas of experience to which Newton’s laws do not apply.” What I do not see is how this makes any difference to my main point, which was that scientific theories are not best understood as sets of universal statements organized into deductive systems. They are better understood as families of models together with claims about the things to which the models apply (Giere 1988, ch. 3). Wise and I seem to agree that Newton’s laws may best be thought of as providing a recipe for constructing models of mechanical systems. I do not see why moving up in generality from Newton’s laws to Hamilton’s equations should move one from mere recipes to fundamental laws. Rather, it seems to me that Hamilton’s equations merely provide a more general recipe for model building. One simply has a bigger family of models. Having said that, I am inclined to agree that there is something to Wise’s preference for the Hamiltonian formulation. The problem is to capture that difference within my particularistic view of scientific theories. Wise refers to the power of theory to “unify experience.” Newton’s theory is traditionally credited with possessing this virtue. It unified terrestrial and celestial motions, for example. My first inclination is to say that this unity consists primarily in the fact that “Newton’s laws” provide a single recipe for constructing models that succeed in representing both terrestrial and celestial systems. The unity provided by Hamilton’s equations is just more of the same. In short, unity may best be understood simply as scope of application. Yet there still seems something more to the Hamiltonian version of mechanics. The word “foundational,” however, does not seem to capture the difference. As Wise himself notes, what counts as “foundational” at any particular time may depend on whether one is in Britain or France. The difference, I think, is that the Hamiltonian approach is more fundamental. That is to say, energy is more fundamental than force. But what does “fundamental” mean in this context? Here I am tempted to invoke a realist framework. The conviction that energy is more fundamental than force is the conviction that energy is the causally deeper, and more pervasive, quantity in nature. Forces are merely the effects of changes in energy. On this understanding, wide scope of application turns out to be not only pragmatically valuable, but an indicator of something fundamental in nature. Does this mean that I persist in “reducing theory to practice”? And would that be a bad thing? I do not know how to answer these questions. Nor am I convinced it is important to do so. What matters to me is whether this is the right account of the nature and role of theory in science as we now know it.
Is Realism Dead?
289
3. Realism In Explaining Science I distinguish several varieties of empiricism and several varieties of realism (1988, ch. 4). Whether to call my view “realist” or not depends on which variety one has in mind. And whether the label is important depends on the variety of non-realism being denied. Wise’s discussion is informed by nineteenth century intellectual distinctions with which I am only vaguely familiar. When he quotes Boltzmann writing that the “true nature and form” of real systems “must be regarded as absolutely unknown,” I hear echoes of Kant. And I think of Hilary Putnam’s (1981) characterization of “metaphysical realism.” If realism is the view that nature has a definite nature apart from any conceptualization and, moreover, that we could somehow know that nature “directly” without any mediating perceptual and conceptual structures, then I am an anti-realist. Again, if realism is the view that there must be complete isomorphism between model and thing modeled, so that, for example, the ether would have to be regarded as literally containing little wheels, then I am an anti-realist. Our differences come out clearly when Wise writes: Structural analogy, not realism, is the relation of similarity between models and natural phenomena. As I understand it, “structural analogy” is probably the most important kind of similarity between models and real systems. Constructive realism, for me, includes the view that theoretical hypotheses assert the existence of a structural analogy between models and real systems. I might even be pressed into claiming that “similarity of structure” is the only kind of similarity between models and reality that matters. I call this a kind of realism. Wise says that is an empty word; we might as well call it anti-realism. Whether the label is significant depends on the work it does. Wise thinks it is mainly a weapon in the battle against social constructivism. It is that, but much more. It is central to my understanding of a large part of post-positivist philosophy of science. For the moment I will drop the word “realism” in favor of what for me is a more fundamental notion, representation. Logical empiricism was representational in the strong sense that it regarded scientific hypotheses as true or false of the world. Moreover, logical empiricists dreamt of constructing an inductive logic that would provide the rational degree of belief, relative to given evidence, in the truth of any hypothesis. Beginning with Kuhn (1962), a major strain of post-positivist thinking denied the representational nature of science. For Kuhn, science is a puzzle solving activity which provides, at most, a way of looking at the world, but not literally a representation of it – certainly not in the sense that science makes claims about what is true or false of the world. One of the major lines of reaction to Kuhn, that of Lakatos (1970) and Laudan (1977), agrees with Kuhn about the non-representational nature of
290
Ronald N. Giere
science. For Lakatos, progressive research programs are those that generate new empirical (not theoretical) content. Laudan remained closer to Kuhn in maintaining that more progressive programs are those with greater problem solving effectiveness. Both Lakatos and Laudan identified progress with rationality so as to recover the philosophical position that science is rational, in opposition to Kuhn who denied any special rationality for science. There is one more distinction to be made before I can conclude my defense of realism. Laudan, for example, claims that his account of science is representational in the sense that scientific hypotheses are statements that are in fact true or false. He calls this “semantic realism.” But he goes on to argue that there are no, and perhaps can be no, rational grounds for any claims one way or the other. In short, the basis of Laudan’s anti-realism is not semantic, but epistemological. The same is true of van Fraassen’s (1980) anti-realism. My realism has two parts. First, it rejects notions of truth and falsity as being too crude for an adequate theory of science. Taken literally, most scientific claims would have to be judged false, which shows that something is drastically wrong with the analysis. Rather, I regard scientific hypotheses as typically representational in the sense of asserting a structural similarity between an abstract model and some part of the real world. (I say “typically” because I want to allow for the possibility of cases where this is not so. Parts of microphysics may be such a case.) The second part is the theory of scientific judgment, and the theory of experimentation, which Wise, for reasons of exposition, put to one side. As elaborated in Explaining Science, I think there are judgmental strategies for deciding which of several models possesses the greater structural similarity with the world. Typically these strategies involve experimentation. And they are at least sometimes effective in the sense that they provide a substantial probability for leading one to make the right choice (1988, ch. 6).
4. Scientists’ Theories of Science Before turning to social constructivism, I would like to indulge in one methodological aside. Wise suggests “that a naturalistic theory of science, which is taken to mirror the views of theoretical physicists about theory, would do well to consider the history of physicists’ attitudes” (Wise, this volume). That suggestion is ambiguous and can be quite dangerous. Everybody has theories about the world. And most people also have theories about themselves and what they are doing in the world. But one’s theories about oneself are only loosely descriptive of one’s real situation. These theories also, perhaps mainly, serve to rationalize and integrate one’s interests, activities, and ambitions. Thus, as a general rule, actors’ accounts of their own activities
Is Realism Dead?
291
cannot be taken as definitive. These accounts provide just one sort of evidence to be used in the investigation of what the actors are in fact doing. Scientists are no different. The theories scientists propound about their scientific activities do not have a privileged role in the study of science as a human activity. What scientists will say about the nature of their work depends heavily on the context, their interests and scientific opponents, the supposed audience, even their sources of funding. Newton’s claim not to feign hypotheses may be the most famous case in point. The claim is obviously false of his actual scientific practice. It may make more sense when considered in the context of his disputes with Leibniz and the Cartesians. But I am no historian. Let me take a more mundane example from my own experience studying work at a large nuclear physics laboratory (1988, ch. 5). One of my “informants” claimed that physics is like poetry and that physicists are like poets. I don’t know if he ever propounded this theory to his physicist friends. But he told me, most likely because he thought that this is the kind of thing that interests philosophers of science. Well, maybe there is poetry, as well as music, in the hum of a well-tuned cyclotron. But the truth is that this man began his academic life as an English major with strong interests in poetry. I am sure that this fact had more to do with sustaining his theory about the nature of physics than anything going on in that laboratory. I can only speculate about the details of the psychological connections.
5. The Case against Constructivism In my present exposition, the rejection of social constructivist sociologies of science comes out not as a main battle, but as a mopping up operation. Karin Knorr-Cetina (1981), a leading social constructivist, agrees with Laudan that scientists often intend their statements to be true or false of the world. She calls this “intentional realism.” Her argument is that if one examines in detail what goes on in the laboratory, one will find that what might be true or false of the world has little or no effect on which statements come to be accepted as true or false. That is more a matter of social contingency and negotiation than interaction with the world. I do not deny the existence of contingency and social negotiation in the process of doing science. Nor do I deny the power of personal, professional, and even broader social interests. My claim is that experimental strategies can, in some circumstances, overwhelm social interactions and interests, leaving scientists little freedom in their choice of the best fitting model. The final chapter of Explaining Science (1988, ch. 8), which deals with the 1960s revolution in geology, was intended to illustrate this point.
292
Ronald N. Giere
6. Social versus Individual In a review of Explaining Science, the philosopher of science Richard Burian (1988) voices the worry that my account of science leaves too much room for social factors. Wise objects that there is not enough. Actually I think Wise is more nearly correct. But Burian’s worry indicates that there is considerable room for the social in my account. Throughout Explaining Science there are hints of an ecological, or evolutionary, model of the growth of science (1988, pp. 12-26, 133-37, 222, and 24849). At one time I had intended to develop this model in greater detail, but that proved not to be feasible, and was put off, hopefully to be taken up as a later project. It is in this context that I would begin explicitly to model the social processes of science. I am convinced, however, that I have the priorities right. An adequate theory of science must take individuals as its basic entities. This means primarily, but not exclusively, scientists. I even have a theoretical argument for this position (1989). A scientific theory of science must in the end be a causal theory. In the social world the only active causal agents are individuals. This is not to deny that there is a socially constructed social reality. Individuals are enculturated and professionalized. But nothing happens unless individuals act. In particular, nothing changes unless individuals change it. And science is nothing if not changing. So no theory of science that reduces individuals, and their cognitive capacities, to black boxes can possibly be an adequate theory of science. In spite of his social history rhetoric, Wise in practice follows this strategy. His recent work (Wise and Smith 1988) attempts to show how machines like the steam engine and the telegraph provided a means by which the culture of late nineteenth century British commerce and industry became embodied in the content of the physics of the day. But even more important than the machines is the individual scientist, in this case William Thomson, who performed the translation. In the social world of science, as in the social world generally, there are no actions without actors.
7. Is Realism Dead? Since I finished writing Explaining Science, there has been some softening in the sociological position, as Wise notes. Perhaps there is no longer a significant substantive difference between my position and the consensus position among sociologists and social historians of science. Since I am not yet sure that convergence has in fact occurred, let me conclude by stating what I would regard as the minimal consensus on the starting point for developing an adequate theory of science. It is this: We now
Is Realism Dead?
293
know much more about the world than we did three hundred, one hundred, fifty, or even twenty-five years ago. More specifically, many of the models we have today capture more of the structure of various parts of the world, and in more detail, than models available fifty or a hundred years ago. For example, current models of the structure of genetic materials capture more of their real structure than models available in 1950. The primary task of a theory of science is to explain the processes that produced these results. To deny this minimal position, or even to be agnostic about it, is to misconceive the task. It is to retreat into scholasticism and academic irrelevance. If this is the consensus, it marks not the death, but the affirmation of a realist perspective. The good news would be that we could at least temporarily put to rest arguments about realism and get on with the primary task. That would be all to the good because the primary task is more exciting, and more important.* Ronald N. Giere Department of Philosophy and Center for Philosophy of Science University of Minnesota
[email protected] REFERENCES Burian, R. (1988). Review of Giere (1988). Isis 79, 689-91. Giere, R. N. (1988). Explaining Science: A Cognitive Approach. Chicago: University of Chicago Press. Giere, R. N. (1989). The Units of Analysis in Science Studies. In: S. Fuller, M. DeMey, T. Shinn, and S. Woolgar (eds.), The Cognitive Turn: Sociological and Psychological Perspectives on Science. Sociology of the Sciences, vol. 13. Dordrecht: Kluwer Academic. Knorr-Cetina, K. D. (1981). The Manufacture of Knowledge. Oxford: Pergamon Press. Kuhn, T. S. (1962). The Structure of Scientific Revolutions. 2nd edition: 1970. Chicago: University of Chicago Press. Lakatos, I. (1970). Falsification and the Methodology of Scientific Research Programmes. In: I. Lakatos and A. Musgrave (eds.), Criticism and the Growth of Knowledge, pp. 91-196. Cambridge: Cambridge University Press. Laudan, L. (1977). Progress and Its Problems. Berkeley: University of California Press. Putnam, H. (1981). Reason, Truth, and History. Cambridge: Cambridge University Press. Smith, C. and M. N. Wise (1988). Energy and Empire: A Biographical Study of Lord Kelvin. Cambridge: Cambridge University Press. Van Fraassen, B. C. (1980). The Scientific Image. Oxford: Clarendon Press.
⎯⎯⎯⎯⎯⎯⎯ *
The author gratefully acknowledges the support of the National Science Foundation and the hospitality of the Wissenschaftskolleg zu Berlin.
294
Ronald N. Giere
This page intentionally left blank
POZNAŃ STUDIES IN THE PHILOSOPHY OF THE SCIENCES AND THE HUMANITIES
Contents of Back Issues of the Idealization Subseries
VOLUME 1 (1975) No. 1 (Sold out) The Method of Humanistic Interpretation – J. Kmita, Humanistic Interpretation; W. Ławniczak, On a Systematized Interpretation of Works of Fine Arts; J. Topolski, Rational Explanation in History. The Method of Idealization – L. Nowak, Idealization: A Reconstruction of Marx’s Ideas; J. Brzeziński, Interaction, Essential Structure, Experiment; W. Patryas, An Analysis of the “Caeteris Paribus” Clause; I. Nowakowa, Idealization and the Problem of Correspondence. The Application: The Reconstruction of Some Marxist Theories – J. Topolski, Lenin’s Theory of History; J. Kmita, Marx’s Way of Explanation of Social Processes; A. Jasińska, L. Nowak, Foundations of Marx’s Theory of Class: A Reconstruction.
VOLUME 2 (1976) No. 3 (Sold out) Idealizational Concept of Science – L. Nowak, Essence – Idealization – Praxis. An Attempt at a Certain Interpretation of the Marxist Concept of Science; B. Tuchańska, Factor versus Magnitude; J. Brzeziński, Empirical Essentialist Procedures in Behavioral Inquiry; J. Brzeziński, J. Burbelka, A. Klawiter, K. Łastowski, S. Magala, L. Nowak, Law and Theory. A Contribution to the Idealizational Interpretation of Marxist Methodology; P. Chwalisz, P. Kowalik, L. Nowak, W. Patryas, M. Stefański, The Peculiarities of Practical Research. Discussions – T. Batóg, Concretization and Generalization; R. Zielińska, On Inter-Functional Concretization; L. Nowak, A Note on Simplicity; L. Witkowski, A Note on Implicational Concept of Correspondence.
VOLUME 16 (1990) IDEALIZATION I: GENERAL PROBLEMS (Edited by Jerzy Brzeziński, Francesco Coniglione, Theo A.F. Kuipers and Leszek Nowak) Introduction – I. Niiniluoto, Theories, Approximations, and Idealizations. Historical Studies – F. Coniglione, Abstraction and Idealization in Hegel and Marx; B. Hamminga, The Structure of Six Transformations in Marx’s Capital; A.G. de la Sienra, Marx’s Dialectical Method; J. Birner, Idealization and the Development of Capital Theory. Approaches to Idealization – L.J. Cohen, Idealization as a Form of Inductive Reasoning; C. Dilworth, Idealization and the Abstractive-Theoretical Model of Explanation; R. Harré, Idealization in Scientific Practice; L. Nowak, Abstracts Are Not Our Constructs. The Mental Constructs Are Abstracts. Idealization and Problems of the Philosophy of Science – M. Gaul, Models of Cognition or Models of Reality?; P.P. Kirschenmann, Heuristic Strategies: Another Look at Idealization and Concretization; T.A.F. Kuipers, Reduction of Laws and Theories; K. Paprzycka, Reduction and Correspondence in the Idealizational Approach to Science.
VOLUME 17 (1990) IDEALIZATION II: FORMS AND APPLICATIONS (Edited by Jerzy Brzeziński, Francesco Coniglione, Theo A.F. Kuipers and Leszek Nowak) Forms of Idealization – R. Zielińska, A Contribution to the Characteristic of Abstraction; A. Machowski, Significance: An Attempt at a Variational Interpretation; K. Łastowski, On Multi-Level Scientific Theories; A. Kupracz, Concretization and the Correction of Data; E. Hornowska, Certain Approach to Operationalization; I. Nowakowa, External and Internal Determinants of the Development of Science: Some Methodological Remarks. Idealization in Science – H. Rott, Approximation versus Idealization: The KeplerNewton Case; J. Such, The Idealizational Conception of Science and the Law of Universal Gravitation; G. Boscarino, Absolute Space and Idealization in Newton; M. Sachs, Space, Time and Motion in Einstein’s Theory of Relativity; J. Brzeziński, On Experimental Discovery of Essential Factors in Psychological Research; T. Maruszewski, On Some Elements of Science in Everyday Knowledge.
VOLUME 25 (1992) IDEALIZATION III: APPROXIMATION AND TRUTH (Edited by Jerzy Brzeziński and Leszek Nowak) Introduction – L. Nowak, The Idealizational Approach to Science: A Survey. On the Nature of Idealization – M. Kuokkanen and T. Tuomivaara, On the Structure of Idealizations; B. Hamminga, Idealization in the Practice and Methodology of Classical Economics: The Logical Struggle with Lemma’s and Undesired Theorems; R. Zielińska, The Threshold Generalization of the Idealizational Laws; A. Kupracz, Testing and Correspondence; K. Paprzycka, Why Do Idealizational Statements Apply to Reality? Idealization, Approximation, and Truth – T.A.F. Kuipers, Truth Approximation by Concretization; I. Nowakowa, Notion of Truth for Idealization; I. Nowakowa, L. Nowak, “Truth is a System”: An Explication; I. Nowakowa, The Idea of “Truth as
a Process.” An Explication; L. Nowak, On the Concept of Adequacy of Laws. An Idealizational Explication; M. Paprzycki, K. Paprzycka, Accuracy, Essentiality and Idealization. Discussions – J. Sójka, On the Origins of Idealization in the Social Experience; M. Paprzycki, K. Paprzycka, A Note on the Unitarian Explication of Idealization; I. Hanzel, The Pure Idealizational Law – The Inherent Law – The Inherent Idealizational Law.
VOLUME 26 (1992) IDEALIZATION IV: INTELLIGIBILITY IN SCIENCE (Edited by Craig Dilworth) C. Dilworth, Introduction: Idealization and Intelligibility in Science; E. Agazzi, Intelligibility, Understanding and Explanation in Science; H. Lauener, Transcendental Arguments Pragmatically Relativized: Accepted Norms (Conventions) as an A Priori Condition for any Form of Intelligibility; M. Paty, L’Endoreference d’une Science Formalisee de la Nature; B. d’Espagnat, De 1’Intelligibilite du Monde Physique; M. Artigas, Three Levels of Interaction between Science and Philosophy; J. Crompton, The Unity of Knowledge and Understanding in Science; G. Del Re, The Case for Finalism in Science; A. Cordero, Intelligibility and Quantum Theory; O. Costa de Beauregard, De Intelligibilite en Physiąue. Example: Relativite, Quanta, Correlations EPR; L. Fleischhacker, Mathematical Abstraction, Idealization and Intelligibility in Science; B. Ellis, Idealization in Science; P.T. Manicas, Intelligibility and Idealization: Marx and Weber; H. Lind, Intelligibility and Formal Models in Economics; U. Maki, On the Method of Isolation in Economics; C. Dilworth, R. Pyddoke, Principles, Facts and Theories in Economics; J.C. Graves, Intelligibility in Psychotherapy; R. Thom, The True, the False and the Insignificant or Landscaping the Logos.
VOLUME 34 (1994) Izabella Nowakowa IDEALIZATION V: THE DYNAMICS OF IDEALIZATIONS Introduction; Chapter I: Idealization and Theories of Correspondence; Chapter II: Dialectical Correspondence of Scientific Laws; Chapter III: Dialectical Correspondence in Science: Some Examples; Chapter IV: Dialectical Correspondence of Scientific Theories; Chapter V: Generalizations of the Rule of Correspondence; Chapter VI: Extensions of the Rule of Correspondence; Chapter VII: Correspondence and the Empirical Environment of a Theory; Chapter VIII: Some Methodological Problems of Dialectical Correspondence.
VOLUME 38 (1994) IDEALIZATION VI: IDEALIZATION IN ECONOMICS (Edited by Bert Hamminga and Neil B. De Marchi) Introduction – B. Hamminga, N. De Marchi, Preface; B. Hamminga, N. De Marchi, Idealization and the Defence of Economics: Notes toward a History. Part I: General Observations on Idealization in Economics – K.D. Hoover, Six
Queries about Idealization in an Empirical Context; B. Walliser, Three Generalization Processes for Economic Models; S. Cook, D. Hendry, The Theory of Reduction in Econometrics; M.C.W. Janssen, Economic Models and Their Applications; A.G. de la Sienra, Idealization and Empirical Adequacy in Economic Theory; I. Nowakowa, L. Nowak, On Correspondence between Economic Theories; U. Mäki, Isolation, Idealization and Truth in Economics. Part II: Case Studies of Idealization in Economics – N. Cartwright, Mill and Menger: Ideal Elements and Stable Tendencies; W. Balzer, Exchange Versus Influence: A Case of Idealization; K. Cools, B. Hamminga, T.A.F. Kuipers, Truth Approximation by Concretization in Capital Structure Theory; D.M. Hausman, Paul Samuelson as Dr. Frankenstein: When an Idealization Runs Amuck; H.A. Keuzenkamp, What if an Idealization is Problematic? The Case of the Homogeneity Condition in Consumer Demand; W. Diederich, Nowak on Explanation and Idealization in Marx’s Capital; G. Jorland, Idealization and Transformation; J. Birner, Idealizations and Theory Development in Economics. Some History and Logic of the Logic Discovery. Discussions – L. Nowak, The Idealizational Methodology and Economics. Replies to Diederich, Hoover, Janssen, Jorland and Mäki.
VOLUME 42 (1995) IDEALIZATION VII: IDEALIZATION, STRUCTURALISM, AND APPROXIMATION (Edited by Martti Kuokkanen) Idealization, Approximation and Counterfactuals in the Structuralist Framework – T.A.F. Kuipers, The Refined Structure of Theories; C.U. Moulines and R. Straub, Approximation and Idealization from the Structuralist Point of View; I.A. Kieseppä, A Note on the Structuralist Account of Approximation; C.U. Moulines and R. Straub, A Reply to Kieseppä; W. Balzer and G. Zoubek, Structuralist Aspects of Idealization; A. Ibarra and T. Mormann, Counterfactual Deformation and Idealization in a Structuralist Framework; I.A. Kieseppä, Assessing the Structuralist Theory of Verisimilitude. Idealization, Approximation and Theory Formation – L. Nowak, Remarks on the Nature of Galileo’s Methodological Revolution; I. Niiniluoto, Approximation in Applied Science; E. Heise, P. Gerjets and R. Westermann, Idealized Action Phases. A Concise Rubicon Theory; K.G. Troitzsch, Modelling, Simulation, and Structuralism; V. Rantala and T. Vadén, Idealization in Cognitive Science. A Study in Counterfactual Correspondence; M. Sintonen and M. Kiikeri, Idealization in Evolutionary Biology; T. Tuomivaara, On Idealizations in Ecology; M. Kuokkanen and M. Häyry, Early Utilitarianism and Its Idealizations from a Systematic Point of View. Idealization, Approximation and Measurement – R. Westermann, Measurement-Theoretical Idealizations and Empirical Research Practice; U. Konerding, Probability as an Idealization of Relative Frequency. A Case Study by Means of the BTL-Model; R. Suck and J. Wienöbst, The Empirical Claim of Probability Statements, Idealized Bernoulli Experiments and Their Approximate Version; P.J. Lahti, Idealizations in Quantum Theory of Measurement.
VOLUME 56 (1997) IDEALIZATION VIII: MODELLING IN PSYCHOLOGY (Edited by Jerzy Brzeziński, Bodo Krause and Tomasz Maruszewski) Part I: Philosophical and Methodological Problems of Cognition Process – J. Wane, Idealizing the Cartesian-Newtonian Paradigm as Reality: The Impact of New-Paradigm Physics on Psychological Theory; E. Hornowska, Operationalization of Psychological Magnitudes. Assumptions-Structure-Consequences; T. Bachmann, Creating Analogies – On Aspects of the Mapping Process between Knowledge Domains; H. Schaub, Modelling Action Regulation. Part II: The Structure of Ideal Learning Process – S. Ohlson, J.J. Jewett, Ideal Adaptive Agents and the Learning Curve; B. Krause, Towards a Theory of Cognitive Learning; B. Krause, U. Gauger, Learning and Use of Invariances: Experiments and Network Simulation; M. Friedrich, “Reaction Time” in the Neural Network Module ART 1. Part III: Control Processes in Memory – J. Tzelgov, V. Yehene, M. Naveh-Benjamin, From Memory to Automaticity and Vice Versa: On the Relation between Memory and Automaticity; H. Hagendorf, S. Fisher, B. Sá, The Function of Working Memory in Coordination of Mental Transformations; L. Nowak, On Common-Sense and (Para-)Idealization; I. Nowakowa, On the Problem of Induction. Toward an Idealizational Paraphrase.
VOLUME 63 (1998) IDEALIZATION IX: IDEALIZATION IN CONTEMPORARY PHYSICS (Edited by Niall Shanks) N. Shanks, Introduction; M. Bishop, An Epistemological Role for Thought Experiments; I. Nowak and L. Nowak, “Models” and “Experiments” as Homogeneous Families of Notions; S. French and J. Ladyman, A Semantic Perspective on Idealization in Quantum Mechanics; Ch. Liu, Decoherence and Idealization in Quantum Measurement; S. Hartmann, Idealization in Quantum Field Theory; R. F. Hendry, Models and Approximations in Quantum Chemistry; D. Howard, Astride the Divided Line: Platonism, Empiricism, and Einstein's Epistemological Opportunism; G. Gale, Idealization in Cosmology: A Case Study; A. Maidens, Idealization, Heuristics and the Principle of Equivalence; A. Rueger and D. Sharp, Idealization and Stability: A Perspective from Nonlinear Dynamics; D. L. Holt and R. G. Holt, Toward a Very Old Account of Rationality in Experiment: Occult Practices in Chaotic Sonoluminescence.
VOLUME 69 (2000) Izabella Nowakowa, Leszek Nowak IDEALIZATION X: THE RICHNESS OF IDEALIZATION Preface; Introduction – Science as a Caricature of Reality. Part I: THREE METHODOLOGICAL REVOLUTIONS – 1. The First Idealizational Revolution. Galileo’s-Newton’s Model of Free Fall; 2. The Second Idealizational Revolution. Darwin’s Theory of Natural Selection; 3. The Third Idealizational Revolution. Marx’s
Theory of Reproduction. Part II: THE METHOD OF IDEALIZATION – 4. The Idealizational Approach to Science: A New Survey; 5. On the Concept of Dialectical Correspondence; 6. On Inner Concretization. A Certain Generalization of the Notions of Concretization and Dialectical Correspondence; 7. Concretization in Qualitative Contexts; 8. Law and Theory: Some Expansions; 9. On Multiplicity of Idealization. Part III: EXPLANATIONS AND APPLICATIONS – 10. The Ontology of the Idealizational Theory; 11. Creativity in Theory-building; 12. Discovery and Correspondence; 13. The Problem of Induction. Toward an Idealizational Paraphrase; 14. “Model(s) and “Experiment(s). An Analysis of Two Homogeneous Families of Notions; 15. On Theories, Half-Theories, One-fourth-Theories, etc.; 16. On Explanation and Its Fallacies; 17. Testability and Fuzziness; 18. Constructing the Notion; 19. On Economic Modeling; 20. Ajdukiewicz, Chomsky and the Status of the Theory of Natural Language; 21. Historical Narration; 22. The Rational Legislator. Part IV: TRUTH AND IDEALIZATION – 23. A Notion of Truth for Idealization; 24. “Truth is a System”: An Explication; 25. On the Concept of Adequacy of Laws; 26. Approximation and the Two Ideas of Truth; 27. On the Historicity of Knowledge. Part V: A GENERALIZATION OF IDEALIZATION – 28. Abstracts Are Not Our Constructs. The Mental Constructs Are Abstracts; 29. Metaphors and Deformation; 30. Realism, Supra-Realism and Idealization. REFERENCES – I. Writings on Idealization; II. Other Writings.
VOLUME 82 (2004) IDEALIZATION XI: HISTORICAL STUDIES ON ABSTRACTION AND IDEALIZATION (Edited by Francesco Coniglione, Roberto Poli and Robin Rollinger) Preface. GENERAL PERSPECTIVES – I. Angelelli, Adventures of Abstraction; A. Bäck, What is Being qua Being?; F. Coniglione, Between Abstraction and Idealization: Scientific Practice and Philosophical Awareness. CASE STUDIES – D.P. Henry, Anselm on Abstracts; L. Spruit, Agent Intellect and Phantasms. On the Preliminaries of Peripatetic Abstraction; R.D. Rollinger, Hermann Lotze on Abstraction and Platonic Ideas; R. Poli, W.E. Johnson’s Determinable-Determinate Opposition and his Theory of Abstraction; M. van der Schaar, The Red of a Rose. On the Significance of Stout's Category of Abstract Particulars; C. Ortiz Hill, Abstraction and Idealization in Edmund Husserl and Georg Cantor prior to 1895; G.E. Rosado Haddock, Idealization in Mathematics: Husserl and Beyond; A. Klawiter, Why Did Husserl not Become the Galileo of the Science of Consciousness?; G. Camardi, Ideal Types and Scientific Theories.
This page intentionally left blank
Democracy and the Post-Totalitarian Experience. Edited by Leszek Koczanowicz and Beth J. Singer. Frederic R. Kellogg and Łukasz Nysler, Assistant Editors. Amsterdam/New York, NY 2005. XIV, 224 pp. (Value Inquiry Book Series 167) ISBN: 90-420-1635-3
€ 48,-/US $ 67.-
This book presents the work of Polish and American philosophers about Poland’s transition from Communist domination to democracy. Among their topics are nationalism, liberalism, law and justice, academic freedom, religion, fascism, and anti-Semitism. Beyond their insights into the ongoing situation in Poland, these essays have broader implications, inspiring reflection on dealing with needed social changes.
USA/Canada: One Rockefeller Plaza, Ste. 1420, New York, NY 10020, Tel. (212) 265-6360, Call toll-free (U.S. only) 1-800-225-3998, Fax (212) 265-6402 All other countries: Tijnmuiden 7, 1046 AK Amsterdam, The Netherlands. Tel. ++ 31 (0)20 611 48 21, Fax ++ 31 (0)20 447 29 79
[email protected] www.rodopi.nl Please note that the exchange rate is subject to fluctuations
Putting Peace into Practice
Evaluating Policy on Local and Global Levels Edited by Nancy Nyquist Potter Amsterdam/New York, NY 2004. XV, 197 pp. (Value Inquiry Book Series 164) ISBN: 90-420-1863-1
Paper
€ 42,-/US $ 55.-
This book examines the role and limits of policies in shaping attitudes and actions toward war, violence, and peace. Authors examine militaristic language and metaphor, effects of media violence on children, humanitarian intervention, sanctions, peacemaking, sex offender treatment programs, nationalism, cosmopolitanism, community, and political forgiveness to identify problem policies and develop better ones.
USA/Canada: One Rockefeller Plaza, Ste. 1420, New York, NY 10020, Tel. (212) 265-6360, Call toll-free (U.S. only) 1-800-225-3998, Fax (212) 265-6402 All other countries: Tijnmuiden 7, 1046 AK Amsterdam, The Netherlands. Tel. ++ 31 (0)20 611 48 21, Fax ++ 31 (0)20 447 29 79
[email protected] www.rodopi.nl Please note that the exchange rate is subject to fluctuations
Operation Barbarossa
Ideology and Ethics Against Human Dignity André Mineau
Amsterdam/New York, NY 2004. XIV, 244 pp. (Value Inquiry Book Series 161)
ISBN: 90-420-1633-7
€ 52,-/US$ 68.-
This book purports that, given Operation Barbarossa’s concept and scope, it would have been impossible without Nazi ideology, that we cannot understand it in the absence of its reference to the Holocaust. It asks and attempts to answer whether we can describe ideology without reference to ethics and speak about genocide while ignoring philosophy.
The VALUE INQUIRY BOOK SERIES (VIBS) is an international scholarly program, founded in 1992 by Robert Ginsberg, that publishes philosophical books in all areas of value inquiry, including social and political thought, ethics, applied philosophy, aesthetics, feminism, pragmatism, personalism, religious values, medical and health values, values in education, values in science and technology, humanistic psychology, cognitive science, formal axiology, history of philosophy, post-communist thought, peace theory, law and society, and theory of culture.
USA/Canada: One Rockefeller Plaza, Ste. 1420, New York, NY 10020, Tel. (212) 265-6360, Call toll-free (U.S. only) 1-800-225-3998, Fax (212) 265-6402 All other countries: Tijnmuiden 7, 1046 AK Amsterdam, The Netherlands. Tel. ++ 31 (0)20 611 48 21, Fax ++ 31 (0)20 447 29 79 www.rodopi.nl
[email protected] Please note that the exchange rate is subject to fluctuations