Acknowledgments The editors would like to acknowledge the support of NIH/NIA to this project (grants AG09282, AG04581, and AG12263). Also, we would like to thank the departments of psychology at Cleveland State University, The University of Northern Colorado, and The University of Am~erdam for their support of this project. Finally, we would be remiss in failing to acknowledge the contributors' fine work, and the technical assistance of Beth Goldstein (typesetting), Lida Allen (electronic information transfer), Scott Hayer (digital image processing), and Fabian Ferreri (Word for Windows advice).
vii PREFACE Component cognitive processes have played a critical role in the development of experimental aging research and theory in psychology. A quick perusal of the articles published in Journal of Gerontology: Psychological Sciences, Psychology and Aging, and ExperimentalAging Research will confirm this fact. However, in the last five to ten years, there has been a substantial increase in the number of articles attempting to isolate a single factor (or small subset of factors) responsa'ble for age differences in information processing. This view of aging is t~equently termed the complexity model or the generalized slowing model. The primary assumption in this view is that age differences in cognition are due simply to a relatively larger performance decrement on the part of older adults (compared to younger adults) as task complexity increases. Support for generalized slowing has come almost exclusively t~om the results of regression analyses of old and young response latencies. Because generalized complexity theorists have questioned the utility of using component cognitive processes as theoretical constructs, we feel it is time to re-state why component cognitive processes are critical to any t[lorough understanding of age differences in cognition. Thus, the present edited volume represents an attempt to demonstrate the utility of the process-specific approach to cognitive aging. Central to this effort are illustrations of how regression analyses may provide evidence for general slowing by maximizing explained variance while at the same time obscuring important local sources of variance. This volume concentrates on age differences in word and language processing, because these factors relate to reading, and reading is a critical cognitive process used in everyday life. Furthermore, age differences in word and language processing illustrate the importance of taking component cognitive processes into consideration. We feel, though, that the breadth of the coverage of the present book attests to the wide range of cognitive processes involved in word and language processing. In the spirit of cognitive science, this book is split into three different parts: 1) Traditional information processing approaches to age differences in word and language processing (Part I); 2) neuropsychological approaches (Part II); and 3) psychophysiological approaches (Part HI). Traditional information processing approaches use reaction time and accuracy as dependent variables and employ healthy younger and older adults as research participants. Neuropsychological approaches compare healthy aging to abnormal aging (e.g., Alzheimefs disease). Psychophysiological approaches can examine either healthy or abnormal subjects, but psychophysiological dependent variables such as event-related potentials are used in addition to reaction time or accuracy. We hope that this use of converging operations will provide a more complete picture of age differences in word and language processing than simply examining a single approach. Chapters within each section are organized on the basis of an information processing continuum ranging from word sensation and perception, to word cognition, and to language processes. Chapters consist of critical reviews, empirical reports, and position papers that are contemporarily relevant to cognitive aging research and theory. The first chapter (actually the first two chapters in Part I) in each Part consists of a tutorial in that particular area of the literature. We have taken special care to make this book as thorough, yet as readable, as possible. This book will be of interest to researchers and students of gerontology, cognitive psychology, neuropsychology, psychophysiology, cognitive science, and health fields.
ix
ADDRESSES OF SENIOR AUTHORS: Phil Allen, Ph.D. Dept. of Psychology Cleveland State University Euclid Ave. at East 24th St. Cleveland, Ohio 44115 Paul Amrhein, Ph.D. Department of Psychology University of New Mexico Albuquerque, NM 87131 Ted Bashore, Ph.D. Dept. of Psychology University of Northern Colorado Greeley, Colorado 80639
Richard Ferraro, Ph.D. Department of Psychology Box 7187 University of North Dakota Grand Forks, ND 58202-7187 Donald Fisher, Ph.D. 117 Amity St. Amherst, MA 01002 David Friedman, Ph.D. New York State Psychiatric Institute Cognitive Electrophysiology Laboratory 722 West 168th Street New York, NY 10032 Grover Cfilmore, Ph.D. Case Western Reserve University 10900 Euclid Ave. Cleveland, OH 44106-7123 Marilyn Hartman, Ph.D. Psychology, CB #3270 Davie Hall, University of North Carolina Chapel Hill, NC 27599-3270
x
George Kellas, Ph.D. Department of Psychology University of Kansas Lawrence, KS 66045 Johnathan King, Ph.D. Department of Cognitive Science University of California, San Diego D-015 La Jolla, CA 92093-0515 David Mitchell, Ph.D. Psychology Department SMU Dallas, TX 75275 Beth Ober, Ph.D. Applied Behavioral Sciences University of California, Davis Davis, CA 95616 Marian Patterson, Ph.D. 2520 Fairmount Blvd. Cleveland Heights, OH 44106 Richard Ridderinkhof~ Ph.D. Department of Psychology University of Am~erdam Reotersstraat 15 1018 WB Am~erdam The Netherlands Leann Stadtlander, Ph.D. Department of Psychology 300 Traphagen Hall Montana State University Bozeman, MT 59717 Elizabeth Stine, Ph.D. Department of Psychology Conant Hall University of New Hampshire Durham, NH 03824
Addresses of Senior Authors
Age Differences in Word and Language Processing Ph. Allen and Th.R. Bashore (Editors) 9 1995 Elsevier Science B.V. All rights reserved.
Why latent models are needed to test hypotheses about the slowing of word and language processes in older adults* Donald L. Fishera, Arthur D. Fiskb, and Susan A. Duffyc a U n i v e r s i t y of Massachusetts at Amherst
bGoorgia Institute of Technology Mount Holyoke College
1. INTRODUCTION Younger and older adults' performance has been compared on a number of different tasks. Many such tasks use response time as a dependent variable. Older subjects are typically slower than younger subjects. This suggests that the cognitive processes which mediate the behavior of older adults are slowed. A primary goal of studies of age-related slowing has been to determine whether all processes are slowed and, if not, which particular processes are slowed and which are spared. A knowledge of how much each process is slowed bears critically on the evaluation of the various theoretical attempts to relate the slowing of particular cognitive processes to a wide range of performance decrements (Salthouse, 1991) including deficits on tests of explicit and implicit memory (Howard and Wiggs, 1993), attention (Giambra, 1993; Madden and Plude, 1993), and intelligence (Hertzog, 1989; Schaie, 1989). This knowledge is critical because in most cases it is not known whether the particular processes thought to produce a given deficit are slowed or spared. A knowledge of how much each individual process is slowed also bears critically on the more practical attempts to design displays and interfaces for older adults which make it easier for them to perform the ordinary tasks involved in day to day living. Specifically, this knowledge makes it possible to focus on redesigning those aspects of the environment which reduce the slowing of the processes which are most affected by aging (Staplin and Fisk, 1991; Walker, Philbin and Fisk, 1994). Although knowledge of how much each individual process is slowed cannot be obtained with the techniques currently used in aging research, it is generally agreed that these techniques allow one to conclude that all processes are slowed at least some in older adults. However, the agreement ends here. In some studies, the slowing has been reported as constant across several different task domains. The conclusion is drawn that the cognitive processes are all slowed by one and the same function (Cerella, Poon and Williams, 1980; Myerson, Hale, Wagstait~ Pooh and Smith, 1990). In other more recent studies, the slowing is reported as constant within a given domain (e.g.,the lexical domain), but different across domains (e.g., the * Please send reprint requests to Donald L. Fisher, 114 Marston Hall, IEOR, University of Massachusetts, Amherst, MA 01003. Donald L. Fisher was supported during the writing of the chapter by a National Institutes ofHealth (NIA) Grant No. R01AG12461; Arthur D. Fisk was supported during the writing of the chapter by a National Institutes of Health (NIA) Grant No. R01AG07654.
2
D.L. Fisher et al.
lexical and nonlexical domains). The conclusion is drawn that the cognitive processes are slowed by different functions in the separate domains (Cerella, 1985; Lima, Hale and Myerson, 1091; Mayr and Kleigl, 1993). These conclusions rest on two primary and largely untested assumptions. First, it is assumed that older and younger adults process information identically in each of the tasks where their performance has been compared (the assumption of a structural equivalence). Second, it is assumed that if one function relates the observed response times of the older and younger adults in a given set of tasks, then one function relates the unobserved durations of the individual processes which mediate the behavior of the older and younger adults in the same set of tasks (the assumption of a functional equivalence). Although various investigators have acknowledged the importance of these two assmnptions for testing theories of cognitive slowing, they have not been directly examined. The violation of either or both of the above assumptions can lead to potential problems. Specifically, if the first assumption, the assumption of structural equivalence is violated, then the differences in the time it takes younger and older adults to perform various tasks may be due wholly to the use of different processes to perform a given task and not to prolongations of the durations of a common set of cognitive processes. Thus, potentially, none of the processes may be slowed. Psychologists now know that the strategies adults use to perform a given task can vary widely from one individual to the next (e.g., Hunt, 1978). Perhaps the existence of different strategies can explain why older adults are slower than younger adults, at least on some tasks. Problems still remain even if the first assumption is satisfied. Specifically, if the first assumption is satisfied and if the second assumption, the assumption of functional equivalence, is violated then the fact that older adults' response times are well fit by a single slowing function on a given set of tasks need not imply that the slowing of the cognitive processes is identical across processes. In fact, some processes might not be slowed at all. Nevertheless, investigators frequently make the decision to accept a model which assumes that all processes are slowed identically based on the finding that the older adults' response times are well fit by a single function of the younger adults' response times. Such a decision will be called a false positive decision if in fact the individual processes are slowed differentially. Just the opposite problem can arise. Specifically, the fact that older adults' response times are well fit by different slowing functions across separate sets of tasks need not imply that the slowing of the cognitive processes varies among processes. In fact, all processes might be slowed identically. Nevertheless, investigators typically make the decision to reject a model which assumes all processes are slowed identically based on the finding that older adults' response times are best fit by different functions of younger adults' response times across task domains. Such a decision will be called a false negative decision if the in fact the individual processes are slowed identically. In summary, if one or both of the above assumptions are not satisfied, then the techniques that currently are widely used in studies of age-related slowing can easily lead the investigator falsely to accept or reject the theory of general slowing. However, alternative techniques do exist for testing theories of cognitive slowing when these assmnptions are not satisfied. In particular, if the assumption of structural equivalence is violated, then techniques have been described which can be used to identify the exact structure of older and younger adults cognitive networks in a given task (Schweickert, 1978; Schweickert and Townsend, 1989; Steinberg, 1969; Townsend and Schweickert, 1989). Since a detailed overview of these techniques is now available (Schweickert, Fisher and Goldstein, 1994), we will not pursue these particular techniques further. If the assumption of functional equivalence is
Why latent models are needed to test hypotheses
3
violated, then techniques exist which can be used to derive the exact slowing of the individual cognitive processes from the overall response times. Elsewhere we have summarized these techniques (Fisher, 1994; Fisk, Fisher and Rogers, 1992; Fisk and Fisher, 1994). In the remainder of this chapter, we want to describe these latter techniques in more detail. Unlike current techniques, the techniques we will describe require that one have a detailed model of processing for the younger and older adults in each of the tasks on which their performance is compared. 2. GLOBAL AND LATENT MODELS In order to develop techniques which can be used to determine whether and by how much the component cognitive processes are slowed, we will argue that it is important, indeed necessary, to differentiate between latent and global models of slowing (Table 1). We will consider only two types of global models: global models of general slowing (Case I in Table 1) in which the slowing of older adults' response times remains constant across tasks (indexed by i) and domains (indexed by j) and global models of domain-specific slowing (Case II) in which the slowing of older adults' response times remains constant across tasks within a domain, but varies across domains. In both global models of slowing, it is assumed that the older adults' average response time Oo on task i in domain j is a function of the younger adults' average response time Yo on the same task plus some random error E 0 (E is the Greek uppercase epsilon). And in both models, the argument of the global slowing function is the response time of the younger adults. However, in the global models of domain-specific slowing the slowing function is indexed by the domainj whereas no index is needed for this function in the global models of general slowing. (Throughout, random variables will be represented by uppercase letters; constants and variables will be represented by lowercase italicized letters, labels for processes and other components of figures will be represented by lowercase nonitalicized letters) As with global models, we will consider only two types of latent models: latent models of general slowing (Case HI) in which the time on average it takes an older adult to complete each of the latent cognitive processes (indexed by k) is slowed by the same function and latent models of process-specific slowing (Case IV) in which the time on average it takes an older adult to complete each of the latent cognitive processes (also indexed by k) is slowed by a different function. In both of the latent models of slowing, it is assumed that the time on average Ak it takes an older adult to complete a particular latent cognitive process xk is some function of the time on average Ak it takes a younger adult to complete the same latent cognitive process xk. And in both models, the argument of the latent slowing function is the average time it takes younger adults to complete a particular latent process. However, in the latent models of process-specific slowing the slowing function is indexed by the process k. An example can make the above distinctions between global and latent models of slowing more transparent. In particular, consider simple lexical and nonlexical memory search tasks. Subjects are asked to memorize either a list of n words (the lexical task) or n digits (the nonlexical task). A probe word or digit is presented on each trial. Subjects must indicate whether the probe is or is not present in the memory set. Response time and accuracy are the dependent variables. Here, we want to predict the response times of the older adults.
4
D.L. Fisher et al.
First, consider global models of slowing. In such models, we take as a starting point for predicting the older adults' global response times the global response times of the younger Table 1 Models of Slowing I. Global Models of General Slowing I 0o = flYo) + Eo
II. Global Models of Domain-Specific Slowing 2 0o = fJ(Yo) + EO
III. Latent Models of General Slowing 3 Ak =J(Ak) IV. Latent Models of Process-Specific Slowing 4
Ak:A(Ak) 1 O0"is the time on average it takes older adults to respond on task i in domainj; Yo is the time on average it takes younger adults to respond on the same task; E0.is the error; fis the global general slowing function. 2 The global domain-specific slowingfunction is denoted 3 Akis the time on average that it takes older adults to complete latent processXk; Ak is the time on average that it takes younger adults to complete latent processXk; fis the latent general slowingfunction. 4 The latent process-specific slowingfunction is denotedJ~.
adults. Specifically, in a global model of general slowing (Case I), the time on average it takes the older adults to indicate that a probe is present at each memory set size for both the lexical and nonlexical probes can be written simply as: Oo
= f ( Y o ) + Eo .
(1)
Here, i indexes the task (memory set size), j indexes the domain (lexical or nonlexical). In a global model of domain-specific slowing (Case II), we would now need to index the slowing function by j, the domain, so that: Oo
:
f j ( Y o ) + E0.
(2)
There are many different global slowing functions we could substitute into the above equations. One of the simplest such functions (and perhaps the dominant function; see Cerella and Hale, in press) is s i l l y a constant, say 13, times the response time of the younger adults. Thus, we would rewrite Equation 1 as: Oo
= flYo + Eo
(3)
We will refer to the above model as the global multiplicative model of general slowing. Equation 2 is rewritten similarly with [3 now indexed by j. We will also talk about a global #near model of general slowing. Here, we simply add an intercept to the model:
Why latent models are needed to test hypotheses
E[OI
=
5
(4)
a + #lU.
Other global slowing functions have been proposed. For example, more complex power functions will sometimes explain the relation between the older and younger adults' global response times better than the simple multipfieative relations (Hale, Myerson and Wagstafl~ 1987; Myerson, Hale, Wagstafl~ Poon and Smith, 1990). However, the multiplieative model holds up extremely well by itself over a large range of response times (Cerella and Hale, in press). Next, consider latent models of slowing. In such models we take as a starting point for predicting older adults' response times the durations of the latent processes which govern the behavior of younger adults. Generally, in the task described above there are four latent processes: encoding, comparison, decision and response. Let E, C, D and R represent, respectively, the time on average it takes younger subjects to execute each of the above four processes. Assume that these processes are executed in series: subjects first encode the probe word, they then compare this probe with all stimuli in the memory set, they next decide whether the probe is or is not one of these stimuli, and finally they respond (Figure la). To keep the development as simple as possible, assume that the mean durations of these processes do not vary across tasks (memory set size) or domains (lexieal or nonlexieal) except for the mean duration of the comparison process which varies with the memory set size. Then, given that processing in this task is exhaustive (Steinberg, 1966), the younger adults' response times can now be written as a function of the durations of the various latent processes plus some error:
Yo
= E+
C, + D + R + Eo .
(5)
To keep the exposition a straightforward one, assume that the latent slowing function is linear. Then in a latent model of general slowing (Case III), the older adults' response times will depend on the slowing of each of the latent cognitive processes by the same ftmetionf Oo
= fiE) + f ( c , )
+ riD) + fiR) + E0 9
(6)
In a latent model of process-specific slowing (Case IV), the older adults' response times will depend on the slowing of each of the component processes by a different function: Oo
=
fe(E) + fc(C,)
+ f a ( D ) + f , ( / ~ + E0 9
(7)
It will be important to note that the latent and global models of general slowing can sometimes make identical predictions (Fisher, 1994). So, for example, assume that the slowing function was a multiplieative one and remained constant across the above four processes. Then, a latent model of general slowing can be rewritten as: Oo
= f(E+
C, + D + R ) + E0
= f(Yo) + Eo.
This is identical to the expression above for a global model of general slowing.
(8)
6
D.L. Fisher et al.
Encoding
Comparison
Comparison
Decision
Response
0
0 E
v
C.
C~
D
r
(a) p ~ ~ SOA
D1 P2
D2
R2
(b) Figure 1. a) Representation of the network of processes, encoding (e), comparison (c), decision (d) and response (r) processes, governing behavior in a memory scanning task. (Two items are in the memory set; scanning is exhaustive), b) Representation of the double stimulationtask. An understanding of the latent models can now be used to determine whether and by how much each of the latent cognitive processes is slowed. For example, suppose that we assume the four slowing functions in Equation 7 are multiplicative ones so that fe = eE, fc(Ci) = cCi, and so on. Then, we can determine the actual values of the multiplicative constants e, c, d and r. Additionally, we can determine whether these values differ significantly from one another. It follows that we can differentiate between latent models of general slowing and latent models of process-specific slowing and therefore avoid making the assumption of functional equivalence. We now want to describe the methods one uses to identify the separate slowing functions. The methods depend on the type or class of latent network which is used to represent processing in a given task. A number of different classes of latent models have been reported in the literature. We want to describe the two major such classes. In the first section below, we discuss PERT networks. PERT networks can be (and have been) used to represent processing within a broad range of laboratory tasks (Fisher and Goldstein, 1983; Townsend and Schweickert, 1989; Schweickert, 1978; Schweickert and Townsend, 1989). The simple serial and parallel networks described throughout the information processing literature are examples of such networks (Luce, 1986; Townsend and Ashby, 1983). We will discuss in detail the construction and testing of PERT network models of memory search and double stimulation tasks. PERT networks are a special case of a more general network, the Order-ofProcessing (OP) network, which can be used in situations where not all assumptions of the PERT network are satisfied (Fisher, 1985; Goldstein and Fisher, 1991, 1992; Schweickert and Fisher, 1987; Schweickert, Fisher and Goldstein, 1994). In the second section, we discuss interactive inhibition models (McClelland and Rumelhart, 1981). Interactive inhibition models have been used to model a wide range of language tasks among younger adults. Only recently has this modeling effort been extended to older adults (Bowles, 1990, 1993; Laver and Burke, 1993; Salthouse, 1988). We will discuss in detail the construction of an interactive inhibition model which can be used to predict the time it takes older and younger adults to name a target
Why latent models are needed to test hypotheses
7
word in a lexical naming task. Interactive inhibition models are a special case of connectionist networks (McClelland and Rumelhart, 1986). Unfortunately there is not time to discuss either these more general connectionist models or the more general OP networks referred to above. 3. MODELS OF SLOWING: PERT NETWORKS To begin, we want to describe very briefly the class of latent models based on a representation of processing as a PERT network. We mentioned above that simple serial and parallel networks were examples of PERT networks. More generally, a PERT network is a directed, acyclic (no cycles) graph with a single source (start node) and single sink (finish node). Examples are presented in Figure 1. The arcs represent the latent processes. A path is an unbroken sequence of arcs connecting two nodes. For example, in Figure lb the sequence of arcs, SOA, p:, d:, and rE, represents one of the two paths between the start and finish nodes. All nodes are A N D nodes. Thus, all processes which exit from a node cannot begin until all processes which terminate at the node have completed. For example, in Figure lb, process dE cannot begin until both processes d~ and p: have completed. Finally, all processes which can begin do begin. We want to use PERT networks to compute the time that it takes a subject to respond. Define the duration of a path as the sum of the durations of the processes which lie along the path. In Figure la, the response time is simply equal to the duration of the single path from the start to the finish node. In Figure lb, it may not be clear immediately from the definition of a PERT network how one should compute the response time. Recall that process dz cannot begin until both processes dl and p2 have completed (which depend, respectively, on processes pl and SOA completing). If we let P~ be a random variable representing the duration of process pl, D~ be a random variable representing the duration of process dl, and so on, then the time that it takes a subject to respond (measured from the moment the two processes at the source node begin executing) will be equal to the sum of the duration of the longest path between nodes 1 and 2 plus the duration of the path between nodes 2 and 3. Setting T equal to the response time, we obtain: 7'2
-- m a x i m u m { p ~
+ D~, SOA + P2} + 0 2 + R 2 .
(9)
Of course, we want to compute the average or expected response time, not the response time on a particular trial. We can do this analytically or computationally. Once we can predict the response times of the older and younger adults based on the underlying network of processes, we candetermine by just how much each of the latent processes is slowed in the older adults. We now want to compare this technique with current techniques using tasks where the current techniques have led either to false positive or false negative decisions. 3.1. False Positive Decisions
To begin, we compare the construction and evaluation of latent models of slowing with standard tests of slowing using a task where the existing methods of analysis might lead an investigator falsely to accept a model which assumes that all latent processes are slowed identically. The existing methods are confined primarily t o an analysis of the correlation coefficient obtained from the overall regression of the older adults' response times on the younger adults' response times. As Fisk, Fisher and Rogers (1992) note, there exists the real
8
D.L. Fisher et al.
possibility that an investigator will falsely accept a latent model of general slowing when the latent model is really one of process-specific slowing. For example, Fisk et al. construct a latent model of process-specific slowing where the proportionate slowing of older adults in one task is twice their proportionate slowing in a second task. Nevertheless, the global model of general slowing still explains 97.5% of the variability. In order to avoid this particular problem, investigators will sometimes look beyond the single measure of explained variability. For example, Myerson, Wagstaff and Hale (1994) note that a simple visual inspection of the points in a Brinley plot (a cross plot of younger and older adults' response times) can often uncover separate slowing functions across tasks. Specifically, it might well be the case that the points for one set of related tasks visually fell along one line, those for a second set of related tasks visually fell along a second line (Myerson et al., their Figure 2a). And, even when separate slowing functions are visually hidden in the original Brinley plot (Myerson et al., their Figure la), it may still be possible to identify the existence of separate regressions by examining the residual plots (Myerson et al., their Figure lb) since such plots often amplify patterns present in the original data. Examining both the Brinley and the residual plots is certainly important. But, it does not always suffice. Specifically, it is still necessary to construct and then test latent models of general and process-specific slowing. An example can make the point clear (Fisher, 1994). The Brinley plot in Figure 2a represents the results from a study reported by Salthouse and Somberg (1982). In that study, they asked younger and older adults to indicate whether a probe digit was or was not present in a memory set. Salthouse and Somberg varied the number of items in the memory set (1 or 4), the visibility of probe digit (degraded or intact), and the difficulty of the response (simple or complex). Thus, given that both younger and older adults participated in the study, there are a total of 32 conditions (i.e., 2 levels of memory set size x 2 levels of probe visibility x 2 levels of response difficulty x 2 levels of probe presence x 2 levels of age). A linear model of general slowing explains fully 98.2% of the variance (collapsing across the probe present and absent conditions). Visual inspection of the, Brinley plot in Figure 2a does not indicate to us any obvious problem with the model of general slowing. The differences between the predicted and observed response times in each condition are plotted in Figure 2b (computing these differences or residuals required estimating younger and older adults' observed response times in each condition from Figure la in Salthouse and Somberg since the response times were not directly reported in the article; the predicted older adults' response times could then be computed directly fromthe regression weights as reported). There is no clear pattern present in the residuals, that is, the residuals do not lie along two clearly different lines. Thus, there is nothing to suggest the existence of separate slowing fimctions in either the original Brinley plot (Figure 2a) or the plot of the residuals (Figure 2b). Yet, the model of general slowing can be rejected in favor of a model in which it is assumed that the rate parameters which govern the slowing of the latent cognitive processes are not identical (Fisher, 1994). Testing latent models of general and process-specific slowing entails a three step process. First, for the younger adults a latent or network model must be constructed. Second, for the older adults the latent model must be modified, the exact modification depending on the type of slowing-- general or process-specific. Finally, the joint (combined) fit of the younger adults' latent model and the older adults' latent model of general slowing must be compared
9
Why latent models are needed to test hypotheses
(a)
e
a
9
s
s
s
js sj
-J
~ p~ j s J
J
s Old - . ~ 1 9 ~ 1 1
"
(Young)
s
/ s s S jP sS /
f /! 91
,
I
i
.3
I
.S
9
I
.
!
.lr
,
J
!
,
1.1
!
1.3
,
!
11.S
.
I
l
1.7
( ~ ) Youl~
(b) 20( -II
15( -10( m
II
II
i
II
5C-" Residuals 0-"
i
-5s - -
i
-10( - II
-15( - 400
! 500
I 600
I 700
! 800
I 900
I I I i 1000 1100 1200 1300 1400
Y o u n g R e s p o n s e T i m e s (ms)
Figure 2. a) A Brinley plot of younger and older adults response times in the memory search task run by Salthouse and Somberg (1982). b) The residuals (older adults' observed response times minus older adults' predicted response times) plotted as a function of the younger adults response times. statistically with the joint fit of the younger adults' latent model and the older adults' latent model of process-specific slowing. A detailed review of the testing techniques is beyond the scope of the present chapter. However, given our very clear emphasis on the need to construct
10
D.L. Fisher et al.
and fit latent models, we will briefly present some of these details. We will use the memory search task to illustrate just how the process works. First, we need to construct a latent model of processing for the younger adults. Recall here that we assumed that performance in the memory search task could be modeled as a set of four processes arranged in series (Figure la). Thus, the latent model is a PERT network. The younger adults' response time on a given trial can then be written as the sum of the encoding, comparison, decision and response times [Equation 0]. To make the latent model useable, we need to predict the average or expected response times. As noted earlier in the chapter, this can be done either analytically or computationally. Here we take the analytic route since it is the more straightforward one. In particular, the average or expected response time E[Yo] is equal to the sum of the average or expected durations of the component processes, i.e., (10)
E[yo] = E[E] + E[C,] + E[D] + E[R] + E[Eo].
We assume that the expected error, E[E#], is equal to zero. So, we need to know only the otheffour expectations. The expectations in the sum will presumably vary with the condition. Table 2 Predicted Response Times of Younger Adults Predicted Response Time
Target Present
Probe Intact
Response Simple
E[Y] = tae d- tac -~- tad -~- tar
yes
yes
yes
E[Y] = ta~ + tac + tad + tar'
yes
yes
no
E[Y] = ta~,+ ta~ + tad + tar
yes
no
yes
E[Y] = ta~,+ ta~ + tad + tar'
yes
no
no
E[Y] : tae d- tac -~- tad' d- tar
no
yes
yes
E[Y] : tat d- tac d- tad' -~- tar'
no
yes
no
E[Y] : tae' -[- tac -~- tad' -}- tar
no
no
yes
E[Y] = ta~,+ ta~ + tad' + tar'
no
no
no
Specifically, the average encoding time should be longer when the probe is degraded [represented by tae,, i.e., E[E] = tae, in Equation 10] than when the probe is intact (ta~). The average comparison time when there is one stimulus in the memory set (tat) will be exactly four times shorter than this comparison time when four stimuli are in the memory set since the search is exhaustive. The average decision time will be longer when the probe is absent (tad') than when it is present (tad). And the average response time will be longer when the response is a complex one (tar') than when it is a simple one (tar). In this case there are a total of 16 conditions and 7 parameters needed to predict the expected response time of the younger adults. The predictions for all 8 conditions with a memory set size of one are listed in Table 2. The predictions for all 8 conditions with a memory set size of four follow immediately by
Why latent models are needed to test hypotheses
11
substituting 4~tc for ~tc in each equation in Table 2. Second, we need to construct the latent models of general and process-specific slowing for the older adults by modifying appropriately the latent model which describes the behavior of the younger adults. This is simple enough if we assume, not unreasonably, that the latent model of general slowing is a multiplicative one (Cerella and Hale, in press). In this case only one extra parameter is needed to describe the slowing of the older adults in the 16 conditions in which they participated. Thus, a total of 8 parameters are needed to fit jointly the latent model for the younger adults and the latent multiplicative model of general slowing for the older adults to the 32 observations for both groups (16 for the younger adults, 16 for the older adults). Similarly, if we assume that the latent model of process-specific slowing is a multiplicative one, then we need an additional four parameters, one slowing constant for each of the four processes. Thus, a total of 11 parameters are needed to fit jointly the latent model for younger adults and the latent multiplicative model of process-specific slowing for the older adults, again to the 32 observations for both groups. Third, we need to determine whether the joint fit of the latent model for the younger adults and the latent model of process-specific slowing for the older adults explains siL-,nificantly more variance than the joint fit of the latent model for the younger adults and the latent model of general slowing for the older adults. This requires estimating the parameters. In particular, values of the parameters are sought which minimize the error sum of squares, SSE, where SSE is equal to the sum over all 32 conditions of the square of the difference between the observations and predictions in each condition. Set SSE(PSS) equal to the error sum of squares for the process-specific model; set SSE(GS) equal to the error sum of squares for the general model. Set df(PSS) equal to the degrees of freedom for the process-specific model: dJ(PSS) = 32- 11. And set df(GS) equal to the degrees of freedom for the general slowing model: df(GS) = 32 - 8. Then, the statistic,
SSE(GS) - SSE(PSS)] SSE(PSS)]
(11)
has an F distribution with dJ(GS) - df(PSS) degrees of freedom in the numerator and df(PSS) degrees of freedom in the denominator if the model of general slowing is the correct one (assuming that the errors are normally distributed). If this statistic is significant at an appropriate level, then we can reject the model of general slowing. [Note that although this procedure could in principle be used to test the latent models of general and process-specific slowing that we just developed, the above procedure had to be modified so that it applied to the specific set of results reported by Salthouse and Somberg (1982) since they did not report the means in all 32 conditions. Fisher (1994) describes the additional modifications needed to fit their results. In fact, as noted above, when this modified approach is applied to the Salthouse and Somberg data, the latent model of general slowing must be rejected.] In summary, a global model of general slowing can appear to fit the results extremely well, both visually and statistically, as evidenced in the study of memory scanning reported by Salthouse and Somberg (1982). Yet, as Fisher (1994) argues, this does not rule out a latent model of process-specific slowing in which the processes are arranged in a PERT network. One clear way to rule out a latent model of process-specific slowing is to construct the latent
12
D.L. Fisher et al.
representation and then compare the fit of the latent models of general and process-specific slowing. We have described briefly how to do such above using the memory scanning task as an example. The general characteristics of the procedure would not change across tasks: construct a latent model for the younger adults; construct latent models of general and process-specific slowing for the older adults; and then test the two models. Of course, the details of the latent model will change with the task. This would be troublesome if latent models were extremely difficult to construct and test. However, procedures exist which make it relatively easy to do such for even complex latent models which can be represented as PERT networks (Schweickert and Townsend, 1989, Townsend and Schweickert, 1989).
3.2. False Negatives As noted earlier, a problem opposite to the one identified in the above section can arise when a global model of general slowing fails to fit the results across two or more domains. Here, the tendency is to want to reject the associated latent model of general slowing. However, caution is required when one or more of the domains contain tasks where there exist processes which influence the response time of a subject but are not under the direct control of the subject. Such processes, called exogenous processes by Fisher (1994; also see Fisk and Fisher, 1994), are frequently present in the tasks used to study the slowing of language skills (Bowles, 1993; Balota, Black and Cheney, 1992; Balota and Duchek, 1988; Burke, White and Diaz, 1987; Howard, Shaw and Heisey, 1986; Laver and Burke, 1993; Madden, 1989). For example, in a lexical decision task the target is not presented simultaneously with the prime but instead is delayed by some amount (the stimulus onset asynchrony or SOA). This delay is not under the control of the subject, but does influence the time it takes the subject to decide whether or not a word has been presented. Caution is required when attempting to fit a global model of general slowing to tasks which do and do not contain exogenous processes because one can (correctly) reject the global model of general slowing even though a latent model of general slowing is the one actually governing behavior, a situation we have labelled a false negative decision. We now want to compare the construction and evaluation of latent models of slowing with standard tests of slowing in cases where a false negative decision will occur if standard techniques are applied. We need both a task containing exogenous and a task not containing exogenous processes. In this section, the example we use for a task containing exogenous processes is a double stimulation task like that described by Pashler and Johnston (1989) since performance in such tasks is easily modeled by the PERT networks described here (in a later section, the example we use for a task containing exogenous processes is the one described above, the lexical naming task, since such tasks are well modeled by the connectionist networks described in that later section). The example we use for a task which does not contain exogenous processes is a task we have akeady described, the memory search task. For both tasks, we develop latent models of general slowing and predict older and younger adults response times. And for both tasks, we regress the predicted older adults' response times on the predicted younger adults' response times. We then compare the global slowing functions obtained from the above two regressions and show that these functions differ even though the latent model of slowing is a general one. Thus, we will have shown that the standard techniques can easily l"ad an investigator falsely to reject a latent model of general slowing.
Why latent models are needed to test hypotheses
13
To begin, we want to develop the latent model of general slowing in the double stimulation task, the task which contains an exogenous process. To do this, we need to describe the double stimulation task in more detail. In the double stimulation task, two stimuli are presented, one stimulus followed some short time laterby a second stimulus (here, as in the Icxicalnaming task, the time between the presentation of the two stimuli,the SOA, represents the exogenous process). A separate decision and response must be made to each stimulus (these could be Icxical decisions, though typically they are not such). The latent network governing processing in this task can be represented most simply in Figure Ib (the details are described in Schweickert, Fisher and Goldstein, 1994). It is assumed that the peripheral processing of the two stimuli can go on in parallel(pl and p2 in Figure Ib). It is assumed that the decision about the second stimulus (d2) cannot begin until the decision about the first stimulus (d~) has been made, a constraint consistent with the notion that the central processor has a very limited capacity (Broadbcnt, 1958). We now want to predict the average response times to the second stimulus (the only prediction of interest here since this response time is determined, in part, by the duration of the SOA or exogenous process). Above, we said that such predictions could be obtained either analytically or computationally. In the preceding example, we took the analytical approach, deriving closed form expressions for the average response time. Here, we will take the computational approach, estimating the average response time by using the computer to simulate processing. To begin, we need to make an assumption about the distribution of the various process durations and the length of the SOA. To keep things simple, we assume that the time it takes to complete both processes pl and dl is exponentially distributed with a mean of 100, that the time it takes to complete process p2 is exponentially distributed with a mean of 200, and that the time it takes to complete both processes d2 and r2 is exponentially distributed with a mean of 100. Furthermore, we assume that three SOAs are used, 50, 100 and 150. At each SOA, we then generate sample times from the distributions associated with the various processes or sums which appear in Equation 9. We then use a slight modification of this equation to compute the time T2 that it takes subjects to respond to the second stimulus on a particular trial. The modification is required because the response time to the second stimulus is measured from the moment that the second stimulus is presented (at the initiation of p2), not from moment that the first stimulus is presented (at the initiation of p~). Thus, in order to compute T2 We simply need to subtract SOA from the fight hand side of Equation 9. And we average over many trials to get an estimate of the expected response time. Doing such, we predict that the younger adults will take 352, 379 and 415 ms to respond to the second stimulus at, respectively, SOAs of 50, 100 and 150 ms. Now, to simulate general slowing, we increase the mean process duration by say 100% so that processes p~, d~, and r2 and d2 have a mean of 200 and p2 has a mean of 400. Again, averaging over many trials, we predict that the older adults will take, respectively, 683, 704 and 737 ms to respond to the second stimulus. If we regress the older adults' predicted response times on the younger adults' predicted response times, we find a correlation of 0.999 with a y intercept of 375 and a slope of 0.87. We have now calculated the parameters (the slope and intercept) which researchers have used to draw inferences about global slowing. Suppose that we now nm these same simulated subjects in a task inwhich there is no exogenous'process.such asthe~memory search task. Again, we assume that the mean process durations for the older subjects are slowed by 100%. We want to determine whether the slope and intercept remain unchanged in the new simulation. If the task can be represented as a PERT network and there are no exogenous
14
D.L. Fisher et al.
processes, as we are now assuming, then the latent model of general slowing reduces to a global model of general slowing (Fisher, 1994). Thus, in all conditions of the task, the predicted response times of the older adults will be twice as long as those of the younger adults since the mean process durations of the older adults are each slowed by 100%. Now, were we to:regress older adults' predicted response times on younger adults' predicted response times in a task without exogenous processes, we would find a correlation of 1.0, an intercept of 0.0, and a slope of 2.0. This result differs considerably from the above finding of an intercept of 375 and a slope of 0.87 when the task contains an exogenous process. As we have shown, however, this difference in the slope and intercept across the two tasks should not lead us to conclude that the two task domains are governed by different slowing functions. Intuitively, the reason for the different slopes in the tasks which do and do not contain exogenous processes can be made clear if one assumes that the durations of the processes in model of the double stimulation task are constants instead of random variables. If the ~SOA is relatively short, the combined duration of process pl and dl will exceed the combined duration of the SOA and process pz. Thus, the time it takes the younger adults to respond to the second stimulus is given by the sum, pl + d l + d2 + r2- SOA, whereas the time that it takes older adults to respond to this stimulus is given by the sum, ~(pl +dl + d2 + r2)- SOA, assuming that the slowing is a general one. However, when the SOA is relatively long, the combined duration of process pl and dl will be less than the combined duration of the SOA and process p2. Thus, the time it takes the younger adults to respond to the second stimulus is given by the sum, p2 + d2 + r2, whereas the time that it takes older adults to respond to this stimulus is given by the sum, ~(p2 + d2 + r2). Note although the latent slowing function, ~, is a general one, a global multiplicative model with the same slowing function does not fit the results across long and short SOAs for, if such were the case, the ratio of the younger to the older adults' response times when the SOA is short should equal this ratio when the SOA is long Clearly such is not the case, i.e., Pl .+ d l +
fl(p,+
d2 + r 2 - S O A
d, + d, + r, ) - S O A
P2 + d2 + r2 fl(P, + d, + r,)"
(12)
And, it is not the case because the duration of the exogenous process is not lengthened by the slowing function whereas the durations of the endogenous processes are each lengthened by the slowing function. In summary, as claimed, one global slowing function relates the response times of the older and younger adults in tasks with exogenous processes and a second global slowing function relates the response times of the older and younger adults in tasks without exogenous processes, even though the same latent slowing function 13relates the mean duration of each of the processes in the older adults' latent network to the mean duration of the associated process in the younger adults' latent network. Thus, there exists the real potential for falsely rejecting a latent model of general slowing if one inspects only the multiplicative slowing function derived from the fit of a global model of general slowing across different task domains when exogenous processes govern performance in one but not all domains.
Why latent models are needed to test hypotheses
15
3.3. Global Models of General and Domain-Specific Slowing Some readers may wonder at this point whether the above criticisms of global models of general slowing apply equally to global models of domain-specific slowing. After all, the fit of a global model of general slowing can be improved upon by considering the task domain (Cerella, 1985; Lima, Hale and Myerson, 1991; Mayr and Kliegl, 1993). For example, Lima et al. have recently reported a meta-analysis of both lexical and nonlexical tasks. The lexical tasks are defined simply as tasks that use words as stimuli; nonlexical tasks as those that do not use words. Lima et al. find that older adults are slowed significantly more in lexical tasks than they are in nonlexical tasks. Specifically, regressing older adults' responses times on younger adults response times, they find that within the lexical domain the slope is equal to 1,48 whereas within the nonlexical domain the slope is equal to 2.05. Cerella reports that older adults are slowed si~ificantly more in experimental tasks than they are in control tasks (CereUa notes that the definition of which was the experimental and which the control tasks was largely arbitrary and varied from study to study). Finally, Mayr and Kliegl find that older adults are slowed si~ificantly more in tasks which require subjects to coordinate several steps in the task (tasks which they define as high in coordinative complexity) than they are in tasks which do not require such coordination but do require the execution of multiple steps (tasks which they define as high in sequential complexity). Unfortunately, although it is true that global models of domain-specific slowing fit much better than do global models of general slowing, the same criticisms apply to both sets of models. Specifically, false positive decisions remain a real possibility because inspection of a Brinley plot and the associated residuals within a given domain is not necessarily any more instructive than inspection of a Brinley plot and the associated residuals across all domains. False negative decisions remain a real possibility if the test of separate slowing functions within a domain comes from the comparison of the slopes of global models of slowing fit to different sets of tasks within the domain where one set of tasks contains exogenous processes and the other does not. Again, the key to avoiding both false positive and false negative evaluations is the construction and testing of latent models of general and process-specific slowing. 4. MODELS OF SLOWING: CONNECTIONIST NETWORKS Not all behaviors can easily be modelled by PERT networks. In particular, behaviors whose outcome is dependent on the outcome of activity within the lexicon are frequently better described by the interactive inhibition (or more general connectionist) models first used extensively by McClelland and gumelhart (1981, 1986). Recently, connectionist models have been used to study the effects of aging, both generally (Salthouse, 1988) and in specific tasks such as naming (Bowles, 1993; Laver and Burke, 1993). Below, we want to develop a simple latent interactive inhibition model of slowing in a lexical naming task. We will show that unless one specifically tests the latent interactive inhibition model it is possible falsely to accept a latent model of general slowing when the correct model is one of process-specific slowing; or, alternatively, it is possible falsely to reject a latent model of general slowing when it is the correct model.
16
D.L. Fisher et al.
4.1 False Positive Decisions
Here we want to focus on the false positive decisions. That is, we want to show that a global model of general slowing can explain an overwhelming percentage of the variability even when the true model is a latent interactive inhibition model of process-specific slowing. The specific lexical naming task that we want to consider is the one run by Balota and Duchek (1988). Subjects were first given a prime word followed some brief time later by a target word. The time between the presentation of the prime and target words was varied from 200 to 800 ms. In the condition we will model, the prime either was highly related to the target or was a neutral prime (the word blank). The observed response times for younger and older adults in the high related (HR) and neutral (HN) conditions were estimated from Balota and Duchek (their Figure 1) at SOAs of 200, 500 and 800 ms (Table 3). Initially, suppose that we fit a global model of general slowing. Then, if we regress the response times of older adults on those of younger adults, we fmd: O = -201.31 + 1.63Y
(13)
where r 2 = . 9 5 . T h u s , it appears that a global model of general slowing does quite well. However, suppose that we go on to construct a very simple, latent interactive inhibition model of behavior in the lexical naming task for the younger adults and a latent interactive inhibition model of process-specific slowing in this same task for older adults. The network used to represent behavior in the task consists of several nodes and links between these nodes. To keep the discussion a simple one, assume that there is a node in semantic memory which codes the meaning of the prime (node 3 in Figure 3) and a node in semantic memory which codes the meaning of the target (node 4). Similarly, assume that there is an orthographic node (node 1) which codes the spelling of the prime and an orthographic node (node 2) which codes the spelling of the target. The activation level of each node will vary in time. We will assume that the activation ranges between 0 and 1 at each node. Let a~t) represent the activation of node i at time t. Set the activation of the semantic nodes and the target orthographic node to 0 at the start of the trial (when the prime is presented). Set the activation of the prime orthographic node to 1.0 at the start of the trial. Set the activation of the target orthographic node to 1.0 when it is presented. Assume that the prime orthographic node (node 1) is linked with strength ct31 to the prime semantic node (node 3), that the target orthographic node (node 2) is linked with strength ct42 to the target semantic node (node 4), and that the prime semantic node is linked with strength ct43 to the target semantic node. Finally, assume that the link strengths in our simple example take on values between 0 and 1. Now, we know the activation of all nodes at the start of a trial. We want to predict this activation at each point in time after a trial is initiated. We want to do such because we will assume that at some point the activation of the target semantic node will reach a level sufiiciently high that it causes the subject to initiate the naming of the target. We will use a very simple rule to predict the increase in activation at each node over some small interval of time fit. Specifically, the activation a~t + fit) of node i at time t + fit will be set equal to its activation a~t) at time t plus some fraction equal to 1 - a~t) of whatever activation has ~spread from the other nodes which are linked to it during the interval fit. The amount of activation which spreads from nodej to node i will be set equal to a fraction of the activation at node j determined by the link strength between node j and node i, i.e., equal to the product u~j(t). If we assume no decay, then throughout a trial al(t) = 1,
Why latent models are needed to test hypotheses
17
Table 3 Lexical Naming Latencies Observations I SOA
Older Adults
Younger Adults
al~ 2
aN 2
SP 2
I-~
aN
SP
200
674
678
4
537
539
2
500
643
660
17
514
526
12
800
627
645
18
510
523
13
Predictions SOA
Older Adults
Younger Adults
HR
HN
SP
HR
HN
SP
200
664
669
4
538
542
7
500
650
658
11
524
531
9
800
636
646
12
509
519
11
1Taken from Balota and Duchek (1988, Figure 1). 2HR: Highly Related Prime; HN: Neutral Prime; SP: Semantic Priming Effect (the difference between the neutral and highly related latencies).
before the target is presented, a2(t) = 0: after the target is presented a2(t) = 1. Thus, we need to compute only the activation a 3(0 and a4(t) at each of the semantic nodes. It follows from the above, that the activation at each of the semantic nodes is computed is follows: (14) Cl3 a .4.- (~0 --
a3 (O -[- a31Gl (O ( 1 -
a3 (O) ,
Computationally, one starts at time 0. A suitably small increment of time 8t is selected. The activation is computed at t = St. Next the activation is computed at 28t, and so on. Of course, we are not interested in the activation at each of the semantic nodes per se. Rather, we want to know how long it takes a subject to name the target. Let c represent the critical level of activation of the target node above which the subject decides to pronounce the target. A younger subject initiates a verbal response as soon as the activation of the target exceeds the critical activation, i.e., as soon as a4(t + fit) >_c. Set ty(SOA, related) equal to the shortest time t + 8t such that the above condition is met when the SOA is as indicated and the prime is related to the target. Let Vr represent the time it actually takes a subject to select and execute a
18
D.L. Fisher et al.
semantic prime (node 3)
semantic target (node 4)
verbal response (node 5)
(/,43
(/,54
(X31
(/,42
orthographic target (node 2)
orthographic prime (node 1)
Figure 3. Representation of the nodes in the conneetionist network which governs behavior in a lexieal naming tasks.
verbal response. This corresponds to the time that it takes the activation to flow from the target semantic node (node 4) to the decision and response node (node 5). Then, the younger adults' response time y o u n g ( S O A , r e l a t e d ) can be written as: young(SOA, related)
=
t y (SOA, r e l a t e d ) + v , .
(15)
As it stands, this prediction would not change across related and neutral conditions because we have not differentiated above between the link strengths in these two conditions. Of course, such a difference is expected. In particular, in the neutral condition, we cannot assume that the strength ct43(neutral) of the link between the neutral prime semantic node and the target semantic node is zero. However, we can assume that this strength is less than the strength of the same link between the related prime semantic node and the target semantic node. In summary, in order to model the younger adults naming latencies in the related and neutral conditions we need to estimate the link strengths r aaz, cx43(related) and tx43(neutral), the critical threshold c, and the response selection and execution time v,. In order to model the older adults' naming latencies in the related and neutral conditions we need to estimate the above parameters as well as whatever parameters reflect the aging process. We will assume the existence of only one parameter. In particular, we will assume that the activation between any two connected pair of nodes, except for the target semantic and response nodes, is spread less rapidly, being reduced by a factor of re. The exception corresponds to the assumption that the psychomotor portion of the response, whose duration is represented by Vr in the younger adults, remains unchanged in the older adults (Cerella, 1985). Then, we obtain:
Why latent models are needed to test hypotheses
19
(16) Cl3 a "[- ~ 0
=
Cl3 (O "}- I~ a 31C11(~)( 1 -
G 3 (O ) ,
Now, if for the older adults we set to equal to the shortest time such that a4(t + 8 0 >_c, then: old(SOA, related)
=
to(SOA, related) + vr 9
(17)
A similar equation holds in the neutral condition. At this point, we can fit the above model [Equations 15 and l7] to the results from Balota and Duchek (1988). We did not search the entire parameter space systematically since we obtained such a good fit with the initial parameter settings. Specifically, we make the predictions reported in the lower half of Table 3. These predictions correspond to the parameter values: a~l = .005,
a~(related)=
.00025,
a~(neutral)=
.00020,
(18) a,t2 = .005,
Jr = .78, c = .9.
Note that the latent model is a process-specific one since the duration vr of the response selection and execution processes is identical across changes in age whereas the time it takes the target node to reach the critical activation level c is not. This simple latent model of process-specific slowing explains fully 99% of the variance. Thus, we see again the potential for false positive decisions when only the fit of a global model of general slowing is examined. That is, we have shown that a latent model of process-specific slowing can better explain the results from a lexical naming task than can a global model of general slowing even though the global model of general slowing explains most of the variance (95%). Of course, we have only illustrated here the potential for a false positive decision. To show that the latent model of process-specific slowing that we have constructed fits significantly better than a latent model of general slowing, we would need to construct the latter model and then, as we did in the previous section, compare statistically the fit of the two latent models. We should point out that Balota and Duchek did not address specifically the issue of whether the latent slowing was a general or process-specific one. We use their results simply to illustrate the potential for false positive decisions.
4.2 False Negative Decisions We now need to discuss the potential for false negative decisions when connectionist models are used to describe performance both in tasks which do and in tasks which do not contain exogenous processes. To illustrate this potential, we follow the same procedure we used when discussing false negative decisions in cases where the underlying behavior was represented as a PERT network. Specifically, we construct a latent model of general slowing which can be used to predict performance in tasks with and without exogenous processes. We will then regress the predicted response times of the older adults on the predicted response times of the younger adults in each set of tasks and show that the slope differs across tasks even though the latent slowing function did not change.
20
D.L. Fisher et al.
To begin, consider a task with an exogenous process. In this case, we can use the above lexical naming task in the highly related condition. As noted there, the process representing the SOA is an exogenous process. However, now we want a latent model of general slowing, not a latent model of process-specific slowing. Here, we simply assume that the slowing in the spread of activation affects all links, including the link between the target semantic and response nodes. Set this slowing factor to ~ = .78, as above. To keep things simple, we arbitrarily set all link strengths equal to .005. Furthermore, as above, assume that activation does not spread from the target semantic node until its threshold is reached. Finally, assume that a response and decision have completed executing as soon as the critical threshold is reached at node 5, the same threshold as the target semantic node (c = .9). Then the times that it takes younger and older adults at each of the three SOAs to make a response in the highly related (HR) conditions are given in Table 4. Regressing the older adults' response times on the younger adults' response times, we find the correlation is 0.99, the intercept is 302.9 and the slope is .97. Next, consider a task with no exogenous processes. We imagine here a task like visual search or memory search task which requires the repetition of a process, the number of repetitions depending on the number of stimuli which are presented. In particular, assume that one, two or three nodes (processes) are in series. Assume that each node is a threshold node. Assume that the slowing reduction, link strengths and critical thresholds are set at the same values in this task as they were in the above task. The predicted times that it takes younger and older adults to complete one, two and three processes arranged in series are given in Table 4. If we now regress older adults' response times on the younger adults' response times, we find the correlation is 0.99, the intercept is 637.3 and the slope is 1.30. These parameters Table 4 Connectionist Model of Tasks With and Without Exogenous Processes Predictions Naming Latency
Search Latency
SOA
Younger Adults
Older Adults
Set Size
Younger Adults
Older Adults
200
950
1220
1
461
591
500
936
1206
2
973
1247
800
921
1192
3
1485
1963
differ from those we calculated for the task which involved exogenous processes. Thus, we see again that a latent model of general slowing can be the correct one even when a global model of general slowing can be rejected across different domains.
Why latent models are needed to test hypotheses
21
5. DISCUSSION We have spoken above of the methodological, theoretical and practical advantages which derive from using latent models of slowing. We now want to summarize and extend each of these advantages.
5.1 Methodological Importance To begin, we noted at the outset that investigators typically assume that if one function governs the global response times, then one function governs the latent process durations (what we called the assumption of functional equivalence). We have seen that there are two related violations of this assumption. Specifically, we have shown that a global model of general slowing can fit the results from existing experiments extremely well (e.g., Balota and Duchek, 1988; Salthouse and Somberg, 1982) even though a latent model of process-specific slowing is the correct one. Thus, false positive decisions remain a very real possibility. Additionally, we have shown that a global model of general slowing can fail to fit the results across different sets of tasks (i.e., tasks with and without exogenous processes; Lima et al., 1991) even though a latent model of general slowing is the correct one. Thus, false negative decisions remain a real possibility. The latent models of general and process-specific slowing that we constructed were drawn t~om the two major classes of information processing models, PERT networks and connectionist networks. We did this to emphasize the point that the methodological problems do not disappear as the models change. We also noted at the outset that investigators typically assume that the structure which governs the processing of older adults in a given task is identical to the structure which governs the processing of younger adults in the same task (what we referred to as the assumption of structural equivalence). We have ourselves made this assumption throughout. However, ultimately this assumption needs to be tested for each task using the recent techniques like those described by Schweickert (1978; Schweickert and Townsend, 1989; Schweickert, Fisher and Goldstein, 1994). If the assumption is violated, then it would be the case that a change in speed does not directly reflect a slowing of the processes used by the younger adults. For example, although not in the domain of aging, consider Logan's (1988) instance based theory of automaticity. After a great deal of practice, subjects change from a slow algorithmic retrieval to a fast memory-based retrieval. Here, the increase in speed is not produced by an increase in the rate with which the latent processes are executed but rather by a substitution of one set of processes which together take a long time to complete (though each individual process may be executed very quickly) with a second, different and perhaps smaller set of processes which together take less time to execute. Thus, the change in speed is a byproduct of the change in the retrieval mechanisms. In most cases, we do not have direct access to the durations of the latent processes or to the structure of the network and thus must derive both from the pattern of overall response times. However, recent research in cognitive neuroscience suggests that someday soon it may be possible to measure directly the duration of certain groups of cognitive processes. We want now briefly to comment on the relation between latent models of process-specific slowing and the recent research in the cognitive neurosciences (for an extensive commentary on this relation, the interested reader should consult Johnson and gybash, 1993). Our comments take as their starting point a recent review by Bashore (1993). He argues that the evidence from
22
D.L. Fisher et al.
cognitive psychophysiological studies indicates that the response end of processing slows more than the stimulus end. There are two key components to his argument. First, he considers a meta-analysis which used both P300 latencies and response times as the dependent variables in tasks which required only very simple motor responses (Bashore, Osman and Heffley, 1989). The global response times of the older adults were regressed on the global response times of the younger adults. A global multiplicative model of general slowing explained best the results from tasks which required only simple motor responses [[3 = 1.27 in Equation 0]. However, when the P300 latencies of the older adults were regressed on the P300 latencies of the younger adults, a global additive model of general slowing best explained the results [~ : 80.10 ms in Equation 0 with [3 set equal to 1]. The P300 latency is assumed to reflect the time that it takes subjects to evaluate and categorize stimuli quite independently of the time that it takes subjects to make a response. The time that it takes subjects to make a response is then given by the difference between the overall latency and the P300 latency (e.g., Ford, Roth, Mohs, Hopkins, and Kopell, 1979). When this difference for the older adults is regressed on this difference for the younger adults, a global linear model now best explained the results [a - 49.76 and 13= 1.32 in Equation 0]. In short, the meta-analysis ofBashore et al. suggests that the slowing function governing stimulus encoding and categorization is of one form, that governing response selection and organization of another form. The second set of studies reviewed by Bashore (1993) also implicates separate slowing functions in the somewhat more complex memory scanning tasks we described earlier in the article. In particular, a global additive function best explains the relation between the P300 latencies of the older and younger adults whereas a global linear function best explains the relation between the overall response time - P300 difference in the older adults and this same difference in the younger adults. In summary, the identification of the exact slowing of each of the latent processes can in principle avoid whatever problems attend the violation of the assumptions of structural and functional equivalence. In most cases, this identification can be done only indirectly However, at least in some cases, it appears possible to measure the durations of the latent processes themselves. In either case, the latent model of general slowing cannot be mistaken for a latent model of process-specific slowing, or conversely, if it is known by how much each of the latent processes is slowed. Similarly, the latent model describing the behavior of the older adults need not be assumed identical to the latent model describing the behavior of the younger adults if the network is tested explicitly.
5.2 Theoretical Importance In addition to their importance to methodology, the techniques we described for identifying the slowing of each of the latent processes have importance for the development of a unified theory of slowing. Such a unified theory was possible when there existed just one global model of general slowing. And, a unification appeared reachable when the global model of general slowing was replaced by separate models of domain-specific slowing. However, now these domains are multiplying several fold. And the global models of domain-specific slowing may need to be replaced by still more finely tuned global models of task-specific slowing. For example, Laver and Burke (1993) find the reduction in the semantic priming effect (obtained by subtracting related priming latencies ~om unrelated priming latencies) for older adults is not a proportional one. This runs counter to the meta-analysis reported by Lima et al. (1991) where it will be recalled that within the lexical domain a proportionate slowing of
Why latent models are needed to test hypotheses
23
approximately 50% was observed among older adults. Similarly, Fisk, Fisher and Rogers (1992), who reanalyze several conditions from Fisk and Rogers ( 1991), report that older adults are slowed significantly more in the early trials of memory search tasks which use disjoint sets of semantic categories as targets and distractors (called a consistent mapping or CM task, e.g., see Schneider and Shiffdn, 1977) than they are later in these tasks after considerable practice. Fisk and Rogers also found that this was not the case for visual search even though the same type of stimuli were to be searched by the subjects. Such findings are not consistent with the conclusion that one multiplicative factor governs the slowing of all lexical processes in the lexical domain. Contrariwise, Lima et al. (1991) find that evidence that one multiplicative factor probably does not govern slowing in the nonlexical domain. Specifically, they note that the magnitude of the proportional slowing of older adults in a n o n l e x i c a l picture naming study (Bowles, 1990; Thomas, Fozard and Waugh, 1977) resembles more closely the magnitude of the proportional slowing of older adults in lexical studies. Not only does it appear that recent results may require the introduction of separate global models of task-specific slowing, but a more detailed consideration of various tasks suggests that separate global models of condition-specific slowing within a given task may be needed. For example, tasks which measure the retention of skilled performance have recently been conducted within one of our labs (e.g., See Anderson-Garlach, 1994; Fisk, Cooper, Hertzog and Anderson-Garlach, 1994; Fisk, Hertzog, Lee, Rogers and Anderson-Gadach, 1994). We find that: (a) older and younger adults retain an impressive amount of skill even after 16 months without exposure to the task; (b) retention performance declines within a three month period and that decline remains stable between three and six months for both younger and older adults; (c) older and younger adults equally retain general, task-relevant skills; (d) older adults' performance declines more than younger adults' performance for both extensively trained and moderately trained stimuli; (e) when an interfering processing activity is inserted prior to the retention interval, older adults' performance declines disproportionately more than younger adults' performance especially when compared with a task not subjected to such interference; and (f) depending on the type of search task, for both younger and older adults the initial retention deficit is largely attenuated by the end of the retention retraining periods that were used. Since the slowing varies across conditions within the retention task, we can only conclude that still more detailed global models of condition-specific slowing will be needed. The proliferation of slowing functions across domains and now, perhaps, tasks and conditions within tasks is clearly not a desirable development if it is masking a more fundamental simplicity. Cerella (1985) points to the core of the problem with the global models and like approaches when he comments on the differential slowing of older adults in control and experimental tasks: "However important for aging theory, the designation experimental or control is psychologically somewhat arbitrary, because a given task may be introduced as either one or the other depending on the tasks with which it is paired in the context of the study (p. 78)." More generally, any definition of a set of domains, tasks within domains, or conditions within tasks which is based on the common external characteristics of the set (e.g., defining lexical tasks as simply tasks where words are used as stimuli) is problematic if what is ultimately desired is a definition of a set which is based on the common internal or latent cognitive processes operating in those tasks. In summary, there is a growing body of empirical evidence that if global models are pursued, the domains identified so far will need to be divided still more finally. The division
24
D.L. Fisher et al.
increasingly refers back to the latent cognitive processes. For example, Lima et al. (1991) suggest that the proportional slowing of older adults in a nonlexical picture naming study more clearly resembles the proportional slowing of older adults in lexical studies because the picture naming study relies heavily on word retrieval, a process which presumably is a lexical one. Thus, both the empirical evidence and the existence of real methodological problems indicate the importance of testing latent models of general and process-specific slowing.
5.3 Applied Potential A final and perhaps the most important, result of constructing and testing latent models of slowing is the opportunity it provides ultimately to assist older adults performing the various activities that make it possible to lead a relatively independent life. Many examples could be used to illustrate this point. Wewill choose as an illustration the development of a physical interface (e.g., a computer mouse) for an information system or an automatic teller machine which optimizes an older adult's interaction with the device. To produce the optimal design we need to answer questions like the following. Is it better to reduce the gain of the movement control system (the mouse) and if so, what reduction is optimal? Or, is it better to provide "sticky borders" such that when a pointing device gets close to an area, it is attracted and "stuck" to it and, if so, exactly how "sticky" should the borders be? Or are there still other solutions? Walker, Philbin and Fisk (1994) investigated various explanations for differences between younger and older adults in movement control tasks which might be used to answer the above questions. These explanations for age-related differences in movement control include the following: (a) older adults differ in the number and duration of the submovements they make within an overall movement; (b) older adults have similar control and movement structures but are flowed due to an inability to produce high levels of force; (c) older adults are more error aversive than younger adults, therefore generate flower, but more accurate movements; and (d)older adults have a higher noise-to-force ratio (i.e., the ratio of the standard deviation of the force to the force itself) than younger adults. The explanations Walker et al. offered took as their starting point the optimized submovement model (Meyer, Abrams, Komblum, Wright and Smith, 1988; Meyer, Smith, Komblum, Abrams, and Wright, 1990; Walker, Meyer and Smelcer, 1993). The optimized submovement model provides a framework for incorporating the different explanations of age-related differences in movement control. In order to test these various explanations, Walker et al. (1994) ran 16 older (75-70) and 16 younger (18-23) adults in an eight session experiment. The first four sessions were used to gather information about subjects' abilities, to familiarize the subjects with the use of the movement control device, and to familiarize the subjects with the payoff conditions for the speed and accuracy of movement. Also, during these sessions baseline movement times were established for each individual subject for use in manipulating payoff conditions during the remaining experimental sessions. The procedure required subjects to move a mouse-controlled cursor from a starting location to a target box on a video monitor. Subjects pressed a button on the mouse at the start of each trial and then released this button as soon as they felt confident that the cursor had entered the target box. Response time and accuracy were measured on each trial. Manipulation of the payoff structure successfully shifted each subject's relative speed-accuracy trade-off function across sessions. Using the latent optimized
Why latent models are needed to test hypotheses
25
submovement model to analyze the results, it can be determined that older adults have a higher noise-to-force ratio than younger adults. Thus, the older adults frequently require more submovements to position the cursor inside the target box. Reducing the rate of travel of the cursor with respect to the mouse could potentially decrease the total time it takes the older adults to position the cursor since the first submovement would more frequently land the cursor near the target box. The solution arises from an understanding of the underlying processing required for a given task. Obviously, an analysis of global models would be of little use here.
5.4 Summary In summary, many investigators have used regression models in which older adults' response times are predicted from younger adults' response times (what we have called global models) to draw inferences about the slowing of latent processes. In particular, investigators have sought to test the claim that all underlying processes are slowed by the same function or alternatively, that these processes are slowed by different fimctions. We have argued that these global models are not always adequate for testing such claims and that latent models of slowing should be considered. These latter models have three advantages over the currently popular regression models because the latent models, but not the regression models, allow the investigator directly to determine the slowing function for each process involved in a particular task. First, the investigator can avoid what we have termed false positive and false negative errors. Second, the investigator can identify which processes are slowed and which are spared. Theories of the relation between the slowing of particular processes and the production of specific performance decrements can then be more thoroughly tested. Finally, the investigator can identify which processes are slowed the most. The design of better environments for the elderly can then take place in a principled fashion. Arguing that it is important methodologically, theoretically and practically to construct and then test latent models of slowing is one thing. Actually doing such is quite another. In fact, we suspect that a great many readers may find somewhat cumbersome, if not daunting, the actual constructing and testing of models for each of the tasks that they run. Of course, for many tasks the models already exist. Such models have been reviewed in recent texts (Luce, 1986; Townsend and Ashby, 1983). To test the various theories of cognitive slowing, one needs only to modify the existing models by introducing into these models general or processspecific slowing parameters. However, there do exist many tasks, especially applied tasks, for which latent models of slowing still need to be constructed. Numerous techniques exist for identifying the structure of the latent models. The recent techniques described by Schweickert (1978; Schweickert and Townsend, 1989; Schweickert et al., 1994) are very powerful and, we believe, deserve more attention. A number of techniques exist for testing a latent model of slowing once its structure has been identified. We believe that several of the recently described techniques could prove very useful here (Fisher, 1986; Fisher and Goldstein, 1983; Goldstein and Fisher, 1991, 1992; Kliegl, Mayr and Krampe, 1994; Schweickert et al.). We hope that there will arise out of the detailed testing of latent models a much better understanding of just which cognitive processes are and are not slowed. And we hope that this understanding will bring some clarity given the increasing proliferation of global models of domain-, task-, and now condition-specific slowing. Whereas there are potentially an infinite number of domains,
26
D.L. Fisher et al.
tasks and conditions, it has been argued that there are only a finite number of elementary information processes (Newell and Simon, 1972, page 5). REFERENCES
Anderson-Garlach, M. M. and Fisk, A. D. (1994, April). Age-related retention of skilled performance: Within-subject examination of visual search, memory search, and lexical decision. Presented at the Fifth Cognitive Aging Conference, Atlanta. Balota, D. A., Black, S. and Cheney, M. (1992). Automatic and attentional priming in young and old adults: Reevaluation of the two-process model. Journal of Experimental Psychology: Human Perception and Performance, 18, 489-502. Balota, D. A. and Duchek, J. M. (1988). Age-related differences in lexical access, spreading activation, and simple pronunciation. Psychology and Aging, 3, 84-93. Bashore, T. 1L (1993). Differential effects of aging on neurocognitive functions subserving speeded mental processing (pp. 37-76). In J. Cerella, J. Rybash, W. Hoyer, and M. L. Commons (Eds.), Adults information processing: Limits on loss. San Diego: Academic Press. Bashore, T. 1L, Osman, A. and Heffley, E. F. (1989). Mental slowing in elderly persons: A cognitive psychophysiological analysis. Psychology and Aging, 4, 235-244. Bowles, N. L. (1993). Semantic processes serving picture naming (pp. 303-326). In J. CereUa, J. Rybash, W. Hoyer, and M. L. Commons (Eds.), Adults information processing: Limits on loss. San Diego: Academic Press. Bowles, N. L. (1990, April). Semantic activation in young and older adults during early processing in a primed picture naming task. Paper presented at the Cognitive Aging Conference, Atlanta, GA. Brinley, J. F. (1965). Cognitive sets, speed and accuracy of performance in the elderly. In A. T. Welford and J. E. Birren (Eds.), Behavior, aging and the nervous system. Charles C. Thomas, Springfield, IL. Broadbent, D. E. Perception and communication. Elmsford, NY: Pergamon Press. Burke, D., White, H. and Diaz, D. (1987). Semantic priming in young and older adults: Evidence for age constancy in automatic and attentional processes. Journal of Experimental Psychology: Human Perception and Performance, 1 13, 79-88. Cerella, J. (1985). Information processing rates in the elderly. Psychological Bulletin, 98, 6783. Cerella, J. and Hale, S. (in press). The rise and fall of information-processing rates over the fife-span. Acta Psychologica. Cerella, J., Poon, L. W. and Williams, D. M. (1980). Age and the complexity hypothesis. In L. W. Pooh (Ed.), Aging in the 1980s: Psychological issues (pp. 332-340). Washington, DC: American Psychological Association. Fisher, D. L. (1985). Network models of reaction time: The generalized OP diagram In G. d~dewalle (Ed.), Cognition, Information Processing and Motivation (Volume 3). Ammerdam: North-Holland Press, 229-254. Fisher, D. L. (1994). Cognitive aging: Models of general, task-specific and process-specific slowing. Submitted for publication, Psychological Bulletin.
Why latent models are needed to test hypotheses
27
Fisher, D. L. and Goldstein, W. M. (1983). Stochastic PERT networks as models of cognition: Derivation of the mean, variance and distribution of reaction time using orderof-Processing (OP) diagrams. Journal of Mathematical Psychology, 27, 121-151. Fisk, A. D. and Fisher, D. L. (1994). Brinley plots and theories of aging: The explicit, muddled and implicit debates. Journal of Gerontology: Psychological Sciences, 49, P81P89~ Fisk, A. D. and Rogers, W. A. (1991). Toward an understanding of age-related visual search effects, dournal of Experimental Psychology: General, 120, 131-149. Fisk, A. D., Fisher, D. L. and Rogers, W. A. (1992). General slowing alone cannot explain age-related search effects: A reply to Cerella (1991). Journal of Experimental Psychology: General, 121, 73-78. Fisk, A. D., Cooper, B. P., Hertzog, C. and Anderson-Garlach, M. M. (1994a, submitted). Age-related retention of skilled memory search: Examination of associative learning, interference, and task-specific skills. Journal of Gerontology: Psychological Sciences. Fisk, A. D., Hertzog, C., Lee, M. D., Rogers, W. A. and Anderson-Gadach, M. M. (1994b). Long-term retention of skilled visual search: Do young adults retain more than old adults? Psychology and Aging, 9, 206-215. Ford, J. M., Roth, W. T., Mobs, l~ C., Hopkins, W. F. and Kopell, B. S. (1979). Eventrelated potentials recorded from young and old adults during a memory retrieval task. Electroencephalography and Clinical Neurophysiology, 47, 450-459. Giambra, L. M. (1993). Sustained attention in older adults: performance and processes (pp. 259- 272). In J. Cerella, J. Rybash, W. Hoyer and M. L. Commons (Eds.), Adult information processing: Limits on loss. San Diego: Academic Press. Goldstein, W. M. and Fisher, D. L. (1992). Stochastic networks as models of cognition: Deriving predictions for resource constrained mental processing. Journal of Mathematical Psychology, 36, 129-145. Goldstein, W. M. and Fisher, D. L. (1991). Stochastic networks as models of cognition: Derivation of response time distributions using the Order-of-Processing method. Journal of Mathematical Psychology, 35(2), 214-241. Hale, S., Lima, S. D. and Myerson, J. (1991). Global cognitive slowing in the nonlexical domain: An experimental validation. Psychology andAging, 6, 512-521. Hale, S., Myerson, J. and Wagstafl~ D. (1987). General slowing of nonverbal information processing: Evidence for the power law. Journal of Gerontology, 42, 131-136. Hertzog, C. (1989). Influences of cognitive slowing on age differences in intelligence. Developmental Psychology, 5, 636-651. Howard, D. V. and Wiggs, C. L. (1993). Aging and learning: Insights l~om implicit and explicit tests (pp. 512- 528). In J. Cerella, J. Rybash, W. Hoyer and M. L. Commons (Eds.), Adult information processing: Limits on loss. San Diego: Academic Press. Howard, D. V., Shaw, 1L J. and Heisey, J. G. (1986). Aging and the time course of semantic activation. Journal of Gerontology, 41, 195-203. Hunt, E. (1978). Mechanisms of verbal ability. Psychological Review, 85, 109-130. Johnson, S. H. and Rybash, J. M. (1993). A cognitive neuroscience perspective on age-related slowing: Developmental changes in the functional architecture (pp. 143-173). In J. Cerella, J. Rybash, W. Hoyer, and M. L. Commons (Eds.), Adults information processing: Limits on loss. San Diego: Academic Press.
28
D.L. Fisher et al.
Kliegl, P~, Mayr, U. and Krampe, IZ T. (1994). Time-accuracy functions for determining process and person differences: An application to cognitive aging. Cognitive Psychology, 26, 134-164. Laver, G. D. and Burke, D. M. (1993). Why do semantic priming effects increase in old age? A recta-analysis. Psychology and Aging, 8, 34-43. Lima, S. D., Hale, S. and Myerson, J. (1991). How general is general slowing? Evidence from the lexical domain. Psychology and Aging, 6, 416-425. Logan, G. D. (1988). Toward an instance theory of automatization. Psychological Review, 95, 492-527. Luce, It D. (1986). Response times: Their role in inferring elementary mental organization. New York: Oxford University Press. Madden, D. J. (1989). Visual word identification and age-related slowing. Cognitive Development, 4, 1-29. Madden, D. J. and Plude, D. J. (1993). Selective preservation of selective attention (pp. 273 302). In J. Cerella, J. Rybash, W. Hoyer and M. L. Commons (Eds.), Adult information processing." Limits on loss. San Diego: Academic Press. Mayr, U. and Kleigl, 1L (1993). Sequential and coordinative complexity: Age-based processing limitations in figural transformations. Journal of Experimental Psychology: Learning, Memory and Cognition, 19, 1297-1320. McClelland, J. L. an Rumelhart, D. E. (1981). An interactive activation model of context effect in letter perception: Part 1. Psychological Review, 88, 375-407. McClelland, J. L. and Rumelhart, D. E. (1986). Explorations in parallel distributed processing: A handbook of models, programs and exercises. Cambridge: MIT Press. Meyer, D. E., Abrams, 1L A., Kornblum, S., Wright, C. E., and Smith, J. E. K. (1988). Optimality in human motor performance: Ideal control of rapid aimed movements. Psychological Review, 95, 340-375. Meyer, D. E., Smith, J. E. K., Kornblum, S., Abrams, 1L A. and Wright, C. E. (1990). Speedaccuracy tradeoffs in rapid aimed movements: Toward a theory of rapid voluntary action. In M. Jeannerod (Ed.), Attention and performance XIV. Hillsdale, NJ: Erlbaum. Myerson, J., Wagstaff~ D. and Hale, S. (1994). Brinley plots, explained variance, and the analysis of age differences in response latencies. Journal of Gerontology: Psychological Sciences, 49, P72-P80. Myerson, J., Hale, S., Wagstaff, D., Pooh, L. and Smith, G. (1990). The information loss model: A mathematical theory of age-related cognitive slowing. Psychological Review, 97, 475-487. Newell, A. and Simon, H. A. (1972). Human problem solving. Englewood Cliffs, NJ: Prentice-Hall. Pashler, H. and Johnston, J. C. (1989). Chronometric evidence for central postponement in temporally overlapping tasks. Quarterly Journal of Experimental Psychology, 41A, 19-45. Salthouse, T. A. (1988). Resource-reduction interpretations of cognitive aging. Developmental Review, 8, 238-272. Salthouse, T. A. (1991). Theoretical perspectives on cognitive aging. Hillsdale, NJ: Lawrence Erlbaunl Salthouse, T. A. and Somberg, B. L. (1982). Isolating the age deficit in speeded performance. Journal of Gerontology, 37, 349-357.
Why latent models are needed to test hypotheses
29
Schaie, K~ W. (1989). Perceptual speed in adulthood: Cross-sectional and longitudinal studies. Psychology and Aging, 4, 443-453. Schneider, W. and Shill'in, 1L (1977). Controlled and automatic human information processing: I. Detection, search and attention. Psychological Review, 84, 1-66. Schweickert, 1L (1978). A critical path generalization of the additive factor method: Analysis of a Stroop task. Journal of Mathematical Psychology, 18, 105-139. Schweickert, 1L and Fisher, D. L. (1987). Stochastic network models. In G. Salvendy (Ed.), Handbook of human factors. New York: Wiley. Schweickert, 1L and Townsend, J. T. (1989). A trichotomy: Interactions of factors prolonging sequential and concurrent mental processes in stochastic discrete mental (PERT) networks. Journal of Mathematical Psychology, 33, 328-347. Schweickert, 1L, Fisher, D. L. and Goldstein, W. M. (1994). General latent network theory: Structural and quantitative analysis of networks of cognitive processes. Revision requested, Psychological Review. Staplin, L. and Fisk, A. D. (1991). A cognitive engineering approach to improving signalized left turn intersections. Human Factors, 33, 559-571. Sternberg, S. (1966). High-speed scanning in human memory. Science, 153, 652-654. Sternberg, S. (1969). The discovery of processing stages: Extension of Donders' method. In W. G. Koster (Ed.), Attention and Performance//(pp. 276-315). Amsterdam: NorthHolland. Thomas, J. C., Fozard, J. L. and Waugh, N. C. (1977). Age-related differences in naming latency. American Journal of Psychology, 90, 499-509. Townsend, J. T. and Ashby, F. G. (1983). Stochastic modeling of elementary psychological processes. Cambridge: cambridge University Press. Townsend, J. T. and Schweickert, 1L (1989). Toward a trichotomy method of reaction times: Laying the foundation of stochastic mental networks. Journal of Mathematical Psychology, 33, 309-327. Walker, N., Meyer, D. E., and Smelter, J. B. (1993). Spatial and temporal characteristics of rapid cursor-positioning movements with electromechanical mice in human-computer interaction. Human Factors, 35, 431-458. Walker, N., Philbin, D. and Fisk, A. D. (1994, April). Optimization and movement control among older adults. Presented at the Fifth Cognitive Aging Conference, Atlanta.
30
Age Differences in Word and Language Processing Ph. Allen and Th.R. Bashore (Editors) 9 1995 Elsevier Science B.V. All rights reserved.
Visual w o r d encoding and the effect of adult age and w o r d frequency* Philip A. Allen a, David J. Madden b, and Steve Slane a aCleveland State University bCenter for the Study of Aging and Human Development, Duke University Medical Center
The visual processing of print is a fundamental human skill. Indeed, reading represents one of the primary forms of human communication. The first step in the process of reading-word encoding--is the topic of the present paper. We are particularly interested in the effect of increased adult age on visual word encoding. By word encoding, we mean the transduction of fight information from a word stimulus into a code that is then used for lexical access. This specialized form of pattern perception involves both the formation of an input code and lexical access. In order to address this issue of age differences in visual word encoding, we will first review the literature on basic word encoding (both the transduction and the lexical access components). After developing a general framework with which to conceptualize the processes of word encoding, we will then review theories of aging with regard to visual word recognition. Finally, the aging literature on visual word recognition and word frequency will be reviewed and integrated with general theories of word recognition and cognitive aging. There are two primary goals for this paper. First, we wanted to provide a relatively indepth review of the basic literature in visual word encoding and lexical access so as to familiarize aging researchers with this vast and daunting body of work. Secondly, after describing the component processes invOlved in visual word recognition, we wanted to determine if age affected these processes equivalently across task complexity, or whether certain processes were affected more by increased adult age than were others. This is equivalent to raising the issue of whether a single (presumably biological) process can account for age differences in information processing (e.g., Cerella, 1990, 1991; Myerson, Hale, Wagestaff~ Poon,& Smith, 1990; Salthouse, 1991), or whether local cognitive processes (i.e., different stages of processing) modulate age differences due to general biological processes (e.g., Allen, Madden, Weber, & Groth, 1993; Allen, Madden, Cerella, Jerge, & Betts, 1994; Amrhein & Theios, 1993; Balota & Ferraro, 1993; Bashore, Osman, & Hefley, 1989). 1. PART I: BASIC THEORIES OF WORD ENCODING Word recognition research in psychology has a long history (e.g., Cattell, 1886; Pillsbury, 1897). Nevertheless, over the past century, researchers have failed to resolve a fimdamental issue in word recognition. Namely, what is the basic unit of analysis in the pattern perception process in word recognition? Cattell (1886) found that subjects could identify component letters presented in a word context more readily than letters presented in a *This research was supported in part by NIH grant AG09282to the first author. Email: p.allen @ csuohio.edu
Visual word encoding and the effect of adult age and word frequency
31
nonword context when using tachistoscopic presentation (i.e., a word superiority effect). This finding led researchers to propose that the word encoding process was analytic in nature. That is, these researchers believed that words were formed from their component letters (e.g., Adams, 1979, McCldland & Rumelhart, 1981). We will term this view of word encoding "analytic." However, Pillsbury (1897) found for a reaction time (RT) task that words relative to nonwords tended to conceal their component letters (i.e., a word inferiority effect). These results led Pillsbury to propose that individual words were encoded as a single unit independent of letters. We will term this view of word encoding "holistic." (Ironically, Cattdl personally believed that the basic unit of analysis was the "total word picture," see Woodworth, 1938, even though more recent theorists interpreted his data as being supportive of an analytic basic unit of analysis, e.g., Krueger, 1989.) For the last century, experimental psychologists have been debating whether initial visual word encoding is analytic (letter-level encoding, e.g., Adams, 1979; Humphreys, Evett, & Quinlan, 1990; McClelland & gumdhart, 1981; Paap, Newsome, & Noel, 1984) or holistic (word-levd encoding, e.g., Johnson, 1975; Masson, 1986; Wheeler, 1970). Furthermore, there is evidence that under certain conditions individuals encode words as component syllables (e.g., Hansen & godgers, 1968; Lima & Pollatsek, 1983) or morphemes (e.g., Forster, 1976; Taft, 1979). Clearly, however, it is not possible for each of the three or four different proposed encoding schemes to be the single "basic unit of analysis" in visual word recognition. In order to resolve this dilemma, some word recognition theorists have proposed hybrid models (e.g., Allen & Madden, 1990; Allen & Emerson, 1991; Allen, Wallace & Weber, in press; Besner & Johnston, 1989; Carr & Pollatsek, 1985; Coltheart, Curtis, Atkins, & Hailer, 1993; Healy, Oliver, & McNamara, 1987; Healy, Conboy, & Drewnowski, 1987). These hybrid models can predict both a word superiority effect and a word inferiority effect under the appropriate task conditions (Allen & Emerson, 1991; Allen et al., in press; Besner & Johnston, 1989). However, the controversy concerning holistic versus analytic encoding is still not completely resolved, because hybrid models still tend to be either holistically (e.g., Allen et al., in press) or analytically (e.g., Besaer & Johnston, 1989) biased. Hybrid models of visual word recognition predict that word-levd (i.e., the word as a single unit), syllable-level (i.e., the word as component syllables), and letter-level (i.e., the word as component letters) representations of words are formed separately, but in parallel. These different levels of representation are assumed to be involved in a processing horse race to the central executive. Depending on the task (e.g., letter identification, lexical decision, or naming), a given level of representation may be optimal. For example, when subjects are required to identify letters within words or nonwords (a letter identification task, Allen & Madden, 1990), the letter-level code is optimal, because word-levd and syllable-levd codes tend to conceal component letters (Allen & Emerson, 1991). Alternatively, the word-levd channel tends to be optimal for a lexical decision task (a task in which subjects are required to determine whether a letter string forms a real word, or not--see Allen et al., in press), and the syllable-level channel tends to be optimal for a naming task (a task in which subjects are required to pronounce letter strings, see Allen, Madden, Cerella, Jerge, & Betts, in press). It should be noted, though, that word-levd and letter-levd input codes are traditionally thought to be orthographic in nature (e.g., Allen & Madden, 1990; Healy et al., 1987), whereas the syllable-level input code is frequently assumed to be phonological in nature (e.g., Coltheart et al., 1993). Thus, in many hybrid models of visual word recognition, there is a phonological channel in which codes are formed using grapheme-to-phoneme correspondence rules (GPC
32
Ph.A. Allen et al.
rules, e.g., Coltheart, Davelaar, Jonasson, & Besaer, 1977) and there is at least one orthographic channel (letter-level and/or word-level). An extended version of this "dua! route" model is the hybrid model of Allen et al. (in press; Allen et al., 1994). This model contains both orthographic and phonological pathways. Furthermore, both pathways receive input from up to three different input channels (or modules, see Fodor, 1983): word-level, syllable-level, and letter-level (refer to Figure 1). A more detailed discussion of these "dual-route" models will be presented in the upcoming section on lexical access.
Word-Level ~ Orthographic
Pathway
Letter-Level ~
Word
Identification Superposition 1
(Addressed)
Syllable-Level ~ VisualInput Phonological Pathway
Syllable-Level ~
Pronunciation
Letter-Level ~
Pronunciation t
(Assembled)
Figure1 All of the models of visual word recognition discussed so far are nile-based, "nondistributed" models. That is, they are really information processing variants of a "Turing machine" (Turing, 1936). A Turing machine is a symbol-manipulating system (Forster, 1989). The goal of a Turing machine is to emulate human information processing with a system (e.g., algorithms) consisting of a relatively small number of properties and capabilities. One important aspect of a Turing machine is that the state of the machine is known at any given point in time. However, another fundamentally different form of information processing architecture that is becoming increasing more prevalent today is the connectionist, or parallel distributed processing (PDP) framework (e.g., Rumelhart & McClelland, 1986; McClelland & Rumelhart, 1986). While Turing machine-based architectures are based upon symbolic, rule-based representations, PDP architectures are based upon sub-symbolic models in which information is represented as a pattern of activation across a set of processing "nodes." These nodes can be conceptualized as neurons, thus, PDP models really contain theoretical entities (i.e., nodes) that resemble the "cell assemblies" referred to by Hebb (1949). Hence, although it is still a controversial issue as to whether PDP architectures will perform visual word recognition as well as rule-based models (e.g., Besner, Twilley, McCann, Seergobin, 1990; Joordens & Besner, 1994), one positive aspect of PDP models is that they do appear to mimic biological
Visual word encoding and the effect of adult age and word frequency
33
brain processes more closely than do rule-based models. Because there are hidden layers in PDP models, though, one cannot be sure of the state of the machine at any given point in time.
WORDS
SYLLABLES
LETTERS
FEATURES
IMAGE Figure 2
The interactive activation model (IAM) of McClelland and Rumelhart (1981) and the developmental model of Seidenberg and McClelland (1989) are two examples of PDP models of visual word recognition (although the IAM uses local lexical representation, so this model is not a truly "distributed" PDP model). Ironically, although these models include the term "parallel" in their names, neither of the models use parallel encoding schemes. That is, the IAM always encodes words as component letters (McClelland & Rumelhart, 1981), and the developmental model always encodes words as "Wickelgraphs" (a Wickelgraph is a three-letter sequence, or triplet, Seidenberg & McClelland, 1989). Thus, these models do not use multiple levels of encoding initially, and this design characteristic can be problematic (see Allen et al., in press). However, there is no intrinsic reason that a PDP model cannot encode words using multiple input channels. The PDP models are parallel in the sense that letter-level activation feeds forward to the word level before the letter level has completed processing (i.e., a cascaded system, see Figure 2). The feedback between the word and letter levels is the interactive, or connectionist, aspect of the models. This allows the word level to affect processing at the lower tier of the letter level. For example, this sort of model predicts a word superiority effect, because the word level would facilitate processing at the letter level, however, nonwords (especially
34
Ph.A. Allen et al.
random letter strings) would not provide this top-down facilitation of the letter level. It is unclear, though, how the aforementioned PDP models could predict a word inferiority effect. One seemingly simple solution would be to develop a PDP model in which there were separate letter-level and word-level (and, perhaps, syllable-level) input channels that interacted horizontally, but such a model has yet to be formalized. 1.1. The Nature of the Functional Stimulus
We have noted so far that even though more sophisticated hybrid nile-based modds and interactive PDP models have been formulated to explain visual word recognition, it is still unclear how individuals actually visually encode words during reading. Allen et al. (in press) have argued that from a systems design standpoint, holistic encoding is inherently more efficient than analytic encoding because holistic encoding requires less processing capacity. That is, if one encodes a six-letter word holistically, then one represents the word as a single unit (or one total word picture). However, if one encodes a six-letter word analytically, then one represents a word as six separate units (or letter pictures), and this would be equivalent to encoding six objects instead of one. Furthermore, the letter units are smaller, and therefore, they require a more detailed "grain" of processing resolution than required by a single wordlevel unit. Thus, it would seem to be more efficient to encode words holistically (see Allen et al., in press). Holistic models have frequently been assumed to use template matching as a pattern recognition mechanism (a replica of the actual visual stimulus is stored in memory). That is, words act like "pictures" that directly access the mental lexicon in order to be reco~ized or identified. The inherent problem with this approach, though, is that there is considerable stimulus variability in patterns (in this case, words) that are reco~aized as being identical (Neisser, 1967). For example, we recognize "word," "WORD," and "wOrD" as being the same word. To overcome this problem, some researchers have proposed that individuals use a small subset of visual features to recognize visual objects (e.g., Gibson, 1965; SelfiJdge, 1959). However, such models (e.g., the "Pandemonium" model of Selfridge, 1959) have limited generalizability in a manner not unlike template matching models. That is, the Pandemonium model can recognize component letters, but the model cannot recognize objects in a painting. Furthermore, in is unclear how models of this type would predict empirically obtained results for a mixed-case disadvantage for visual word recognition (when it takes longer to reco~ize a word presented in mixed case than in consistent case on a lexical decision or naming task). The Pandemonium model cannot parsimoniously predict a mixed-case disadvantage that varies as a function of exposure duration and lexicality (i.e., word vs. nonword) because it has a single analytic input channel (Allen et al., in press). A more general method of recognizing a wide variety of objects is termed "identification by components" (Biederman, 1987). This approach uses a primitive set of visual, geometric components to account for all possible objects (Biederman, 1987; Marr, 1982). These simple visual geometric properties tend to remain constant in both static and dynamic perceptual processing. The geometric components are all variations of a cylinder, and Biederman (1987) estimated that selection from a set of no more than 36 different geons (using three-dimensional processing) could be used to form all known objects. The identification by components approach, though, would still have difficulty in predicting a mixed-case
Visual word encoding and the effect of adult age and word frequency
35
disadvantage unless geons were used as building blocks for multiple input channels (e.g., Allen & Emerson, 1991). It is quite likely that PDP theorists used the theoretical construct of visual features to form letters, because they felt that this avoided the potential problems inherent in some template matching models--which are that they require an unlimited supply of templates, and they have extremely limited generalizability. Note that models of reading that use letters to form words need only to be able to recognize the 26 different letters of the alphabet in order to recognize every possible English word. Unfortunately, as noted earlier, constraining models to recognize words using only component letters slows down the potential encoding speed of words, and also makes such models (e.g., the IAM ofMcClelland & Rumelhart, 1981) unable to account for data in a straightforward manner in which a word context inhibits the recognition of component letters relative to nonwords (i.e., a word inferiority effect, e.g., Allen & Emerson, 1991; Healy et al., 1987). An alternative approach to visual pattern recognition other than identification-bycomponents is termed "model-based identification" (e.g., Brooks, 1981). Applying this general information processing architecture to visual word recognition, word stimuli are encoded simultaneously as both words (holistically) and letters (analytically) using an encoding process termed "normalization" (e.g., Allen, Wallace, & Weber, in press). By doing so, the modelbased identification can overcome much of the bottleneck problems of models that form words solely from component letters. By normalization we mean a "transformation which would give the same output for every member of certain well-defined categories" (Neisser, 1967, p. 63). By applying normalization to visual word recognition, we need at least two input channels-word level and letter level--and we need to assume that input channels encode letters and words using the spatial frequency pattern of the stimuli (although the latter assumption is not required, the model could theoretically encode stimuli as visual features rather than spatial frequency patterns). The inpm channels then normalize slight variations in the spatial frequency pattern of word-consistent font type or case type so that the same code is formed for "word" and "WORD." This can be accomplished by using a normalization algorithm such as Fourier synthesis (Allen & Emerson, 1991; Allen, Wallace, & Weber, in press). Although Fourier synthesis uses a linear algorithm and the human visual system is clearly non-linear at times, it appears that Fourier synthesis is still the best norma~ation option available to visual recognition researchers at the present time (Graham, 1981). It should be noted that using a model-based approach (spatial frequency encoding) rather than an identification-by-components approach (encoding by visual features) to visual word encoding is a marked departure from the models of pattern perception that are typically used in psychology. This is probably because of the influence of Mart (e.g., 1982) who persuasively argued for the visual feature detector approach. However, more recent evidence suggests that the brain initially uses spatial frequency filtering to encode visual information (see Van Essen, Anderson, & Felleman, 1992). It now appears that what researchers think of as feature detection occurs much later in neural processing than does spatial filtering (Van Essen et al., 1992). In both models of visual encoding (Mart, 1982, and Van Essen et al., 1992), it is important to keep in mind that this process involves multiple stages. Initially, visual stimuli are encoded as a set of primitive features (Mart, 1982) or broad spatial frequency tuning (Derrington & Lennie, 1984; DeValois & DeValois, 1987; Van Essen et al., 1992). From a psychological perspective, we might term this first stage of encoding the detection stage. After
36
Ph.A. Allen et al.
detection, the second stage of visual encoding involves interpretation of the "blob-level" representation (either by edge detection, Mart, 1982, or by spatial frequency filtering, DeValois & DeValois, 1987). For a discussion of the neural pathways involved in visual word recognition, see Cart (1992), Cart and Posner (1992), or Allen, Wallace, and Weber (in press). 1.2. A Hybrid Horse Race Model
A model based upon the spatial filtering design principles mentioned in the preceding paragraphs has been proposed by Allen and Emerson (1991) and Allen et al. (in press). In the most basic form, the model assumes the existence of two orthographic input channels used for visual word recognition/identification. Specifically, this hybrid model uses a fast word-level input channel and a typically slower letter-level channel to recognize and identify words. These input channels (or modules, Fodor, 1983) are involved in a stochastic (i.e., probabilistic) horse race to the central processor. The word-level channel encodes words using the spatial frequency pattern of entire words (holistic encoding) and the letter-level channel encodes words using the spatial l~equency pattern of component letters (analytic channel). In this sort of encoding architecture, the words "word" and "WORD" are actually normalized to be recognized as the concept "word." However, Fourier synthesis would not be successful in forming a holistic code of "word" from "wOrD" because mixing case results in a pattern that cannot be normalized by Fourier synthesis (Allen et al., in press). Thus, words presented in mixed case must be processed using the typically slower letter-level input channel. The empirical finding of a mixed-case disadvantage for words on a lexical decision task (e.g., Allen, Madden, Weber, & Groth, 1993; Allen et al., in press) or on a naming task (Allen, Madden, Cerella, Jerge, & Betts, 1994) is consistent with the aforementioned hypothesis that consistentcase words tend to be encoded first by the word-level channel (assuming that they are familiar), but that mixed-case words t a d to be encoded by the letter-level channel. The hybrid model (Allen et al., in press) further assumes that the word-level input channel accesses the holistic-semantic lexicon. (The holistic-semantic lexicon is actually a submodule of the word-level input channel.) In this manner, the model predicts word t~equency advantages for words processed using the word-level channel. However, the hybrid model recognizes unfamiliar words by using the typically slower letter-level channel, because the spatial frequency pattern of a new word would be definitionally unfamiliar (thereby preventing the word-level channel from producing a code). Since all English words can be formed from the component spatial frequency patterns of the 26 letters of the English alphabet (albeit at a slower rate than encoding words using holistic methods), the model can create new entries in the mental lexicon through the use of "superposition." That is, the central processor "superposes" the letter-level code into a pseudo-word-level code. The model assumes that once an individual has encountered a new word a sufficient number of times, the letter-level representation of that word is "superposed" into a word-level template so that this word can now be directly recognized holistically (Allen & Emerson, 1991). It should be noted, though, that the superposition process is rather time consuming and requires a considerable allocation of processing resources. Therefore, lexical access using a superposition process is only used when such access cannot be accomplished via the word-level input channel (Allen & Emerson, 1991). In this manner, the hybrid model can account for how children and adults learn to read new words holistically.
Visual word encoding and the effect of adult age and word frequency
37
After the word-level and letter-level (and probably syllable-level) input channels output their respective codes, this information is sent to the central processor (Allen & Madden, 1990). The central processor then selects (i.e., switches attention to) the first available code in an attempt to solve the task at hand (and the other codes that were output more slowly are stored temporarily in an output buffer). For a letter identification task, this allows the hybrid model to predict a word frequency disadvantage (when higher-frequency words conceal their component letters relative to lower-frequency words, Allen & Madden, 1990; Healy et al., 1987) and a word inferiority effect (e.g., Allen & Emerson, 1991). That is, for a letter identification task, subjects decide whether a target letter matches the initial letter of a subsequently presented word or nonword. The hybrid model hypothesizes that the word-level channel wins the race to the central processor for very-high and medium-high frequency words, but that the letter level channel wins the race to the central processor for lowerfrequency words (Allen & Madden, 1990). Furthermore, very-high-frequency words are output so rapidly that the central processor has time to determine that this code is the wrong one for the task at hand and to switch attention to the letter-level input channel before that code has been output (i.e., letter identification requires letter-level information that is not readily available from a word-level code). However, for medium-high-frequency words, the word-level channel just barely outputs its code before the letter-level channel. Thus, by the time the central processor determines that it cannot use the word-level code to solve the task at hand, the letter-level code has been output into its buffer, and this results in longer access time. Because the letter-level channel wins the race to the central processor for lowerfrequency words, there is no delay in central-processor access to the letter-level code for these data. Therefore, the model predicts longer letter identification RT for medium-high frequency words than for high-frequency or lower-frequency words (Allen & Emerson, 1991; Allen & Madden, 1989, 1990; Johnson et al., 1989). For lexical decision and naming tasks, the hybrid model predicts a word frequency advantage and a mixed-case disadvantage (Allen, Wallace, & Weber, in press). That is, the model predicts that the word-level code is needed to solve the task at hand (or a pseudo-wordlevel code for words presented in mixed case), and that the lexical access component of wordlevel processing results in a word frequency advantage. However, it should be noted that there are different types of word-level codes depending upon the processing pathway. For example, the hybrid model would typically use the rule-based phonological pathway to conduct a naming task, whereas the model would use the orthographic pathway to conduct a lexical decision task (Allen et al., in press). There is considerable evidence for this dual-route architecture (e.g., Carr & Pollatsek, 1985; Coltheart et al., 1993). Finally, the dual encoding architecture within a given pathway (e.g., word-level and letter-level) of the hybrid model allows it to account for the differential pattern of mixed-case disadvantage effects found across different stimulus exposure durations for words and nonwords on a lexical decision task (Allen et al., in press). It is not clear how models that always encode words from component letters (e.g., McClelland & Rumelhart, 1981; Paap, Newsome, McDonald, & Schvaneveldt, 1982; Paap, Newsome, & Noel, 1984) can account for this finding. 1.3. R u l e - B a s e d versus P D P models
Now we are ready to re-examine the issue of whether rule-based or sub-symbolic, PDP models are preferable processing architectures for the optimal model of visual word
38
Ph.A. Allen et al.
recognition. It was noted earlier that the IAM (McClelland & Rumelhart, 1981) was a cascaded, PDP model that always encoded words initially as component letters, but that allowed the word-level and the letter-level tiers to interact via excitatory and/or inhibitory feedback (vertical interaction). Alternatively, we also described a rule-based, hybrid horse race model (Allen et al., in press) that encoded words initially as both words and component letters. Research by Allen et al. (in press) indicated that the dual encoding mechanism of the rule-based model was better able to account for certain aspects of the mixed-case disadvantage and the word inferiority effect. However, there is no good design reason for why a PDP model cannot encode words holistieally as well as analytically (this would result in horizontal interaction between the word and letter levels). Thus, in terms of implementing a model of visual word encoding, it appears that both rule-based models and sub-symbolic PDP models can account for a good deal of available data--as long as the sub-symbolic model uses at least two different types of initial encoding (e.g., word level and letter level, see Allen et al., in press), and as long as there is some sort of local representation of words (i.e., a lexicon, see Coltheart et al., 1993). In summary, the strength of rule-based models is that we know the state of the machine at all times (this increases model precision) (Turing, 1936). This class of model can emulate all known logical and mathematical operations (Minsky, 1967), and human cognition appears to be rule based (e.g., Coltheart et al., 1993; Pinker & Prince, 1988). Alternatively, the strengths of sub-symbolic PDP models are that they more closely resemble neural processing than do rule-based models (Rumelhart & McClelland, 1986), and that one can observe how new patterns are learned (Seidenberg & McClelland, 1989). 2. PART H: BASIC THEORIES OF LEXICAL ACCESS Up to this point, we have been primarily concerned with examining how word and letter stimuli are transduced into a form that the human visual information processing system can use. However, this initial transduction process is only the first step of the pattern perception of words. That is, to be of functional significance, the transduced codes need to be recognized and/or identified. By recognition, it is meant that the transduced code representing a word, syllable, or letter is familiar (Besner & Johnston, 1989; Besaer & McCann, 1987). Alternatively, identification refers to the process of determining the name of the transduced word, syllable, or letter (Besner & Johnston, 1989). With regard to words, a further assumption of many models of visual word recognition is that identification not only allows individuals to name words, but also to know their meaning (see Cart & Pollatsek, 1985). This is especially the case for what are termed "lexical instance models" of visual word recognition, because these models assume that lexical access occurs. Lexical access using the orthographic route(s) (word-leveL syllable-level, or letter-level) is the process of addressing the mental lexicon, or mental dictionary. That is, individuals code the transduced word stimulus (regardless of whether it was initially encoded using word-leveL syllable-level, or letter-level units of analysis) into a prototype representation, and then use this representation to access the name and meaning of that word in the mental lexicon. Alternatively, lexical access using the phonological route is the process of using grapheme-to-phoneme correspondence (GPC) rules (or more holistic encoding schemes, see Allen et al., 1994) to generate, or assemble, a phonological representation of a word that is then used to access the lexicon.
Visual word encoding and the effect of adult age and word frequency
39
We will discuss three classes of lexical access models: rule-based, lexical-instance models; rule-based, dual-route models; and sub-symbolic, distributed models. The first two classes of models use rules to reco~mfize and identify words and have "local" representations of words in a mental lexicon, whereas the latter class of models does not use rules to recognize and identify words and does not possess a real lexicon (i.e., words are represented in a distributed manner). 2.1. Rule-Based Lexical-lnstance Models
The lexical-instance models all assume the existence of a memory store termed a lexicon, and that individual words, or prototypes, are represented within this store. This class of models assumes that a visually encoded word addresses an entry (or entries) in the mental lexicon with the outcome being either recognition or identification of a given word. Contrary to the general rules used in an assembled, phonological system, lexical instance models assume that individuals know many instances of words from experience, and that these experiences can be described by rules (Cart & Pollatsek, 1985). There are three sub-types of lexical-instance models: logogen models, lexical search models, and verification models (Cart & Pollatsek, 1985). Logogen models (e.g., Morton, 1969) assume that an encoded stimulus addresses all items stored in the lexicon in parallel. The words stored in the mental lexicon have different threshold levels. Specifically, as words become increasingly more familiar, their activation threshold decreases. This allows logogen models to account for word frequency advantages fotmd for lexical decision (e.g., Allen, McNeal, & Kvak, 1992; Dobbs, Friedman, & Lloyd, 1985) and naming tasks (Monsell, Doyle, & Haggart, 1989), because the activation threshold for higher-frequency words is lower than the activation threshold for lower-frequency words. This sort of lexical architecture can produce the partial activation of multiple lexical entries. However, to be tenable, this sort of model must typically produce activation of the appropriate entry in the lexicon. Probably the most widely cited example of a logogen model of visual word recognition is the IAM of McClelland and Rumelhart (1981; Rumelhart & McClelland, 1982). Interestingly, the IAM is typically classified as a PDP model. However, words are individually represented in the lexicon of the IAM, so representation in this model is local rather than distributed (i.e., in a truly distributed model, all words are distributed simultaneously as a pattern of activation across a set of processing nodes, Seidenberg & McClelland, 1989). Another subtype of lexical-instance models is the lexical search model (e.g., Forster, 1976; TaR, 1979). This model assumes that the lexicon is organized as a function of word frequency so that higher-frequency words tend to have higher priority access than do lowerfrequency words. The model uses an access code made up of either morphemes (Tat~, 1979) or syllables (Lima & Pollatsek, 1983) to activate a candidate set of word representations stored in the lexicon. The recoded input stimulus is then compared serially to the candidate set of words that contains similar or equivalent syllables or morphemes, and the model then selects the best-fitting candidate to recognize or identify. The design goal of forming a candidate set is to limit the size of the search set, because the search process is serial in nature (Carr & Pollatsek, 1985). A slight variation of the lexical search model is the verification class of models. According to verification models (e.g., Becker, 1976; Paap et al., 1982), there is a top-down
40
Ph.A. Allen et al.
selection process that verifies that the word initially picked during lexical access is, in fact, the correct word. Contrary to the lexical search model in which word frequency effects result directly from lexical access, word frequency effects in verification models result solely from order of verification effects (e.g., Paap et al., 1982). Indeed, Paap et al. (1982) proposed that no word frequency effects should occur when verification is prevented. However, it is clear that word frequency effects continue to occur even in situations in which the activationverification model ofPaap et al. (1982) predicts that verification is impossible (see Allen et al., 1992; Dobbs et al., 1985). Although verification models can account for a wide variety of phenomena (e.g., Becker, 1976; Paap et al., 1982), it appears that the activation-verification version of the model (Paap et al., 1982) cannot account for replicated word frequency experiments that manipulate exposure duration. 2.2. Rule-Based Dual-Route Models
Another class of lexical access models are the dual-route models (e.g., Carr & Pollatsek, 1985; Coltheart et al., 1977, 1993). Dual route models provide a mechanism that allows the prommciation (or naming) of both familiar and unfamiliar words (or nonwords), while still providing a visually accessible route to the lexicon (Carr & Pollatsek, 1985). These models propose that visual word recognition is accomplished through the use of two processing pathways. One pathway is orthographic in nature, and the lexicon is addressed using an orthographic code. This pathway of the dual-route model is equivalent to the lexicalinstance models. The other pathway uses GPC rules (or, perhaps, multiple grapheme-tosyllable or multiple grapheme-to-whole word correspondence rules, Allen et al., 1994), is phonological in nature, and accesses the lexicon with a phonological code (Coltheart et al., 1993). The original versions of the dual-route model did not provide the phonological channel with the ability to access a lexicon (Carr & Pollatsek, 1985), however, more recent versions of phonological processing models do assume that this channel involves lexical access (e.g., Coltheart et al., 1993; Levelt, Schiefers, Vorberg, Meyer, Pechmann, & Havinga, 1991). Because of their inclusion of multiple pathways, dual-route models can account for a multitude of phenomena (Carr & Pollatsek, 1985). However, because of their complexity, these models are painfiflly difficult to understand, at times. Unfortunately, this complexity aspect underscores a major reason why after over 100 years of scrutiny we still do not completely understand visual word recognition. Clearly, there is no simple explanation for the phenomena. 2.3. Revisiting the Hybrid Horse Race Model
Earlier in the paper, we outlined a hybrid, horse race model of visual word recognition (and identification) that was holistically biased (e.g., Allen & Emerson, 1991; Allen & Madden, 1990; Allen et al., in press). Much of the initial encoding architecture of the model was discussed earlier, however, we are now ready to address how the model performs lexical access. Although the model is architecturally dual-route in nature (see Allen et al., in press, 1994), the orthographic pathway of the model has been tested much more extensively than the phonological route (Allen et al., 1994). Thus, we need to address the lexical-instance characteristics of this model. This is no simple task, because the hybrid model includes both logogen and lexical search characteristics.
Visual word encoding and the effect of adult age and word frequency
41
The lexical access characteristics of the hybrid model resemble a logogen model because lexical entries are assumed to have progressively lower activation thresholds as the frequency of the word entry increases (Allen, McNeal, & Kvak, 1992). However, the hybrid model assumes that the coded spatial frequency pattern of the word stimulus (the holistic code) is used as an access code. This considerably lessens the number lexical entries that need to be partially activated per encoded word. This aspect of the hybrid model resembles lexical search models because an access code is used to limit the number of comparisons that need to be made between the encoded representation of the spatial frequency pattern of a word and similar lexical entries. Note that words in this model are encoded in the lexicon as "pictures." Once a picture, or icon, is accessed, then the name and/or meaning of the word can be determined (actually, meaning is probably represented in a sub-module of the orthographic lexicon). However, the lexical access process is parallel in the hybrid model, and word frequency is encoded by assuming that higher frequency words are actually closer in psychological space to the lexical accessing retrieval "pointer" (and this assumption is consistent with Forster's, 1976, 1989, lexical search model). When multiple entries, or icons, in the lexicon have been partially activated because they are visually similar to the word stimulus, then this is resolved using normalization procedures until a lexical entry is activated sufficiently to surpass its threshold. Note that the hybrid model avoids many of the problems associated with having a lexicon represented as component letters. Instead of representing words as component letters, words are represented holistically. Such an architecture can account for why orthographic neighborhood effects (see Stadtlander's chapter in the present volume) tend to be occur primarily for nonwords (e.g., Coltheart et al., 1977). That is, nonwords must be superposed in order to be input into the lexicon, and the superposition process is analytic. Thus, letter-level characteristics such as orthographic neighborhood effects should affect performance for nonwords (or, perhaps, very low frequency words that are not sufficiently activated to pass firing threshold). Finally, the hybrid model rejects nonwords when the input stimulus fails to adequately activate any items in the lexicon (Allen et al., in press). 2.4. Sub-Symbolic Distributed Models Our final class of lexical access models is quite different from the first two classes, because the final class of models uses distributed representation rather than the local representation used by the two previously mentioned classes of models. As noted earlier, this sort of connectionist (or distributed) model resembles a neural net. We will limit our discussion to the developmental model of Seidenberg and McClelland (1989), because this model has been tested more extensively than any other distributed PDP model of visual word recognition (e.g., Besner et al., 1990; Fera & Besner, 1992; Joordens & Besner, 1994; Coltheart et al., 1993). Every time the neural net is activated, the orthographic layer (i.e., letter detectors) is activated across its complete set of units. The model encodes letter strings as Wickelgraph triplets. This pattern of activation from the orthographic layer is then sent (or "spread") to the next layer--the hidden units. The hidden units are also connected to all the orthographic layer units in a feedback loop. The activation strength of the connections between the orthographic and the hidden units is then modified across processing cycles so that the hidden layer units can learn the input pattern. An orthographic error score is then computed by comparing the pattern of activation across the orthographic units after feedback
42
Ph.A. Allen et al.
from the hidden layer units to the pattern of activation across the orthographic units when the letter string was originally presented to the orthographic layer. The smaller the error score, the better the hidden units have learned the pattern. To perform a lexical decision task, the developmental model (Seidenberg & McClelland, 1989) uses orthographic error scores to index the level of familiarity. If the orthographic error score is lower than some preset criterion, then this letter string is assumed to be familiar, and a word decision is made. However, if the orthographic error score is greater than some preset criterion, then this letter string is classified as being unfamiliar, and a nonword decision is made. This model is essentially a connectionist revision of the decision model proposed by Balota and Chumbley (1984). Note that this model does not actually conduct lexical access in order to perform a lexical decision task. Indeed, the model does not even contain the local representations that define a lexicon. In order to name words, the developmental model projects the activation across units from the hidden layer to a phonetic layer (Seidenberg & McClelland, 1989). This allows individuals to name input letter strings. Phonological error scores are formed by comparing the pattern of activation across phonetic units as the result of letter string stimulation to the pattern of activation across phonetic units as the result of direct phonemic specification (when the correct pronunciation is input into the net) (Seidenberg & McClelland, 1989). The model then assumes that this pattern of activation across the phonological units is used as input to form an articulatory-motor program which can then be implemented by the motor system-resulting in naming. It is important to note that the developmental model uses the same network to process words both orthographically and phonologically (Seidenberg & McClelland, 1989). Thus, this distributed model is at odds with dual-route models, because the dual-route models use separate orthographic and phonological routes (Coltheart et al., 1993). Although the parsimony and apparent precision of the distributed, developmental model are impressive (the developmental model settled on the correct pronunciation for 2,820 out of 2897 words which were input into the model, Seidenberg & McClelland, 1989), the lack of a local lexicon presents some serious problems for the model. In general, the distributed model makes more errors under certain conditions than do humans. For example, the developmental model has difficulty in recognizing exception words not included in the original set of words used by Seidenberg and McClelland (see Besner et al., 1990; Coltheart et al., 1993; Fera & Besner, 1992; Joordens & Besner, 1994). However, distributed models with local lexical representation are not susceptible to this same criticism (e.g., Coltheart et al., 1993). If such models were to employ more than one encoding route (e.g., word level and letter level, or syllable level and letter level), then these PDP models might be able to account for human visual word recognition performance as well as rule-based models (see the chapter in the present volume by Kellas, Paul, & Vu for an example of the IAM applied to aging). Forster (1989), though, has noted that distributed models with local lexical representation still have problems in resolving orthographic neighborhood effects when the lexical entries in the same neighborhood have different frequencies. 2.5. Post-Lexical Effects of Word Frequency? One final issue concerning lexical access is that of the locus of word frequency effects. The finding that subjects respond more rapidly to higher-frequency words than to lower-
Visual word encoding and the effect of adult age and word frequency
43
frequency words for both lexical decision and naming tasks has provided fundamental support for idea that individuals possess at least one mental lexicon. The aforementioned rule-based models all assume that word frequency effects are the result of faster lexical access for higherfrequency words. However, Balota and Chumbley (1984, 1985) proposed that a substantial portion of the word frequency effect was due to decision processes that occurred aider lexical access. Although the work of Balota and Chumbley does suggest that s o m e of the word frequency effect is not the result of lexical access processes, considerable research suggests that a large portion of the word frequency effect for lexical decision and naming tasks is the result of lexical access processes (Allen, McNeal, & Kvak, 1992; Connine, Mullennix, Shernofl~ & Yelen, 1990; Monsell et al., 1989). 2.6. Empirical Measures of Lexical Access
In the basic visual word recognition literature, lexical access speed has been examined by manipulating two different factors: word frequency and semantic priming. Word frequency refers to how common a word is in written American English (e.g., Kucera & Francis, 1967). Semantic priming refers to the facilitation of a target word by a semantically related prime word relative to a semantically unrelated prime word (e.g., Neely, 1990). With regard to word frequency, the assumption is that the mental lexicon is organized so as to allow more speedy access to higher-frequency words than to lower-frequency words (e.g., Allen, McNeal, & Kvak, 1992; Carr & Pollatsek, 1985; Dobbs, Friedman, & Lloyd, 1985) (although as noted earlier, connectionist, or distributed, models implement a lexicon somewhat differently). With regard to semantic priming, the relation to lexical access speed is somewhat more circuitous. That is, the activation of the prime word is assumed to spread to other semantically related words stored in the mental lexicon (i.e., "spreading activation") so that target words that are semantically related to prime words are already partially activated (relative to target words that are not semantically related to the prime word) (Collins & Quillian, 1969; Nelson, Schreiber, & McEvoy, 1992). It is important to remember, though, that semantic priming was designed primarily to measure how meaning affects lexical access. Lexical access speed is a secondary concern for semantic priming studies. 2.7. Summary of the Word Encoding Tutorial
In the preceding sections, we have reviewed theories of how words and letters are coded by the information processing system for later use (i.e., the transduction process). We came to the conclusion that it is probably necessary to assume that visual word encoding uses multiple input channels (e.g., Allen & Emerson, 1991; Allen et al., in press; Healy et al., 1987; Johnson et al., 1989). Also, we reviewed the literature on lexical access and concluded there is good evidence for the existence of a mental lexicon (Allen, McNeal, & Kvak, 1992; Cart & Pollatsek, 1985; Forster, 1989; Monsell et al., 1989). It appears that this lexicon has both logogen (Morton, 1969) and lexical search (Forster, 1976, 1989) characteristics. We also discussed in some detail the difference between symbolic (role-based) and subsymbolic (distributed) architectures of visual word recognition/identification. We noted that rule-based models are based on a Turing machine and seem to capture the human predilection to mimic (or actually use) role-based information processing (e.g., Pinker & Prince, 1988). Alternatively, we mentioned that distributed models are based upon Hebb's (1949) notion of a
44
Ph.A. Allen et al.
"cell assembly" and are particularly effective at illustrating how a system learns a new task (McClelland & Rumelhart, 1986). However, even though we readily admit that distributed models have some interesting attributes, there is still reason for concern that they may not work without adding a local lexicon (e.g., Besner et al., 1990; Fera & Besner, 1992; Joordens & Besner, 1994; Coltheart et al., 1993) as well as an attentional selection mechanism (Forster, 1989). When these attributes are added to a distributed model, then this model is for all intents and purposes an embellished rule-based model (Forster, 1989). Consequently, In the next section, we will use the hybrid model (i.e., rule based) as a guide to interpreting the aging literature because (1) it has multiple input - channels not contained by extant distributed models (and these multiple input channels appear to be necessary to account for visual word recognition data, Allen et al., in press), (2) it has been applied to all three of tasks discussed below, and (3) the model can account for age differences in t~equency effects for a letter identification task that distributed models without a selective attention mechanism (i.e., a central processor) cannot account for. 3. PART IH: THE EFFECT OF AGING ON VISUAL WORD RECOGNITION Now that we have reviewed the literature on how word stimuli are coded into the visual system and then how these words are reco~ized and/or identified through lexical access (or some other analogous process for distributed models), we are ready to address how increased adult age affects visual word recognition for studies that have manipulated word t~equency. To begin, we will discuss some models of cognitive aging. Next, the literature on age differences in visual acuity will be briefly examined. Finally, we will review the aging literature on letter identification, lexical decision, and naming tasks, and examine whether these results can be accounted for by the hybrid model applied to aging, and whether these data require the inclusion of cognitive constructs (localized models) or simply require a behavioral description bereft of any such cognitive constructs (generalized models).
3.1. General Models of Cognitive Aging As noted earlier, a major goal of the present paper was to determine whether age differences in visual word recognition are the result of a single general factor (e.g., Allen, 1990, 1991; Salthouse, 1991), a small number of general factors (e.g., Cerella, 1985, 1990; Lima, Hale, & Myerson, 1991; Myerson, Ferraro, Hale, & Lima, 1992), or whether some stages of processing are more affected by aging than others (e.g., Allen, Madden, & Crozier, 1991; Allen, Madden, Weber, & Groth, 1993; Allen, Madden, Cerella, Jerge, & Betts, 1994; Amrhein, this volume; Bashore et al., 1989; Fisher, Fisk, & Duffy, this volume; Fisk & Rogers, 1990; Fisk & Fisher, 1991; Balota & Ferraro, 1993; Madden et al., 1993). This is a crucial issue for the psychology of aging, because the class of models used by researchers and theoreticians will have a major impact upon whether cognitive processes remain relevant constructs in experimental aging research. Process-based models emphasize how aging has differential effects across processing stages (e.g., Fisk & Fisher, 1994). This sort of model assumes that we cannot understand how aging affects information processing without alluding to component cognitive processes (Allen et al., 1993). However, aging models that emphasize a single factor (e.g., Allen, 1991, Salthouse, 1991) or just a few general factors (e.g., Cerella, 1985; Myerson et al., 1990) do not emphasize the importance of cognitive constructs. Indeed,
Visual word encoding and the effect of adult age and word frequency
45
entropy-based models of aging (e.g., Allen, 1990, 1991; Myerson et al., 1990) are really neurologically based. 3.2. Generalized Models
In order to address this issue of generalized models of aging (single-factor models and models that include a small number of factors) versus localized modds of aging (models that emphasize local cognitive processes), we need to develop these two views of aging in more detail. Generalized models of aging assume that both young and older adults use the same processing stages, but that older adults take longer to process information at each stage than do young adults (Cerella, 1985, 1991). In order to test this view of aging, investigators frequently plot older adults' data across task conditions along the y axis and plot younger adults' data across the same task conditions along the x axis. Then, the best-fitting slowing function is determined by using least-squares regression (or some other curve-fitting routine). This approach to examining age differences is typically termed a Brinley plot (Brinley, 1965). In Brinley-plot analyses, the single best-fitting "slowing function" frequently accounts for over 90% of the experimental variance. Investigators have found linear, multiplicative, additive, and non-linear functions that resulted in the best-fitting function for a Brinley plot (e.g., Cerella, 1985, 1991; Myerson et al., 1990). Cerella (1985) proposed that a linear function would best predict older adults' RT from young adults' analogous RT. This linear function was based upon a multiplicative component for central processes and an additive component for peripheral processes. Advocates of Bfinley plots have emphasized how such regression techniques are particularly effective at finding commonalties in a data set (e.g., Cerella, 1994; Myerson, Wagsta~ & Hale, 1994). It should be noted, though, that such regression methods are not overly sensitive in detecting interactions compared to an analysis of variance (e.g., McClelland & Judd, 1993; Kliegl, 1994; Perfect, 1994). Thus, one should take particular care in interpreting Brinley plots as conclusive evidence for a single factor (or small subset of factors) accounting for age differences in a given task or group of tasks (see Fisher et al, in the present volume, or Fisk & Fisher, 1994). Also, it is important to remember that the proportion of variance accounted for is not the only relevant factor. For example, a basic assumption of the experimental method is that interactions should be interpreted before main effects (e.g., Keppel, 1991). Thus, a situation could occur in which an interaction accounted for less variance than a main effect (and this is typically the case for each separate interaction), but that in which the interaction qualified the main effect. On the other hand, one should also take particular care in concluding that localized age differences are present simply on the basis of interactions present in an analysis of variance, because the presence of an interaction with age, by itself; does not preclude localized age differences (e.g., Cerella, 1991). Another method of testing generalized models of cognitive aging is to use partial correlation (Madden, 1992; Salthouse, 1985) or hierarchical regression (Salthouse, 1991) methods. Typically, these models assume that age differences in information processing speed are to a large extent the result of a single processing speed factor (Salthouse, 1985, 1991). In order to test for this possibility, advocates of this method examine age differences in task performance on at least two tasks or multiple levels of a single task. For example, Madden (1992) had subjects participate in both a lexical decision/priming task and the Wechsler Adult Intelligence Scale-Revised (WAIS-R) digit symbol task. Digit symbol task performance was
46
Ph.A. Allen et al.
used as a general measure of processing speed, lexical decision performance was assumed to measure lexical access time, and stimulus degradation (placing asterisks between letters of words on some trials) was assmned to measure word encoding time. When digit symbol task performance was partialed out in the Madden (1992) study, the correlation between age and mean RT decreased from .49 to .29. Also, the correlation between age and stimulus degradation decreased from .57 to .49 when digit symbol task performance was partialed. This suggested that the effect of age was partially due to processing speed. However, even after digit symbol performance was partialed out, the aforementioned correlations still remained reliable. Furthermore, the correlation between age and stimulus degradation after digit symbol performance had been extracted (r = .49) appeared to be greater than the correlation between age and overall mean RT after digit symbol performance had been extracted (r = .29). This finding implied that age differences in encoding (as measured by stimulus degradation effects) were greater than overall age differences (encoding, lexical access, response selection, response execution--as measured by mean RT over all task conditions). This suggests that there are localized (at encoding) age differences in addition to generalized age differences. Although the preceding paragraphs have described two methods of testing a generalized model of aging (Brinley plots and partial correlation), these are methods of analysis rather than theoretical frameworks. Indeed, without additional theoretical development, the concept of generalized slowing reduces to a tautology--that older adults slow down relative to young adults because older adults are slower. Thus, in order for generalized slowing to be a tenable scientific model of cognitive aging, it is necessary for the model to explain what general factor (or small subset of factors) causes age-related slowing. The first attempt to do this was the complexity hypothesis (e.g., Birren, 1965; Cerella, Pooh, & Williams, 1980; Salthouse, 1985). The complexity hypothesis predicts that older adults are slower than young adults at information processing tasks because of age-related changes in the central nervous system (Birren, 1965). The complexity model predicts that as task complexity increases, age differences should become progressively larger. However, the complexity hypothesis, alone, would appear to be a post-hoc explanation. That is, this hypothesis predicts that a given task condition is more complex because it takes longer to complete, and a given task condition takes longer to complete because it is more complex. What is necessary is an additional factor that can be used to define complexity independent of post-hoc task performance. Thus, such a framework must provide a construct that predicts an antecedent cause. The extensive research on cognitive processes provides aging researchers with a rich source of potential candidates for such a factor that predicts task complexity. For example, Salthouse (1991) has proposed that age differences in processing speed are the result of an age difference in working memory (also see Stine's contribution to the present volume). Because much is known about the processing resource limitations of working memory (e.g., Baddeley & Hitch, 1974), researchers can define task condition complexity before an experiment is conducted. Another method of defining processing complexity independent of examining task performance RT in a post hoc manner is to use the concept of entropy as an antecedent cause (e.g., Allen, Patterson, & Propper, 1994). The concept of entropy is based upon a scientific law--the Second Law of Thermodynamics. The basic assumption is that entropy must increase across the lifespan, and that this should increase levels of neural (or internal) noise (Allen, 1990, 1991; Allen & Coyne, 1988; Allen, Madden, Cerella, Jerge, & Betts, 1994, Experiment
Visual word encoding and the effect of adult age and word frequency
47
4; Allen, Patterson, & Propper, 1994; Krueger & Allen, 1987; Welford, 1958). For example, from the predictions of entropy, we know that information should be represented in memory more variably with increases in adult age, and this predicted effect has been empirically verified (e.g., Allen, 1991; Allen, Katffman, & Propper, 1994, Part I & Part II). Furthermore, we can predict levels of entropy across age using methods borrowed from statistical dynamics in physics. These levels of entropy can then be compared to behavioral data such as RT (e.g., Allen, Kaufman, & Propper, 1994; Parts I & II). Interestingly, entropy data suggest that processing variability may be an even more general factor than processing speed in accounting for age differences (Allen, Kaufinan, & Propper, 1994, Part II).
3.3. Localized Slowing In addition to generalized models of aging, there is another class of models that is localized, or process-specific, in nature. As a simplifying assumption, these models also assume that both young and older adults process information using the same number and order of processing stages (although see Fisher et al., in this volume, for a cautionary note on this assumption). However, an important assumption of localized models is that some cognitive processes are affected more than others by increases in adult age (e.g., Allen, Madden, Cerella, Jerge, & Betts, 1994; Allen et al., 1993; Balota & Ferraro, 1993; Bashore et al., 1989; Fisk & Fisher, 1994; Madden et al., 1993). One manner of conceptualizing localized models of cognitive aging is to assume that local cognitive processes modulate (i.e., qualify) the effect of relatively general biological processes. For example, a general factor such as entropy may not impact uniformly upon different levels of information processing, because these different levels of information processing may require different levels of processing resources (and processes that require more processing resources would be particularly affected in a deleterious manner by entropy). Thus, age-related increases in entropy may affect visual encoding more than lexical access in word recognition tasks (Allen et al., 1993; Madden, 1992). A fundamental assumption of localized models is that one must examine component cognitive processes in order to understand how aging affects information processing. Hence, this approach holds that general curve-fitting descriptions of age differences in information processing across a wide variety of tasks (e.g., Cerella, 1985, 1991) will not tell us the whole story of cognitive aging, because they do not adequately address the fundamental cognitive processes (or internal computations) that are involved (e.g., Allen et al., 1993; Balota & Ferraro, 1993). For example, Allen, Madden, Cerella, Jerge, and Betts (1994) have demonstrated using localized curve-fitting procedures that one needs two different slowing factors to describe age differences in visual word encoding but only one slowing factor to describe age differences in lexical access. As noted earlier, one prediction of localized models of cognitive aging is that a single psychological or biological cause cannot directly explain age differences in information processing across different processing stages and different tasks. This is because cognitive constructs define task difficulty in an a priori manner that cannot be accomplished parsimoniously without alluding to such cognitive constructs. Thus, the emphasis in localized models is placed upon illustrating how component cognitive processes (or the different processes involved in different cognitive tasks) qualify the effect of general factors on aging. However, the localized view does not necessarily assume that no single law exists that can explain age differences in information processing. Instead, the localized view of aging
48
Ph.A. Allen et al.
assumes that there is a qualitative difference in representational levels between biological processes and cognitive processes. This difference in representational levels results in emergent characteristics of cognition such that there is no simple one-to-one mapping between biological levels and cognitive levels. (This is similar to the Gestalt dictum that "the whole is different from the sum of its parts.") Indeed, one could argue that physics, chemistry, biology, and experimental psychology all study the same basic phenomena--but at different levels on the molecular-to-molar continuum Thus, according to the localized framework, cognitive constructs are meaning~l independent of biological constructs even though neural (biological) processes clearly have a major impact upon cognitive processes. Consequently, localized models of aging predict that cognitive aging cannot be reduced to a single biological (or psychological) factor--unless that biological factor can explain behavior (or emergent characteristics) at different levels of the cognitive factor or factors. In order to test for localized age differences, Madden, Pierce, and Allen (1992) proposed a method that combined Brinley plot and ANOVA procedures. First, an ANOVA is used to determine if there are Age x Task interactions. Next, the task condition RT means for young adults were used as an independent variable and older adults' task condition RT means formed the dependent variable using linear regression methods (i.e., a Brinley plot). The bestfitting slowing function derived from the Brinley plot was then used to transform young adults' raw latencies. Finally, the transformed latencies of the young adults and the untransformed latencies of the older adults were used as the dependent variable in an ANOVA. Transforming young adults' data using the best-fitting slowing function results in young adults being transformed into predicted "older adults." If any interactions between age and task remain in the transformed ANOVA, this indicates that a single general slowing model cannot adequately predict the observed age differences (i.e., an interaction exists for the Bfinley plot data). Localized models of cognitive aging predict that Age x Task interactions should remain in the transformed analysis. Alternatively, general slowing models of cognitive aging predict that all Age x Task interactions should be eliminated in the transformed analysis. Although this transformed analysis does eliminate some Age x Task interactions, the procedure typically does not eliminate all Age x Task interactions (e.g., Allen et al., 1993; Allen, Madden, Cerella, Jerge, & Betts, 1994; Madden et al., 1992, 1993). Furthermore, the transform procedure can also form new Age x Task interactions (e.g., Allen et al., 1993). Such results imply that although some generalized effects are present, localized age differences do exist. Of course, it should be noted that multiple-factor (i.e., process-specific) slowing functions such as those used in Allen, Madden, Cerella, Jerge, and Betts, (1994) will eliminate all interactions with age if these models can account for all interactions present in the Brinley plot.
3.4. Visual Acuity It is important to consider whether age differences in other factors such as visual acuity might affect our understanding of age differences in visual word encoding. This is because age differences in visual acuity could affect stimulus encoding processes across age. Previous research has demonstrated that older adults have poorer visual acuity than do young adults (e.g., Owsley, Sekular, & Siemson, 1983; Pitts, 1982; Weale, 1986). A major problem is that older adults typically have only one-third as much retinal illumination as young adults (Weale, 1986). Thus, one potential explanation for why older adults take longer to recognize words
Visual word encoding and the effect of adult age and word frequency
49
than do young adults (e.g., Allen, Madden, & Crozier, 1991, 1993; Cerella & Fozard, 1984; Madden, 1992) is that older adults have poorer visual acuity (Owsley et al., 1983, Pitts, 1982). However, Allen, Madden, Cerella, Jerge, & Betts (1994, Experiment 3) recently tested this hypothesis using a naming task that varied the case type and the exposure duration of word and pronounceable nonword stimuli. Previous research has documented that subjects take longer to recognize/identify words presented in mixed case than in consistent case (i.e., the mixed-case disadvantage), because mixed-case stimuli force subjects to encode stimuli using a slower input channel (e.g., Allen et al., 1993, in press). Allen, Madden, Cerella, Jerge, and Betts (1994) proposed that brief exposure durations should be especially difficult for older adult subjects who had poorer visual acuity than did young adults (e.g., it would take more time for enough visual information to accumulate for older adults if'less light reached the retina within each sample, Madden & Allen, 1994). Thus, if older adults' typically larger mixed-case disadvantage (Allen et al., 1993; Allen, Madden, Cerella, Jerge, & Betts, 1994) was the result of a lack of rentinal illumination, then older adults should have revealed a relatively larger age difference for the mixed-case disadvantage for briefer exposure durations than for presentation-until-response. The Allen et al. experiment, though, did not find an Age x Case Type x Exposure Duration interaction even though the Age x Case Type interaction was reliable. This result suggests that older adults' larger mixed-case disadvantage was the result of an encoding decrement that occurred after initial sensory transduction. Indeed, several recent studies have found that at best, age differences in visual acuity can account for main effects, but not for Age x Task interactions (e.g., Allen, Weber, & Madden, 1994; Long & Crambert, 1990). Consequently, although older adults' poorer visual acuity may slow down their accumulation rate thereby affecting main affects for age, it does not appear that age differences in visual word encoding can be accounted for by simply alluding to age differences in visual acuity. However, even though age differences in visual acuity do not adequately account for age differences in encoding, there is evidence that a later stage(s) of encoding may be relevant to this issue. Earlier in the present paper, it was proposed that visual stimulus encoding involves more than one stage of processing. A recent study by Allen, Weber, and Madden (1994) illustrates the distinction between two such encoding stages. These researchers hypothesized that visual stimulus encoding involves both an initial detection stage and a subsequent identification stage. Initially, a "blob-level" representation is formed that is used to detect the presence of a stimulus (e.g., DeValois & DeValois, 1987; Mart, 1982; Pashler, 1987). In order to identify objects within the blob-level representation (e.g., words), though, it is necessary to focus attention upon different sections of the blob-level representation (i.e., attentional selection) so that this information can be interpreted (Allen & Madden, 1990; Pashler, 1987). Allen, Weber, and Madden (1994) using a visual search task involving letters as stimuli found evidence that older adults had difficulty in selecting stimuli during the identification stage of processing. However, when illumination was varied (40 cd/m: vs. 18 cd/m:) between two groups of young adults in an attempt to simulate older adults' performances by decreasing retinal illumination in younger adults, this reduction in illumination did not result in the young adults in the lower illumination condition (18 cd/m2) resembling older adults in the higher illumination condition (40 cd/m:). Consequently, the Allen et al. (1994) study suggested that age differences at the identification stage of encoding rather than the detection stage of encoding accounted for age differences in visual search.
50
Ph.A. Allen et al.
4. PART IV: A REVIEW OF THE AGING AND VISUAL WORD RECOGNITION LITERATURE W H E N WORD FREQUENCY IS VARIED
In order to determine if age differences in visual word encoding are generalized or localized, experiments necessarily must manipulate at least two different stages of the word recognition/identification process. These stages include: detection (forming the initial bloblevel representation), selection (focusing on a particular portion of the blob-level representation), lexical access (looking a word up in the "mental dictionary"), decision (e.g., "word" or "nonword" for a lexical decision task), response selection (mapping the decision on to the appropriate response), and response execution (directing muscles to carry out the appropriate response). Word encoding involves the first three of the six listed stages of word recognition/identification. The implicit assumption of this stages framework is that additive factors logic holds (e.g., Sternberg, 1966). This simplifying assumption allows one to assume, for example, that case mixing effects and word frequency effects measure different processing stages (e.g., Allen et al., 1993). 4.1. Age differences in Lexical Access?
As noted earlier, lexical access is typically indexed by manipulating word frequency or semantic priming. Researchers have examined age differences in lexical access using both indices. With regard to semantic priming, considerable aging research has been conducted (e.g., Burke, White, & Diaz, 1987; Howard, Shaw, & Heisey, 1986; Madden, 1989, 1992; Madden et al., 1993). In the present volume, Kellas, Paul, and Vu (normal aging) and Ober and Shenaut (Alzheimer's disease) will address semantic priming. The gist from these studies suggest that there are few appreciable age differences in semantic priming (at least for healthy older adults). Because of the already existing literature reviewing semantic priming, the present examination of visual word processing will focus upon aging studies that have measured lexical access by manipulating word frequency. In reviewing aging studies that have manipulated word frequency, it is important to address the factors of education and vocabulary ability (frequently measured using WAIS-R Vocabulary subscale scores). For example, Tainturier, Tremblay, and LeCours (1989, 1992) found that individuals showed a progressively larger word frequency effect as their number of years of education decreased. With regard to age, the controversy concerns whether researchers should match their young adult and older adult groups on educational and verbal ability levels. This is an issue, because older adults tend to score higher than young adults on vocabulary tests. Balota and Ferraro (1993, 1994) found that when verbal ability was matched across age, older adults showed slightly larger word frequency effects than did young adults. This suggested that older adults actually took longer than young adults to perform lexical access for low-frequency words. However, Allen, Madden, Cerella, Jerge, and Betts (1994) and Tainturier et al. (1989, 1992) found no age differences in word frequency when verbal ability was controlled. An assumption of aging research is that older adults will tend to have more exposure to words than will young adults because older adults have lived longer. In addition to focusing our review on research that has manipulated word frequency, the present review will also focus on three common word processing tasks: letter identification, lexical decision, and pronunciation (naming). Letter identification consists of determining whether a visually present letter matches the initial letter of a subsequently visually
Visual word encoding and the effect of adult age and word frequency
51
presented word (i.e., a pre-cue task) or whether the first letter of a word matches a subsequently presented letter (i.e., a post-cue task) (see Allen & Madden, 1990). Letter identification involves word processing because word frequency effects are present for this task (e.g., Allen & Madden, 1990; Johnson, Allen, & Strand, 1989). A lexical decision task requires subjects to decide whether visually presented letter strings form actual English words (subjects either respond "word" or "nonword," see Carr & Pollatsek, 1985). Finally, a naming task requires subjects to pronounce visually presented words or pronounceable nonwords (see Carr & Pollatsek, 1985). Note that the naming task requires individuals to form a specific speech production output code for each named stimulus, but, contrary to a lexical decision task, individuals do not have to determine if a visual stimulus forms an actual English word on a naming task. 4.2. Letter Identification Studies.
There have been two letter identification studies conducted that examined aging and word frequency. Allen and Madden (1989) used a pre-cue task in which a letter target was presented, followed by the presentation of a word. Subjects were instructed to decide whether the pre-cued letter matched the initial letter of the subsequently presented word. As would be expected, Allen and Madden (1989) found that older adults took longer, in general, than young adults to perform the task (644 ms vs. 412 ms, respectively). However, this study revealed that the pattern of word frequency effects varied across age. The word frequency factor was based on the Kucera and Francis (1967) norms, very-high-frequency (VHF) = 240-660 instances, medium-high-frequency (MHF) = 155-235, low-frequency (LF) = 40-54, and verylow-frequency (VLF) = 1-5. For the young adults, there was a non-monotonic function across word frequency (VHF = 406 ms, MHF = 429 ms, LF = 408 ms, VLF = 405 ms), but older adults showed a monotonic function across word frequency (VHF = 636 ms, MHF = 633 ms, LF = 638 ms, VLF = 671 ms). That is, young adults showed a partial word frequency disadvantage for a letter identification task that is consistent with horse race models of visual word recognition (e.g., Allen & Madden, 1990; Allen & Emerson, 1991; Johnson et al., 1989). Older adults, though, showed a partial word frequency advantage for a letter identification task. Clearly, these data indicate that young and older adults were processing the letter identification stimuli in a different mariner (see Fisher et al., in the present volume). In the next section (Part V), we will argue that both young and older adults used the same basic processing architecture, but that encoding difficulties for the letter-level input channel forced older adults to use the rather circuitous word-level input channel for conducting a letter identification task. These letter identification data provided support for a localized model of cognitive aging. To illustrate this point, one needs only to test a single slowing function for these data. When the best-fitting slowing function is computed for these data using a Brinley plot [Y = 1.39(X) + 72], the amount of explained variance is only 36 percent. Allen, Madden, and Crozier (1991) replicated the letter identification results of Allen and Madden (1989). However, Allen and colleagues tested subjects on both a pre-cue letter identification task and a lexical decision task. In this letter identification study, Allen et al. (1991) used lower-frequency words for the VLF category (1-5 instances using the Kucera & Francis, 1967, norms) than did Allen and Madden (1989, 40-55 instances). Thus, the results of Allen et al. (1991) could not have been due to using words that were too high in word frequency. These letter identification data from Allen et al. (1991) replicated the results of
52
Ph.A. Allen et al.
Allen and Madden (1989). That is, the pattern of word frequency effects were different for young (VHF = 414 ms, MI-IF = 450 ms, LF = 407 ms, VLF = 412 ms) than for older adults (VHF = 724 ms, MHF = 735 ms, LF = 766 ms, VLF = 775 ms). Indeed, young adults revealed a significant quadratic trend (but no linear trend), however, older adults revealed a siL,nificant linear trend (but no quadratic trend) (Allen et al., 1991). The best-fitting slowing function for these letter identification data [Y = 1.36(X) + 178] (r' = .38) was also similar to that found by Allen and Madden (1989). Consequently, both of the available aging studies conducted on a letter identification task found a quadratic trend across word frequency for young adults (also see Allen & Emerson, 1991; Allen & Madden, 1990; Johnson et al., 1989) but found a linear trend across word frequency for older adults. Given the poor fit of the bestfitting slowing functions for these experiments (r 2 = .36, .38, respectively), and the different trend functions for young and older adults across word frequency, the letter identification data are more consistent with a localized model of age differences than with a generalized model. 4.3. Lexical Decision Studies.
Instead of requiring an individual to decide whether a letter matches the initial letter of a subsequently presented word (i.e., a letter identification task), a lexical decision task requires an individual to decide whether a letter string forms an English word. The Allen et al. (1991) study involved a within subjects manipulation of both a letter identification and a lexical decision task. However, for the lexical decision task, both young and older adults showed a linear trend across word frequency (young: VHF = 485, MHF = 517, LF = 572, VLF = 613; Older = VHF-- 852, MHF = 878, LF = 923, VLF = 1031). When these lexical decision data were plotted, the best-fiRing slowing function was Y = 1.70(X) -6 (r 2 = .83). When both the lexical decision and the letter identification data from Allen et al. (1991) were plotted, the bestfitting slowing function was Y = 1.48(X) + 118 (r 2 =. 84). The overall slowing function for Allen et al. (1991) could account for 84 percent of the experimental variance for both letter identification and lexical decision tasks. However, since Brinley plots are not particularly sensitive in detecting interactions (Perfect, 1994), it was necessary to test for interactions. In order to do so, we used the procedure suggested by Madden et al. (1992). In this procedure, the task condition RT means for young adults are used as an independent variable and older adults' task condition RT means form the dependent variable using linear regression methods (i.e., a Brinley plot). Next, the best-fiRing slowing function derived from the Brinley plot is used to transform young adults' raw latencies. Finally, the transformed latencies of the young adults and the untransformed latencies of the older adults were used as the dependent variable in an ANOVA. When this transformed analysis was conducted on the lexical decision and letter identification data of Allen et al. (1991) using the best-fitting slowing function to transform young adults' data, the results demonstrated that the Age x Task x Word Frequency interaction still remained significant [F(3, 138) = 3.59, p < .02]. Given the presence of this interaction in the transformed data, it was necessary to assume the existence of a different slowing function for a lexical decision task than for a letter identification task. One interpretation o f the aforementioned finding of different slowing functions for letter identification and lexical decision tasks is that these tasks simply map onto different slowing domains (e.g., Lima et al., 1991, Myerson et al., 1992). Indeed, Lima et al. (1991) classified letter processing as a non-lexical domain task but classified a lexical decision task as
Visual word encoding and the effect of adult age and word frequency
53
a lexical domain task. However, the present letter identification task clearly involves lexical access as is indicated by the presence of the word frequency effect for this task (see Allen & Emerson, 1991; Allen & Madden, 1990, for an in-depth discussion ofthis topic). Although the letter matching portion of the letter identification task does not involve lexical access, the word-level input channel that interferes with the letter-level input channel's access to the central processor does involve lexical access. Thus, both the letter identification and the lexical decision task data from Allen et al. (1991) are clearly affected by lexical access. Hence, it would seem necessary to classify both of these tasks within the lexical domain--even though both tasks reveal different slowing functions. Consequently, the present letter identification data provide a disconfirmation of the notion that a single slowing function can adequately describe age differences in the lexical domain. In another quasi-lexical decision experiment, Bowles and Poon (1981) also manipulated word frequency. In their study, however, subjects keep their fight and left index fingers on separate response keys. If both of the two letter strings that were presented simultaneously were real words, then subjects were instructed to lift a finger from one key. However, if one, or both, of the letter strings did not form real English words, then the subjects were instructed to lift the other finger from the response key. Thus, for "word" trials, there could be two "high-frequency" words (HH), two "low-frequency" words (LL), or one high-frequency word and one low-frequency word (ILL). For the "nonword" trials, there could be one nonword and one high-frequency word (HN), one nonword and one low-frequency word (LN), or two nonwords (NN). The results indicated that there was a word frequency advantage for both age groups, but age did not interact with word frequency. When a Brinley plot was computed for these data, the best-fitting least-squares regression equation was Y = 1.65(X)-396 (r 2 = .97). Based upon the slowing function data, the Bowles and Poon (1981) data are consistent with a generalized slowing model. However, the age differences for stimulus type (word vs. nonword) appear to be larger than the age differences for word frequency in the Bowles and Poon experiment (at least, based upon cell means). Although the ANOVA data were not available for additional analysis, it would have been useful to determine whether the transform analysis used above would have resulted in an interaction (i.e., evidence for localized slowing). Allen, Madden, Weber, and Groth (1993) reported three aging experiments that involved a lexical decision task and that manipulated word frequency. All three of the Allen et al. (1993) experiments presented a single letter string on each trial, and subjects were instructed to decide whether this letter string formed a real English word. In the first experiment, case type, stimulus type, and word frequency were manipulated. It was assumed that mixed-case presentation would handicap the word-level channel (because the mixed-case stimulus would result in an unfamiliar spatial frequency pattern), resulting in indirect lexical access occun~g indirectly through the letter-level input channel (Allen & Emerson, 1991). If age differences in visual word recognition occurred during the transduction stage of processing, then age should have interacted with case type. Alternatively, if age differences in visual word recognition occurred at lexical access, then age should have interacted with word frequency. Also, for lowercase presentation, it was hypothesized that the word-level channel would typically be used to achieve lexical access. The results indicated that older adults showed a larger mixed-case disadvantage than did young adults (an Age x Case Type interaction), but that both groups showed comparable word frequency advantages. Thus, these data suggested that older adults took longer to visually encode words, but that there were no appreciable age differences in lexical access. The best-fitting slowing function for these data
54
Ph.A. Allen et al.
was Y - 2.08(X) - 5 3 0 ( r 2 = .94). When the transform analysis was conducted on the data from Experiment 1, the Age x Case Type x Word Length interaction was still present. In Experiment 2 of Allen et al. (1993), word frequency, case type and spacing (0 vs. 1 space between adjacent letters)were manipulated to further examine whether age differences in visual word recognition performance were the result of encoding or of lexical access effects. There was a significant Age x Case Type x Spacing Type interaction but age did not interact with word frequency. Again, these results implied that age differences were occurring during the initial transduction process rather than during the later lexical access process. The bestfitting slowing function for Experiment 2 was Y = 1.93(X) - 495 (r 2 = .91). When this slowing function was used to carry out the Madden transform of these data, this analysis revealed that the Age x Case Type x Spacing Type interaction was eliminated, but that an Age x Word Length interaction was created. In Experiment 3 of Allen et al. (1993), case type was not manipulated. Instead, response selection load (go/no-go vs. two-choice tasks) and word frequency were manipulated. For the "go/no-go" portion of the task, subjects responded "word" if the stimulus formed a real word (a "go" trial). However, if the stimulus did not form a word, then subjects did not respond (a "no-go" trial). Alternatively for the two-choice trials, subjects always responded whether a letter string was a "word" or a "nonword." The results showed that older adults exhibited a relatively larger response selection load effect (two-choice - go/no-go RT) than did young adults, although both age groups showed the same magnitude of word frequency advantage. The best-fitting slowing function for Experiment 3 was Y = 1.60(X) 269 (r 2 = .83). When the best-fitting slowing function was used to conduct the transform analysis, the Age x Task Type interaction was eliminated, but an Age x Word Frequency interaction was created. This Age x Word Frequency interaction was especially surprising, because older adults actually showed a smaller word frequency advantage for the transformed analysis than did young adults. Such a result indicates that there was an interaction across task conditions and age in the Brinley plot. That is, more than one slowing function was needed to adequately describe age differences in Experiment 3 of Allen et al. (1993). Overall, the results of the three lexical decision experiments of Allen et al. (1993) indicated that a single-factor slowing function could not account for the greater age-related slowing found for stimulus transduction and response selection stages than for the lexical access stage of visual word recognition. It should be noted, though, that this finding is not necessarily inconsistent with a slightly less strict interpretation of generalized slowing. Namely, a domain-specific slowing model could predict a multi-factor slowing function that included different components for lexical and non-lexical domains (e.g., Lima et al., 1991; Myerson et al., 1992; Myerson et al., 1994). Although domain-specific slowing is not particularly generalized in the sense that it uses more than one slowing factor, this sort of approach is generalized in the sense that the same slowing functions should work for lexical stimuli across different tasks. The Allen et al. (1993) data, however, appear to be inconsistent with even domain-specific slowing. This is because subjects were processing words in all three of the Allen et al. (1993) experiments (i.e., within the lexical domain), yet multiple slowing functions were still needed to adequately describe the data (as demonstrated by the existence of task interactions with age for all three experiments when using the transform procedure of Madden et al., 1992). In a recent lexical decision experiment, Balota and Ferraro (1994) manipulated word frequency and repetition for young and older adults. (Very old adults as well as Alzheimer's
Visual word encoding and the effect of adult age and word frequency
55
patients were also tested, but we will limit our present examination to the young adult and the older adult age groups). This study matched young and older adults on Vocabulary subscores of the Wechsler Adult Intelligence Scale-Revised (WAIS-R) (although the Vocabulary scores of both young and older adults reported in this experiment were considerably lower than the WAIS-R Vocabulary scores reported by Allen et al., 1991, 1993). Balota and Ferraro found a slightly larger word frequency effect for healthy older adults (compared to young adults), and this effect was even larger for non-repeated than for repeated words. It is important to note, though, that this effect was relatively small--older adults' word frequency effect was only 19 ms larger than young adults' word frequency effect. When the Balota and Ferraro (1994) data are analyzed using a Brinley plot, the best-fitting, least-squares regression function is Y 1.3(X) - 36 (r 2 - .99). Tainturier et al. (1992) conducted a lexical decision task that manipulated word frequency, but these researchers also manipulated years of education. Tainturier et al. (1992) found a larger word frequency effect for individuals with fewer years of education, but age did not interact with word frequency. That is, young and older adults showed the same word frequency advantage, but it took individuals with fewer years of education longer to access their mental lexicon (compared to individuals with more years of education). When the Tainturier et al. (1992) data are analyzed using a Brinley plot, the best-fitting slowing function is Y = 1.72(X) - 269 (r 2 = .91). Although it appears that the results of Allen, Madden, and Crozier (1991, Allen et al., 1993) and Tainturier et al. (1992) are inconsistent with those of Balota and Ferraro (1994), the discrepancy may actually be the result of differences in reading level and/or verbal ability. For example, Frederiksen (1978) found that poor readers showed larger word frequency effects than did average or superior readers. We suspect that, in general, healthy older adults tend to have larger vocabularies than do healthy young adults. Thus, if the Balota and Ferraro (1994) sample of young and older adults were relatively poor readers (especially the older adults who had WAIS-R Vocabulary scores that were one-third lower than those reported by Allen et al., 1991, 1993, in four different samples of participants), then this might explain the discrepancy across studies. 4.4. Pronunciation Studies.
Balota and Ferraro (1993) reported an aging study that examined pronunciation onset (or naming) performance. Word frequency and regularity of orthographic-phonological (O-P) correspondence were manipulated (e.g., "pint" is irregular whereas "mint" is regular). The Balota and Ferraro naming study found that older adults showed a slightly larger word frequency effect than did young adults, but that there were no age differences in O-P correspondence (regularity). Balota and Ferraro (1993) did not report RT means so it was not possible to conduct a Bfinley analysis or a Madden transform analysis. However, these authors used a similar approach to that used by Allen et al. (1993). Namely, Balota and Ferraro hypothesized that the regularity effect measured a different stage of processing than did the word frequency effect. Thus, these data appear to provide yet another example in which agerelated slowing apparently occurred during one stage of processing, but not at another. The Balota and Ferraro (1993) study matched young and older adults on Boston Naming Test performance, so it is difficult to compare subjects' verbal ability in this study to subjects in other studies that used the WAIS-R Vocabulary subscale test. It is probably a fair
56
Ph.A. Allen et al.
generalization, though, that, older adults tend to have larger vocabularies than young adults because of older adults' greater experience with words (as determined by the WAIS-R Vocabulary subscale scores in numerous aging studies). Thus, if we compared a sample of older adults at the 50th percentile rank for WAIS-R vocabulary scores for the population of older adults to an equally educated sample of young adults at the 50th percentile rank for WAIS-R vocabulary scores for the population of young adults, it is quite likely that there would be no age differences in word frequency effects, or that older adults would even show slightly smaller word frequency effects. However, if we match for vocabulary performance, it is quite likely that we are dealing with a less-intelligent sample of older adults than young adults in terms of their percentile ranks within their respective age groups (i.e., a regression artifact). Thus, although aging studies that match on the basis of vocabulary scores are useful, one should keep in mind that such matching will probably unfairly penalize older adults. Allen, Madden, Cerella, Jerge, and Betts (1994) recently reported the results of four naming experiments that varied case type, word frequency, and number of perturbed pixels. In all four experiments, subjects pronotmced 5-letter and 6-letter, two-syllable words and pronounceable nonwords (although subjects also performed a lexical decision task in Experiment 2). In Experiment 1, letter strings were presented in consistent lowercase (LC, e.g., "hello"), mixed-case by syllables (SYL, e.g., "HELlo"), and mixed-case by adjacent letters (MC, e.g., "hElLo"), and the same four word frequency categories were used as were used in Allen et al. (in press). The results revealed a larger mixed-case disadvantage for older adults than for young adults, and this was especially so for MC versus LC trials (LC and SYL RTs were similar). However, there was no age difference for the word frequency advantage. When these data were analyzed using Brinley plots, it was determined that two slowing functions were necessary--one for the MC condition [Y = 1.23(X) - 78], and one for the combined LC and SYL conditions [Y = 1.23(X) - 135]. When this three-factor model (i.e., one slope, but two intercepts) was used in the transform analysis, there were no Age x Task interactions. Allen et al. (1994) hypothesized that the slope parameter measured central (or more cognitive) processes, but that the intercepts measured perceptual-motor processes (in this case, the amount of perceptual normalization that needed to be conducted). Experiment 2 of Allen et al. (1994) was the same as Experiment 1, except that Experiment 2 contained both naming and lexical decision tasks (this was a within subjects factor). There was an Age x Case Type interaction, but age did not interact with either word frequency or task type. Because a naming task and a lexical decision task have different decision stage demands (e.g., Balota & Chumbley, 1984), the finding that the Age x Case Type interaction did not further interact with task type indicated that the locus of the age difference in the mixed-case disadvantage was not at the decision stage but, rather, at the hypothesized encoding stage (e.g., Allen et al., 1993). Furthermore, it was necessary to use a separate slowing function for the MC condition than for the combined LC and SYL conditions for both the naming [MC: Y = 1.30(X)- 8); LC & SYL: Y - 1.30(X)- 119] and the lexical decision tasks [MC: Y = 1.28(X) + 106); LC & SYL = Y = 1.28(X) - 56]. When three-factor models were used to transform both the naming and lexical decision data of young adults, the transform analyses revealed no Age x Task interactions. Consequently, the results of Experiment 2 of Allen et al. (1994) required an interpretation based upon the use of processspecific slowing because two different slowing functions were required to adequately describe the data. Experiment 3 of Allen et al. (1994) was the same as Experiment 1, except Experiment 3 used both brief presentation and presentation-until-response exposure durations
Visual word encoding and the effect of adult age and word frequency
57
(Experiments 1, 2, and 4 used all presentation-until-response exposure durations). The results from Experiment 3 revealed an Age x Case Type interaction, but exposure duration did not further interact with these variables. Also, age did not interact with word frequency. The finding that exposure duration did not affect the magnitude of the age difference for the mixedcase disadvantage suggested that this mixed-case disadvantage effect was not the result of age differences in retinal illuminance, but rather due to some latter-occurring perceptual process (Allen et al., 1994). As was the case for earlier experiments, a separate slowing function was required for the LC condition and for the MC and SYL conditions for both exposure durations [100 ms exposure duration: LC: Y = 0.92(X) + 222; MC & SYL: Y = 0.92(X) + 157; presentation-until-response: LC: Y = 0.92(X) + 229; MC & SYL: Y = 0.92(X) + 162]. When these slowing functions were used to transform the young adults' RT from Experiment 3, the transformed ANOVA eliminated all Age x Task interactions for the brief exposure duration and all but one Age x Task interaction for the longer exposure duration condition (this lone interaction was the result of a non-monotonic cell for older adults in the presentation-untilresponse condition). Thus, the data from Experiment 3 of Allen et al. (1994) also suggested that older adults snow greater slowing at the encoding stage (i.e., the perceptual-motor parameters) than at the lexical access stage (i.e., the central stage). Finally, Experiment 4 of Allen et al. (1994) manipulated the number of perturbed pixels (0, 2, or 4) rather than case type. The notion was that this procedure would result in the wordlevel channel being able to output a code even as the number of perturbations increased. That is, case mixing prevents the hybrid model from outputting a word-level code because the holistic pattern is non-linear, and, therefore, Fourier synthesis would not be successful (Allen et al., in press). However, pixel perturbation would result in a linear transformation, thus, the word-level channel would be able to form a code even when four pixels per letter were perturbed. The results from Experiment 4 revealed an Age x Pixel Perturbation Type interaction, but no Age x Word Frequency interaction. These findings suggested that even when just the word-level channel was used for processing, there were still age differences in encoding, but not age differences in lexical access. It was necessary to use separate slowing functions for the four-pixel perturbation condition [Y = 0.98(X) + 294] and for the combined zero-pixel and 2-pixel perturbation conditions [Y = 0.98(X) + 147]. As was the case for all three earlier experiments, when a three-factor slowing function based upon these two twofactor slowing functions was used to transform the young adults' data, the resulting transformed ANOVA eliminated all interactions with age. In summary, all four of the Allen et al. (1994) naming experiments required more than one slowing fimction to account for the encoding data. Allen et al. used the F-test suggested by Fisk, Fisher, and Rogers (1991) to test for whether a given model accounted for more variance than another. Specifically, in all four experiments, the three-factor model (i.e., two two-factor models that shared the same slope but required different intercepts) accounted for significantly more variance than did the best-fitting single two-factor model. However, the best-fitting four-factor model (i.e., two two-factor models--each with different slopes and intercepts) did not account for significantly more variance than did the three-factor model. These data indicated that a single factor can account for lexical access speed differences across age, but that two factors are necessary in order to account for encoding performance across age.
58
Ph.A. Allen et al.
5. PART V: W H E R E DO WE STAND? After having reviewed the literature on age differences in visual word recognition when word frequency was manipulated, we now can ask two central questions. First, can we account for the letter identification, lexical decision, and naming data using a single framework? Secondly, can these data be described adequately with a single factor, or is it necessary to allude to process-specific factors? The answer to the first question appears to be "yes." That is, the age differences data from letter identification (Allen & Madden, 1989; Allen et al., 1991), lexical decision (e.g., Allen et al., 1993; Balota & Ferraro, 1993; Tainturier et al., 1992), and naming (e.g., Allen et al., 1994; Balota & Ferraro, 1994) tasks can be accounted for using the hybrid model of Allen et al. (in press) (refer to Figure 1). 5.1. Letter Identification
For a letter identification task using visual presentation, the model can account for older adults showing a word frequency advantage and the young adults showing a nonmonotonic effect across word frequency by assuming that the older adults have difficulty in forming letter-level codes (Allen, Madden, & Crozier, 1991). This is because letter-level codes require more processing resolution that do word-level codes (Allen & Emerson, 1991). Thus, older adults must use the segmentation process (Allen & Emerson, 1991) via the central processor to break down a word-level code into a pseudo-letter-level code so that the letter matching process of the letter identification task can be conducted. Because older adults initially form a word-level code (and this involves lexical access), the model predicts a word frequency advantage for older adults. Indeed, Allen, Madden, and Crozier found that older adults showed approximately the same linear trend across word frequency for a letter identification task as for a lexical decision task. However, young adults revealed a quadratic trend for a letter identification task (as would be predicted if the word-level and the letter-level input channels were involved in a processing horse race), but a word frequency advantage for a lexical decision task (Allen, Madden, & Crozier, 1991). Hence, older adults tend to rely upon the word-level input channel information for both letter identification and lexical decision tasks. For young adults, though, the word-level input channel interfered with the letter-level input channel for a letter identification task, although the letter-level channel information was typically used to conduct this task (Allen & Emerson, 1991; Allen & Madden, 1990). Alternatively for a lexical decision task, young adults typically used the word-level input channel to conduct the task (Allen et al., 1992, 1993) unless case mixing or some other stimulus manipulation made the word-level (holistic) stimulus unfamiliar (Allen et al., in press). 5.2. Lexical Decision and Naming Tasks
~he hybrid model predicts a larger mixed-case disadvantage for older adults than for young adults on a lexical decision task because the a mixed-case word must be initially encoded using the letter-level input channel (Allen et al., 1993, in press). This letter-level code of the word must then be superposed into a pseudo-word-level code by the central processor so that lexical access can occur. However, the superposition task is processing resource intensive, and this is particularly difficult for older adults who already have fewer processing
Visual word encoding and the effect of adult age and word frequency
59
resources than young adults (e.g., Allen et al., 1993; Madden, 1986, 1990). Note that the superposition task is really a normalization process conducted on the letter-level code. When pixels of letters are perturbed (e.g., Allen, Madden, Cerella, Jerge, & Betts, 1994, Experiment 4), the hybrid model predicts that this sort of stimulus can be normalized by the word-level channel. However, in both cases (Age x Case Type or Age x Perturbation Type), the hybrid model accounts for task interactions with age by assuming that older adults are especially affected by the processing resource-intensive normalization procedures. For a naming task, both young and older adults typically used the phonological pathway rather than the orthographic pathway used for a lexical decision task (e.g., Allen et al., 1994; Balota & Ferraro, 1993). The syllable-level input channel typically wins the race to the central processor (rather than the letter-level, GPC channel, see Figure 1), and this code is used to form a motor code for naming. However, this holistic input channel can be handicapped by presenting an unfamiliar syllable-level stimulus (e.g., by mixing case within a syllable). Under such circumstances, the letter-level input channel is used (which uses grapheme-to-phoneme correspondence, or GPC rules) to perform a naming task (Allen, Madden, Cerella, Jerge, & Betts, 1994). As was the case for the lexical decision data, the hybrid model accounts for age differences in naming as a function of stimulus case type by alluding to an age difference in normalization performance. Namely, the model predicts that older adults will be adversely affected by increased normalization requirements of a task. Thus, when case is mixed within a word, the syllable-level, phonological channel is handicapped, and the input code from the slower GPC channel must be superposed so that lexical access can occur (i.e., GPC rules allow naming but not direct lexical access). The GPC channel deals with smaller pieces of information than does the syllable-level channel, thus, the GPC channel requires greater processing resources in order to normalize a stimulus. When the stimulus is particularly unfamiliar as is the situation for nonwords, the age difference in encoding should even be larger. Also, when pixels are perturbed within letters of a word on a naming task, then a normalization process must be carried out on the input stimulus pattern in order to reco~ize the word. That is, the syllable-level channel is not prevented from outputting a code, but larger levels of normalization are required in order to cohere the percept. Once again, an age difference in processing resources would result in a larger normalization cost for older adults. 5.3. General or Process-Specific Slowing?
The aging literature on visual word recognition for which word frequency is manipulated provides a solid argument for the idea that process-specific slowing occurs (Allen & Madden, 1989, 1991; Allen et al., 1993; Allen, Madden, Cerella, Jerge, & Betts, 1994; Balota & Ferraro, 1993). However, there is evidence suggesting that there exists a generalized slowing component, as well (Allen, Madden, Cerella, Jerge, & Betts, 1994). In the present section, we will outline a theory that can account for this seeming paradox. However, before we present this theoretical account, the basic findings will be summarized from the three tasks (letter identification, lexical decision, and naming tasks) using the three methods of analysis for determining whether localized or generalized slowing is present (i.e., Brinley plots, the transform proposed by Madden et al., 1992, and hierarchical regression).
60
Ph.A. Allen et al.
5.4. Letter Identification
The letter identification data reviewed (Allen & Madden, 1989; Allen, Madden, & Crozier, 1991) are qualitatively different from the lexical decision and naming data because there is strong evidence indicating that older adults typically used a different processing channel (word-level) to conduct a letter identification task than did young adults (letter-level). This is the case even though both age groups possessed the same basic processing architecture (i.e., both word-level and letter-level input channels and a central processor, see Allen et al., 1991). The word frequency effects found for this task demonstrated that lexical access did occur (Allen & Emerson, 1991; Allen & Madden, 1990), therefore, the task falls within the lexical domain (Lima et al., 1991). The Brinley plots for both of the reported experiments (Allen & Madden, 1989; Allen et al., 1991) accounted for less than 40% of the total variance, and older adults showed a significant linear trend while young adults revealed a significant quadratic trend for the letter identification task (Allen et al., 1991). Consequently, the most parsimonious manner to interpret these data is to assume that older adults tend to use topdown processing via the word-level input channel to conduct letter identification, whereas YOung adults tend to use bottom-up processing via the letter-level input channel for a letter identification task. Clearly these data support the concern of Fisher et al. (the present volume) that young and older adults may not always use the same type of processing for a given task. 5.5. Lexical Decision Task
The seven experiments reviewed that examined age differences for a lexical decision task that manipulated word frequency (Allen, Madden, & Crozier, 1991; Allen et al., 1993, Experiments 1-3; Balota & Ferraro, 1994; Bowles & Pooh, 1981; Tainturier et al., 1992) suggested that older adults and younger adults did use the same processing stages. For example, Allen et al., 1991, found significant linear trends for both age groups for a lexical decision task. Also, there were no appreciable age differences in word frequency effects-especially if the older adult sample was not limited to individuals with rather low vocabulary scores (based upon percentile rank within an age group). These lexical decision data did reveal Brinley plots for individual experiments with relatively high r-squared values (typically < .90). Also, when the transformed analysis suggested by Madden et al. (1992) was conducted on the data from Allen et al. (1991, 1993), some (but not all) task interactions with age were eliminated. (It should be noted that the necessary data was not available in order to conduct the transformed analyses on the other experiments.) These results suggest that there is a component of generalized slowing that does occur for older adults on a lexical decision task. However, the transformed analyses still revealed Age x Task interactions for all four experiments. This finding indicated that there were significant interactions for these Brinley plots within the same lexical processing domain, thus, these data indicated that process-specific slowing occurred for these lexical decision data (in addition to the previously mentioned generalized slowing). 5.6. Naming The five reviewed naming studies (Allen, Madden, Cerella, Jerge, & Betts, 1994, Experiments 1-4; Balota & Ferraro, 1993) showed quite similar results as did the previously
Visual word encoding and the effect of adult age and word frequency
61
discussed lexical decision experiments. Both age groups showed word frequency advantages, and these word frequency effects were quite similar. The Brinley plots for the individual naming experiments revealed r-squared values that tended to account for at least 90% of the total variance (excluding the Balota & Ferraro, 1993, study for which no word frequency means were reported). However, the Allen et al. (1994) Experiments (1-4)required a threefactor model (two different slowing functions with the same slopes but different intercepts). Based on Cerella (1985), slopes are assumed to measure age differences in central (lexical access, retrieval, and decision) processes and y-intercepts are assumed to measure peripheral (encoding, response selection, and response execution) processes. A slope of approximately 1.00 suggests little age appreciable differences at the central processing stage, and a positive intercept suggests age-related slowing at the peripheral stage (Cerella, 1985). Therefore, the present naming data from Allen et al. (1994, Experiments 1-4) suggested slight general slowing for the central processes (i.e., a slope of slightly greater than 1.00), but differential slowing for the peripheral processing stages depending upon processing load. These data provided concrete support for the idea that there is process-specific age-related slowing for naming tasks. Finally, hierachical regression conducted on Experiments 1-2 of Allen et al. (1994) continued to find that age was a significant predictor of mixed-case performance even when the variance associated with lowercase performance was extracted first (i.e., before the variance associated with age was entered into the model). However, age no longer si~ificantly predicted mixed-case performance (or the 4-pixel perturbation condition of Experiment 4) for Experiments 4 of Allen et al. (1994). (When SYL RT instead ofLC RT was used in Experiment 3, age continued to predict MC RT even when SYL RT variance was extracted first. However, the aberrant cell for older adults' presentation-until-response, MHF, LC condition prevented age from predicting MC RT when LC RT variance was extracted first.) The hybrid model (Allen et al., in press) predicted the obtained results for Experiment 4 (because the model predicts that all levels of pixel perturbation will be encoded by the same channel, but that lowercase and mixed-case words will be encoded by different channels). Thus, the finding that the age variable's ability to predict mixed-case performance was attenuated (Experiments 1-3) suggests that there was some general factor accounting for a si~ificant portion of the total variance (Salthouse & Coon, 1994). However, the finding from Experiments 1-3 that age still predicted mixed-case performance even when the variance associated with lowercase or mixed-case by syllable performance was extracted first provided evidence that a generalized factor, alone, was not sufficient to account for these results (also see Madden, 1992, for a similar finding). Consequently, the extant naming data that include a manipulation of word frequency suggest that there are both generalized and process-specific components to age differences in visual word processing. 5.7. The "Big Picture" So far, we have discussed only Brinley plots for individual experiments. Lima et al. (1991) and Myerson et al. (1992) have proposed that all lexical domain tasks can be described
using a single slowing function of the form: Y = 1.50(X) - 68 (in which "Y" refers to older adults' predicted RT and "X" refers to young adults' actual RT). To test this prediction, we entered the data from all 13 presently reviewed experiments that contained cell means into a single Brinley plot (see Figure 3) (two letter identification experiments: Allen & Madden,
62
Ph.A. Allen et al.
1989; Allen et al., 1991; seven lexical decision experiments: Allen et al., 1991, 1993, Experiments 1-3; Balota & Ferraro, 1994; Bowles & Pooh, 1981; Tainturier et al., 1992; and four naming studies: Allen et al., 1994, Experiments 1-4). The best-fitting linear slowing function for all 13 data sets (190 data points) was: Y = 1.12(X) + 158 (r2 = .79). However, the six cell means from the Bowles and Pooh (1981) experiment contained RTs that were 700 ms longer than any of the latencies from other experiments. Both Fisk, Fisher, and Rodgers (1991) and Perfect (1994) have argued that such extreme outliers can result in artificially high r-squared values on Brinley plots. Therefore, we re-ran the overall Brinley analysis with 12 data sets (i.e., excluding the Bowles & Pooh, 1981, data). The resulting best-fitting linear slowing fimction (184 data points) was: Y = .96(X) + 250 (r2 = .67). The finding that the single best-fitting slowing function could account for only 67% of the total variance of 12 aging experiments that manipulated word frequency is clearly inconsistent with the results ofLima et al. (1991) and Myerson et al., (1992) who found that a single slowing function could account for over 90% of the total variance of lexical domain aging
Word Frequency Meto-Anolysis
1 600
-
1 5O0 1400
OmoO ~
1300 _ 1200
CL~ L. (D -(3
0
~176 ~oo o'~l 9
llO0 1000 900 8OO 7O0 6OO 5OO 3OO
,100
500
600
700
800
900
1000
1100
1200
1300
1400
Young RT
Figure 3. data. With 33% of the variance unexplained, one would be hard-pressed to claim that a single two-factor model (i.e., a model with one slope and one intercept) adequately accounted for these data. Also, both of the overall Brinley plots (using 13 and 12 data sets, respectively) revealed slowing fimctions with slopes approaching 1.00 and large, positive intercepts. Thus, these data examining age differences in visual word recognition for which word frequency was manipulated suggest that there is little or no age-related slowing for central processes, but that
Visual word encoding and the effect of adult age and word frequency
63
there is considerable slowing occurring at the peripheral stages of processing (e.g., encoding, response selection, and response execution). The results of Allen et al. (1993; Allen, Madden, Cerella, Jerge, & Betts, 1994), in particular, suggest that age differences are especially pronounced for visual word encoding. Again, note that the presently obtained slopes and intercepts are inconsistent with slowing function for the lexical domain proposed by Lima et al. (1991). That is, the present slopes were approximately 1.00 whereas the slope found by Lima et al. (1991) and Myerson et al. (1992) was 1.50. Also, the present intercepts for both the 12and 13-experiment data sets were positive and greater than 150 ms, whereas the intercept for the Lima et al. (1991) slowing function was slightly negative (-68 ms). The differences between the present overall Brinley plot and that proposed by Lima et al. (1991) and Myerson et al. (1992) suggest that a single slowing fimction cannot account for all lexical domain data for age differences. Indeed, the overall Brinley plot data combined with the previously discussed Madden et al. (1992) transform data from Allen et al. (1991, 1993, 1994), and the hierarchical regression data from Allen et al. (1994) all indicate the process-specific age-related slowing occurs for visual word recognition studies in the lexical domain. 5.7. A General Model
We now have evidence from the data of nearly 500 subjects from visual word recognition studies that manipulated word frequency (i.e., the present set of experiments) that age differences are essentially additive in nature (slopes of close to 1.00 and large, positive intercepts). However, this finding seems to be at odds with the Cerella (1985) meta-analysis, because Cerella found larger age differences for central processing stages than for peripheral processing stages. This seeming paradox can be accounted for, though, if we assume that the 35 experiments analyzed in the Cerella meta-analysis (i.e., none of these were lexical domain tasks) examined episodic and procedural memory tasks, but that the present experiments examined semantic memory processing. We propose that tasks that involve lexical access (e.g., letter identification, lexical decision, and naming) or the retrieval of highly memorized arithmetic facts (e.g., Allen, Ashcrafi, & Weber, 1992; Geary & Wiley, 1991; Geary, Frensch, & Wiley, 1993) will tend to show little age differences in lexical access or arithmetic fact retrieval, but that there will still be substantial age differences at the encoding and/or response selection stages of processing. Alternatively, in tasks for which decision-stage processing does not involve semantic memory (e.g., visual search, memory search, or letter matching), we propose that there will be larger decision stage age decrements than encoding and/or response selection age decrements (e.g., Cerella, 1985). Of course, this is a preliminary formulation (but also see Allen et al., 1992). Furthermore, it does appear that semantic priming may complicate the issue (the Myerson et al., 1992, results that found a slope of 1.50 were based upon a metaanalysis of semantic priming studies--although also see Laver & Burke, 1993, for a processspecific interpretation of the semantic priming data on aging). However, this basic model of semantic memory expertise nullifying most age differences in decision stages does have considerable heuristic appeal, and can also account for results such as those reported by Stine (in the present volume) in which older adults apparently encode words more slowly yet still evidence comparable reading times to young adults.
64
Ph.A. Allen et al.
REFERENCES
Adams, M.J. (1979). Modds of word recognition. Cognitive Psychology, 11, 133-176. Allen, P.A. (1990). Influence of processing variability on adult age differences in memory distribution of order information. Cognitive Development, 5, 177-192. Allen, P.A. (1991). On age differences in processing variability and scanning speed. Journal of Gerontology: Psychological Sciences, 46, P19 l-P201. Allen, P.A., & Coyne, A.C. (1988). Age differences in primary organization or processing variability? Part II: Evidence for processing variability. Experimental Aging Research, 14, 150-156. Allen, P.A., & Madden, D.J. (1989). Adult age differences in the effects of word t~equency during visual letter identification. Cognitive Development, 4, 283-294. Allen, P.A., & Emerson, P.L. (1991). Holism revisited: Evidence for parallel independent word-level and letter-level processors during word recognition. Journal of Experimental Psychology: Human Perception and Performance, 17, 489-511. Allen, P.A., & Madden, D.J. (1990). Evidence for a parallel input serial analysis model of word processing. Journal of Experimental Psychology: Human Perception and Performance, 16, 48-64. Allen, P.A., Madden, D.J., & Crozier, L.C. (1991). Adult age differences in letter-level and word-level processing. Psychology and Aging, 6, 261-271. Allen, P.A., Wallace, B, & Waag, E. (1991). Effect of imagery ability on letter-level and wordlevel processing. Perception & Psychophysics, 49, 295-300. Allen, P.A., Ashcrafi, M.H., Weber, T.A. (1992). On mental multiplication and age. Psychology and Aging, 7, 536-545. Allen, P.A., McNeal, M., & Kvak, D. (1992). Perhaps the lexicon is coded as a function of word frequency. Journal of Memory and Language, 31, 826-844. Allen, P.A., Madden, D.J., Weber, T.A., & Groth, I~E. (1993). Influence of age and processing stage on visual word recognition. Psychology and Aging, 8, 274-282. Allen, P.A., Kaufman, M., & Propper, 1~ (1994). The psychophysics of aging and entropy Part I: Very-short-term memory. Submitted manuscript. Allen, P.A., Kaufman, M., & Popper, 1~ (1994). The psychophysics of aging and entropy Part II: Memory storage and retrieval. Submitted manuscript. Allen, P.A., Patterson, M.B., Propper, R.E. (1994). Influence of letter size on age differences in letter matching. Journal of Gerontology: Psychological Sciences, 49, P24-P28. Allen, P.A., Wallace, B., & LoSchiavo, F. (1994). Influence of imaging ability on word transformation. Memory & Cognition, 5, 565-574. Allen, P.A., Weber, T.A., & Madden, D.J. (1994). Adult age differences in attention: Filtering or selection? Journal of Gerontology: Psychological Sciences, 49, P213-P222. Allen, P.A., Cerella, J., Madden, D.J.,Jerge, K., & Betts, L. (1994). Age differences in naming as a function of wordfrequency and case type. Unpublished manuscript. Allen, P.A., Wallace, B., & Weber, T.A. (in press). Influence of case type, word t~equency, and exposure duration on visual word recognition. Journal of Experimental Psychology: Human Perception and Performance. Amrhein, P.C., & Theios, J. (1993). The time it takes elderly and young individuals to draw pictures and to write words. Psychology and Aging, 8, 197-206.
Visual word encoding and the effect of adult age and wordfrequency
65
Amrhein, P.C. (in press). Evidence for task specificity in age-related slowing: A review of speeded picture-word processing studies. In P. Allen and T. Bashore (Eds.), Age differences in word and language processing. New York: North-Holland. Baddeley, A.D., & Hitch, G. (1974). Working memory. In G. Bower (Ed.), Recent advances in learning and motivation (Vol. 8). New York: Academic Press. Balota, D.A., & Chumbley, J.I. (1984). Are lexical decisions a good index of lexical access? The role of word frequency in the neglected decision stage. Journal of Experimental Psychology: Human Perception and Performance, 10, 340-357. Balota, D.A., & Chumbley, J.I. (1985). The locus of word frequency effects in the pronunciation task: Lexical access and/or production? Journal of Memory and Language, 24, 340-357. Balota, D.A., & Duchek, J.M. (1988). Age-related differences in lexical access, spreading activation, and simple pronunciation. Psychology and Aging, 3, 84-93. Balota, D.A., & Ferraro, F.1L (1993). A dissociation of frequency and regularity effects in pronunciation performance across young adults, older adults, and individuals with senile dementia of the Alzheimer's type. Journal of Memory and Language, 32, 573-592. Balota, D.A., & Ferraro, F.1L (1994). Lexical, sublexical, and implicit memory processes in healthy young, healthy older adults and in individuals with senile dementia of the Alzheimer's type. Manuscript submitted for publication. Bashore, T.1L, Osman, A., & Heffley, E.F. (1989). Mental slowing in elderly persons. A cognitive psychophysiological analysis. Psychology and Aging, 4, 235-244. Becker, C.A. (1976). Semantic context effects in visual word recognition: An analysis of semantic strategies. Journal of Experimental Psychology: Human Perception and Performance, 2, 556- 566. Besaer, D., & McCann, 1LS. (1987). Word frequency and pattern distortion in visual word identification and production: An examination of four classes of models. In M. Coltheart (Ed.), Attention and Performance XII: The psychology of reading. Hillsdale, N.J.: Erlbaum. Besner, D., & Johnston, J.C. (1989). Reading and the mental lexicon: On the uptake of visual information. In W. Marslen-Wilson (Ed.), Lexical representation and process. (pp. 291316). Cambridge, MA: MIT Press. Besner, D., Twilley, L., McCann, R.S., & Seergobin, K. (1990). On the association between connectionism and the data: Are a few words necessary? Psychological Review, 97, 432446. Biederman, I. (1987). Recognition-by-components: A theory of human image understanding. Psychological Review, 94, 115-147. Birren, J.E. (1965). Age changes in the speed of behavior: Its central nature and physiological correlates. In A.T. Welford & J.E. Birren (Eds.), Behavior, aging, and the nervous system. Springfield, IL: Charles C. Thomas. Bowles, N.L., & Poon, L.W. (1981). The effect of age on speed of lexical access. Experimental Aging Research, 7, 417-426. Brinley, J.F. (1965). Cognitive sets, speed and accuracy of performance in the elderly. In A.T. Welford & J.E. Birren (Eds.), Behavior, aging, and the nervous system. Springfield, IL: Charles C. Thomas. Brooks, 1LA. (1981). Symbolic reasoning among 3-D models and 2-D images. Artificial Intelligence, 17, 205-244.
66
Ph.A. Allen et al.
Burke, D.M., White, H., & Diaz, D.L. (1987). Semantic priming in young and old adults: Evidence for age constancy in automatic and attentional processes. Journal of Experimental Psychology: Human Perception and Performance, 13, 79-88. Cart, T.H. (1992). Automaticity and cognitive anatomy: Is word recognition "automatic?" American Journal of Psychology, 105, 201-237. Cart, T.H., & Pollatsek, A. (1985). Recognizing printed words: A look at current models. In D. Besner, T.G. Waller, & G.E. MacKinnon (Eds.), Reading research: Advances m theory and practice (vol. 5). New York: Academic Press. Cart, T.H., & Posner, M. (1992). The impact of learning to read on the functional anatomy of language processing. In B. de Gelder and J. Morais (Eds.), Language and literacy: Comparative approaches. Cambridge, MA: MIT Press. Cattell, J.M. (1886). The time taken up by cerebral operations. Mind, 11, 220-242. Cerella, J. (1985). Information processing rates in the elderly. Psychological Bulletin, 98, 6783. Cerella, J. (1990). Aging and information processing rate. In J.E. Bitten and ICW. Schaie (Eds.) Handbook of the psychology of aging (3rd ed., pp. 201-221). San Diego, CA: Academic Press. Cerella, J. (1991). Age effects may be local, not global: Comment on Fisk and Rogers (1991). Journal of Experimental Psychology: General, 120, 235-243. Cerella, J. (1994). Generalized slowing in Brinley plots. Journal of Gerontology: Psychological Sciences, 49, P65-P71. Cerella, J., & Hale, S. (1994). The rise and fall of information processing rates over the life span. Acta Psychologica, 86, 109-197. Cerella, J., & Fozard, J.L. (1984). Lexical access and age. Developmental Psychology, 20, 235-243. Cerella, J., Poon, L., & Williams, D.M. (1980). A quantitative theory of mental processing time and age. In L.W. Poon (Ed.), Aging in the 1980's: Psychological issues. Washington, D. C.: American Psychological Association. Collins, A.M. & Quillian, M.R. (1969). Retrieval time from semantic memory. Journal of Verbal Learning and Verbal Behavior, 8, 240-247. Coltheart, M., Davelaar, E., Jonasson, J., & Besner, D. (1977). Access to the internal lexicon. In S. Dornic (Ed.), Attention and performance V1. London: Academic Press. Coltheart, M., Curtis, B., Atkins, P., & Hailer, M. (1993). Models of reading aloud: Dualroute and Parallel-distributed-processing approaches. Psychological Review, 100, 589-608. Connine, C.M., Mullennix, J., Shemoff~ E., & Yelen, J. (1990). Word familiarity and frequency in visual and auditory word recognition. Journal of Experimental Psychology: Learning Memory and Cognition, 16, 1084-1096. Derrington, A., & Lennie, P. (1984). Spatial and temporal contrast sensitivities of neurons in the lateral geniculate nucleus of the macaque. Journal of Physiology, 357, 219-240. DeValois, ILL., & DeValois, I~I~ (1987). Spatial vision. New York: Oxford University Press. Dobbs, A.R., Friedman, A., & Lloyd, J. (1985). Frequency effects in lexical decisions: A test of the verification model. Journal of Experimental Psychology: Human Perception and Performance, 11, 81-92. Fera, P., & Besner, D. (1992). The process of lexical decision: More words about a parallel distributed processing model. Journal of Experimental Psychology: Learning, Memory, and Cognition, 18, 749-764.
Visual word encoding and the effect of adult age and wordfrequency
67
Fisher, D.L., Fisk, A.D., & Duffy, S.A. (in press). Why latent models are needed to test hypotheses about the flowing of word and language processes in older adults. In P. Allen and T. Bashore (Eds.), Age differences in word and language processing. New York: North-Holland. Fisk, A.D., & Fisher, D.L. (1994). Brinley plots and theories of aging: The explicit, muddled, and implicit debates. Journal of Gerontology: Psychological Sciences, 49, P81-P89. Fisk, A.D., Fisher, D.L., & Rodgers, W.A. (1992). General slowing alone cannot explain agerelated search effects: A reply to Cerella (1991). Journal of Experimental Psychology: General, 121, 73-78. Fisk, A.D., & Kodgers, W.A. (1991). Toward an understanding of age-related memory and visual search effects. Journal of Experimental Psychology: General, 120, 131-149. Fodor, J. A. (1983). The modularity of the mind. Cambridge, MA: MIT Press. Forster, I(I. (1976). Accessing the mental lexicon. In 1LJ. Wales & E. Walker (Eds.), New approaches to language mechanisms (pp. 89-109). Amsterdam: North-Holland. Forster, I~I. (1979). levels of processing and the structure of the language processor. In W.E. Cooper & E. Walker (Eds.), Sentence processing: Psycholinguistic studies presented to Merrill Garrett (pp. 27-85). Hillsdale, NJ: Erlbaum. Forster, I~I. (1989). Basic issues in lexical processing. In W. Marslen-Wilson (Ed.), Lexical representation andprocess (pp. 75-107). Cambridge, MA: MIT Press. Frederiksen, J.1L (1978). Assessment oflexical decoding and lexical skills and their relation to reading proficiency. In A.M. Lesgold, J.W. Pellegrino, S.D. Fokkema, and 1L Glaser (Eds.), Cognitive psychology and instruction (pp. 153-169). New York: Plenum Geary, D.C., & Wiley, J.G. (1991). Cognitive addition: Strategy choice and speed-ofprocessing differences in young and elderly adults. Psychology and Aging, 6, 474-483. Geary, D.C., Frensch, P., & Wiley, J.G. (1993). Simple and complex mental subtraction: Strategy choice and speed-of-processing differences in younger and older adults. Psychology and Aging, 8, 242-256. Gibson, E. (1965). Learning to read. Science, 148, 1066-1072. Graham, N. (1981). The visual system does a crude Fourier analysis of patterns. SIAM-AMS Proceedings, 13, 1-16. Hansen, D., & Rodgers, T.S. (1968). An exploration ofpsycholinguistic units in initial reading. In I~S. Goodman (Ed.), The psycholinguistic nature of the reading process. Detroit: Wayne State University Press. Healy, A.F. (1976). Detection errors on the word "the": Evidence for reading units larger than letters. Journal of Experimental Psychology: Human Perception and Performance, 2, 235-242. Healy, A. F., Oliver, W. L., & McNamara, T. P. (1987). Detecting letters in continuous text: Effects of display size. Journal of Experimental Psychology: Human Perception and Performance, 13, 279-290. Healy, A.F., Conboy, G.L., Drewnowski, A. (1987). Characterizing the processing units of reading: Effects of intra-and interword spaces in a letter detection task. In B. Britton and S. Glynn (Eds.), Executive control processes in reading (pp. 279-296). Hillsdale, NJ: Erlbaum. Hebb, D.O. (1949). The organization of behavior. New York: Wiley. Howard, D.V., Shaw, R.J., & Heisey, J.G. (1986). Aging and the time course of semantic activation. Journal of Gerontology, 41, 195-203.
68
Ph.A. Allen et al.
Humphreys, G.W., EveR, L.J., & Quinlan, P.T. (1990). Orthographic processing in visual word identification. Cognitive Psychology, 22, 517-560. Johnson, N.F. (1975). On the function of letters in word id~atification: Some data and a preliminary model. Journal of Verbal Learning and Verbal Behavior, 14, 17-29. Johnson, N. F., Allen, P. A., & Strand, T. L. (1989). On the role of word frequency in the detection of component letters. Memory & Cognition, 17, 474-482. Joordens, S., & Besner, D. (1994). When banking on meaning is not (yet) money in the bank: Explorations in connectionist modeling. Journal of Experimental Psychology: Learning, Memory, and Cognition, 20, 1051-1062. Kellas, G., Paul, S.T., & Vu, H. (in press). Aging and language performance: From isolated words to multiple sentence contexts. In P. Allen & T. Bashore (Eds.), Age differences in word and language processing. New York: North-Holland. Keppel, G. (1991). Design and analysis: A researcher's handbook. Englewood Cliffs, NJ: Prentice-Hall. Kliegl, 1L (1994). Individual differences in cognitive aging: An experimental perspective. Paper presented at the 1994 Cognitive Aging Conference. Krueger, L.E. (1989). Detection of intraword and interword letter repetition: A test of the word unitization hypothesis. Memory & Cognition, 17, 48-57. Krueger, L.E., & Allen, P.A. (1987). Same-different judgments of foveal and parafoveal letter pairs by older adults. Perception & Psychophysics, 41, 329-334. Kucera, H., & Francis, W. N. (1967). Computational analysis of present-day American English. Providence, RI: Brown University Press. Laver, G.D., & Burke, D.M. (1993). Why do semantic priming effects increase in old age? Psychology and Aging, 8, 34-43. Levelt, W., Schiefers, H., Vorberg, D., Meyer, A., Pechman, T., & Havinga, J. (1991). The time course of lexical access in speech production: A study of picture naming. Psychological Review, 98, 122-142. Lima, S.D., & Pollatsek, A. (1983). Lexical access via an orthographic code? The Basic Orthographic Syllabic Structure (BOSS) reconsidered. Journal of Verbal Learning and Verbal Behavior, 22, 310-332. Lima, S.D., Hale, S., & Myerson, J. (1991). How general is general slowing? Evidence from the lexical domain. Psychology and Aging, 6, 416-425. Long, G.M., & Crambert, R.F. (1990). The nature and basis of age related changes in dynamic visual acuity. Psychology and Aging, 5, 138-143. Madden, D.J. (1986). Adult age differences in the attentional capacity demands of visual search. Cognitive Development, 1, 335-363. Madden, D.J. (1989). Visual word identification and age-related slowing. Cognitive Develpoment, 4, 1-29. Madden, D.J. (1990). Adult age differences in attentional selectivity and capacity. European Journal of Cognitive Psychology, 2, 229-252. Madden, D.J. (1992). Four to ten milliseconds per year: Age-related flowing of visual word recognition. Journal of Gerontology: Psychological Sciences, 47, P59-P68. Madden, D.J., Pierce, T.W., & Allen, P.A. (1992). Adult age differences in attentional allocation. Psychology and Aging, 7, 594-601. Madden, D.J., Pierce, T.W., & Allen, P.A. (1993). Age-related flowing and the time course of semantic priming in visual word recognition. Psychology and Aging, 8, 490-507.
Visual word encoding and the effect of adult age and wordfrequency
69
Madden, D.J., & Allen, P.A. (in press). Aging and the speed/accuracy relation in visual search: Evidence for an accumulator model. Optometry and Vision Science. Marr, D. (1982). Vision. San Francisco: Freeman. Masson, M.E.J. (1986). Identification of typographically transformed words: Instance-based skill acquisition. Journal of Experimental Psychology: Learning, Memory, and Cognition, 12, 479-488. McClelland, G.H., & Judd, C.M. (1993). Statistical difficulties of detecting interactions and moderator effects. Psychological Bulletin, 114, 376-390. McClelland, J.L., & Rumelhart, D.E. (1981). An interactive activation model of context effects in letter perception: Part I. An account of basic findings. Psychological Review, 88, 375407. McClelland, J.L., & Rumelhart, D.E. (1986). Parallel distributed processing: Explorations in the microstructure of cognition, VoL 2: Psychological and biological models. Cambridge, MA: MIT Press. Minsky, M. (1967). Computation: finite and infinite machines. Englewood Cliffs, NJ: Prentice-Hall. MonselL S., Doyle, M.C., & Haggard, P.N. (1989). Effects of frequency on visual word recognition tasks: Where are they? Journal of Experimental Psychology: General, 118, 43-71. Morton. J. (1969). Interaction of information in word recognition. Psychological Review, 76, 165-178. Myerson, J., Hale, S., Wagstaff~ D., Poon, L.W., Smith, G.A. (1990). The information loss model: A mathematical theory of age-related cognitive slowing. Psychological Review, 97, 475-487. Myerson, J., Ferraro, K, Hale, S., Lima, S. (1992). General slowing in semantic priming and word recognition. Psychology and Aging, 7, 257-270. Myerson, J., Wagsta~ D., & Hale, S. (1994). Brinley plots, explained variance, and the analysis of age differences in response latencies. Journal of Gerontology: Psychological Sciences, 49, P72-P80. Neely, J. (1991). Semantic priming effects in visual word recognition: A selective review of current findings and theories. In D. Besner and G.W. Humphreys (Eds.), Basic processes in reading: Visual word recognition. (pp. 264-336). Hillsdale, NJ: Erlbaunl Neisser, U. (1967). Cognitive psychology. New York: Appleton-Century-Crofts. Nelson, D., Schreiber, T., & McEvoy, C. (1992). Processing implicit and explicit representations. Psychological Review, 99, 322-348. Ober, B.A., & Shenaut, G.I~ (in press). Semantic priming in Alzheimer's disease: Metaanalysis and theoretical evaluation. In P. Allen & T. Bashore (Eds.), Age differences in word and language processing. New York: North-Holland. Owsley, C., Sekular, K, & Siemsen, D. (1983). Contrast sensitivity throughout adulthood. Vision Research, 23, 689-699. Paap, I~K, Newsome, S.L., & Noel, KW. (i984). Word shape's in poor shape for the race to the lexicon. Journal of Experimental Psychology: Human Perception and Performance, 10, 413-428. Paap, K.K, Newsome, S.L., McDonald, J.E., Schvaneveldt, KW. (1982). An activationverification model for letter and word recognition. Psychological Review, 89, 573-594.
70
Ph.A. Allen et al.
Pashler, H. (1987). Target-distractor discriminability in visual search. Perception & Psychophysics, 41, 285-291. Perfect, J.T. (1994). What can Brinley plots tell us about cognitive aging? Journal of Gerontology: Psychological Sciences, 49, P60-P64. Pillsbury, W.B. (1897). A study in apperception. American Journal of Psychology, 8, 315-393. Pinker, S., & Prince, A. (1988). On language and connectionism: Analysis of a parallel distributed processing model of language acquisition. Cognition, 28, 73-193. Pitts, D.G. (1982). Visual acuity as a function of age. Journal of the American Optometric Association, 53, 117-124. Rumelhart, D.E., & McClelland, J.L. (1982). An interactive activation model of context effects in letter perception: Part I. An account of basic findings. Psychological Review, 89, 60-94. Rumelhart, D.E., & McClelland, J.L. (1986). Parallel distributed processing: Explorations in the microstructure of cognition, Vol. 1: Foundations. Cambridge, MA: MIT Press. Salthouse, T.A. (1985). Speed of behavior and its implications for cognition. In J.E. Birren and & K.W. Schaie (Eds.), Handbook of the psychology of aging (2nd Ed., pp. 400-426). New York: Van Nostrand Reinhold. Salthouse, T.A. (1991). Mediation of adult age differences in cognition by reductions in working memory and speed of processing. Psychological Science, 2, 179-183. Salthouse, T.A., & Coon, V. (1994). Interpretation of differential deficits: The case of aging and mental arithmetic. Journal of Experimental Psychology: Learning, Memory, and Cognition, 20, 1172-1182. Seidenberg, M.S., & McClelland, J.L. (1989). A distributed, developmental model of word recognition and naming. Psychological Review, 96, 523-568. Selfridge, O.G. Pandemonium: A paradigm for learning. (1959). In Symposium on the mechanization of though processes. London: HM Stationary Office. Stadtlander, L. (in press). Age differences in orthographic and frequency neighborhoods. In P. Allen and T. Bashore (Eds.), Age differences in word and language processing. New York: North-Holland. Steinberg, S. (1966). High-speed scanning in human memory. Science, 153, 652-654. Stine, E. (in press). Aging and the distribution of resources in working memory. In P. Allen & T. Bashore (Eds.), Age differences in word and language processing. New York: NorthHolland. Taft, M. (1979). Lexical access via an orthographic code: The Basic Orthographic Syllable Structure (BOSS). Journal of Verbal Laerning and Verbal Behavior, 18, 21-39. Taft, M., & Forster, K.I. (1976). Lexical storage and retrieval of polymorphemic and polysyllabic words. Journal of Verbal Learning and Verbal Behavior, 15, 607-620. Tainturier, M-J., Trembley, M., & Lecours, A.R. (1989). Aging and the word frequency effect. Neuropsychologica, 27, 1197-1203. Tainturier, M-J., Trembley, M., & Lecours, A.R. (1992). Educational level and the word frequency effect: A lexical decision investigation. Brain and Language, 43, 460-474. Turing, A.M. (1936). On computable numbers, with an application to the Entscheidungsproblen~ Proceedings of the London Mathematics Society (Series 2), 230265. Van Essen, D., Anderson, C., & Felleman, D. (1992). Information processing in the primate visual system: An integrated systems perspective. Science, 419-423. Weale, R.A. (1986). Aging and normal vision. Vision Research, 26, 1507-1512.
Visual word encoding and the effect of adult age and word frequency
Wheeler, (1970). Processes in word recognition. Cognitive Psychology, 1, 59-85. Welford, A. (1958). Ageing and human skill. Oxford: Oxford University Press. Woodworth, 1LS. (1938). Experimental Psychology. New York: Holt.
71
72
Age Differences in Word and Language Processing Ph. Allen and Th.R. Bashore (Editors) 9 1995 Elsevier Science B.V. All rights reserved.
Age differences in orthographic and frequency neighborhoods Leann M. Stadtlander* Montana State University, Bozeman, MT
Previous researchers have examined young adults' responses in a lexical decision task to words and nonwords which are similar to other words (i.e., the "neighborhood" of the word). They have also explored the effect of high- and low-frequency stimulus words and high- and low-frequency in the neighborhood of the target word. The present study extended this work by examining differences in the responses ofyotmger (age 18-32) and older adults (age 60-75) to items which varied in the frequency of the target word as well as the size and frequency of the words in the target's neighborhood. Older adults were found to respond in a qualitatively different manner than the younger adults. This finding was interpreted through an internal noise model of aging. 1. INTRODUCTION The average adult reader of English, has at least 50,000 words in their vocabulary (Monsell, Doyle & Haggard, 1989). These 50,000 words are constructed from the set of 26 letters in the alphabet. Further, all of the 26 letters are constructed from a small set of features (Andrews, 1992; Gibson, 1969). Logically, many words must share features and letters with other words. How is it that the reader selects the correct word (i.e., lexical entry) from the many possible candidates in memory? Selection of the correct lexical entry, from all other possible candidates, is one of the most fimdamental issues in models of word recognition. The selection process provides an insight into how lexical memory (i.e., the lexicon) is structured and organized. All models or theories of the recognition process speculate on the way that the lexicon is organized (typically, by visual or phonological features), and in turn, suggest a method of finding a specific item that has been "filed" in the lexicon. Such theoretical frameworks provide hypothetical points in the selection process that may be tested experimentally. For example, if the mental lexicon is organized by the visual or phonological features of the lexical entries, then an experimenter should be able to influence a subject's response to a stimulus word which has similar features to many lexical entries in memory. This logic was applied in an experiment by Coltheart, Davelaar, Jonasson, and Besner (1977, Experiment 2). Coltheart et al. controlled the similarity of the features of a series of word and nonword stimuli which varied by what they called "N" or "orthographic neighborhood" size. "N" was defined as the number of English words that could be produced
* Author Notes: The author would like to offer her appreciation to Lester E. Krueger for his contributionto the design of this experiment. My thanks also to Susan Boardman and Emily Hoffman for assistance in the testing of subjects.
Age differences in orthographic and frequency neighbourhoods
73
by replacing one letter, while preserving letter position for all other letters. The final stimulus list consisted of words and nonwords with either a high or low value of N. Coltheart et al. (1977) used a lexical decision task (LDT). In this task the subject is required to decide if a stimulus item, shown on a computer screen, is a word or a nonword. The authors found an effect of N only for nonword trials; that is, having many similar (word) neighbors considerably slowed the "nonword" response. ~There was no effect for words, which led Coltheart et al. to conclude that the number of items physically similar to a given word is not related to the response time (RT) to reco~ize it as a word. However, new evidence suggests that other properties of a neighborhood do affect visual word recognition.* Andrews (1989; 1992) reported a study in which not only neighborhood size was controlled, but also, the frequency of the stimulus word, i.e., the "target" word. Target words were either common (i.e., high-frequency) words or were less common (i.e., low-frequency) words. Andrews, using a LDT, found that low-frequency target words with large neighborhoods resulted in a faster RT than those from small neighborhoods. No neighborhood size effect was evident when the target items were highfrequency words. Additional data of Grainger (1990; Grainger, O'Regan, Jacobs, & Segui, 1989), are suggestive of an even more complicated picture. Grainger (1990), using large neighborhoods with low- and medium-frequency target words in a LDT, developed four classes of words: (1) words with zero neighbors; (2) words with neighbors of only lower frequencies; (3) words with one neighbor of higher frequency; (4) words with more than one neighbor of higher frequency. Grainger found that when the target had one or more neighbors of higher frequency, RT was slowed. Although these studies used different types of stimuli, some consistencies between the experiments are apparent. Consistent with both experiments are the class of stimuli "large neighborhoods with low-frequency target words." In Andrews' (1989; 1992) experiments these items were found to result in a faster RT when compared to similar target words with small neighborhoods. Grainger (1990), using this class of items, found that items within the target word's neighborhood could be an influencing factor. If a single neighbor was of higher frequency than the target word, RT was slowed compared to a neighborhood of all low-frequency words. These findings are important in how we conceptualize visual word recognition. The data indicate that at some point in the search for a low-frequency word in the mental lexicon, many similar low-frequency words do not hinder the recognition process while words of higher frequency slow the process. It would seem appropriate at this point to examine the locus of the "typical" frequency effect in models of visual word recognition. 1.1. Frequency Effects in the Mental Lexicon Many theories of word recognition assume that the mental lexicon is organized as a function of word frequency (e.g., Cart & Pollatsek, 1985; Forger, 1976, 1979; Monsell et al., 1989; Morton, 1969). Such models assume that visual features of stimuli are matched to representations stored in the mental lexicon. Further, the models suggest that word Neighborhood effects have also been reported in studies of: accuracyof identification of masked words (Luce, 1986), children's naming (Laxon, Coltheart, & Keating, 1988), and naming latencyfor German words (Gunther & Greese, 1985; Scheerer, 1987).
74
L.M. Stadtlander
frequency influences access to the items in the lexicon, by permitting easier access to more common words (i.e., ones of high-frequency in a language) than to less common words. Thus, it is generally assumed that the locus of the word frequency effect is at the lexical access stage in processing (though see Balota & Chumbley, 1984, for a contrary view). Since the first use of the LDT, RT has been shown to be a sensitive measure of frequency (Forster & Chambers, 1973; Rubenstein, Garfield, & Millikan, 1970; Taft, 1979; Whaley, 1978). Thus, common words (i.e., ones of high-frequency in a language), such as door, are responded to more quickly in the LDT than are uncommon words (i.e., ones of low-frequency in a language), such as cask. Taking into account the information gained from Andrews' (1989; 1992) and Grainger's (1990) studies reported above, in which it became apparent that at some point in the search for a low-frequency word in the mental lexicon, similar words of higher frequency slow the process, a clearer picture of the recognition process evolves. We can conceptualize the recognition process, through a "genetic" model, as one in which a continuum of activation is present. First, the stimulus word is encoded into a set of visual features. Second, words composed of similar features are activated, with more common items more strongly activated than less common ones. Based upon the Grainger (1990) work, it appears that the high-frequency neighbors, in some way interfere with the recognition of a lower frequency word. Let us now take this knowledge and apply it to the issue of age-related changes in lexical processing.
1.2. Age-Related Changes in Processing Several LDT studies, in which word frequency was manipulated as a function of adult age, have been previously reported (e.g., Allen, Madden, & Crozier, 1991; Allen, Madden, Webber, & Groth, 1993; Bowles & Pooh, 1981). These studies found no age differences in word frequency; i.e., both age groups showed a frequency effect, whereby, high-frequency words were recognized faster than low-frequency. Allen et al. (1991) also reported a longer response time but less errors for older adults, as compared to younger adults, in a standard LDT (see also, Bowles & Pooh, 1985; Cerella & Fozard, 1984; Mueller, Kausler, & Faherty, 1980; Waugh & Barr, 1982). However, Allen et al. found no qualitative difference in response to different frequency items between the two age groups. Although older adults appear to show similar frequency effects when compared with younger adults, they nonetheless, consistently take longer to respond than do younger adults. Two theoretical views have been suggested to explain age-related differences in response speed and accuracy. The first view, is that with age there is an increase in neural (internal) noise (Allen, Namazi, Patterson, Crozier, & Groth, 1992; Cremer & Zeet~ 1987; Welford, 1958). Many previous researchers have proposed that aging increases the level of spurious neural activity or internal noise, thereby reducing the signal-to-noise ratio and producing agerelated deficits (Crossman & Szafran, 1956; Gregory, 1959; Krueger & Allen, 1987; Layton, 1975; Vickers, Nettlebeck, & Willson, 1972; Welford, 1965, 1977). Internal noise is assumed to randomly change visual features during perception (Krueger, 1978; Krueger & Allen, 1987). Krueger (1978) postulated three dements in a theory of internal noise. We can apply these dements to the issue of processing words which differ in the size of their orthographic neighborhoods.
Age differences in orthographic and frequency neighbourhoods
75
1. The perceiver bases the decision on how many mismatching features a difference counter records. Internal noise should affect input from the stimulus item by introducing random features in the perceptual image of the stimulus item A theory of internal noise suggests that the perceiver must compare differences (i.e., a difference counter) between the lexical entry of the item in memory and the developing perceptual image. 2. Internal noise affects the comparison process. This element suggests that internal noise affects the process of selecting the correct lexical entry from similar entries, by increasing the potential candidates or neighbors. This leads to the hypothesis that as the size of the neighborhood increases (i.e., for a large neighborhood as compared to a small neighborhood) it should increase the difficulty of selecting the correct entry. 3. I f the difference count is not sufficiently high or low, the decision is postponed so that a second or third glance may be made. Krueger (1978) theorized that the perceiver continually resamples or rechecks the stimulus until sufficient evidence has been acquired. This element suggests that if internal noise does increase with age, then older individuals should on average, take longer to respond (which as discussed earlier, is indeed the case) in order to recheck the stimulus. Further, although older adults may be slower in their responses, they may also be more accurate due to the rechecking step. The second theoretical view, is that with age there is a general slowing in processing (Birren, 1974; Cerella & Fozard, 1984; Salthouse, 1985; also see Hasher & Zacks, 1988). There are two subtly different versions of cognitive slowing, which can be labeled the strong and weak versions (Hartley, 1992). A strong theory specifies a physiological mechanism that is responsible for the slowing, such as uniform slowing of synaptic transmission or information loss at each transmission (Myerson, Hale, Wagstafl~ Poon, & Smith, 1990). This theoretical viewpoint would suggest that there should be no qualitative change with age. In effect, the responses have been mathematically transformed resulting in an overall slower response (Allen et al., 1993). According to this view, errors should be equivalent in the two age groups. By contrast, a weak theory of slowing does not specify a mechanism, but rather simply attempts to identify a function (i.e., a mathematical transformation) that provides a good fit to the relationship between the younger and older adults' performance. Thus, it provides a description, not an explanation. 1.3. The Present Experiment
In the present experiment, a series of stimulus items were developed which varied in the l) frequency of the target items (high- or low-frequency), 2) frequency of the neighborhood (high-frequency or low-frequency), and 3) size of the neighborhood ( large or small; and is equivalent to Coltheart's "N"). These categories of items should permit a finer examination of the interactions of age, frequency, and size of the neighborhood in a LDT. Predictions. The internal noise model suggests that internal noise affects the process of selecting the correct lexical entry from similar entries by increasing the potential candidates or neighbors. This suggestion leads to the hypothesis that as the size of the neighborhood increases (i.e., for a large neighborhood as compared to a small neighborhood) the difficulty of selecting the correct entry should also increase. This would predict that older adults should have a great deal of problems with large neighborhoods, particularly for the conditions with a low-frequency target word and high-frequency neighbors. It should be difficult to resolve these items, suggesting a longer RT should be
76
L.M. Stadtlander
evident. Nonword trials with large neighborhoods should also be difficult to resolve, particularly for older adults due to the presence of many similar words requiring the opposite response, again, suggesting a longer RT and perhaps more errors. The slowing model suggests that older adults should be considerably slower than the younger adults as predicted by a single best-fitting slowing fimction. However both age groups should be approximately equivalent in errors. In order to examine the effect of slowing, a method proposed by Madden, Pierce and Allen (1992) will be used. In this method a best-fitting regression function is determined between the older and younger adults. Assuming a si~ificant r 2 is found, the regression ftmction is then used to transform the younger adults' raw data. An analysis of variance (ANOVA) is then performed on the data (transformed for younger adults but untransformed for older adults). Because the younger adults' raw latencies have been transformed by the function representing generalized slowing, siL_,nificant Age x Condition interactions for this ANOVA would represent those effects beyond a slowing model (Allen et al., 1993). 2. M E T H O D
Subjects. The younger subjects were volunteers from Introductory Psychology classes at Montana State University. Older subjects were volunteers from the Bozeman, Montana community, contacted through ads in newsletters and the local Senior Citizen Center. All subjects were required to have visual acuity of at least 20/40 (corrected). Fiiteen older (M age = 67.0 years; 7 males and 8 females) and 15 younger (M age = 19.2 years; 8 males and 7 females) subjects participated. Older subjects received an average of 65 on the Vocabulary Subscale and 17 on the Information Subscale of the WAIS-R and had completed an average of 16 years of schooling. The younger subjects averaged 43 on the Vocabulary Subscale and 12 on the Information Subscale and had completed an average of 13 years of schooling. Stimulus materials. Lowercase letters were presented on a 486 DEC color computer screen. The letters, were presented as thin, illuminated lines on a dark screen. All of the 240 words contained five letters and all were presented once to each subject. The word lists contained 43 instances of items with repeated letters (e.g., guess, sales); the nonword list was equated for this factor. The word lists were devised relative to three criteria: 1) frequency (based upon Kucera & Francis, 1967, and Thorndike & Lorge, 1944, norms) of the target word (high-frequency target [HFT] vs. low-frequency target [LFT]); 2) frequency of the neighborhood members of the target word (two or more neighbors of higher-frequency [HFN], all low-frequency neighbors [LFI~); 3) size of the target word's neighborhood (large or small). High-frequency words occurred more than 50 times per million words in printed English ("A" in Thorndike & Lorge norms), whereas, low-frequency words occurred less than 15 per million. The mean frequency for high-frequency words was 213.48 per million, and for low-frequency words was 6.05 per million. The criterion for large-neighborhood items was that at least five different words could be formed by changing single letters; the mean size for large neighborhoods was 8.09. For small-neighborhood items the criterion was that no more than three such alternatives could be formed; the mean size for small neighborhoods was 1.83. The neighborhoods were constructed primarily through the use of the WordPerfect spell checker. Specifically, in this word-processing package, a "wild card" character can be used to replace a letter in a word. The spelling-check program then lists all words from its
Age differences in orthographic and frequency neighbourhoods
77
20,000-word lexicon whose letters are positionally consistent within the word. The total number of consistent words across all positions constitutes the neighborhood for a particular word, as defined in this experiment. For example, the target word blank would generate a neighborhood consisting of." clank, flank, plank, blink, black, and bland. Nonword lists were similarly devised, based upon the frequency of the (word) neighborhood of the nonword (high-frequency neighborhood [HFN] or low-frequency neighborhood [LFI~), and the size of the (word) neighborhood (large or small). As an example, the target nonword bason would generate a neighborhood consisting of jason,
mason, bison, bacon, baron, baton, and basin. The 480 regular trials (240 words, 240 nonwords) were randomly intermixed, and formed into 16 blocks of 30 trials each. Two practice trials preceded each block, and there was an initial practice block, resulting in a total of 544 trials. Four different random orderings of trials were used in the experiment. Procedure. On each trial a fixation mark appeared alone for .7 sec, and then the word or nonword appeared just above the fixation mark until a response was made. Subjects were instructed to respond as rapidly as possible without sacrificing accuracy. Half of the subjects pressed a right-hand button if the target was a word, and the leit-hand button if it was a nonword. The other half of the subjects had the reverse hand assignment. 3. RESULTS As is typical in lexical decision tasks, subjects responded faster to words (M = 720 ms) than to nonwords (M = 912 ms; F[1,28] = 20.95, p < .0001). Words also resulted in fewer errors (M = 7.0%) than nonwords ( M = 12.72%; F[1,28] = 33.49, p < .0001).
3.1. Analyses of Words As is consistent with Allen et al. (1991), older adults had a longer response time but had less errors. In a 2 x 2 x 2 x 2 Analysis of Variance (ANOVA), which examined age (young vs. old) by size of neighborhood (large vs. small) by target word frequency (low vs. high) by frequency of the word neighborhood (low vs. high); a main effect of age was evident for RT (F[1,28] = 6.379,/2 < .01]. Older subjects responded slower (M = 766 ms) than younger subjects (M = 673.5). Older subjects were also more accurate than younger subjects (M's : 5.4% and 18.8% respectively; F[1,28] = 7.31,p < .01). There was also an effect of target frequency, whereby, low-frequency target items were responded to slower (M = 747 ms) than high-frequency target items (M = 693 ms; F[1,28] = 91.59, p < .0001). A (non-significant) interaction with age was evident for RT (older adults / high frequency target = 735 ms; older adults / low frequency target = 784 ms; younger adults / high frequency target = 627 ms; younger adults / low frequency target = 701 ms; F[1,28] = 3.53,p = .07).
78
L.M. Stadtlander
Additionally, low-frequency (M = 10.1%) items generated more errors than did high-frequency (M = 4.05%; F[1,28] = 33.49, p < .0001) items. However, contrary to Allen et al. (1991) who found no qualitative difference between the two age groups for target frequency, in the present experiment an interaction was evident for errors (F[ 1,28] = 6.26,
Figure high).
1.
Percent Errors (PE) as a function of age (young vs. old) and frequency of target word (low vs.
p < .01. As shown in Figure 1, a greater difference was present between the two age groups for low-frequency target items as opposed to high-frequency targets. There was no main effect of neighborhood size (large vs. small; p > . 10) for RT or errors, nor an interaction with age for RT. However, an interaction was present for errors between neighborhood size and age (F[ 1,28] = 7.61, p < .01). Whereby, older adults were more accurate on small neighborhoods (4.9% vs. 6.2% for large neighborhoods) as compared to young adults (7.5% for small neighborhoods vs. 7.8% for large neighborhoods). A priori analyses, based upon Andrews' (1992) work, were conducted for the two age groups which examined neighborhood size and frequency of the target word and neighborhood (see Tables 1 and 2). Differences were found for RT and errors between large and small neighborhoods on items with low-frequency target words and low-frequency neighborhoods for the older adults (RT: large neighborhoods = 792.35 ms, small neighborhoods = 756.63 ms; t[14] = 2.39, p < .05; PE: large neighborhoods = 1.67%, small neighborhoods = 1.07%; t[14] = 2.08, p < .05). Note, however, that these findings are in the opposite direction from Andrews' data with young adults. Andrews reported large neighborhoods resulted in faster RT.
Age differences in orthographic and frequency neighbourhoods
79
For young adults, differences were found for RT between large and small neighborhoods on items with high-frequency target words and low-frequency Table 1. Response Time (RT) and Percent Errors (PE) for Older Adults by Neighborhood Size and Frequency of Target Word and Neighborhood; Standard Error in Parentheses. Older Adults Large neighborhood RT:792.34ms (34.00)** Low-Frequency Target / PE: 8.3 (2.8)* Low-Frequency Neighborhood RT: 781.29 ms (35.42) Low-Frequency Target / PE: 7.65 (1.8) High-Frequency Neighborhood High-Frequency Target / RT: 730.06 ms (37.19) Low-Frequency PE: 3.3 (1.5) Neighborhood High-Frequency Target / RT: 732.44 ms (28.58) PE: 5.65 (1.6) High-Frequency Neighborhood * p < .05 **p,, o
c
(D
1500--
0 ___! cO
1000
-
73 cO
o L.. &) -(3 Ld
500 -
~.~t~~ .#,
mosk Type
/
* Comporison 9 Production
I I
I
Urea et al., 1991)
o
o
I
I
i
500
1000
1500
!
2000
Young Condition Lotency ( m s ) Figure 4. Scatter plot and best-fitting lines for comparison tasks and production tasks from the aging and picture-word processing literature. Also, included in the production task data are the three word-naming studies from the Lima et al. (1991) meta-analysis.
As can be seen in Figure 4, adding these three word-naming studies to the production task data set alters the regression line little and actually strengthens the fit. The production task now has a slope of .93 with an intercept of 167.4 ms, and accounts for 87.6% of the condition mean variance. Also, the slope of .93 is si~ificantly less than the slope of 1.47 of the comparison task line, t(139) = 5.22, p < .001. As an additional indicator that the production and comparison tasks continue to represent two distinct slowing functions, the regression line fitting latencies across task (where RTELDmLY= 1.05 RTvotmG+ 204.1 ms) now accounts for only 34.8% of the elderly condition mean variance. Finally, for sake of completeness, I again recomputed this regression line including two other word-naming studies in the aging literature (see Table 1). The resulting regression line now has a slope of. 86 with an intercept of 222.2 ms, and accounts for 86.4% of the condition mean variance. The slope of .86 for this production task line is again si~ificantly different from the slope of 1.47 for the comparison task line, t(177) = 6.91, p < .001. And again, as an
164
P. C. Amrhein
additional indicator that the production and comparison tasks continue to represent two distinct slowing functions, the regression line fitting latencies across task (where RTELDF~LY= 1.13 RTyousG+ 133.4 ms) now accounts for only 45.6% of the elderly condition mean variance. The message here is clear: Production tasks involving picture or word stimuli do not undergo proportional, age-related slowing, whereas comparison tasks do. However, both tasks share an additive increase likely due to sensory-motor deficits in the elderly subjects (see Botwinick, 1984). This suggests that lexical and pictorial retrieval are spared in aging (see also Bowles, 1993; Duchek & Balota, 1993), but aspects of decision processing are not. By design, all comparison-task studies in Table 1 have a decision component (e.g., responding 'yes' or 'no'). This is also the case for the non-production lexical tasks (i.e., lexical decision, judgment, categorization, etc.) analyzed by Lima et al. (1991) which also exhibit proportional slowing. Indeed, if latencies from such tasks were included in the regression analysis of the picture-word comparison tasks, the goodness of fit (r 2) should increase similarly to that seen for the regression analysis of the production tasks when the word-naming studies were included. Thus, at least within the response range analyzed here and by Lima et al. (1991), stimulus modality (i.e., lexical or nonlexical) does not impact on the nature of age-related slowing, but task type does. The reader should also note that in the first recta-analysis, only two within-mode condition latencies (i.e., from Amrhein & Theios, 1993: draw picture from picture stimulus; write word from word stimulus) were included among the predominant cross-modality condition latencies (i.e., from Amrhein & Theios, 1993: draw picture from word stimulus; write word from picture stimulus; and from the remaining studies: name picture stimulus). However, even though cross-modality latencies are longer than within-mode latencies, both condition types should fall on the same regression line because the additional latency increment for cross-modality transfer has been found to be age-, stimulus modality-, and production taskindependent (Amrhein & Theios, 1993; Amrhein, 1994). This claim is especially supported by the minimal impact that the latencies of the word-naming (i.e., a within-mode condition) studies had on the slope of the regression line in the two subsequent recta-analyses. In stun, this demonstration indicates task specificity in age-related slowing, and thus the need to explain task performance differences at the underlying process level. In other words, a fixed proportional slowing account based on overall task performance does not explain aging effects and non-effects in speeded lexical and nonlexical tasks. 4. FUTURE DIRECTIONS As was discussed earlier, the drawing-writing task used here offers a balanced solution to the incompleteness of the traditional naming-reading task. The drawing-writing task can also be implemented to more comprehensively address issues such as picture-word priming and Stroop-like interference effects. Only a handful of findings have been reported in the literature concerning these issues in their relation to aging (and then using only the picture-naming task; e.g., Bowles, 1994; Mitchell, 1989). Implicit in Equations 1-6 presented earlier is the assumption that within-modality conditions (i.e., draw a picture given a picture stimulus, write a word given a word stimulus, read aloud a word stimulus) do not involve semantic memory access whereas cross-modality conditions (draw a picture given a word stimulus; write a word given a picture stimulus, name
Evidence for task specificity in age-related slowing
165
a picture stimulus) do involve it. Incorporation of a picture-word priming paradigm into the drawing-writing task would provide a critical test of this assumption. For young subjects, the evidence from naming tasks suggests support for the Amodal models, where pictures and word share a common semantic store (e.g., Bajo, 1988). Specifically, Bajo (1988) found that conceptually related word and picture primes facilitate naming picture stimuli but not reading word stimuli. Accordingly, if her study was conducted using the drawing-writing task, the expectation is that conceptually related word and picture primes would also facilitate writing names for picture stimuli but not writing down word stimuli. Moreover, if picture-word processing is balanced temporally as it appears to be given the studies of Amrhein (1994), Amrhein and Theios (1993) and Theios and Amrhein (1989) then it should also be found that conceptually related word and picture primes facilitate drawing pictures from word stimuli but not drawing pictures from picture stimuli. From the aging and word priming literature, it appears that prime facilitation may actually be greater for elderly over young subjects (although the rate of corresponding spreading activation may be constant, see Balota & Duchek, 1988, but also see Howard, Shaw & Heisey, 1986). This difference may be due to slower word encoding and response processes which give semantic priming mechanisms (i.e., spreading activation) additional time to function (see e.g., Balota & Duchek, 1988; Bowles & Pooh, 1985; Burke, White & Diaz, 1987; Howard, McAndrews, & Lasaga, 1981). Indeed, the data of Bowles (1994) suggest that when age-based perceptual differences among the primes are accounted for, prime facilitation at long SOAs (500-700 ms) is equivalent for elderly and young subjects. By using the drawing-writing task, issues of age-related slowing in semantic priming can be comprehensively addressed by testing pictures and word as primes, target stimuli and output productions. In addition to determining similarities or differences in semantic priming effects due to modality of prime, target and output production, the model approach presented in Equations 1-6, when applied to these data, allows the determination of the specific subprocesses which are influenced (and to what extent) by subject age, prime-target relatedness, as well as prime, target, and output production modality (see Amrhein, 1994; Amrhein & Theios, 1993). Noticeably absent from the aging and picture-word studies reviewed using comparison tasks was the manipulation of a semantic variable, notably category membership. This variable has received substantial attention in the general picture-word processing literature because of its importance in testing theoretical models (e.g., Harris, Morris & Bassett, 1977; Pellegrino, et al., 1977; Potter & Faulconer, 1975; Snodgrass & McCuUough, 1986; te Linde, 1982). In a categorization tasks, where two stimuli are presented for a binary category membership decision, proportional slowing for elderly subjects should be found that is consistent with that found for picture-word comparison tasks reviewed here. Moreover, because I am arguing that it is the decision subprocesses underlying comparison tasks, in general, that exhibit age-related proportional slowing, the slowing for picture-word categorization tasks should be the same as that seen for categorization tasks (actually all comparison tasks) using only lexical stimuli. Finally, one obvious limitation of this (and most other) meta-analyses is that the production and comparison task data included were from different experiments (i.e., different subjects appeared across the two task types). A critical test of the task specificity shown here would be to assess elderly and young groups on a set of stimuli for production and comparison tasks within-subjects. Indeed, given a stimulus set which satisfies the conditions of size, featural similarity and familiarity specified by Theios and Amrhein (1989) and Snodgrass and
P. C. Amrhein
166
McCullough (1986), and a comprehensive set of conditions for each task (concerning crossmodality and within-modality representation retrieval), the three hypotheses (elderly slowing, elderly spatial deficit, and specific picture-word processing model assumptions) could be tested collectively. REFERENCES
Allen, P.A., Madden, D.J. & Crozier, L.C. (1991). Adult age differences in letter-levd and word-levd processing. Psychology and Aging, 6, 261-272. Allen, P.A., Madden, D.J., Weber, T.A. & Groth, ICE. (1993). Influence of age and processing stage on visual word recognition. Psychology and Aging, 8, 274-282. Amrhein, P.C. (1994). Temporal invariance for picture-word translation: Evidence from drawing-writing and naming-reading tasks, Memory & Cognition, 22, 442-454. Amrhein, P.C., & Theios, J. (1993). The time it takes for elderly and young individuals to draw pictures and write words, Psychology and Aging, 8, 197-205. Are~berg, D. (1978). Differences and changes with age in the Benton Visual Retention Test. Journal of Gerontology, 33, 534-540. Bajo, M.T. (1988). Semantic facilitation with pictures and words. Journal of Experimental Psychology: Learning, Memory, and Cognition, 14, 579-589. Balota, D.A. & Chumbley, J.I. (1984). Are lexical decisions a good measure oflexical access? The role of word frequency in the neglected decision stage. Journal of Experimental Psychology: Human Perception and Performance, 10, 340-357. Balota, D.A. & Chumbley, J.I. (1985). The locus of word-frequency effects in the pronunciation task: Lexical access and/or production? Journal of Memory and Language, 24, 89-106. Balota, D.A. & Duchek, J.M. (1988). Age-related differences in lexical access, spreading activation, and simple pronunciation. Psychology and Aging, 3, 84-93. Botwinick, J. (1984). Aging and Behavior. New York: Springer. Bowles, N.L. (1993). Semantic processes that serve picture naming. In J. Cerdla, J. Rybash, W. Hoyer, & M.L. Commons (Eds.), Adult information processing: Limits on loss. (pp. 303-328). San Diego, CA: Academic Press. Bowles, N.L. (1994). Age and rate of activation in semantic memory. Psychology and Aging, 9, 414-429. Bowles, N.L. & Pooh, L.W. (1981). The effect of age on speed oflexical access. Experimental
Aging Research, 7, 417-425. Bowles, N.L. & Pooh, L.W. (1985). Aging and retrieval of words in semantic memory. Journal of Gerontology, 40, 71-77. Canestrari, 1LE. (1968). Age changes in acquisition. In G.A. Talland (Ed.), Human aging and behavior (pp. 169-188). New York: Academic Press. Cattdl, J.M. (1886). The time it takes to see and name objects. Mind, 11, 63-65. Cerdla, J. (1990). Aging and information-processing rate. In J. Birren & I~W. Schaie (Eds.), Handbook of the psychology of aging (pp. 201-221). San Diego: Academic Press. Cerella, J. (1994). Generalized slowing in Brinley plots. Journal of Gerontology, 49, P65-71. Cerdla, J., & Fozard, J.L. (1984). Lexical access and age. Developmental Psychology, 20, 235-243.
Evidencefor task specificity in age-relatedslowing
167
Cerella, J., & Hale, S. (1994). The rise and fall of information processing rates over the life span. Acta Psychologwa, 86, 109-198. Dell, G.S., & O'Seaghdha, P.G. (1991). Mediated and convergent lexical priming in language production: A comment on Levelt et al. (1991). Psychological Review, 98, 604-614. Duchek, J.M., & Balota, D.A. (1993). Sparing activation processes in older adults. In J. Cerella, J. gybash, W. Hoyer, & M.L. Commons (Eds.), Adult reformat~on processing: Limits on loss. (pp. 384-406). San Diego, CA: Academic Press. Elias, M.F., & Kinsbourne, M. (1974). Age and sex differences in the processing of verbal and nonverbal stimuli. Journal of Gerontology, 29, 162-171. Falmagne, J.C. (1965). Stochastic models for choice reaction time with application to experimental results. Journal of Mathematical Psychology, 12, 77-124. Farah, M.J., & Kosslyn, S.M. (1981). Structure and strategy in image generation. Cognitive Science, 4, 371-383. Fisk, A.D., & Fisher, D.L. (1994). Brinley plots and theories of aging: The explicit, muddled, and implicit debates. Journal of Gerontology, 49, P81-89. Fisk, A.D., Fisher, D.L. & Rogers, W.A. (1992). General slowing alone cannot explain agerelated search effects: Reply to Cerella (1991). Journal of Experimental PsychologyGeneral, 121, 73-78. Fraisse, P. (1969). Why is naming longer than reading? Acta Psychologica, 30, 96-103. Geary, D., Frensch, P., & Wiley, J. (1993). Simple and complex mental subtraction: Strategy choice and speed-of-processing differences in younger and older adults. Psychology and Aging, 8, 242-256. Glaser, W.1L (1992). Picture naming. Cognition, 42, 61-105. Glaser, W.1L, & Dungelhofl~ F.-J. (1984). The time course of picture-word interference. Journal of Experimental Psychology: Human Perception and Performance, 10, 640-654. Glaser, W.IL, & Glaser, M.O. (1989). Context effects in Stroop-like word and picture processing. Journal of Experimental Psychology: General, 118, 13-42. Goulet, P., Ska, B., & Kahn, H.J. (1994). Is there a decline in picture naming with advancing age? Journal of Speech and Hearing Research, 37, 629-644. Hale, S., Lima, S.D., & Myerson, J. (1991). General cognitive slowing in the noniexical domain: An experimental validation. Psychology and Aging, 6, 512-521. Hale, S., Myerson, J., & Wagstafl~ D. (1987). General slowing of nonverbal information processing: Evidence for a power law. Journal of Gerontology, 42, 131-136. Halpem, D.F. (1984). Age differences in response time to verbal and symbolic traffic signs. Experimental Aging Research, 10, 201-204. Harker, J,O. & Riege, W.H. (1985). Aging and delay effects on recognition of words and designs. Journal of Gerontology, 40, 601-604. Harris, P.L., Morris, P.E., & Bassett, E. (1977). Classifying pictures and words: Implications for the dual-coding hypothesis. Memory & Cognition, 5, 242-246. Hertzog, C. (1992). Aging, information processing speed, and intelligence. In I~W. Schaie (Ed.), Annual review of gerontology and germtries (Vol. 11, pp. 55-79). New York: Springer. Howard, D.V., Shaw, R,J., & Heisey, J.G. (1986). Aging and the time course of semantic activation. Journal of Gerontology, 41, 195-203. Hulicka, I.M. & Grossman, J.L. (1967). Age group comparisons for the use of mediators in paired-associate learning. Journal of Gerontology, 22, 46-51.
168
P.c. Amrhein
Kausler, D.H. (1991). Experimental psychology, cognition, and human aging. New York: Springer-Verlag. Kosslyn, S. M. (1980). Image andMind. Cambridge, MA: Harvard University Press. La Heij, W. (1988). Components of Stroop-like interference in picture naming. Memory & Cognition, 16, 400-410. Laver, G.D., & Burke, D.M. (1993). Why do semantic priming effects increase in old age? A Meta-Analysis. Psychology and Aging, 8, 34-43. Levelt, W.J.M., Schriefers, H.,Meyer, A.S., Pechman, T., Vorberg, D., & Havinga, J. (1991). The time course of lexical access in speech production: A study of picture naming. Psychological Review, 98, 122-142. Levelt, W.J.M., Schriefers, H., Vorberg, D., Meyer, A.S., Pechman, T., & Havinga, J. (1991). Normal and deviant lexical processing: Reply to Dell and O'Seaghdha (1991). Psychological Review, 98, 615-618. Lima, S.D., Hale, S., & Myerson, J. (1991). How general is general slowing? Evidence from the lexical domain. Psychology andAging, 6, 416-425. Lupker, S., & Theios, J. (1975). Tests of two classes of models for choice reaction time. Journal of Experimental Psychology: Human Perception and Performance, 104, 137-146. McCauley, C., Parmalee, C.M., Sperber, 1LD., & Carr, T.H. (1980). Early extraction of meaning from pictures and its relation to conscious identification. Journal of Experimental Psychology: Human Perception and Performance, 6, 265-276. McEvoy, C.L. (1988). Automatic and strategic processes in picture naming. Journal of Experimental Psychology: Learning, Memory and Cognition, 4, 618-626. Mergler, N.L., & Zandi, T. (1983). Adult age differences in speed and accuracy of matching verbal and pictorial signs. Educational Gerontology, 9, 73-85. Mitchell, D.B. (1989). How many memory systems? Evidence from aging. Journal of Experimental Psychology: Learning, Memory, and Cognition, 15, 31-49. Molenaar, P.C.M., & van der Molen, M.W. (1994). On the discrimination between global and local trend hypotheses of life-span changes in processing speed. Acta Psychologica, 86, 273-293. Myerson, J., & Hale, S. (1993). General slowing and age invariance in cognitive processing: The other side of the coin. In J. Cerella, J. Kybash, W. Hoyer, & M.L. Commons (Eds.), Adult information processing: Limits on loss. (pp. 115-141). San Diego, CA: Academic Press. Myerson, J., Ferraro, F.1L, Hale, S., & Lima, S.D. (1992). General slowing in semantic priming and word recognition. Psychology and Aging, 7, 257-270. Myerson, J., Hale, S., Wagstaff~ D., Pooh, L.W., & Smith, G.A. (1990). The information loss model: A mathematical theory of age-related cognitive slowing. Psychological Review, 97, 475-487. Nebes, 1LD. (1976). Verbal-pictorial recoding in the elderly. Journal of Gerontology, 31, 421427. Nebes, 1LD., Boiler, F., & Holland, A. (1986). Use of semantic context by patients with Alzheimer's Disease. Psychology and Aging, 1, 261-269. Nebes, 1LD., Brady, C.B., & Hut~ F.J. (1989). Automatic and attentional mechanisms of semantic priming in Alzheimefs Disease. Journal of Clinical and Experimental Neuropsychology, 11, 219-230.
Evidencefor task specificity in age-relatedslowing
169
Paivio, A. (1966). Latency of verbal associations and imagery to noun stimuli as a function of abstractness and generality. Canadian Journal of Psychology, 20, 378-387. Paivio, A. (1971). Imagery and verbal processes. New York: Holt, Rinehart & Winston Paivio, A. (1975). Perceptual comparisons through the mind's eye. Memory & Cognition, 3, 635-648. Paivio, A. (1983). The empirical case for dual coding. In J.C. Yuille (Ed.), Imagery, memory and cognition (pp. 307-332). Hillsdale, NJ: Lawrence Erlbaum Associates. Paivio, A. (1986). Mental representations. New York: Oxford University Press. Paivio, A., Clark, J.M., Digdon, N., & Bons, T. (1989). Referential processing: Reciprocity and correlates of naming and imaging. Memory & Cognition, 17, 163-174. Park, D.C., & Puglisi, J.T. (1985). Older adult's memory for the color of pictures. Journal of Gerontology, 40, 198-204. Pellegrino, J.W., gosinski, 1L1L, Chiesi, H.L., & Siegel, A. (1977). Picture-word differences in decision latency: An analysis of single and dual memory models. Memory & Cognition, 5, 383-396. Pezdek, I~ (1983). Memory for items and their spatial locations by young and elderly adults. Developmental Psychology, 19, 895-900. Poon, L.W., & Fozard, J.L. (1978). Speed of retrieval from long-term memory in relation to age, familiarity, and datedness of information. Journal of Gerontology, 33, 711-717. Potter, M.C., & Faulconer, B.A. (1975). Time to understand pictures and words. Nature, 253, 437-438. Seymour, P.H.K. (1973). A model for reading, naming, and comparison. British Journal of Psychology, 64, 35-49. Seymour, P.H.I~ (1974). Generation of a pictorial code. Memory & Cognition, 2, 224-232. Seymour, P.H.K. (1979). Human visual cognition. New York: St. Martin's Press. Smith, A.D., & FuUerton, A.M. (1980). Age differences in episodic and semantic memory: Implications for language and cognition. In L.W. Poon, J.L. Fozard, L.S. Cermak, D. Arenberg & L.W. Thompson (Eds.), New directions in memory and aging (pp. 139-155). Hillsdale, NJ: Lawrence Erlbaum Associates. Smith, A.D., Park, D.C., Cherry, I~, & Berkowsky, K. (1990). Age differences in memory for concrete and abstract pictures. Journal of Gerontology: Psychological Sciences, 45, 205209. Smith, M.C., & Magee, L.E. (1980). Tracing the time course of picture-word processing. Journal of Experimental Psychology: General, 109, 373-392. Snodgrass, J.G, (1980). Towards a model for picture and word processing. In P. Kolers & M. Wrolstad (Eds.), Processing of visible language (Vol. 2, pp. 565-584). New York: Plenum Publishing. Snodgrass, J.G. (1984). Concepts and their surface representations. Journal of Verbal Learning and Verbal Behavior, 23, 3-22. Snodgrass, J.G., & McCullough, B. (1986). The role of visual similarity in picture categorization. Journal of Experimental Psychology: Learning, Memory, and Cognition, 12, 147-154. Snodgrass, J.G., & Vanderwart, M. (1980). A standardized set of 260 pictures: Norms for name agreement, image agreement, familiarity, and visual complexity. Journal of Experimental Psychology: Human Learning and Memory, 6, 174-215.
170
P.c. Amrhein
Theios, J., & Amrhein, P.C. (1989). Theoretical analysis of the cognitive processing oflexical and pictorial stimuli: Reading, naming, and visual and conceptual comparisons. Psychological Review, 96, 5-24. Thomas, J.C., Fozard, J.L., & Waugh, N.C. (1977). Age-related differences in naming latency. American Journal of Psychology, 90, 499-509. Trahan, D.E., Larrabee, G.J., & Levin, H.S. (1986). Age-related differences in recognition memory for pictures. Experimental Aging Research, 12, 147-150. Treat, N.J., & Reese, H.W. (1976). Age, pacing, and imagery in paired-associated learning. Developmental Psychology, 12, 119-124. Tubi, N., & Calev, A. (1989). Verbal and visuospatial recall by younger and older subjects: Use of matched tasks. Psychology and Aging, 4, 493-495. Waugh, N.C., Thomas, J.C., & Fozard, J.L. (1978). Retrieval from different memory stores. Journal of Gerontology, 33, 718-724. Wheeldon, L.I~, & Monsell, S. (1992). The locus of repetition priming of spoken word production. The Quarterly Journal of Experimental Psychology, 44A, 723-761. Winograd, E., & Simon, E.W. (1980). Visual memory and imagery. In L.W. Pooh, J.L. Fozard, L.S. Cermak, D. Arenberg & L.W. Thompson (Eds.), New directions in memory and aging (pp. 485-506). Hillsdale, NJ: Lawrence Erlbaum Associates. Winograd, E., Smith, A. & Simon, E.W. (1982). Aging and the picture superiority effect in recall. Journal of Gerontology, 37, 70-75.
Age Differences in Word and Language Processing Ph. Allen and Th.R. Bashore (Editors) 9 1995 Elsevier Science B.V. All rights reserved.
171
Aging and the distribution of resources in w o r k i n g memory* Elizabeth A. L. Stine University of New Hampshire
In a 1988 address at the Cognitive Aging Conference, Tim Salthouse diagnosed the Cognitive Aging community with a terrible disease, RORAP Syndrome, an acronym for an insidious pathology in which its victims are compelled to provide vacuous accounts for agerelated changes in cognition, and in particular, to Rely On Resources As a Pseudo-explanation .(Salthouse, 1988). He helped those of us suffering t~om this syndrome to identify ourselves, and then went on to provide a 12-step program to help us avoid the circular reasoning that commonly strikes those afflicted with the syndrome. In this talk and subsequent "pamphlets" .(Salthouse, 1988a; Salthouse, 1988b; Salthouse, 1988c; Salthouse, 1991a; Salthouse, 1991b; Salthouse & Babcock, 1991), he has discussed three dominant metaphors used to conceptualize an age-related reduction in processing resources, alternatively, as a reduction in processing speed (slowing), a reduction in mental energy (attention), and finally (which brings us to the topic of this chapter), a reduction in working memory capacity. The working memory (WM) construct is an intellectual descendent of William James' .(1890) notion of "primary memory." Essentially, the world is bigger than we are, and our capacity for constructing knowledge is virtually infinite, but mediating the two is a buffer that is extremely limited in its capacity for storing and manipulating information. The assumption is that this buffer varies between as well as within age groups, and that individual differences in this capacity affect a wide range of cognitive abilities. The years since the diagnosis of RORAP Syndrome have seen a plethora of studies in which recovering RO1LAPs have worked to operationalize this elusive construct. So time as a resource is the change in reaction time in response to conditions varying in complexity; mental energy is the ability to effectively carry out multiple activities at once (as in divided attention tasks), and working memory capacity is measured in a loaded span task as number of items that can be recalled after they are encoded during some concurrent activity. Interestingly, these investigations have revealed an homology among these metaphors. Salient examples are McDowd and Craik's .(1988) demonstration that divided attention increases processing time, especially for older adults, Salthouse's numerous demonstrations (e.g., .Salthouse, 1991a ) that perceptual-motor speed can account for loaded span performance, Tun et al.'s .(1991) demonstration that older adults with higher spans are better at divided attention tasks, and data t~om our lab .(Stine & Hindman, 1994) showing that readers with greater working memory capacity are faster at encoding the idea units from text.
AUTHOR NOTE: Address correspondence to: Department of Psychology, Conant Hall, University of New Hampshire, Durham, NH 03824, (603) 862-3806. Email:
[email protected] This chapter is based on a presentation at the Fifth Cognitive Aging Conference in Atlanta, GA on April 7, 1994. The research described from our lab was supportedby grant R29 AG08382 from the National Institute on Aging. I am grateful to Dan Morrow and Phil Allen for helpful commentson earlier drafts of this manuscript.
172
E.A.L. Stine
So energy is time, and time is space, and space is energy, and I suppose we are left with if the reader will excuse the expression - - "resources." In fact, Salthouse .(1991b) has articulated the interchangeable nature of these constructs very nicely: Although it is convenient to categorize speculations about processing resources in terms of metaphors of time, space, and energy, these conceptualizations are not necessarily distinct and independent. Not only are the arguments for the each resource based on similar reasoning, but the same results are sometimes interpreted as evidence for different resource conceptualizations .... [T]he different metaphorical interpretations may be interrelated .... [I]mpairments in divided attention might be a consequence of slowness in alternating between processes .... reductions in attention would lead to increase delays between successive processing operations .... Slower rates of processing have additionally been linked to reduced capacity of working memory .... (Salthouse, 1991, pp. 346-348) In this paper, I would like to do two things: First, I'd like to give a brief historical overview of the working memory concept, considering some recent findings in the cognitive aging literature implicating the role of WM resources in age-related changes in performance. Second, I'd like to describe yet another affliction from which those in my lab and elsewhere are trying to recover, WM UPROAR Syndrome: the Woefully Misbegotten Underestimate of the Partitioning of Resources to Overcome Age-Related deficiencies. That is, even though many of the age-related differences we observe may well be explicable in terms of a reduction in working memory resources, I'd like to argue that there is some flexibility in the system in the manner in which resources are allocated, and such allocation can go a long way toward maintaining a high level of performance in later adulthood. 1. AN A C C E L E R A T E D HISTORY OF THE W O R K I N G M E M O R Y CONCEPT AND ITS APPLICATION TO COGNITIVE A G I N G R E S E A R C H As noted earlier, WM can be traced back to James' (1890) discussion of primary memory, our consciousness of the "specious present" (Vol 1, pp. 641-642). With over a hundred years of rumination and empirical contortions of this concept behind us, James' description of the ephemeral nature of primary memory and the consequent quest for coherence is still compelling: [For] a state o f mind to survive in memory it must have endured f o r a certain length o f time .... Any state of mind which is shut up to its own moment and fails to become an object for succeeding states of mind, is as if it belonged to another stream of thought. Or rather, it belongs physically, not intellectually, to its own stream, forming a bridge from one segment of it to another, but not being appropriated inwardly by later segments or appearing as part of the empirical self.... All the intellectual value for us of a state of mind depends on our after-memory of it .... Only then does it count for us. (pp. 643-644, italics are James') That metaphorical single step into the stream of consciousness, that ever-present "now," was incorporated into experimental psychology as a limited-capacity bottleneck in the cognitive system Waugh and Norman .(1965) drew heavily on James' presentation in their seminal paper on primary memory (PM), but specifically conceptualized this bottleneck as a passive store:
Aging and the distribution of resources in working memory
173
An event in PM has never leit consciousness and is part of the psychological present .... PM is a faithfifl record of events just perceived .... James believed that PM extends over a fixed period of time. We propose instead that it encompasses a certain number of events regardless of the time they take to occur" (pp. 92-93). Their goal was to distinguish PM from a more enduring secondary memory store. As a mathematical function relating the number of items recalled to the number of intervening items, the emphasis was on storage. This was true even as the model was elaborated into a full multistore model (with a "short-term store" and a "long-term store") and granted "control processes" that managed the contents of the short-term store .(Atkinson & Shiffdn, 1968): the primary, short-term memory was regarded as a number of slots that were continually filled and replaced. The multistore approach was criticized on many counts, but most saliently on the grounds that capacity and duration of the short-term store was not, in fact, independent of the effects of knowledge, a rather fundamental assumption of the model.(Craik & Lockhart, 1972). In spite of the fact that short-term memory is periodically buried in the literature, the core conceptualization behind James' primary memory has remained with us. To accommodate findings that capacity depended on the nature of what was being done with items as well as the nature of concurrent processing, the passive store was endowed with a capacity for processing which competed with storage demands and re-christened "working memory." A primary architect has been Alan Baddeley: The core of the working memory system consists of a limited capacity 'work space' which can be divided between storage and processing demands (Baddeley & Hitch, 1974, pp. 75-76). [This] model subdivided WM into three components, the Central Executive, which formed the control centre of the system, was assumed to select and operate various control processes. It was assumed to have a limited amount of processing capacity, some of which could be devoted to the short-term storage of information. It was able to ottload some of the storage demands to subsidiary slave systems, ... the Articulatory Loop, which was able to maintain verbal material..., and the Visuo-Spatial Scratch P a d [responsible for] the visualization of spatial material .(Baddeley, 198L, p.
18). Some theorists now reserve the term "short-term memory" to refer to the acoustically-based storage buffer needed for the preliminary analysis of language and use the term "working memory" to refer to the processing component that manipulates information (Graesser, Singer, & Trabasso, 1994), but the point is that concern with James' "specious present" has largely shifted from issues of storage properties to issues of processing function. With Daneman and Carpenter's .(1980) demonstration that the sentence span task (which requires subjects to process sentences while holding their final words in memory) was highly predictive of verbal SAT scores, working memory became entrenched in the literature as the individual difference that made the difference in language processing, in memory, in problem solving. The trade-off between processing and storage seems like a potential source of individual differences in reading comprehension .... The better reader might have more efficient processes so that he/she effectively would have more capacity for storing and maintaining information. (p. 451)
174
E.A.L. Stine
Table 1. Studies testing the prediction of an Age by Complexity interaction. Study Babcock & Salthouse (1990)
Dependent variable Digit and location recall
Difficulty manipulation Concurrenttask/ retrieval demands
AX C?
Crossley & Hiscock (1992)
Tapping speed
Reading normal vs rotated text Speech repetition vs. fluency Maze solving vs. tracking
Yes
Salthouse et al. ( 1 9 9 0 )
Accuracyin numerical and spatial tracking
Number variables tracked Numberof operations
No
Salthouse (1992)
Reasoning accuracy
Number of premises
Yes
No
(omitting Ss at chance) Salthouse & Skovronek (1992)
Accuracy in matching rotated cube
Angle of cube
Yes
Tun et al. (1991)
Speech recall
Concurrent RT
No
Tun et al. (1992)
RT during speech recall
Numberof RT choices Propositional density
No
Wiegerson & Meertse (1990)
Digit recall
Digit span vs. missing span
Yes
This research was important in expanding the focus of working memory from experimental work illustrating the general principles of memory in the generic college sophomore to a consideration of the differences between people who were more or less successful in their cognitive accompli~qhments. The divergent approaches of Baddeley and Daneman offer an interesting contrast that has set the tone for much of the cognitive aging work in this area. In the former, support for the model is garnered by charting the deterioration of performance as demands for storage and processing are increased. In the latter, support rests on intercorrelations between a presumed estimate of working memory and criterial measures of cognitive performance. Given the success of the working memory construct in explaining cognitive performance among the young, the extension to aging was obvious. It could be that there is a reduction in working memory capacity or resources, and that this is responsible for much of the observed cognitive declines. Consistent with the approaches just outlined, this has been addressed in two ways. The first examines the differential effects of increasing storage and processing demands on older adults. If aging brings a decrease in WM resources, then elderly
Aging and the distribution of resources in working memory
175
performance should be disproportionately depressed by these demands (in terms of requiting more time or engendering lower accuracy). The meta-analytic approaches of Hale, Myerson, Cerella, and others .(Cerella, 1990; Myerson, Hale, Wagstait~ Poon, & Smith, 1990) have for the most part supported the resource deficit hypothesis. Manipulations within individual studies, however, have met with mixed success. Table 1 summarizes several studies which have addressed this question. This table is by no means exhaustive but illustrates what I think are some representative studies in this area. There is about a 50:50 hit rate in obtaining the critical Age X Complexity interaction. While there have been some interesting post hoc attempts to explain what kind of difficulty exacerbates age differences and what kind does not, we really do not have a good theory of the nature of the complexity that taxes elderly resources. For example, the distinctions between storage capacity and processing capacity, or between structural capacity and operational capacity, while intuitively appealing, don't seem to account for the difference. In a recent paper, Salthouse, Babcock, and Shaw (1991) has suggested that what may be required is a transformation of the essential nature of the representation to a more abstract form This remains to be tested. In any case, we still don't have a theory of what kind of empirical difficulty strains the resource capacity of older adults. Even though the Age by Complexity interaction has been elusive, the individual differences approach has been somewhat more successful in supporting working memory as a mediator of age differences in cognitive performance. Table 2 summarizes representative studies using this approach. Again, this list is not exhaustive but (I hope) fair in showing that using a wide variety of indices of working memory m measures derived out of the context of criterial performance, like loaded span or perceptual speed, and measures derived from within the context of criterial performance, like repeating requests for information and a diverse set of measures of cognition, age-related variance in performance can often be substantially accounted for by a decline in working memory resources. These approaches are, of course, predicated on the assumption that there is an undifferentiated pool of resources in working memory. Some theorists, however, have argued otherwise, that is, that working memory is best thought of as a set of distributed capacities of different modalities. This is exemplified by Monsell .(1984), who after a thorough review of patterns of selective interference, argued: "The simplest conclusion is that WM is no more (or less) than a heterogeneous array of independent temporary storage capacities intrinsic to various subsystems specialized for processing in specific domains" (p. 344). A similar conclusion was reached by Daneman and Tardif.(1987). Creating a set of working memory measures, some verbal and some spatial, they also obtained separate measures for processing performance on these tasks with and without the storage component. Contrary to the notion that it is a generalized capacity for simultaneous processing and storage that is responsible for individual differences in cognitive outcome, spatial tasks were more predictive of math SAT, verbal tasks were more predictive of verbal S A T - and simultaneous storage didn't make a difference! That is, the processing component of the task was just as predictive of performance as was the processing plus storage component. They argued: We think we now have a wider range of measures of working memory capacity and that the picture suggests the need for abandoning the notion of a "general and central limitation on information processing .... " .... At the very least, we may have to posit two separate processors, one for representing and manipulating verbal-symboli
176
E.A.L. Stine
Table 2. Do individual differences in an index of working memory account to some extent for age-related change in cognition? Study
WM Index
Context
Criterion
Hultsch et al. (1992)
Reading span
Out
Text recall Word recall
No
Morrow et al. (1992)
Reading span
Out
Text recall
Yes
Salthouse (1992)
Reading and computation span Perceptual speed Recognition of premises
Out
Reasoning accuracy
Yes
Salthouse & Skovronek (1992)
Line span Repeated requests for infn Recognition of infn
Out
Accuracy in matching rotated cube
Yes
Stine et al. (1993)
Average loaded span
Out
Troyer et al. (1994)
Executive function (concept formation, flexibility)
Out Word recall Visualmemory
Yes
Tun et al. (1991)
Average loaded span expository texts
Out
Yes
Out In
In In Recall of spoken narrative
Recall of spoken
Yes
information, and a second for representing and manipulating spatial information" (p.
502). While there seems to be some consensus that verbal and non-verbal processing represent different factors in working memory, issues of the structure of WM remain unresolved. The "distributed capacities" approach is increasingly incorporated into cognitive aging research. For example, even the generalizability of "general slowing" seems to depend on whether the task is in the verbal or nonverbal domain .(Hale, Lima, & Myerson, 1991; Myerson, Ferraro, Hale, & Lima, 1992). 2. W O R K I N G M E M O R Y , L A N G U A G E UNDERSTANDING, AND A G E One outcome of the distributed capacities notion appears to be that the literature on working memory has become more insular to particular domains, and it is particularly flourishing in the domain of language. It is to the role of working memory in language
Aging and the distribution of resources in working memory
177
processing that I now turn. In fact, language provides an interesting laboratory for the study of working memory. That is, one way to study the effects of straining working memory resources is to have subjects simultaneously encode word lists, monitor a CRT for pink elephants, pat their heads, and rub their bellies. Another way is to have them comprehend language. In so doing, they must access word meanings, create a text-based representation of propositional content, and construct a broader representation of the situation at the mental model l e v e l - as when text understanding entails encoding relative spatial arrangement, or the emotional reaction, predisposition, or goals of characters. The application of working memory models in language has been a major force in driving the dominant metaphor from what was once a box of items to what is now more often conceptualized as the amount of activation available in a knowledge net. Good examples of this perspective may be found in recent papers by Just and Carpenter .(1992) and by Engle and colleagues.(1992): [Our] purpose ... is to present a theoretical integration of the storage and processing functions of working memory in language comprehension .... In this framework, capacity can be expressed as the maximum amount of activation available in working memory to support either of these functions .... We propose that individuals vary in the amount of activation they have available for meeting the computational and storage demands of language" (Just & Carpenter, 1992, pp. 123-4). Working memory consists of those knowledge units that have recently been activated either from objects in the environment or as a result of productions and are in various states of loss of activation through either decay or inhibition" (Engle, Cantor, & Carullo, 1992, p. 990). This metaphor accommodates a number of interesting findings. For example, Gernsbacher's .(1991) work showing that high-span subjects are more likely to suppress the irrelevant meanings of words suggests that high-span readers have more resources because they are effective in selectively activating appropriate knowledge nodes, i.e., distributing their resources to relevant information. Just and Carpenter's (1992) work showing that the impenetrability of syntactic analysis by semantic constraints is not true for high-span readers is explicable in terms of the greater activation they have available; thus, high-span readers are better able to simultaneously process the multiple levels of discourse. Finally, there are many findings with respect to reference and coherence in discourse that suggest that information that is not currently active in working memory can, nevertheless, be readily accessible under some circumstances. Notions invoked to explain these findings must abandon simple assumptions about information either being "in" or "not in"working memory.(Kintsch, 1988); for example, Sanford and Garrod .(1981) make a distinction between information in an explicit focus in working memory that is highly active (and can therefore be unambiguously referenced with pronouns) and information in an implicit focus that is less active but relatively available because of inferential connections entailed by what is active; Glenberg and colleagues .(1987) describe "discourse pointers" in which information that is active in working memory points to (or makes more available) relevant information that is not active; Graesser et al..(1994) have argued that a "search for meaning" in discourse drives production rules in WM that guide longterm memory searches; finally, Hintzman .(1986) and O'Brien .(1995) have argued for the existence of an automatic resonance process invoked by the featural overlap between the active nodes in working memory and knowledge net.
178
E.A.L. Stine
This shiR in the foundational metaphor for representing the contents of working memory may prove to be important for understanding age differences in language processing. Let me give you an example of some recent findings from our lab that would be hard to explain with the WM "box" metaphor but are well accommodated by the "activated nodes" metaphor (Hakala, RizeUa, Stine, & O'Brien, in preparation). In this study, young and older adults read narratives in which a protagonist is initially introduced as having a given predisposition, for example, Andy is described as being a vegetarian. The narrative line proceeds with filler material without again making reference to this fact. At a later point, the character performs an act that is inconsistent with the earlier characterization, for example, Andy eats a fivecourse meal including steak tartare, lobster medallions, and raspberries with whipped cream When readers encounter this information, their reading times increase (rdative to control passages in which these statements represent no inconsistency, for example, if Andy were described as being a gourmet), suggesting that they, in fact, reco~ize that (in the context of this narrative anyway), this action is inconsistent with the nature of the protagonist. Now this is hard for a traditional WM store model to explain. Since the description about Andy is no longer "in working memory," readers should be oblivious to the break in coherence. Thus, it must be that there is some mental model level of representation that is at least in implicit focus to guide comprehension, enabling the reader to reactivate critical information into explicit focus so that the inconsistency can be resolved; presumably the increase in reading time in the inconsistent condition represents the allocation of resources needed to search, reactivate, and perform the elaborative processing needed to achieve coherence. In support of this notion, recall is actually higher in the inconsistent condition than in the consistent one .(Albrecht & O'Brien, 1993). Two results, however, suggest that the identical increment in reading time for younger and older adults (i.e., no Age X Consistency interaction) was not sufficient to complete coherence processing among the old when the pieces of information to be reconciled were far apart in the text. First, under these conditions, older adults did not show the improvement in recall that younger adults did .(Hess & Tate, 1991). In addition, there was evidence that the average older adult did not reactivate into explicit focus the information needed to resolve the inconsistency. In another experiment relying on a technique developed by Jerry Myers and colleagues, subjects responded to probes testing the availability of the protagonist's features. So for example, on target trials subjects verify statements like "Andy is a vegetarian." While this information was less available aider the filler information for both younger and older adults, it was only the younger readers who reactivated the target in the face of inconsistency presumably to resolve it and update the mental model in light of this new information, e.g., OK, so Andy is a vegetarian with occasional lapses. Thus, it would appear the average older adult in our sample did not do this, and to the extent that it is generally the case that elders do not allocate resources in working memory to such reintegration, this might contribute to agerelated declines in discourse processing. This explanation is consistent with Craik's argument .(Craik & Jennings, 1992) that elderly adults are less likely to self-initiate processing. This account has potentially profound implications for understanding the phenomenological experience of language comprehension in later life. In James' terms, if each segment were "shut up in its own moment," then the "intellectual value" of the discourse would be diminished. This resource allocation hypothesis must be examined in light of the fact that older adults did indeed slow their reading when faced with the inconsistency. On the average, older
Aging and the distribution of resources in working memory
179
adults were as responsive to the break in global coherence as were the young - - assuming that the allocation of reading time is equally effective for young and old. But perhaps that assumption is not warranted. This is an important issue and it is to the interpretation of reading time allocation that I would like to direct the argument. In a recent study, Hartley, Stojack, Mushaney, Atmon, and Lee .(1994) have shown across a variety of methods that when young and older adults are matched on reading time, older adults recall systematically less, and that that difference increases as more time is allocated (see their Figure 1). In addition, evidence for cognitive slowing in a variety of tasks (see Salthouse, 1991b)would also suggest that the allocation of equivalent amounts of time by younger and older readers would not be expected to yield the same outcome. A Brinley plot of our reading times makes this point. With a correlation of .98 between the reading times of young and old, and a slope of .94, the Brinley analysis supported the contention that the older adults were responding j u s t as the young. There were two things striking about this analysis. First, the extremely high correlations between the reading times of the young and the old in a domain that does not involve discrete trials, but rather self-paced reading, is noteworthy since it suggests a qualitative similarity in resource allocation across a range of text demands. In addition, the fact that the slope of the fimction relating reading times of the old to those of the young was just about unity suggests a curious exception to the now familiar Brinley plot with a slope of about 1.5. In fact, we have collected a lot of reading time data in my lab lately, and we have been consistently struck by the fact that these Brinley plots never conform to the predictions of generalized slowing. Note we are not concerned with discriminating degrees of slowing .(Perfect, 1994); we simply never observe it at all in reading time! Now it could be that such results suggest a preservation of p r o c e s s i n g - a failure to find age-related slowing in the domain of language, as suggested by the meta-analysis oflexical decision times by Laver and Burke .(1993). This seems unlikely if these are the reading times that ultimately yield recall or comprehension d e f i c i t s - as is the case with the criterion measure of this task, which were the probe verification times. A perusal of recent literature on text processing in which reading time was measured suggests that it is not atypical for the average adult reader to fail to accommodate for cognitive slowing. For example, Hartley .(1993) found that within-group variability overshadowed between group differences in reading speed. Similarly Hartley et al. (1994) did not find age differences in self-paced reading speed. A Brinley analysis of the mean reading times in different conditions by Hamm and Hasher .(1992) yield a slope of.84 (r=.70). For the domain experts in Morrow et al..(1992), the Brinley slope is 1.18 (r=.90) and for the domain novices, it is .84 (r=.82). The one reading time data set for which a Brinley analysis produces a high, positive slope is that of Connelly, Hasher, and Zacks .(1991) in which the slope was 2.86 (r=.99) (numerical values were estimated from the data in Figure 3). In this study, however, differences among conditions were created by adding interfering material. The variation in difficulty in this study then may not reflect natural variation in the demands of reading. In any case, the bulk of the data suggest that older adults are not particularly slower at reading in terms of overall reading time nor in terms of how they keep pace with text difficulty. Thus, there is now considerable evidence for age-constancy in reading time allocation. This ageequivalence would presumably not accommodate cognitive slowing and is furthermore often coupled with age differences at retrieval. Another study from our lab has begun to address more specifically how these resources are allocated on-line .(Stine, Loveless, & Soederberg, in preparation). Subjects read texts
180
E.A.L. Stine
sector-by-sector (sectors were groups of words that were syntactically well-formed and presented in response to a button press). Again, we measured reading time (this time for each sector), and across three conditions the slope ranged from.86 to 1.08. Thus, as in the Hakala study, the reading times did not conform to the predictions of cognitive slowing. In this case, however, there was no group age difference in subsequent recall performance. (This appeared to be an artifact of the higher verbal ability of our elderly group: in a regression analysis of recall, vocabulary level was a positive predictor and age was a si,~m~ificantnegative predictor.) Using techniques pioneered by Doffs Aaronson and Karl Haberlandt (of. Aaronson & Scarborough, 1977; Haberlandt, 1984), we used regression analyses to decompose the reading times for each subject to reflect the allocation of time to word-level (i.e., length in syllables, word frequency (f)), text-level (i.e., number of propositions, number of new concepts introduced, syntactic complexity as measured by Yngve depth), and discourse-level (i.e., serial position) features. Figure 1 shows the average values of these regression coefficients for younger and older adults. Overall, the qualitative way in which resources are allocated in working memory to process these texts is quite similar for young and old: both groups slowed down for longer words (Syll), informationally dense sectors (Props), the introduction of new concepts (NewConc), and complex syntax (Yngve); both groups read more quickly when the words in the sector were more familiar (M log f) and when the sectors were later in the text (SerPos). In spite of this qualitative similarity, there are some subtle age differences in the extent to which the two age groups responded to features that represent the formation of a text-based level of meaning. Older adults allocated less time to process the text-based meaning of the passage, spending less time per proposition, p 80 years) .................. 243 Very Mild A D ............................ 235 Mild/Moderate A D ..................... 378 4. Nebes, Brady & H u f f ( 1 9 8 9 ) - LDT, PWs, w o r d = asssociated/unassociated prime Y o u n g .......................................... 32 Old .............................................. 47 A D ............................................ 238
234
F.R. Ferraro
5. Ober & Shenaut (1988) - LDT, PWs (misspelled words), NWs (random letters) - words = high-freq. + low-freq.
Old AD
PW lex.
NW lex.
114 600
-86 -55
Note: AD signifies Alzheimer's disease; Freq. signifies Frequency; HF si~ifies High Frequency; ISI signifies Inter-Stimulus Interval; lex. si~ifies Lexicality Effect; LDT si~ifies Lexical Decision Task; LF si~ifies Low Frequency; ms signifies milliseconds; NW si~ifies NonWord; PWs si~ifies Pseudowords; R si~ifies Related; SOA si~ifies Stimulus Onset Asynchrony; U si~ifies Unrelated
An immediate observation from Tables 1 and 2 is the fact that across the vast majority of these studies the pattern of the lexicality effect that emerges is one in which young adults have a smaller lexicality effect than healthy older adults, who in turn have a smaller lexicality effect than the demented individuals. Error rates are also consistent with this pattern, with young adults typically being much more accurate than healthy older adults, who in turn are typically much more accurate than the individuals with Alzhdmer's disease. Ferraro and Balota (1993) have interpreted such increases in the lexicality effect as supportive of recent arguments regarding breakdowns in inhibitory processes in both healthy older adults and individuals with AD. The reasoning behind this is as follows: Because primes are always words, subjects need to suppress their word response to the prime items on pseudoword target trials. This suppression ability (or inability) appears most difficult for older adults (as compared to younger adults) and for demented individuals (as compared to older non-demented adults), leading to increases in both response latencies as well as error rates. This particular interpretation of the lexicality effect results provides converging evidence regarding other recent investigations into the increasing breakdown in the ability of these subject groups (old adults, demented adults) in their ability to inhibit partially activated (but inappropriate) information. Recent investigations with older adults (e.g., Hasher & Zacks, 1988; Hasher, Stoltzfus, Zacks, & Rypma, 1991; McDowd & Oseas-Kreger, 1991; Tipper, 1991) have revealed that this subject group does not appear to inhibit irrelevant information as much as healthy young adults. Similarly, the same pattern has emerged in investigations that have tested older adults and demented adults (Balota & Duchek, 1991; Duchek, Balota, Ferraro, Gernsbacher, Faust, & Conner, 1992; Ferraro, Balota, & Connor, 1991). Since these various studies have addressed a variety of different tasks, it may be the case that this failure to inhibit irrelevant information is a general characteristic of both older adults and, to a greater extent, demented individuals. It is precisely this lexicality effect difference across older and demented individuals that could potentially be used as a clinical diagnostic marker. This particular interpretation of the lexicality effect results will be taken up in the Discussion. The point, however, is that the lexicality effect appears to be a potentially-relevant diagnostic marker for cognitive declines evident in older and demented populations.
Aging, Alzheimer's disease, and word recognition
235
While the majority of word recognition studies performed on older and dementing adults are visual, several more recent attempts have been made to investigate the role of aging and AD on auditory word recognition as well. Given the compensatory nature of many older and dementing adults cognitive perfo~ance, it is important to detail how these populations perform word recognition processing when the information arrives within a different modality. Such results have long-term ramifications for how communication processes may be better served in these populations.
4.3. Auditory Word Recognition: Models, Theories and Applications to Healthy Aging and Alzheimer's Disease Like visual word recognition, auditory word recognition also relies on the passage of time and a variety of bottom-up and top-down processes interacting to arrive at a suitable final word candidate. However, that is (primarily) where the similarity between these two versions of word recognition end (although see Johnson, 1992; Johnson & Pugh, 1994, for their Cohort model of visual word recognition). In general, spoken language typically has more demands placed upon its comprehension, and these demands would appear to be more compromising for older and demented adults. For instance, spoken language arrives much faster than written language, and the listener has much less control over the input, unlike in reading (Wingfidd & Stine, 1991). There is also the problem of information overload resulting from such rapid information processing, which would tend to further compromise the elderly and demented with regard to attentional resource capacity (Kellas, Simpson, & Ferraro, 1988) as well as short-term working memory capabilities (Morris, Crick, & Craik, 1988). Various models of auditory word recognition have been advanced with the last 10-15 years, and the similarity to some of the models of visual word recognition detailed earlier is striking. Grosjean (1980) developed the Cohort model of auditory word recognition, and the Gating paradigm has been to auditory word recognition what the lexical decision task has been to visual word recognition. In the Gating paradigm~ participants are presented with short (i.e., 25-50 ms) segments of individual words and must attempt to decide what word they are hearing. If the participant is not successful on the initial segment, or gate, successive gates (usually of 100 ms durations) are presented until the correct response is made. Thus, over real time, increasing amounts of the stimulus is presented until it is identified. The initial gate creates what Grosjean termed the Word-Initial Cohort. That is, a relatively large set of possible word candidates are assembled based on this initial 50 ms segment. With the passage of time and any additional gating information, this cohort is substantially reduced in size, and those words the subjects knows that are not consistent with the presented information are eliminated from the cohort. As additional sensory information is presented, the cohort eventually has only one member remaining, and the individual then reco~izes the word (Marslen-Wilson & Welsh, 1978; Tyler, 1984; Wayland, Wingfield, & Goodglass, 1989). Depending upon the context within which the word is presented, the time needed to identify a word ranges from approximately 200 ms (in context) to 330 ms (no context) (G-rosjean, 1980; Marslen-Wilson & Welsh, 1978; Tyler, 1984). Several researchers in a variety of fields have applied the gating technique to real-time estimations of word recognition processing and have included children (Elliott, Hammer, & Evan, 1987; Walley & Metsala, 1990), young adults (Salasoo & Pisoni, 1985), aphasics (Wingfield, Goodglass, & Smith, 1990), and older adults (Bell, 1989; Craig, 1992; Wingfidd,
236
F.R. Ferraro
Aberdeen, & Stine, 1991; Wingfield & Stine, 1991). As with the visual word recognition literature detailed earlier, results from auditory word recognition tasks must also be tempered by the fact that there are specific deficits in the auditory functioning of both elderly adults and AD individuals (e.g., Sinha, Hollen, Rodriguez, & Miller, 1993). These deficits are noted simply so that the reader is aware of them With effective screening procedures, results obtained can still be of theoretical value. In surveying the auditory word recognition literature, only a handful of studies could be located that directly tested younger adults and older adults on auditory word recognition (Bell, 1989; Craig, 1992; Elliott, Hammer, & Evan, 1987; Wingfield, Aberdeen, & Stine, 1991; Wingfield & Stine, 1991). Furthermore, no studies directly testing for auditory word recognition performance could be located that involved individuals with Alzheimer's disease, although AD patients have severe dysfunction when attempting to comprehend the speech of other individuals (Appell, Kertesz, & Fisman, 1982; Kaszniak & Wilson, 1985). Furthermore, several deficits exist in a variety of auditory fimctions in AD (Cummings & Benson, 1989; Kurylo, Corkin, Allard, Zatorre, & Growdon, 1993; Margolis, Taylor, & Dunn, 1985). However, despite the paucity of research within this area with these populations, important findings have been obtained and it is possible to make for healthy older adults and specific predictions can, nonetheless, be made for how Alzheimer individuals would likely perform in such situations. Of the five research reports that have investigated auditory word recognition in healthy older adults, the results appeared mixed at best. Bell (1989) found that older adults benefitted more from semantic context, especially with regard to the word-frequency effect, as compared to younger adults. He reasoned that the elderly adults' performance may be the result of an increased reliance on semantic and lexical information as compensation for degraded peripheral and central encoding (i.e., Stanovich, 1980). In Bell's experiment, young and elderly adults were compared on auditory word identification performance in noise as a function of target word frequency, phonemic similarity neighborhood, and degree of semantic context provided by the carder sentence. Craig (1992) studied real-time isolation monosyllabic word recognition performance in younger and older individuals. Subjects were asked to listen to words, guess what they were, and write down their answer as well as indicate (using a 5-point Likert scale) how confident they were in their decisions. Results revealed that major events in the real-time understanding process of spoken word identification occurred at a slower rate for older, as compared to younger, adults. In other words, the older adults were less able to identify target words at earlier gates and took longer to isolate words, as compared to their younger counterparts. Craig speculated that this dysfunction could be the result of aging, a loss of peripheral sensitivity, more central-type auditory differences and changes, or perhaps due to an interaction of the aging process with both central or peripheral processes. EUiott, Hammer, and Evan (1987) tested 5-7 year-old children, 17-year-olds, and adults aged 70-85 years on their auditory word identification performance. These authors also had subjects rate their confidence in their identification performance. In general, teenagers performed better on the gating task than did the young children and the older adults. Older adults tended to provide more phonetic guesses than either of the other age groups. Furthermore, both the teens and the children displayed better performance regarding their average total acceptance point (i.e., the minimum time of stimulus presentation needed to identify the particular word) than did the older adults. The conclusion was that although the older adults (presumably) had greater experience with the words over the course of their lives, this experience was not
Aging, Alzheimer's disease, and word recognition
237
stdticient to counterbalance the inherent difficulties in processing briet~ temporally altered word stimuli. Wingfield and his colleagues, however, have revealed an opposite pattern to the reports listed above. In particular, these authors have revealed an age constancy with regard to auditory word recognition performance (Wingtield, Aberdeen, & Stine, 1991; Wingtield & Stine, 1991). These authors have found that healthy elderly adults are not compromised in the least in their ability, compared to younger adults, in auditory word recognition experiments. Wingfield et al. (1991) presented subjects with 18 sentence contexts (6 high context, 6 low context, 6 neutral) and the task was to identify the final word (i.e., target) of the sentence. Each sentence context and target word were presented over headphones at varying (50 ms) gates. Results revealed the expected main effect of age (young faster than old) and the expected main effect on context (recall better in high context sentences, followed by low context, followed by neutral context sentences). However, the age by context interaction was not simaificant, suggesting that both young and old adults can identify auditorily presented words with little more than the first half of the word's full acoustic duration. This performance increased for both groups when the context became more constraining. Thus, healthy elderly adults can use context effectively in an on-line experimental situation. These results are similar to those of Kinsbourne (1973), as well as the results offered by Humes, Nelson, and Pisoni (1991) and Humes, Nelson, Pisoni, and Lively (1993). Thus, it appears that this particular research area is ripe for further study, especially given the fact that the handful of research reports examining auditory word recognition in healthy elderly individuals is basically split down the middle. The next question concerns how individuals with Alzheimer's disease would perform in similar auditory word recognition tasks. There is ample evidence that Alzheimer's disease produces substantial auditory system degeneration (Esiri, Pearson, & Powell, 1986; Sinha, Hollen, Rodriguez, & Miller, 1993) which can disrupt additional cognitive performance in these individuals. There is also evidence from a longitudinal study (The Chicago Study) that auditory comprehension of single words declines quite rapidly over a longitudinal time period (Kaszniak & Wilson, 1985), suggesting that auditory word recognition performance in a gating situation would likely evidence a similar sort of pattern. 5. SUMMARY & FUTURE DIRECTIONS The present chapter has attempted to summarize the recent literature pertaining to word recognition processes (both visual and auditory) in older adults and individuals with Alzheimer's disease. The vast majority of the studies reviewed suggests that these very basic cognitive processes are not totally spared in these individuals (e.g., see also Martin, 1992). While some breakdowns exist in the sub-processes that influence word recognition performance (i.e., dysfunctions in visual and auditory functioning for instance), older adults and adults with AD are very adept at performing these and similar tasks (i.e., Parasuraman & Nestor, 1993). It appears very promising to include measures of visual and auditory word recognition as regular components ofneuropsychological/assessment batteries. Given the paucity of recent research regarding auditory word recognition, this area appears especially relevant for further investigation. This enthusiasm stems from the fact that, in the literature reviewed for this chapter, several studies have shown that the word recognition paradigms can be a very
238
F.R. Ferraro
important and sensitive diagnostic tool. Of course, results from a visual or auditory word recognition experiment could not be the sole defining criteria for cognitive dysfunction. However, the diagnostic values of these paradigm~ is related to the diagnostic value reaction time paradimns (which visual and auditory word recognition fall) have had in both older adults and adults with a variety of age-related diseases (Gordon & Carson, 1990; Mahurin & Pirozzolo, 1986; 1993; Muller, Richter, Weisbrod, & Klingberg, 1991). Furthermore, Ferraro and Balota (1993) and Ferraro and Sturgill (1994) have revealed how the lexicality effect (difference between pseudowords and words) increases with age and disease status, and how this fits nicely with the currently-popular theoretical mechanism of a failure to inhibit irrelevant information in both healthy aging and dementia of the Alzheimer type. Although it could be said that very little progress has been made in these areas with regard to aging and AD, given the absololute number of studies reviewed here, it does seem correct in saying that the studies that have been performed nicely indicate the validity and reliability of these paradimns in studying basic, elementary cognitive processes in these populations. Future work can only build on the nice foundation already constructed. REFERENCES Albert, M., & Milberg, W. (1989). Semantic processing in patients with Alzheimer's disease. Brain & Language, 3 7, 163-171. Allen, P. A., Madden, D. J., & Crozier, L. C. (1991). Adult age differences in letter-level and word-level processing. Psychology & Aging, 6, 261-271. Allen, P. A., Madden, D. J., Weber, T. A., & Groth, K. E. (1993). Influence of age and processing stage on visual word recognition. Psychology & Aging, 8, 274-282. Alzheimer, A. (1907). A characteristic disease of the cerebral cortex. In I~ Bick, L. Ammaduci, & G. Pepeu (Eds., & Trans.) (1986). The early story of Alzheimer's disease. Padua, Italy: Liviana Press. Alzheimer's Disease & Related Disorders Association (1987). Chicago, IL. American Psychiatric Association (1994). Diagnostic and statistical manual of mental disorders (4th Ed.), Washington, DC: American Psychiatric Association. Appell, J., Kertesz, A., & Fisman, M. (1982). A study of language functioning in Alzheimer disease. Brain & Language, 17, 73-91. Balota, D. A. (1993). Visual word recognition: The journey from features to meaning. In M. A. Gemsbacher (Ed.), Handbook ofpsycholinguistics. NY: Academic Press. Balota, D. A., & Chumbley, J. I. (1984). Are lexical decisions a good measure of lexical access? The role of word frequency in the neglected decision stage. Journal of Experimental Psychology: Human Perception & PerformanJe, 10, 340-357. Balota, D. A., & Duchek, J. M. (1988). Age-related differences in lexical access, spreading activation, and simple pronunciation. Psychology & Aging, 3, 84-93. Balota, D. A., & Duchek, J. M. (1991). Semantic priming effects, lexical repetition effects, and contextual disambiguation effects in healthy aged individuals and individuals with senile dementia of the Alzheimer's type. Brain & Language, 40, 181-201. Balota, D. A., & Ferraro, F. 1~ (1993). A dissociation of frequency and regularity effects in pronunciation performance across young adults, older adults, and individuals with senile dementia of the Alzheimer type. Journal ofMemory & Language, 32, 573-592.
Aging, Alzheimer's disease, and word recognition
239
Balota, D. A., Ferraro, F. 1L, & Connor, L. T. (1991). On the early influence of meaning in word recognition: A review of the literature. In P. J. Schwanentlugel (Ed.), The psychology of word meanings. NJ: LEA. Balota, D. A., & Lorch, 1~ F. (1986). Depth of automatic spreading activation: Mediated priming effects in pronunciation but not in lexical decision. Journal of Experimental Psychology: Learning, Memory, & Cognition, 12, 336-345. Baron, J., & Thurston, I. (1973). An analysis of the word-superiority effect. Cognitive Psychology, 4, 207-228. Becker, C. A. (1979). Semantic context and word frequency effects in visual word recognition. Journal of Experimental Psychology: Human Perception & Performance, 5, 252-259. Bell, T. S. (1989). Age-differences in spoken word identification: Effects oflexical density and semantic context. Research on Speech Perception Progress Report # 15, 283-302. Indiana University Speech Research Laboratory. Berg. L. (1984). Clinical dementia rating. British Journal of Psychiatry, 145, 339. Berg, L. (1988). Clinical dementia rating (CDR). Psychopharmacology Bulletin, 24, 637639. Berg, L., McKeel, D. W., Miller, J. P., Baty, J., & Morris, J. C. (1993). Neuropathological indexes of Alzheimer's disease in demented and nondemented people aged 80 years and older. Archives of Neurology, 50, 349-358. Berg, L., Miller, J. P., Storandt, M., et al. (1988). Mild senile dementia of the Alzheimer type: 2. Longitudinal assessment. Annals of Neurology, 23, 477-484. Berg, L., Smith, D. S., Morris, J. C., et al. (1990). Mild senile dementia ofthe Alzheimer type. 3. Longitudinal and cross-sectional assessment. Annals of Neurology, 28, 648-652. Bowles, N. L., & Poon, L. W. (1981). The effect of age on speed of lexical access. Experimental Aging Research, 7, 417-425. Bowles, N. L., Obler, L. K., & Albert, M. L. (1987). Naming errors in healthy aging and dementia of the Alzheimer type. Cortex, 23, 519-524. Bowles, N. L., & Pooh, L. W. (1988). Age and context effects in lexical decision: An age by context interaction. ExperimentalAging Research, 14, 201-205. Burke, D. M., & Yee, P. L. (1984). Semantic priming during sentence processing by young and older adults. Developmental Psychology, 20, 903-910. Burke, W. J., Miller, J. P., Rubin, E. H., et al. (1988). Reliability of the Washington University Clinical Dementia Rating (CDR). Archives of Neurology, 45, 31-32. Burke, D. M., White, H., & Diaz, D. L. (1987). Semantic priming in young and older adults: Evidence for age constancy in automatic and attentional processes. Journal of Experimental Psychology: Human Perception & Performance, 13, 79-88. Cantor, M. H. (1991). Family and community: Changing roles in an aging society. The Gerontologist, 31, 337-346. Cerella, J. (1985). Information processing rates in the elderly. Psychological Bulletin, 98, 6783. Cerella, J. (1990). Aging and information-processing rate. In J. E. Birren & I~ W. Schaie (Eds.), Handbook of the psychology of aging (pp. 201-221). NY: Academic Press. Cerella, J., & Fozard, J. L. (1984). Lexical access and age. Developmental Psychology, 20, 235-243.
240
F.R. Ferraro
Clark, E. O. (1980). Semantic and episodic memory impairment in normal and cognitively impaired elderly adults. In L. K. Obler & M. L. Albert (Eds.), Language and communication in the elderly. MA: Lexington Books. Cohen, G., & Faulkner, D. (1983). Word recognition: Age differences in contextual facilitation effects. British Journal of Psychology, 74, 239-251. Coltheart, M., Davelaar, E., Jonasson, J., & Besaer, D. (1977). Access to the internal lexicon. In S. Dornic (Ed.), Attention & performance VI. NJ: LEA. Costa, P. T., Whitfield, J. 1L, & Stewart, D. (Eds.) (1989). Alzheimer's disease: Abstracts of the psychological and behavioral literature. Washington, DC" American Psychological Association. Craig, C. H. (1992). Effects of aging on time-gated isolated word-reco~iiton performance. Journal of Speech & Hearing Research, 35, 234-238. Cronin-Golomb, A., Corkin, S., Rizzo, J. F., Cohen, J., Growdon, J. H., & Banks, K. S. (1991). Visual dysfunction in Alzheimer's disease: Relation to normal aging. Annals of Neurology, 29, 41-52. Cummings, J. L. (1988). Dementia of the Alzheimer type: Challenges of definition and clinical diagnosis. In H. Whitaker (Ed.), Neuropsychological studies in nonfocal brain damage (pp. 86-107). NY: Springer. Cummings, J. L., & Benson, D. F. (1989). Speech and language alterations in dmentia syndromes. In A. Ardila & F. Ostrosky-Solis (Eds.), Brain organization of language and cognitive processes (pp. 107-120). NY: Plenunl DeGroot, A. M. B. (1989). Representational aspects of word imageability and word fxequency as assessed through word associations. Journal of Experimental Psychology: Learning, Memory, & Cognition, 15, 824-845. Duchek, J. M., Balota, D. A., Ferraro, F. 1L, Gemsbacher, M. A., Faust, M. A., & Connor, U T. (1992). The inhibition of irrelevant information in younger and older adults. Presented at the XXV International Congress pfPsychology, Brussels, Belgium. Elliott, U L., Hammer, M. A., & Evan, K. E. (1987). Perception of gated, highly familiar spoken monosyllabic nouns by children, teenagers, and older adults. Perception & Psychophysics, 42, 150-157. Esiri, M. M., Pearson, 1L C. A., & Powell, T. P. S. (1986). The cortex ofthe primary auditory area in Alzheimer's disease. Brain Research, 366, 385-387. Ferraro, F. 1L & Balota, D. A. (1993). Semantic and identity priming in senile dementia of the Alzheimer type (SDAT). Paper presented at the 65th annual meeting of the Midwestern Psychological Association, Chicago, IL. Ferraro, F. IL, & Kellas, G. (1992). Age-related changes in the effects of target orientation on word recognition. Journal of Gerontology: PSYCHOLOGICAL SCIENCES, 47, 279-280. Ferraro, F. IL, & Stur~ll~ D. (1994). Lexical properties and lexical effects associated with National Adult Reading Test (NART) stimuli in healthy younger adults and healthy older adults. Presented at the 5th Biennial Cognitive Aging Conference, Atlanta, GA. Ferraro, F. 1L (1994). Word unit analysis during visual word recognition in young and elderly adults. Developmental Neuropsychology, 1O, 13-17. Forger, K. I. (1976). Accessing the mental lexicon. In 1L J. Wales & E. C. T. Walker (Eds.), New approaches to language mechanisms. Am~erdam: North-Holland. Friedland, 1L P. (1993). Alzheimer's disease: Clinical features and differential diagnosis. Neurology, 43 (Suppl 4), $45-$51.
Aging, Alzheimer'sdisease, and word recognition
241
Friedman, 1L B., Ferguson, S., Robinson, S., & Sunderland, T. (1992). Dissociation of mechanisms of reading in Alzheimer's disease. Brain & Language, 43, 400-413. Gemsbacher, M. A. (1984). Resolving 20 years of inconsistent interactions between lexical familiarity and orthography, concreteness, and polysemy. Journal of Experimental Psychology: General, 113, 256-280. Cfilmore, G. C., Groth, K. E., & Thomas, C. W. (under review). Stimulus contrast and word reading in Alzheimefs disease. Submitted for publication. Gordon, B., & Carson, I~ (1990). The basis of choice reaction time slowing in Alzheimefs disease. Brain & Cognition, 13, 148-166. Grainger, J. (1990). Word t~equency and neighborhood ~equency effects in lexical decision and naming. Journal of Memory & Language, 29, 228-244. Grosjean, F. (1980). Spoken word recognition processes and the gating paradigag Perception & Psychophysics, 28, 267-283. Hart, S., & Semple, J. M. (1990). Neuropsychology and the dementias. London: Taylor & Francis. Hasher, L., Stoltzfus, E. 1L, Zacks, 1L T., & gypma, B. (1991). Age and inhibition. Journal of Experimental Psychology: Learning, Memory & Cognition, 17, 163-169. Hasher, L., & Zacks, 1L T. (1988). Working memory, comprehension, and aging: A review and a new view. In G. Bower (Ed.), The psychology of learning and motivation (pp. 193225). San Diego, CA: Academic Press. Henderson, L. (1982). Orthographyand word recognition in reading. NY: Academic Press. Howard, D. V. (1983). The effects of aging and degree of association on the semantic priming oflexical decisions. Experimental Aging Research, 9, 145-151. Howard, D. V., McAndrews, M. P., & Lasaga, M. I. (1981). Semantic priming of lexical decisions in young and old adults. Journal of Gerontology, 36, 707-714. Howard, D. V., Shaw, 1L J., & Heisey, J. G. (1986). Aging and the time course of semantic activation. Journal of Gerontology, 41, 195-203. Hufl~ F. J., Corkin, S., & Growdon, J. (1986). Semantic impairment and anomia in Alzheimer's disease. Brain & Language, 28, 235-249. Hughes, C. P., Berg, L., Danziger, W., Coben, L. A., & Martin, ILL. (1982). A new clinical scale for the staging of dementia. British Journal of Psychiatry, 140, 566-572. Humes, L. E., Ndson, I~ J., & Pisoni, D. B. (1991). Recognition of synthetic speech by hearing-impaired elderly listeners. Journal of Speech & Hearing Research, 34, 1180-1184. Humes, L. E., Nelson, I~ J., Pisoni, D. B., & Lively, S. E. (1993). Effects of age on serial recall of natural and synthetic speech. Journal of Speech & Hearing Research, 36, 634639. HuRon, J. T., Morris, J. L., Elias, J. W., & Poston, J.N. (1993). Contrast sensitivity dysfunction in Alzheimer's disease. Neurology, 43, 2328-2330. James, C. T. (1975). The role of semantic information in lexical decisions. Journal of Experimental Psychology: Human Perception & Performance, 1, 130-136. Jastrzembski, J. E. (1981). Multiple meanings, number of related meanings, frequency of occurrence, and the lexicon. Cognitive Psychology, 13, 278-305. Johnson, N. F. (1992). On the role of cohorts or neighbors in visual word recognition. In 1L Frost, & L. Katz (Eds.), Orthography, phonology, morphology, & meaning (pp. 147-164). NY: Elsevier Science Publishers.
242
F.R. Ferraro
Johnson, N. F., & Pugh, I~ 1L (1994). A cohort model of visual word recognition. Cognitive Psychology 26, 240-330. Jorm, A. F., Fratiglioni, L., & Winblad, B. (1993). Differential diagnosis in dementia: Principle components analysis of clinical data t~om a population survey. Archives of Neurology, 50, 72-77. Kaszniak, A. W., & Wilson, 1L S. (1985). Longitudinal deterioration of language and cognition in dementia of the Ahheimer's type. Presented as part of a symposium at the annual meeting of the International Neuropsychological Society, San Diego, CA. Kellas, G., Simpson, G. B., & Ferraro, F. IL (1988). Aging and performance: A mental workload analysis. In P. Whitney & 1L Ochsman (Eds.), Psychology & productivity (pp. 35-49). NY: Plenum Publishing Co. Kinsboume, M. (1973). Age effects on letter span related to rate and sequential dependency. Journal of Gerontology, 28, 317-319. Knoll, J. F., & Merves, J. S. (1986). Lexical access for concrete and abstract words. Journal of Experimental Psychology: Learning, Memory, & Cognition, 12, 92-107. Kurylo, D. D., Corkin, S., AUard, T., Zatorre, 1L J., & Growden, J. H. (1993). Auditory function in Al~eimer's disease. Neurology, 43, 1893-1899. Light, L. L. (1991). Memory and aging: Four hypotheses in search of data. Annual Review of Psychology, 42, 333-376. Lopez, O. L., Swihart, A. A., Becker, J. T., Reinmuth, O. M., Reynolds, C. F., Rezek, D. L., & Daly, F. L. (1990). Reliability of NINCDS-ADRDA clinical criteria for the diagnosis of Alzheimer's disease. Neurology, 40, 1517-1522. Madden, D. J., & Greene, H. A. (1987). From retina to response: Contrast sensitivity and memory retrieval during visual word recognition. Experimental Aging Research, 13, 1521. Madden, D. J. (1986). Adult age differences in visual word recognition: Semantic encoding and episodic retention. Experimental Aging Research, 12, 71-78. Madden, D. J. (1988). Adult age differences in the effects of sentence context and stimulus degradation during visual word recognition. Psychology & Aging, 3, 167-172. Madden, D. J. (1989). Visual word recognition and age-related slowing. Cognitive Development, 4, 1-29. Madden, D. J. (1992). Four to ten milliseconds per year: Age-related slowing of visual word identification. Journal of Gerontology: PSYCHOLOGICAL SCIENCES, 47, 59-68. Madden, D., Nebes, R. D., & Allen, P. A. (1992). Cognitive slowing in Alzheimer's disease as a fimction of task type and response type. Developmental Neuropsychology, 8, 459-471. Madden, D. J., Pierce, T. W., & Allen, P. A. (1993). Age-related slowing and the time course of semantic priming in visual word recognition. Psychology & Aging, 8, 490-507. Mahurin, 1L IC, & Pirozzolo, F. J. (1986). Chronometric analysis: Clinical applications in aging and dementia. Developmental Neuropsychology, 2, 345-362. Mahurin, 1L I~, & Pirozzolo, F. J. (1993). Application of Hick's law of response speed in Alzheimer and Parkinson's disease. Perceptual & Motor Skills, 77, 107-113. Malec, J. F., Invik, 1L J., & Smith, G. E. (1993). Neuropsychology and normal aging: The clinician's perspective. In 1L W. Parks, 1L E. Zec, & 1L S. Wilson (Eds.), Neuropsychology of Ahheimer's disease and other dementias (pp. 81-111). NY: Oxford University Press.
Aging, Alzheimer's disease, and word recognition
243
Marcel, A. J., & Patterson, K. E. (1980). Word recognition and production: Reciprocity in clinical and normal studies. In J. gequin (Ed.), Attention & performance (Vol 7). NY: Halstead Press. Margolin, D. I. (1992). Probing the multiple facets of human intelligence: The cognitive neuropsychologist as clinician. In D. I. Margolin (Ed.), Cognitive neuropsychology in clinicalpractice (pp. 18-40). NY: Oxford University Press. Margolis, 1L B., Taylor, J. M., & Dunn, E. J. (1985). An abbreviated Speech Sounds Perception Test with a geriatric population. International Journal of Clinical Neuropsychology, 7, 167-169. Matin, O. S. M. (1987). Dementia and visual agnosis. In G. W. Humphreys & M. J. giddoch (Eds.), Visual object processing (pp. 261-280). NJ: LEA. Marslen-Wilson, W. D., & Welsh, A. (1978). Processing interactions and lexical access during word recognition in continuous speech. Cognitive Psychology, 10, 29-63. Martin, A. (1990). Neuropsychology of Alzhdmer's disease: The case for subgroups. In M. F. Schwartz (Ed.), Modular deficits in Alzheimer's disease (pp. 143-175). Boston: MIT Press. Martin, A. (1992). Semantic knowledge in patients with Alzheimer's disease: Evidence for degraded representations. In L. Backmann (Ed.), Memory functioning in dementia (pp. 119-134). NY: Elsevier Science Publishers, B. V. Martin, A., & Fedio, P. (1983). Word production and comprehension in Alzheimer's disease: The breakdown of semantic knowledge. Brain & Language, 19, 124-141. Max, W. (1993). The economic impact of Alzheimer's disease. Neurology, 43 (Suppl. 4), S6S10. McClelland, J. L., & Rumelhart, D. E. (1981). An interactive activation model of context effects in letter perception: Part 1. An account of basic findings. Psychological Review, 88, 375-407. McCulla, M. M., Coats, M., VanFleet, N., Duchek, J. M., Grant, E., & Morris, J. C. (1989). Reliability of clinical nurse specialists in the staging of dementia. Archives of Neurology, 46, 1210-1212. McDowd, J. M., & Oseas-Krueger, D. M. (1991). Aging, inhibition processes, and negative priming. Journal of Gerontology: PSYCHOLOGICAL SCIENCES, 46, 340-345. McKeel, D. W., Ball, M. J., Price, J. L., Smith, D. S., Miller, J. P., Berg, L., & Morris, J. C. (1993). Interlaboratory histopathologic assessment of Al~eimer neuropathology: Different methodologies yield comparable diagnostic results. Alzheimer's Disease and Associated Disorders, 7, 136-151. McKhann, G., Drachman, D., Folstein, M., Katzman, 1L, Price, D., & Stadlan, M. (1984). Clinical diagnosis of Alzheimer's disease: Report of the NINCDS-ADRDA Group under the auspices of the Department of Health and Human Service Task Force on Alzheimer's disease. Neurology, 34, 39-44. Morris, J. C. (1993). The Clinical Dementia Rating (CDR): Current version and scoring rules. Neurology, 43, 2412-2414. Morris, J. C., McKeel, D. W., Fulling, I~, Torack, 1L, & Berg, L. (1988). Validation of the clinical diagnostic criteria for Alzheimer's disease. Annals of Neurology, 24, 17-22. Morris, J. C., McKeel, D. W., Price, J. L., et al. (1988). Very mild senile dementia of the Alzheimer type (SDAT). Neurology, 38, 227.
244
F.R. Ferraro
Morris, 1L G., Gick, M. L., & Craik, F. I. M. (1988). Processing resources and age differences in working memory. Memory & Cognition, 16, 362-266 Morrison, M. H. (1982). Economics of aging: The future of retirement. NY: Van Nostrand Reinhold. Morton, J. (1969). The interaction of information in word recognition. Psychological Review, 76, 165-178. Muller, G., Richter, R. A., Weisbrod, S., & Klingberg, F. (1991). Reaction time prolongation in the early stage of presenile onset of Alzheimer's disease. European Archives of Psychiatry and Clinical Neuropsyehology, 241, 46-48. Myerson, J., Ferraro, F. R., Hale, S., & Lima, S. D. (1992). General slowing in semantic priming and word recognition. Psychology & Aging, 7, 257-270. Nebes, IL D. (1992). Cognitive dysfimetion in Al~eimel~S disease. In F. I. M. Craig & T. A. Salthouse (Eds.), Handbook of cognition and aging (pp. 373-445). NJ: LEA. Nebes, 1L D., & Brady, C. B. (1992). Generalized cognitive slowing and severity of dementia in Alzheimer's disease: Implications for the interpretation of response-time data. Journal of Clinical and Experimental Neuropsychology, 14, 317-326. Nebes, 1L D. (1989). Semantic memory in Alzheimer's disease. Psychological Bulletin, 106, 377-394. Nebes, 1L D., Brady, C. B., & HutZ F. J. (1989). Automatic and attentional mechanisms of semantic priming in Alzheimer's disease. Journal of Clinical and Experimental Neuropsychology, 11, 219-230. Neely, J. H. (1991). Semantic priming effects in visual word recognition: A selective review of the current findings and theories. In D. Besner & G. Humphreys (Eds.), Basic processes in reading: Visual word recognition (pp. 264-336). NJ: LEA. Ober, B. A., & Shenaut, G. IC (1989). Lexical decision and priming in Alzheimer's disease. Neuropsychologia, 26, 273-286. Owsley, C., & Sloane, M. E. (1990). Vision and aging. In. F. Boller & J. Grafman (Eds.), Handbook ofneuropsychology, Vol. 4. (pp. 229-249). NY: Elsevier Publishers. Parasuraman, R., & Nestor, P. G. (1993). Preserved cognitive operations in early Alzheimer's disease. In J. Cerella, J. Rybash, W. J. Hoyer, & M.L. Commons (Eds.), Adult information processing: Limits on loss (pp. 77-111). NY: Academic Press. Pate, D. S., Margolin, D. I., Friedrich, F. J., & Bentley, E. E. (1994). Decision-making and attentional processes in aging and dementia of the Alzheimer's type. Cognitive Neuropsychology, 11, 321-339. Patterson, I~, Graham, N., & Hodges, J. 1L (1994). Reading in Alzheimer's type dementia: A preserved ability? Neuropsychology, 8, 395-407. Pirozzolo, F. J., Nolan, B. H., Kushkowski, M., Mortimer, J. A., & Maletta, G. J. (1988). Latency and accuracy of word recognition in dementia of the Alzheimer type. Alzheimer's Disease and Associated Disorders, 2, 337-341. Pooh, L. W., Messner, S., Martin, P., Noble, C. A., Clayton, G. M., & Johnson, M. A. (1992). The influences of cognitive resources on adaptation and old age. In L. W. Poon (Ed.), The Georgian centenarian study (pp. 31-46). Amityville, NY: Baywood Publishing. Powell, A. L. (1994). Senile dementia of extreme aging: A common disorder of centenarians. Dementia, 5, 106-109. Powell, D. H., & Whitla, D. K. (1994). Normal cognitive aging: Towards empirical perspectives. Current Directions in Psychological Science, 3, 27-31.
Aging, Alzheimer's disease, and word recognition
245
Price, J. L., Davis, P. B., Morris, J. C., & White, D. L. (1991). The distribution of tangles, plaques, and related immuniohistochemical markers in healthy aging and Alzheimer's disease. Neurobiology of Aging, 12, 295-312. Rocca, W. A., Amaducci, L. A., & Schoenberg, B. S. (1986). Epidemiology of clinically diagnosed Alzheimer's disease. Annals of Neurology, 19, 415-424. Rubenstein, H., Garfield, L., & Millikan, J. A. (1970). Homographic entries in the internal lexicon. Journal of Verbal Learning & Verbal Behavior, 9, 487-494. Salasoo, A., & Pisoni, D. (1985). Interaction of knowledge sources in spoken word recognition. Journal of Memory & Language, 24, 210-231. Salthouse, T. A. (1982). Adult cognition: An experimental psychology of human aging. NY: Springer. Seidenberg, M. S., Andersen, E. S., Kempler, D., & Jackson, C. (1994). Do word frequency and consistency effects dissociate with aging? Paper presented at the 5th Biennial Cognitive Aging Conference, Atlanta, GA. Seidenberg, M. S., & McCleHand, J. L. (1989). A distributed, developmental model of word recognition and naming. Psychological Review, 96, 523-568. Shuttleworth, E. C., & Huber, S. J. (1988). The naming disorder of dementia of Alzheimer type. Brain & Language, 34, 222-234. Simpson, G. B. (1984). Lexical ambiguity and its role in models of word recognition. Psychological Bulletin, 96, 316-340. Sinha, U. I~, Hollen, I~ M., Rodriguez, 1L, & Miller, C. A. (1993). Auditory system degeneration in Alzheimer's disease. Neurology, 43, 779-785. Stanovich, I~ E. (1980). Toward an interactive-compensatory model of individual differences in the development of reading fluency. Reading Research Quarterly, 16, 32-71. Storandt, M., Morris, J. C., Rubin, E. H., Coben, L. A., & Berg, L. (1992). Progression of senile dementia of the Alzheimer type on a battery of psychometric tests. In L. Backman (Ed.), Memory functioning in dementia (pp. 207-226). NY: Elsevier Publishers. Terry, 1L, & Katzman, 1L (1983). Senile dementia of the Alzheimer type: Defining a disease. In 1L Katzman & R. Terry (Eds.), The neurology of aging (pp. 51-84). Philadelphia, PA: F. A. Davis Co. Tipper, S. P. (1991). Less attentional selectivity as a result of declining inhibition in older adults. Bulletin of the Psychonomic Society, 29, 45-47. TippeR, L. J., & Farah, M. J. (1994). A computational model of naming in Alzheimer's disease: Unitary or multiple impairments. Neuropsychology, 8, 3-13. Trobe, J. D., & Butter, C. M. (1993). A screening test for integrative visual dysfunction in Al~eimer's disease. Archives of Ophthalmology, 111, 815-818. Tomlinson, B. E. (1982). Plaques, tangles, and Alzheimer's disease. Psychological Medicine, 12, 449-459. Tomlinson, B. E., & Henderson, G. (1976). Some quantitative cerebral findings in normal and demented old people. In R. D. Terry & S. Gershon (Eds.), Neurobiology of aging (pp. 183-204). NY: Raven. Tyler, L. (1984). Structure of the initial cohort: Evidence form gating. Perception & Psyehophysics, 36, 417-427. Tyro, E. L. (1989). Diagnostic assessment in dementia. In Katona, C. (Ed.), Dementia disorders (pp. 18-43). London: Chapman & Hill.
246
F.R. Ferraro
Walley, A. C., & Metsala, J. L. (1990). The growth of lexical constraints on spOken word recognition. Perception & Psychophysics, 47, 267-280. Wayland, S. C., Wingfield, A., & Goodglass, H. (1989). Recognition of isolated words: The dynamics of cohort reduction. Applied Psycholinguistics, 10, 475-487. Williamson, G. M., & Schulz, K (1993). Coping with specific stressors in Alzheimer's disease caregiving. The Gerontologist, 33, 747-753. Wingtield, A., & Stine, E. A. L. (1991). Expert systems in nature: Spoken language processing and adult aging. In J. D. Sitmott & J. C. Cavanaugh (Eds.), Bridging paradigms: Positive development in adulthood and cognitive aging (pp. 237-258). NY: Preager. Wingfield, A., Aberdeen, J. S., & Stine, E. A. L. (1991). Word onset gating and linguistic context in spoken word recognition by young and elderly adults. Journal of Gerontology: PSYCHOLOGICAL SCIENCES, 46, 127-129. Wingfield, A., Goodglass, H., & Smith, I~ L. (1990). Effects of word-onset cuing on picture naming in aphasia: A reconsideration. Brain & Language, 39, 373-390. Wisniewski, H. M., & Merz, G. S. (1985). Neuropathology ofthe aging brain and dementia of the Alzheimer's type. In C. M. Gaitz & T. Samorajski (Eds.), Aging 2000: Our health care destiny: Vol. 1. Biomedical issues (pp. 231-243.). NY: Springer-Verlag.
Age Differences in Word and Language Processing Ph. Allen and Th.R. Bashore (Editors) 9 1995 Elsevier Science B.V. All rights reserved.
247
Semantic priming in Alzheimer's disease: Meta-analysis and theoretical evaluation* Beth A. Ober a'b and Gregory I~ Shenaut a'b aDivision of Human Development, University of California, Davis bVA Northern California System of Clinics, Pleasant Hill, CA
Whether semantic memory is intact in probable Alzheimer's disease (AD) has been the subject of lively debate since the mid-1980's. There is ample evidence of impaired performance on a variety of tasks requiring the access and utilization of semantic knowledge (world knowledge), including knowledge of words and objects. However, the extent to which this impaired performance is due to semantic memory deficits per se, as opposed to deficits in retrieval mechanisms, strategy implementation, deployment of attention, etc., has become the focus of much theoretical discussion and research activity. (For a thorough review of the literature on semantic memory and AD, see Nebes, 1989; for an update, see Nebes, 1992). The lexical priming paradigm has frequently been used as a tool for assessing the intactness of semantic memory in AD. In this paradigm, the effect of related context on reaction time (RT) to pronounce a word or recognize a word (in a mixed list of word and nonword targets) is measured. The reduction in target RT produced by a preceding related context word (related prime) compared to an unrelated context word (unrelated prime) is known as the semantic priming effect. The most widely accepted explanation of the semantic priming effect is as follows: the spread of activation in the semantic memory network, from the prime's concept node to the related target's concept node, increases the activation level of the target node, allowing more rapid access, matching, and/or retrieval processes for that node. If associative connections between related concepts have been weakened or eliminated as a by-product of the neuropathology associated with AD, then one would expect to see siL-,nificantly less semantic priming in AD than in elderly normal (EN) subjects. (Unless otherwise indicated, the term "normal" in this chapter refers to EN subjects.) It~ however, the semantic representations for the concepts have become degraded in some way by AD, the result would be either: (1) less-than-normal or even zero priming effect (if the concept nodes are so badly degraded that spreading activation has a less-than-normal or null effect on level of activation for that concept node), or (2) greater-than-normal priming (if partially degraded concept nodes have "moreto gain" via spread of activation).
AUTHORNOTES: Correspondenceconcerningthis chapter shouldbe sent to Beth ~ Ober, HumanDevelopment, Department of Applied Behavioral Sciences, U.C. Davis, Davis, CA, 95616 or to the following e-mail address:
[email protected]. Portionsof this meta-analysiswerepresentedat the AmericanPsychologicalSocietymeetingin San Diego, June, 1992. B. A. O ~ s research is supported by the Medical Research Service of the Veterans Administration and by the National Institute on Aging (Grant #R29-AG10848 to B. A~ O~r and Grant #P30-AG10129to W.J. Jagust).
248
B.A. Ober and G.K. Shenaut
The results of semantic priming experiments with AD and EN subjects are mixed. A number of experiments have shown equivalent-to-normal AD priming; investigators in these cases have generally concluded that semantic memory structures and processes are not significantly disrupted in AD (e.g., Nebes, Martin, & Horn, 1984; Ober, Shenaut, Jagust, & Stillman, 1991). However, there are quite a few studies which show si,.~nificantly greater-than-normal AD priming (i.e., hyperpriming). Investigators do not agree in their interpretation of these results. Hartman (1991) for example, concluded that attentional abnormalities which affect the utilization of semantic knowledge are responsible for greater-than-normal AD priming. Chertkow, Bub, and Seidenberg (1989) on the other hand, argued that hyperpriming is indicative of degraded representations for those concepts in semantic memory that are used as primes and targets in the lexical priming paradigms. In this chapter, we present a critical review of theory and methods relevant to research on lexical, semantic priming in AD compared to EN subjects. We also present several types of meta-analyses on the data from all available AD-EN semantic priming experiments. 1. STUDIES S U R V E Y E D We made every effort to find all of the lexical, semantic priming experiments with AD and EN subjects, through June 1993. Both PsychINFO and Medline searches were conducted to supplement our '~oy hand" literature searches. We also sent letters of inquiry to 20 of our colleagues who had authored or co-authored published papers, or delivered papers at conferences, on semantic priming (or closely related topics) in AD and normal aging in an attempt to obtain any new AD priming data sets that we would not otherwise be aware of. Sixteen of these letters resulted in a written response, and we obtained several "in press" or unpublished data sets in response to these letters. A total of 22 AD lexical, semantic priming experiments were available to us for the meta-analysis. One of these experiments--Ober and Shenaut---(1988), showed pronounced negative priming for the AD (but not the EN subjects) and was an outlier (more than 2.00 SD below the mean of all experiments) on each of the four priming effect (PE) measures involving the AD subjects: PE for AD subjects, the difference in PE between AD and EN subjects, percent PE (PE divided by unrelated RT x 100) for AD subjects, and the difference in percent PE between AD and EN subjects. Therefore, this experiment was dropped from further analyses.* The remaining 21 experiments represent seven different research laboratories, and include 13 independent samples of AD and EN subjects. Table 1 provides a detailed summary of these experiments, including information about dementia severity for the AD subjects (all AD samples were mild-to-moderate in dementia severity, with only the highest fimctioning of The dropped experiment involved continuous presentation of stimuli, with a low proportion of related pairs, and would therefore have been categorized as an "automatic" priming experiment, if it had been included in Table 1. The mean AD PE was -59 ms and the mean EN PE was 25; the difference in PE between the two groups (9 AD subjects
and 15 EN subjects) was significant, but in the opposite direction from the hyperprimingexperiments in our survey. The AD subjects in this experiment comprised an independent sample from any other studies in the Ober & Shenaut laboratory, they were mild-to-moderatelyimpaired, with an average score of 110 (out of a maximum score of 144) on the Mattis DementiaRating Scale (Mattis, 1976). Nebeset al. (1989, Experiment 1, row 1 of Table 1) was a positive outlier (greater than 2 SD) on absolute AD PE and on the difference in absolute PE between the AD and EN groups; however, this experiment was well within 2 SD of the mean on percent AD PE and on the difference in percent PE between the AD and EN groups. Therefore,this experimentwas retained. None of the other experiments in this survey were outliers on any of the four PE measures.
Table 1 S u m m a r y o f M e t h o d s and Findings for 21 Semantic Priming E x p e r i m e n t C o n d i t i o n s with Alzheimer's Disease ( A D ) and Elderly N o r m a l ( E N ) Subject G r o u p s I
Citation 2
Dementia 3 Index
N Paradigm 4
s o A5
R.p. 6
#Tfifls 7
AD
EN
16 6 11 16 14 24 12 48 10 32 14 17 17 17 16 16 16 20 17 17 17
16 10 36 16 22 31 21 25 10 32 24 19 20 20 21 15 19 20 19 17 20
uRT AD EN
I
1. 2. 3. 4. 5. 6. 7. 8. 9. 10. 11. 12. 13. 14. 15. 16. 17. 18. 19. 20. 21.
N e b e s e t a l . (1989, Exp.1) Chertkow et al. (1989) Margolm (1987/1988) Nebes et al. (1989, Exp.2) Chertkow et al. (1993, Exp. 1) Hartman (1991) Chertkow et al. (1993, Exp.2) Chertkow et al. (1994) Albert & Milberg (1989)9 B.a]ota etal. (~1991~) Chertkow et al. (1993, Exp.3) Ober et al. (1991 a, Exp.4) Ober et al. (1991a, Exp.2) Ober et al. (1991a, Exp.1) Ober & Shenaut (1990, Exp. 1) Ober & Shenaut (1990, Exp.2) Ober et al. (1991a, Exp.6) Nebes et al. (1984) Ober et al. (1991a, Exp.5) Ober et al. (1991b) Ober et al. (1991a~ Exp.3)
MMS=20.0 MMS=17.5 CDR=.95 MMS=20.0 MMS=25.6 MMS=19.5 MMS=25.6 M2VIS=23.0 Mattis=120 CDR=.92 MMS=25.6 MMS=19.3 MMS=19.3 MMS=19.3 MMS=20.4 MMS=20.3 MMS=19.3 MMS=19.0 MMS=19.3 MMS=21.7 MMS=19.3
L-pairs L-pairs L-palrs P-parrs L-parrs P-parrs L-parrs L-parrs L-parrs P-parrs L-parrs L-parrs L-parrs P-pmrs L-cont. L-cont. L-pairs P-cont. P-pairs L-cont. P-pairs
750 1502 600 750 I186 1397 500 1158 1500 1119 250 250 250 250 1500 2000 250 2000 250 1500 250
.50 .50 .67 .30 .33 .67 .33 .33 .50 .63 .33 17 17 17 .36 .36 17 .25 .17 .27 .17
15 50 20 15 25 20 25 25 15 96 25 20 12 12 48 48 40 20 40 72 20
ii
1374 1255 1147 874 838 1065 845 773 987 698 866 773 739 692 529 644 754 709 685 561 694
> > > > > > > > > > > > > > = > > > > > >
727 831 698 574 595 785 607 582 622 556 645 665 635 599 505 523 651 596 600 463 616
PE AD
EN
I
200 141 121 118 94 74 66 54 43 23 53 51 48 26 26 26 24 22 21 10 -1
> > > > > > > > < > = = = = = = = = = = =
65 25 21 22 29 40 31 27 40 11 29 28 25 22 24 25 7 19 12 18 20
C/A 8
e3
~.~.
t,..,. ~
t~
bO
250
B.A. Ober and G.K. Shenaut
N refers to number of subjects, uRT refers to unrelated prime condition RT, PE refers to priming effect, AD refers to Mzheimer's disease subjects, and EN refers to elderly normal subjects. The >, 0.7 Hz.) 'l/"/ ''~1'' t /'' ~/J I//"
-,
,,Yl.
,/I
,j'~. _
, _~
v
Low Frequency ..............)..............I............."F.............I.............. (< 0.7 Hz.) " 1 -
3~v
,i
i
,./2 ,/3 ,,'4 .s
i
6
I
s~o 11'00 ldSO 22'00 n~,~c
Figure 8. Panel A and Panel B display multi-word ERP data at frontal sites for Young and Elderly subjects, showing the recorded ERP and its High and Low frequency components
334
J.w. King and M. Kutas
processing load increases over time, then the slow potential should track the resource demands of the process, and may not necessarily reach an asymptotic voltage. In a sentence comprehension task, the integration of the linguistic input into a discourse level representation requires continual evaluation and constant linking between the current content and previous knowledge about the topic. This ongoing, accretion of processing might therefore be expected to result in a cumulative effect on slow potentials across the sentence. Conversely, one could argue that, as a sentence progresses, its representation becomes more consolidated and subsequent content becomes more predictable. On this view, the processing load at an integrative level would progressively decrease as more input arrives and it is this that is reflected in the ramp-like slow potentials observed. Whichever interpretation of the slow frontal positivity is correct, the pattern we observed in the elderly was very similar to that seen in the younger subjects. While this is hardly conclusive proof~ it does suggest that integrative processing in structurally simple sentences, as verified by comprehension, is little affected by the aging process. However, we would expect to see a clear effect of aging on integration for sentences whose structure imposes a heavy burden on working memory. This was, in fact, a primary motivation for comparing the performance of young and elderly subjects during the processing of sentences known to tax the limits of working memory and thereby lead to increased comprehension difficulty in all readers, but especially in the elderly. From the work of Kemper and her colleagues (e.g. Kemper, 1988), we know that older adults change both their use and comprehension of various syntactic structures as they grow older. Further, the structures most likely to cause difficulties in either production or comprehension are precisely those that are generally argued to make the greatest demands on working memory capacity. Investigating these ideas requires sentence types that differ in their WMC demands but are otherwise similar enough to allow comparisons between individual critical words and between the sentences themselves. For these reasons, psycholinguistic investigations have frequently concentrated on two sentence types that contain relative clauses but which differ subtly in their structure: (la) The reporter who harshly attacked the senator admitted the error. (lb) The reporter who the senator harshly attacked admitted the error. Both sentences (la) and (lb) contain a relative clause modifying the subject of the sentence, but differ in the role that the main subject noun phrase ('~he reporter") plays in the relative clause; in (la), the main-clause subject is also the subject (and agent) o f th e verb in the relative clause, while in (lb), it is the object (and patient). Accordingly, sentences like (la) are known as subject-subject relative (SS) sentences, while those like (lb) are known as subject-object relative (SO) sentences. As any reader can readily attest to, SO sentences (lb) are generally more difficult to process than SS sentences (la), although even SS sentences are more difficult than sentences without relative clauses. A long history of linguistic argument starting with work by Chomsky and Miller (1963) suggests that SO sentences tax working memory to a greater extent, and that this load becomes especially acute at and just following the relative clause verb of SO sentences; it is here where, in more modem theories, two separate thematic role assi,~nments
D o the waves begin to waver ?
YOUNG
335
ELDERLY
4Jl
Frontal Anterior ~,j~"~~ Temporal
~,,^..,.~
~ ~
.__n~
~7
Central m
I i I ! I i 0 200 400
2"n,d V e r b F~ller
I i I | I i 0 2 0 0 400
- - -
3 3 Mas
I , I I I i 0 200 400
Verb
............
I i I i I 7 r 0 200 400muc
3 0 Mai,r~
Verb
Figure 9. Grand Average ERPs from six anterior electrode sites to SO, SS, and non-relative clause control verbs for Young (n=24) and Elderly (n=18) subjects. The difference between the control verbs and the two relative clause types is shaded dark grey, while the differencebetween SO and SS verbs is shaded light grey.
must be carried out. That is, it is here that readers encounter the first verb of the sentence and must determine which noun phrase is indeed the subject. Note that, with these materials, neither semantic nor pramnatic information can be used to make this choice. King and Just (1991) verified that the greatest reading time differences are found at this point, and that these differences were larger for readers with relatively small working memory capacities. While some effect of carrying two (rather than one) noun phrases in working memory might be expected before the end of the relative clause, such effects are generally not obtained in reading times (e.g. King & Just, 1991; Ford 1983; Holmes and O'Regan, 1981). Perhaps under these circumstances, RT measures are not sensitive enough to maintaining a load in WM, or, alternatively, are sensitive to a number of different counteracting effects which therefore yields a null effect. We thus chose to examine the processing of SS and SO sentences in young and elderly subjects by recording ERPs during their extent. In so doing, we uncovered ERP effects that covary with differences in working memory use during parsing, and that also seem to distinguish youngreaders from elderly readers as well as better comprehenders from poorer comprehenders, presumably in part due to WMC limitations. The sentence location immediately following the end of the relative clause ("admiRed" in (la) and lb)) where the
336
J.W. King and M. Kutas
greatest RT effects between SO and SS sentences are generally found is also a site of large ERP effects (see Figure 9). In both young and elderly subjects, not only do the ERPs to main clause verbs from the SS and SO sentences differ from each other, but, as expected, both of these differ from comparable verbs in filler sentences that do not contain relative clauses at all. This is consistent with the suggestion that even SS sentences tax WM relative to sentences without relative clauses, albeit in different ways than SO sentences tax WM. In the case of SS sentences, a WM load may arise because of the greater temporal separation between the subject noun phrase and the (main) verb relative to sentences with simpler structures, rather than because of any difficulties in determining which NP is the true subject. For both the young and elderly subjects, the difference between SO and SS verbs is larger over anterior relative to posterior sites and larger at left (than right) hemisphere sites. The difference between SS and filler sentence verbs is also left lateralized in the young subjects, but not in the elderly subjects; older subjects exhibit a more bilateral and distinctly more frontal difference. We still need to see whether this particular aging difference is a replicable finding. This difference notwithstanding, the overall pattern of ERP to the verbs from the various sentence types is quite similar in the young and elderly subjects. In both age groups, a greater load on working memory at the verb seems to be associated with a larger frontal, slightly lefi-lateralized negativity. By contrast, much greater age-related differences are revealed by the across-sentence ERP data seen in Figure 10. In the younger subjects, the ERPs to both relative clause sentence types are characterized by a positive frontal drLft that is larger for Good compared with Poor comprehenders; likewise, the difference between the two relative clause types is larger for the better comprehenders. This pattern is consistent with the notions that the good comprehenders integrate the content of both sentence types more easily, and that they find the working memory demands made by the two types (relative to their capacity) to be dissimilar. Poor subjects, on the other hand, seem to experience difficulty with SS sentences so that they must stretch their processing capacity even with these "simpler" loads. In briet, at the sentence level, good and poor comprehenders differ in their treatment of SS sentences; their ERPs to SO sentences are roughly similar. Turning to the ERPs of the elderly subjects, we note that the Good comprehenders do exhibit slightly more frontal positivity (i.e. below the baseline) for both sentences types than the Poor comprehenders. However, neither group of elderly subjects shows as much difference between SO and SS sentence types as was present among even the poorer young comprehenders. Two other features of these data deserve brief mention. First, the ERP data of both the young and elderly subjects show a very clear difference between the SO and SS sentences much earlier in the sentence than is typically observed in RT studies; specifically, this difference occurs at the sentence location where the second noun phrase of the SO sentences must be loaded into working memory. Such memory-loading negativities have been seen in non-linguistic tasks as well (e.g., Ruchkin et al., 1990). Another feature of the data from the elderly is that the end of the SS relative clause is marked by a noticeable negative peak (around 3000 msec or word 7). Closer inspection reveals that there is a similar relative negativity for the younger subjects, albeit smaller. We have also observed this clause-ending negativity (CEN), with its fronto-central and left-lateralized distribution in simple declarative sentences (Kutas and King, in press). Thus, the CEN may be an ERP feature of wider interest given the
337
Do the waves begin to waver?
YOUNG
ELDERLY
V
w'l ;2 w3 w4 w5 w6 w7 w8 vl9wi0 . . . . . . . . . . . . . . . . .
_
"~ "~ ~:-~"
0 ' 10'00' 20'00' 30'00'40'00
msec
30: ~uv relporter tulm tim senator harshl~l orltazked ~ e d
the error...
83: The r ~ o r t l n " who harshly r
tim error...
|Ira senator ~ e 4
Figure 10. Grand average multi-word ERPs from leR frontal sites for Young (n=24) and Elderly (n=18) subjects in response to SO and SS sentences. Good comprehenders in each group (n=12 and n=9 respectively) are shown in the top row and the Poor comprehenders shown in the bottom row. Word labels indicated on the left scale correspond to the onset of words 1 through 10 in the the words in the example sentences given below the waveforms.
known importance that clause endings have both in theoretical models of parsing (e.g., Frazier & Fodor, 1978) and in RT and eye movement data (e.g., Just & Carpenter, 1980). Of greatest relevance here, however, is that these processes, too, are intact in the elderly. 3. CONCLUSIONS Like too many other topics within the field of cognitive aging research, not enough is known about how language processes change as people age, let alone about the electrophysiology related to these processes. What we do know is restricted to circumscribed situations, and concerns mostly reading rather than listening or language production. Fortunately, we can leverage this relatively scant information with the greater body of information we have about ERPs and language processing in young adults to reach some tentative conclusions and generate testable hypotheses for future research. From the single word data we report, it appears that the ERPs prominent over the back of the head such as P l and N1, which presumably reflect primarily early visual processing, are quite similar across the lifespan. Indeed, N1-P2 amplitudes varied with comprehension status in both young and elderly subjects alike. In contrast, both the temporal-parietal N1 and the centro-ffontal P2 component were notably (and reproducibly) different in the older subjects, at least under conditions where words were presented at relatively fast rates (i.e. with stimulus onset asynchronies of either 500 or 550 msec. in these studies). While neither of these
338
J.W. King and M. Kutas
YOUNG
"J"-~
ELDERLY
',~ x.f
./% A
]5~zv
1;o 2;o 3;o 4~o
6 1Ao 2;o 3Ao ~Ao I
msec
Figure 11. ERPs to all open class words atthe left anterior temporal site for Young and Elderly subjects. The top row shows traces for all individual subjects overlapped, while the other rows showtraces for approximately matched pairs of Young and Elderly subjects.
components has been studied systematically in language tasks, the localization of an important P2 generator to the basal temporal lobe area suggests that the marked reduction in its size may be related to known reductions in grey matter in that region of the brain. Subcomponents of the P2, likewise, have been implicated in studies of visual working memory, a process also known to be affected adversely by aging. Later components such as the N400 have been better in both young and elderly subjects and show the typical trend of becoming smaller and later with advancing age. These changes in semantic analyses (contextual integration) are dearly quantitative rather than qualitative in nature. The elderly are slower and more variable in their registration of meaning. Exactly what mechanism is at the core of N400 generation remains unclear, although both attentional and inhibitory processes have been suggested. Data from
Do the waves begin to waver?
339
our recordings of longer epochs suggest that much of the normal, sustained processing during reading is essentially unchanged in the elderly, except when their reduced working memory capacity impacts their efficiency at parsing linguistic input and at integrating the results into their ongoing discourse representations. Our observations on the general consequences of aging on reading notwithstanding, we think it important to emphasize that these effects of aging are neither categorical nor absolute. While we have taken care to exclude subjects whose physical or mental health was in question, what we portray here as the result of '~he aging process" is, naturally, the net sum of many influences that differ from individual to individual. As the waveforms in Figure 11 suggest, individual variability is great even at those frontal sites where many aging-related changes are evident on the average; taken one by one, some young and some old subjects look more alike than one would have predicted from examining the averages alone. The grand mean is never the grand meaning. In the future, we expect to see much more work in the field of geriatric psycholinguistics, not only to understand normal developmental changes in language processing, but also to understand changes caused by diseases such as dementia of the Alzheimer's type, Parkinson's dementia, and strokes that effect both the traditional and nontraditional language areas (Ojemann, 1991). ERP-based research promises to be on the forefront of such research efforts, especially if the ecological validity of ERP paradigms can be increased by technological advances in the presentation of auditory stimuli, and in the use of saccade-related potential research in reading paradimns (e.g. Marton & Szirtes, 1988). The increasing availability of high quality anatomical MRI scans should also be crucial, not only to allow the measurement of age-related changes in the brain (e.g., Jernigan et al., 1991) but also as a way to facilitate the identification of the neural generators of ERPs (e.g., Dale & Sereno, 1993). In the end, however, it will take the efforts of more than just neuroscientists to answer the mysteries of what it means to become older. When that story has been told, we should expect to know more about the brain, but also more about story-telling. REFERENCES Allison, T., Ginter, H., McCarthy, G., Nobre, A. C., Puce, A., Luby, M., & Spencer, D. D. (1994). Face recognition in human extrastriate cortex. Journal of Neurophysiology, 71, 821-825. Anderson, J. K (1983). The architecture of cognition. Cambridge, MA: Harvard University Press. Baddeley, A. D. (1986). Working Memory. Oxford: Clarendon Press. Becket, C. A. (1980). Semantic context effects in visual word recognition: An analysis of semantic strategies. Memory & Cognition, 8, 493-512. Becker, C. A. (1982). The development of semantic context effects: Two processes or two strategies?. Reading Research Quarterly, 17, 482-502. Berger, B. (1992). Dopaminergic innervation of the frontal cerebral cortex: evolutionary trends and functional implications. Advances in Neurology, 57, 525-544. Botwinick, J. (1984). Aging and behavior: a comprehensive integration of research findings (3rd edition). New York: Springer. Brown, A. S. (1991). A review of the tip-of-the-tongue experience. Psychological Bulletin, 109, 204-223.
340
J.w. King and M. Kutas
Callaway, E., Tueting, P., & Koslow, S. (Eds.) (1978). Event-Related Brain Potentials in Man. New York: Academic Press. Chapman, R. M., McCrary, J. W., & Chapman, J. A. (1978). Short-term memory: The storage component of human brain responses predicts recall. Science, 202, 1211-1214. Chapman, R. M., McCrary, J. W., & Chapman, J. A. (1981). Memory processes and evoked potentials. Canadian Journal of Psychology, 35, 201-212. Chase, W. G., & Simon, H. A. (1973). Perception in chess. Cognitive Psychology, 4, 55-81. Cohen, G. (1988). Age differences in memory for texts: Production deficiency or processing limitations?. In L. L. Light & D. M. Burke (Eds.), Language, memory, and aging. New York: Cambridge University Press. Cohen, G. (1990). Recognition and retrieval of proper names: Age differences in the fan effect. European Journal of Cognitive Psychology, 2, 193-204. Dale, A. M., & Sereno, M. I. (1993). Improved localization of cortical activity by combining EEG and MEG with MRI cortical surface reconstruction: a linear approach. Journal of Cognitive Neuroscience, 5, 162-176. de Groot, A. M. B. (1990). The locus of the associative-priming effect in the mental lexicon. In D. A. Balota, G. B. Flores d'Arcais & K. Rayner (Eds.), Comprehension processes in reading (pp. 101-123). Hillsdale, NJ: Lawrence Erlbaum Associates. Donchin, E., & Coles, M. G. H. (1988). Is the P300 component a manifestation of context updating?. Behavioral Brain Sciences, 11, 357-374. Eslinger, P. J., & Grattan, L. M. (1993). Frontal lobe and frontal-striatal substrates for different forms of human cognitive flexibility. Neuropsychologia, 31, 17-28. Ford, M. (1983). A method of obtaining measures of local parsing complexity throughout sentences. Journal of Verbal Learning and Verbal Behavior, 22, 203-218. Frazier, L., & Fodor, J. D. (1978). The sausage machine: A new two-stage parsing model. Cognition, 6, 291-325. Fuster, J. M. (1989). The prefrontal cortex: anatomy, physiology, and neuropsychology of the frontal lobe. New York: Raven Press. Gerard, L., Zacks, R. T., Hasher, L., & Radvansky, G. A. (1991). Age deficits in retrieval: the fan effect. Journals of Gerontology, 46, P131-P136. Gemsbacher, M. A., & Faust, M. F. (1991). The mechanism of suppression: A component of general comprehension skill. Journal of Experimental Psychology: Learning, Memory, & Cognition, 17, 245-262. Goldman-Rakic, P. S. (1987). Circuitry of primate prefrontal cortex and regulation of behavior by representational memory. In F. Plum & V. Mountcastle (Eds.), Handbook of Physiology: The Nervous System. Bethesda, MD: American Physiological Society. Graybiel, A. M., Aosaki, T., Flaherty, A. W., & Kimura, M. (1994). The basal ganglia and adaptive motor control. Science, 265, 1826-1831. Gunter, T. C., Jackson, J. L., & Mulder, G. (1992). An electrophysiological study of semantic processing in young and middle-aged academics. Psychophysiology, 29, 38-54. Hamberger, M., & Friedman, D. (1992). Event-related potential correlates of repetition priming and stimulus classification in young, middle-aged, and older adults. Journals of Gerontology, 47, P395-405. Harbin, T. J., Marsh, G. R., & Harvey, M. T. (1984). Differences in the late components ofthe event-related potential due to age and to semantic and non-semantic tasks. Electroencephalography & Clinical Neurophysiology: Evoked Potentials, 59(6), 489-496.
Do the waves begin to waver?
341
Hasher, L., Stokzfus, E. R., Zacks, 1L T., & Rypma, B. (1991). Age and inhibition. Journal of Experimental Psychology: Learning, Memory, & Cognition, 17, 163-169. Henderson, G., Tomlinson, B. E., & Gibson, P. H. (1980). Cell counts in human cerebral cortex in normal adults throughout life using an image analysing computer. Journal of the Neurological Sciences, 46(1), 113-136. Hillyard, S. A. (1985). Electrophysiology of human selective attention. Trends in Neurosciences, 8, 400-405. HiUyard, S. A., & Picton, T. W. (1987). Eleotrophysiology of cognitive processing. Annual Review of Psychology, 34, 33-61. Holcomb, P. J., Coffcy, S. A., & Neville, H. J. (1992). Visual and auditory sentence processing: A developmental analysis using event-related brain potentials. Developmental Neuropsychology, 8(2-3), 203-241. Holcomb, P. J., & Neville, H. J. (1991). Natural speech processing: An analysis using eventrelated brain potentials. Psychobiology, 19, 286-300. Holmes, V. M., &, Rcgan, J. K~ (1981). Eye fixation patterns during the reading of relativeclause sentences. Journal of Verbal Learning & Verbal Behavior, 20, 417-430. Hunt, E. B., Lunneborg, C., & Lewis, J. (1975). What does it mean to be high verbal?. Cognitive Psychology, 2, 194-227. Jeifreys, D. A., & Tukmachi, E. S. (1992). The vertex-positive scalp potential evoked by faces and by objects. Experimental Brain Research, 91(2), 340-50. Jemigan, T. L., Press, G. A., & Hesselink, J. IL (1990). Methods for measuring brain morphologic features on magnetic resonance images. Validation and normal aging. Archives of Neurology, 47, 27-32. Just, M. A., & Carpenter, P. A. (1980). A theory of reading: From eye fixations to comprehension. Psychological Review, 87, 329-354. Just, M. A., & Carpenter, P. A. (1992). A Capacity theory of comprehension: Individual differences in working memory. Psychological Review, 99, 122-149. Karayanidis, F., Andrews, S., Ward, P. B., & McConaghy, N. (1993). Event-related potentials and repetition priming in young, middle-aged and elderly normal subjects. Cognitive Brain Research, 1, 123-134. Kemper, S. (1986). Imitation of complex syntactic constructions by elderly adults. Applied ERP studies of language in aging. Psycholinguistics, 7, 277-287. Kemper, S. (1987). Constraints on psychological processes in discourse production. In H. W. Dechert & M. Raupach (Eds.), Psycholinguistic models of production (pp. 185-188). Norwood, NJ: Ablex Publishing Corp. Kemper, S. (1987). Syntactic complexity and elderly adults' prose recall. Experimental Aging Research, 13, 47-52. Kemper, S. (1988). Geriatric psycholinguistics: Syntactic limitations of oral and written language. In L. L. Light & D. M. Burke (Eds.), Language, memory, and aging (pp. 5876). New York, NY: Cambridge University Press. Kemper, S., Rash, S., Kynette, D., & Norman, S. (1990). Telling stories: The structure of adults' narratives. European Journal of Cognitive Psychology, 2, 205-228. King, J., & Just, M. A. (1991). Individual differences in syntactic processing: The role of working memory. Journal of Memory and Language, 30, 580-602. King, J. W., & Kutas, M. (in press). Who did what and when? Using word- and clause-related ERPs to monitor working memory usage in reading. Journal of Cognitive Neuroscience.
342
J.W. King and M. Kutas
Kutas, M. (1993). In the company of other words: Electrophysiological evidence for singleword and sentence context effects. Language & Cognitive Processes, 8, 533-572. Kutas, M., & Hillyard, S. A. (1980). Event-related brain potentials to semantically inappropriate and surprisingly large words. Biological Psychology, 11, 99-116. Kutas, M., & Hillyard, S. A. (1980). Reading senseless sentences: brain potentials reflect semantic incongruity. Science, 207, 203-205. Kutas, M. Marta, & Hillyard, S. A. (1982). The lateral distribution of event-related potentials during sentence processing. Neuropsychologia, 20, 579-590. Kutas, M., & Hillyard, S. A. (1983). Event-related brain potentials to grammatical errors and semantic anomalies. Memory & Cognition, 11, 539-550. Kutas, M., & Hillyard, S. A. (1984). Brain potentials during reading reflect word expectancy and semantic association. Nature, 307, 161-163. Kutas, M., Lindamood, T., & Hillyard, S. A. (1984). Word expectancy and event-related brain potentials during sentence processing. In S. Kombhm & J. Requin (Eds.), Preparatory states and processes (pp. 217-238). Hillsdale, NJ: ErlbauI~ Kutas, M., Neville, H. J., & Holcomb, P. J. (1987). A preliminary comparison of the N400 response to semantic anomalies during reading, listening and signing. Electroencephalography and Clinical Neurophysiology, Supplement, 39, 325-330. Kutas, M., & Van Petten, C. (1994). Psycholinguistics Electrified. In M. A. Gemsbacher (Ed.), Handbook ofpsycholinguistics (pp. 83-143). San Diego: Academic Press. Kutas, M., Van Petten, C., & Besson, M. (1988). Event-related potential asymmetries during the reading of sentences. Electroencephalography and Clinical Neurophysiology, 69, 218233. Kynette, D., & Kemper, S. (1986). Aging and the loss of grammatical forms: A cross-sectional study of language performance. Language & Communication, 6, 65-72. Light, L. L., & Capps, J. L. (1986). Comprehension of pronouns in young and older adults. Developmental Psychology, 22, 580-585. Lima, S. D., Hale, S., & Myerson, J. (1991). How general is general slowing? Evidence from the lexical domain. Psychology and Aging, 6, 416-425. Mangun, G. R., & Hillyard, S. A. (1991). Modulation of sensory-evoked brain potentials indicate changes in perceptual processing during spatial priming. Journal of Experimental Psychology: Human Perception and Performance, 17, 1057-1074. Mangtm, G. R., Hillyard, S. A., & Luck, S. J. (1993). Electrocortical substrates of visual selective attention. In D. E. Meyer & S. Kornbhm (Eds.), Attention and performance 14: Synergies in experimental psychology, artificial intelligence, and cognitive neuroscience (pp. 219-243). Cambridge, MA: MIT Press. Marton, M., & Szirtes, J. (1988). Context effects on saccade-related brain potentials to words during reading. Neuropsychologia, 26, 453-463. McCallum, W. C., Farmer, S. F., & Pocock, P. V. (1984). The effects of physical and semantic incongruities of auditory event-related potentials. Electroencephalography & Clinical Neurophysiology: Evoked Potentials, 59, 477-488. McGeer, P. L., McGeer, E. G., & Suzuki, J. S. (1977). Aging and extrapyramidal function. Archives of Neurology, 34(1), 33-35. Miller, E. I~, & Desimone, 1~ (1994). Parallel neuronal mechanisms for short-term memory. Science, 263, 520-522.
Do the waves begin to waver?
343
Miller, G. A., & Chomsky, N. (1963). Finitary models of language users. In D. Luce, 1L Bush, & E. Galanter (Eds.), Handbook of mathematical psychology. New York: John Wiley. Mtmte, T. F., Heinze, H-J., & Mangun, G. It (1993). Dissociation of brain activity related to syntactic and semantic aspects of language. Journal of Cognitive Neuroscience, 5, 335344. Naatanen, 1L, & Michie, P. T. (1979). Early selective-attention effects on the evoked potential: A critical review and reinterpretation. Biological Psychology, 8, 81-136. Neville, H. J., Mills, D. L., & Lawson, D. S. (1992). Fractionating language: different neural subsystems with different sensitive periods. Cerebral Cortex, 2, 244-58. Nunez, P. L. (1981). Electric Fields of the Brain. New York: Oxford University Press. Ojemann, G. A. (1991). Cortical organization of language. Journal of Neuroscience, 11, 22812287. Perfetti, C. A., & Lesgold, A. M. (1977). Discourse comprehension and sources of individual differences. In M. A. Just & P. A. Carpenter (Eds.), Cognitive processes in comprehension. Hillsdale, NJ: Erlbaunl Pfefferbaum, A., Mathalon, D. H., Sullivan, E. V., Rawles, J. M., Zipursky, 1L B., & Lim, K. O. (1994). A quantitative magnetic resonance imaging study of changes in brain morphology from infancy to late adulthood. Archives of Neurology, 51, 874-87. Puce, A., Nobre, A. C., Allison, T., Luby, M. L., McCarthy, K. M., & McCarthy, G. (1994). Event-related potential correlates of visual recognition modules in extrastriate cortex. Cognitive Neuroscience Society Meeting, 1, 93-93. Pujol, J., Junque, C., Vendrell, P., Grau, J. M., & Capdevila, A. (1992). Reduction of the substantia nigra width and motor decline in aging and Parkinson's disease. Archives of Neurology, 49, 1119-1122. Renault, B., Kutas, M., Coles, M. G. H., & Gaillard, A. W. K., (Eds.) (1989). Event-related potential studies of cognition. Am~erdam, Holland: Elsevier. Rosier, F. (1993). Beyond reaction time and error rate: Monitoring mental processes by means of slow event-related brain potentials. In W. C. McCallum & S. H. Curry (Eds.), Slow potential changes m the human brain (NATO ASI series. Series A: Life sciences, Volume 254, pp. 105-119). New York, NY: Plenum Press. Ruchkin, D. S., Johnson, 1L, Canoune, H., & Ritter, W. (1990). Short-term memory storage and retention: an event related brain potential study. Electroencephalography and Clinical Neurophysiology, 76, 419-439. Rugg, M. D., Pearl, S., Walker, P., Roberts, 1L C., & Holdstock, J. S. (1994). Word repetition effects on event-related potentials in healthy young and old subjects, and in patients with Alzheimer-type dementia. Neuropsychologia, 32, 381-398. Salthouse, T. A. (1985). Speed of behavior and its implications for cognition. In J. E. Birren & I~ W. Schaie (Eds.), Handbook of the psychology of aging. New York: Van Nostrand Reinhold Co, Inc.. Sawaguchi, T., & Goldman-Rakic, P. S. (1991). D 1 dopamine receptors in prefrontal cortex: involvement in working memory. Science, 251, 947-950. Tipper, S. P. (1985). The negative priming effect: Inhibitory priming by ignored objects. Quarterly Journal of Experimental Psychology: Human Experimental Psychology, 37A, 571-590. Tipper, S. P. (1991). Less attentional selectivity as a result of declining inhibition in older adults. Bulletin of the Psychonomic Society, 29, 45-47.
344
J.w. King and M. Kutas
Van Petten, C., & Kutas, M. (1991). Influences of semantic and syntactic context on open and closed class words. Memory and Cognition, 19, 95-112. Verleger, K (1988). Event-related potentials and cognition: A critique of the context updating hypothesis and an alternative interpretation of P3. Behavioral & Brain Sciences, 11(3), 343-356. Wechsler, D. (1987). Wechsler Memory ScalewRevised. New York: Psychological Corporation. Woodward, S. H., Ford, J. M., & Hammett, S. C. (1993). N4 to spoken sentences in young and older subjects. Electroencephalography & Clinical Neurophysiology, 87, 306-320. Zelinski, E. M. (1988). Integrating information from discourse: Do older adults show deficits? In L. L. Light & D. M. Burke (Eds.), Language, memory, and aging. New York: Cambridge University Press. Zola-Morgan, S., & Squire, L. 1L (1993). The memory system damaged in medial temporal lobe amnesia: Findings from humans and nonhuman primates. In O. Taketostfi, L. 1L Squire, M. E. Raichle, D. I. Perrett & M. Fukuda (Eds.), Brain mechanisms of perception and memory: From neuron to behavior (pp. 241-257). New York, NY: Oxford University Press.
Age Differences in Word and Language Processing Ph. Allen and Th.R. Bashore (Editors) 9 1995 Elsevier Science B.V. All rights reserved.
345
Memory and Aging: An Event-Related Brain Potential Perspective* David Friedman and Monica Fabiani Cognitive Electrophysiology Laboratory, New York State Psychiatric Institute
In this chapter we will review, from an event-related brain potential (ERP) perspective, the results of studies on the cognitive aging of memory-related phenomena. This review will not cover studies of "short-term memory," i.e., "primary" memory (see Waugh & Norman, 1966), as typically assessed via digit span or the Steinberg memory scanning paradi,~mn. Older adults appear to have similar-capacity primary memory stores compared to young adults (see Poon, 1980 for a review) and, aside from longer reaction times and P3 latencies in the memory scanning procedure, show similar scan times per item (as assessed by P3 latency) to those of their young adult counterparts (see Ford et al., 1979). Comprehensive reviews of this literature have been published elsewhere (Bashore, 1990; Friedman, in press; Ford & Pfefferbaum, 1985; Polich, 1991), and the interested reader is referred to these sources for detailed treatments. Working memory (Baddeley, 1986) will also not be reviewed, due to a paucity of studies dealing with ERPs and working memory (for a recent review of behavioral studies, see Salthouse, 1994). This chapter will concern itself primarily with the explicit/implicit memory distinction that has recently become one of the most intensively researched areas within cognitive neuroscience in general and the cognitive aging field in particular. In keeping with the scope of the current volume, virtually all of the ERP studies of explicit and implict memory-related phenomena have employed words as stimuli. Thus, the conclusions we reach at the end of this chapter apply almost exclusively to verbal memory. We will first introduce the explicit/implicit memory distinction and briefly review the major behaviorally-based findings with respect to cognitive aging. We will then briefly discuss the neuropatholo~cal data, considering the extent to which these findings can explain the age-related dissociation between performance on direct
Address correspondence and requests for reprints to: Dr. David Friedman, Cognitive Electrophysiology Laboratory, Unit 58, New York Psychiatric Institute, 722 West 168th Street, New York City, New York 10032. Phone: (212):960-2476. Fax: (212):781-2661. E-mail:
[email protected] ACKNOWLEDGEMENTS: The authors express their deep appreciation to the collaborators on the various projects described in this review, Drs. Steven Berman, Maria Hamberger, Victoria Kazmerski, Walter Ritter, Gregory Simpson, and Joan G. Snodgrass. In addition, we thank Mr. Charles L. Brown for preliminary data reduction and computer programming, and Ms. Charlotte Trott, Ms. Blanca Rincon, Mr. Sean Hewitt, and Mr. Jeff Cheng for their aid in the collection and analysis of data resulting from several of the studies reported here. We thank Ms. Rachel Yarmolinsky and Ms. Eve Vaag for the construction and photo reproduction of figures. Many thanks to Dr. Ted Bashore for providing us with critical commentary and editorial assistance. Thanks also to Dr. Ray Johnson, Jr. for his criticial commentary. Preparation of this chapter was supported in part by grants AG05213 and AG09988 from the NIA, and by the New York State Department of Mental Hygieffe: The Computer Center at New York State Psychiatric Institute is supported in part by a grant (MH-30906) from the National Institute of Mental Health. David Friedman is supported by Research Scientist Development Award #K02 MH00510 from NIMH.
346
D. Friedman and M. Fabiani
and indirect tests of memory. A second intensive area of investigation within cognitive aging, in which traditional psychometric and experimental neuropsychological, as well as psychophysiological techniques have been brought to bear, is the extent to which memory deficits can be explained by changes in frontal lobe function in older individuals. The performance data will be considered first, followed in a later section by ERP findings from this laboratory. After these introductory sections concerned primarily with behaviorally-based data are presented, we will review memory-related ERP phenomena in young adults, following this with a review of the extant memory-related ERP aging literature. 1. DIRECT (EXPLICIT) AND INDIRECT (IMPLICIT) MEMORY AND AGING. It is, by now, amply clear, that older individuals demonstrate deficits, relative to young adults, on traditional episodic or explicit tests of memory, such as free recall, cued recall, and recognition, tests which require the conscious accessing of previously experienced events (for reviews see Light, 1991, and Moscovitch & Winocur, 1992). The evidence for this "ageassociated memory impairment" (or AAMI) for explicit recollection in otherwise normally aging individuals is so overwhelming that a diagnostic category with this label has been proposed (Crook et al., 1986). On the other hand, a reasonably large and burgeoning literature has documented the relative insensitivity of aging on performance during indirect or implicit tests of memory (for reviews, see Davis & Bemstein, 1992 and Howard & Wiggs, 1993). Memory is demonstrated indirectly when the subject's task does not make reference to previously experienced events, but performance shows the facilitating effect of having experienced those events. For example, subjects might be asked to study a fist of words during an initial study phase. This is followed after a delay by a test phase in which subjects are exposed to a series of three-letter stems, some of which formed parts of the words they saw during the study phase, while some were not previously presented (these are baseline or foil stems). They are asked to complete the stem with the "first word that comes to mind." The benefit of previous experience is demonstrated by the typical finding that more "old" stems are completed correctly than foil stems. This performance benefit or priming on this test of word stem completion appears to be preserved with aging. A similarly constructed direct test, stem cued recall, can also be administered. In this case, subjects study a highly similar list of items and are then given the test of stem cued recall. During the test phase the only difference between word stem completion and stem cued recall is the instructional set. For word stem completion the memorial nature of the test is "disguised" by telling subjects they are participating in a "word puzzle" or that they are "aiding in a normative study of word completions" and, as stated above, are to supply the first word that pops into mind. That is, subjects do not necessarily need to consciously access a representation of the previous event in order to show the benefit of previous experience. By contrast, in the direct test, subjects are told that they are to complete the stem only if they specifically remember having seen it during the study phase. In this case, subjects must consciously access their memory for the previous episode, determine whether or not the stem was seen during the study phase, and then supply the appropriate response. Impetus for the investigation of performance on these two types of memory tests was provided by studies of densely amnestic individuals who were severely impaired on direct tests of memory but nonetheless performed as well as normals on indirect tests. Implicit performance was preserved for the same stimuli and episodes for which the explicit deficit had
Memory and aging
347
been demonstrated (e.g., Warrington & Weisenkrantz, 1968; see Cohen & Eichenbaum, 1993, Squire, 1987, and Squire, 1992 for reviews). Moreover, several impressive and highly replicable performance dissociations between these two types of test were published. For example, although the type of orienting activity (semantic versus non-semantic) at study had a dramatic effect on explicit memory performance, it had a smaller or null effect on implicit memory performance. Coupled with the fact that these amnestic patients had well-localized injuries to specific parts of the brain (e.g., medial temporal lobe; diencephalic structures), several theorists proposed that performance on each type of test is subserved by a neuroanatomically and computationally distinct memory system. This pattern of impaired performance on direct and spared performance on indirect memory testing is one of the most powerful pieces of evidence offered in support of the explicit/implidt multiple memory systems approach to human memory. Explicit memory performance is thought to be mediated primarily via the medial temporal lobe system (with the hippocampus playing a pivotal role--see Cohen & Eichenbaum, 1993 for a review and synthesis), whereas performance on indirect tests is thought to be mediated via neural circuits lying outside the medial temporal lobe memory system Presumably, in the case of perceptually-based implicit tests, these computations would be carried out in sensory cortical areas (e.g., Tulving & Schacter, 1990), while in the case of conceptually-based implicit tests, in neocortical association areas (see, for example, Gabfieli et al., in press; see Roediger & McDermott, 1993 for a review). An alternate approach, labeled transfer-appropriate processing, purports that there is no fundamental difference between implicit and explicit memory, and explains dissociations and associations among types of memory tests on the basis of shared processing mechanisms (Roediger, 1990; Roediger & Blaxton, 1987; for a review see Roediger & McDermott, 1993). Data-driven or perceptual processing results when the characteristics of the stimuli (e.g., morality, typography, letter case, surface form) are critical in performing the task. In contrast, conceptually-driven processing is called on when subjects must elaborate, organize or in some fashion meaningfully process the stimuli in order to perform adequately in the task. On this view, the pattern of association and dissociation between direct and indirect tasks (or, for that matter, within direct or indirect tests) will depend critically on the type of processing that subjects must engage in. However, it is difficult to evaluate the tranfer appropriate processing approach with respect to aging, since the majority of studies have not been directed at testing this hypothesis. Another distinction, item versus associative priming, has been used to account for agerelated differences on indirect memory performance (see, for example, Howard et al., 1991). Although some studies of indirect memory for new associations do show the older adult to be impaired relative to the young adult (i.e., associative implicit memory; e.g., Howard et al., 1991), more recent studies have shown that the older adult is also at a disadvantage when implicit memory for items stored in long-term semantic memory is tested (i.e., implicit item memory; e.g., Davis et al., 1990). These latter data suggest that the item versus associative implicit memory distinction may not account for much of the variance in explaining age-related deficits on indirect tests of memory (see also, Friedman et al., in press). Thus, age differences on direct and indirect tests could either be due to a breakdown with aging in strategic processing or in the putative brain system(s) that subserve each form of memory, or in some combination of the two. In addition, several independent and overlapping hypotheses have been advanced to account for memory deficits in the elderly. For example,
348
D. Friedman and M. Fabiani
Light (1991) considered four hypotheses for explaining memory decline in old age: 1) Failures of metamemory: Older subjects have deficient knowledge about memory, do not employ the most efficient strategies, and are poor at monitoring their own memories. These deficiencies might be a consequence of frontal lobe deficits, which have been hypothesized to account for some aspects of cognitive aging--see Moscovitch and Winocur, 1992 for a review; see also frontal lobe section below); 2) Semantic deficit hypothesis: Older subjects do not spontaneously employ elaborative activity to adequately encode to-be-remembered material, leading to poorer quality memory traces that are deficient in richness, extensiveness and depth of encoding; 3) Impairment in deliberate recollection: Older adults have an intact activation mechanism, but show deficiencies in the processing of the spatio-temporal context in which the information was learned, consistent with older individuals showing preserved indirect, but impaired direct memory performance (possibly due to neuropathology in the structures that mediate direct memory); 4) Reduced processing resources: Older subjects display reduced attentional capacity, reduced working memory capacity, and cognitive slowing (see Salthouse, 1991). However, Light (1991) concluded that no one of these hypotheses effectively provides a good explanatory mechanism for the empirical evidence of reduced memory performance with increasing age. 2. NEUROANATOMICAL BASIS OF AGE-RELATED MEMORY DIFFERENCES. Most hypotheses concerning the sources of the cognitive aging of memory have been framed within the brain systems approach, since older subjects appear at first glance to look like less impaired amnestics, in the sense that they are relatively impaired on direct tests of memory, but perform similarly to young adults on indirect tests of memory. Moreover, the extant neuropathological and imaging data appear to support this distinction, pointing to relatively greater neuropathological involvement of structures intimately involved in explicit memory, such as the hippocampus (e.g., Golomb et al., 1994), and relatively less involvement of sensory and neocortical association areas (e.g., Bouras et al., 1994), which are thought to underlie implicit memory performance. However, recent behavioral data obtained during implicit paradigms (e.g., Davis et al., 1990; Hultsch et al., 1991) have begtm to blur this distinction, as these studies have shown poorer implicit memory performance for older relative to younger adults. One possibility that could account for these results might be the extent to which frontal lobe control functions are disturbed with aging (e.g., Stuss et al., 1994; see below). If frontal lobe deficits (which are reported to increase with age), in addition to medial temporal lobe deficits, impact performance on both implicit and explicit memory tasks (which has also been reported in studies of cognitive aging of memory-related phenomena--see frontal lobe section below), this would tend to decrease the age-related implicit/explicit memory dissociative pattern. To what extent do the age-related neuropathological data so far reported map onto those brain areas that are implicated in performance during direct and indirect memory? Although changes with age in indices of what could be termed brain insult (e.g., cell loss; neurofibrillary tangles) appear to occur diffusely throughout the brain, there are some data suggesting that cell loss is more pronounced in the hippocampus than elsewhere (e.g., Dam, 1979). Moreover, Tomlinson et al. (1968) reported that the presence ofneurofibrillary tangles was more prominent in the hippocampus and parahippocampal gyrus than in the neocortex. Similarly, Scheibel et al. (1976), although based on a very small sample of brains, reported
Memory and aging
349
deterioration of hippocampal dendritic systems with increasing age. In a very recent study, Golomb et al. (1994) found a fairly strong relationship between delayed recall and the size of the hippocampal formation (measured via magnetic resonance imaging) in a sample of normally aging older adults, whereas smaller or null relationships were found between the delayed recall measure and non-hippocampal measurements. All of these factors suggest that some of the difficulty the elderly exhibit during direct memory testing could be due to hippocampal malfunction. With respect to those neocortical areas implicated in performance on indirect testing, the data of Terry et al. (1987) suggest that aging has more of an effect on the frontal and temporal lobes than on the parietal lobes, with age-related changes taking the form of a shrinkage of large neurons, and a concomitant increase in the ratio of small to large neurons. Just what cognitive sequelae neuronal shrinkage would have is largely unknown, but the more anterior distribution of the shrinkage is consistent with preserved indirect memory in the elderly being mediated via posterior cortical areas (see Heindel et al., 1989; 1990). Additional support implicating neocortical association areas as putative sites underlying lexical priming performance has been reported by Heindel et al. (1989). The Heindel et al. data make it clear that different indirect tasks (with different task demands and underlying processes--Roediger & McDermott, 1993) are most likely mediated via different brain areas. Since some age-related changes (e.g., cell loss) are not uniform throughout the brain (Haug et al., 1983), just which implicit task performance is preserved with age will depend upon the brain area(s) implicated in task performance and the amount of"damage" they have suffered. 3. FRONTAL LOBE FUNCTION. As previously alluded to, another major hypothesis used to account for cognitive aging phenomena is that the elderly show deficits in frontal lobe function. This evidence comes from a variety of sources, including traditional neuropsychological test performance (see, for example, Albert & Kaplan, 1987), imaging of cerebral blood flow (e.g., Shaw et al., 1984), experimental neuropsychological investigations (e.g., Craik et al., 1990), neuropathological data (Kemper, 1984; Scheibel & Scheibel, 1975), and ERP investigations (reviewed below; e.g., Friedman et al., 1993c; Fabiani & Friedman, in press). For example, Haaland et al. (1987) reported age-related performance deficits in normally aging samples with the Wisconsin Card Sort Test (WCST), a test typically used to assess the extent of frontal lobe dysfunction. Haaland et al's data indicated that deficits in two response categories, number of categories and number of errors were confined primarily to the oldest group of subjects (aged 80 to 87). Albert et al. (1990) reported age-related decrements in performance on three tests of abstraction ability, a function also linked with the frontal lobes. However, care must be exercised when attempting to infer a deficit in frontal lobe function from a psychometric instrument such as the WCST, as there is evidence that questions the sensitivity and specificity of this putative frontal lobe test. Moreover, a so-called "frontal dysfunction" may occur without evidence of focal frontal disturbance, and could result from diffuse cortical damage (Stuss, 1993). Therefore, it is a good idea to employ more than one test of frontal lobe function. Another memory function, memory for source, or the context within which an item is learned, also appears to depend upon the frontal lobes (Schacter et al., 1984). Source amnesia refers to the inability to remember where information was originally acquired, while the ability
350
D. Friedman and M. Fabiani
to recall the information itself is retained. In an experimental neuropsychological study of source amnesia, Craik et al. (1990) reported that, for their sample of older subjects (aged 60-84), the degree of source amnesia was correlated with age and with deficits on measures of frontal lobe fimction. Several other investigators have also replicated this phenomenon in elderly samples (e.g., McIntyre & Craik, 1987; Spencer & Raz, 1994). Thus, contextual information may not be as readily available to the elderly as to their young adult counterparts, and this may account for some of the reported deficits in explicit remembering. Moreover, in support of a frontal lobe role in memory function, frontal lobe contributions to both memory encoding and retrieval have recently been given weight by studies of explicit memory using the PET technique (Kapur et al., 1994; Shallice et al., 1994; Tulving et al., 1994a; 1994b). Indices of frontal lobe dysfunction may also have an impact on implicit memory performance in the elderly. For example, in the results of Davis et al. (1990), stem completion priming performance was negatively correlated with two types of errors (number of trials required to identify the first Wisconsin category; failure to maintain a consistent pattern of response--failure to maintain set) that increased with increasing age on the WCST. Moreover, Parkin & Walter (1992) reported an increase in the proportion of familiarity- compared to contextually-based judgements with age that was correlated with measures of frontal lobe dysfunction. This experiment was motivated by two-process theories of recognition memory (e.g., Mandler, 1980), which postulate that familiarity or perceptual fluency as well as a contextual episodic component contribute to recognition judgements. The fluency or familiarity component is also thought to play a role during implicit memory tests (cf., Gardiner & Java, 1990). Moreover, two-process theories receive some support from the finding that recognition judgements can be partitioned into "know" (i.e., seen before in the experiment and recoanized on the basis of familiarity) and "remember" (i.e., context-related) responses (cf., Gardiner & Java, 1990). Thus, these data begin to suggest the possibility that age-related changes in performance on both implicit and explicit tests could be explained, at least partially, by changes in frontal lobe function. 4. THE ERP ELICITED D U R I N G M E M O R Y TASKS.
In recent years, the ERP has been increasingly employed as a convergent and complementary source of information in studies of memory. Several reviews of the basic findings have been published (Johnson, in press; Kutas, 1988; Rugg, in press; Rugg & Doyle, 1994), and only a brief overview will be provided here. ERPs are time-locked changes (on a millisecond time base) in the brain's ongoing electrical activity that occur in response to sensory, motor, or cognitive events, and can provide information concerning the hierarchy, sequencing and timing of information processing that is typically much more difficult to obtain from behavioral indices alone. In contrast to inferring intervening stages from reaction time, which is the final common pathway for a number of information processing stages, the components (or voltage swings) of the ERP (each presumed to reflect a different aspect of information processing) can be measured relatively directly (see Hillyard & Picton, 1987 for a review). The ERP also provides information on the spatial distribution of electrical activity recorded at the scalp. That is, the distribution of ERP component amplitudes on the scalp surface can be assessed to determine if different conditions, for example, direct versus indirect memory (Kazmerski & Friedman, submitted; Paller, 1990), produce different distributions (or topographies) across the scalp. If two (or more) conditions give rise to different scalp
Memory and aging
351
distributions, one inference that can be made is that the electrical activity elicited by the various conditions is either generated by a different configuration of intracranial sources and/or an amplitude change in a subset of those neural sources (see Johnson, 1993 for a detailed treatment). This kind of information could figure importantly, for example, in inferring whether explicit and implicit memory can be considered to be subserved by functionally and anatomically unique systems, and whether young and older adults differ in the extent to which their ERP distributions can be considered to provide evidence of frontal lobe sources. ERPs have been recorded during direct (e.g., Bentin et al., 1992; Donchin & Fabiani, 1991; Fabiani & Donchin, in press; Friedman, 1990a; Johnson et al., 1985; Neville et al., 1986; Paller, 1990; Rugg & Nagy, 1989; Smith & Halgren, 1989), as well as indirect (e.g., Bentin & Moscovitch, 1990; Friedman et al., 1993a; Paller & Kutas, 1992; Rugg et al., 1988) tests of memory. Immediately below we review the major findings resulting from many of these studies of memory in young adults, separately for measures of encoding and retrieval. This is followed by a review of age-related changes during similar kinds of memory tasks.
4.1. Encoding The first investigators to compute ERP measures recorded during a study phase as a function of subsequent memory test phase performance were Sanquist and his colleagues (1980). Since that seminal study, several investigators have noted an association between electrical activity recorded at study and performance on subsequent direct (e.g., Friedman, 1990b; Paller, 1990; Paller & Kutas, 1992; Paller et al., 1987a; Karis et al., 1984; Fabiani et al., 1986, 1990) as well as indirect (e.g., Paller et al., 1987b) tests of memory (for a comprehensive review of these studies, see Johnson, in press). In these investigations, the ERPs elicited at study (i.e., to "new" or first presentation items) are averaged according to whether those items were or were not subsequently correctly recalled, recognized, or primed at test. The difference between these two ERPs is then evaluated, either by measuring separately the two waveforms, or by subtracting the subsequently incorrect from the subsequently correct ERPs. The resulting difference is considered to represent the ERP sign of those processes that led to successful memory test performance (presumably related to encoding activity). Figure 1 depicts a typical subsequent memory effect recorded to new or first presentation words during a continuous recognition memory paradigm (Friedman, 1990b). Some authors have interpreted the subsequent memory effect in terms of the modulation of the P3b* component (e.g., Donchin & Coles, 1988). Other authors, however, have pointed out that the distribution of the difference between these two classes of ERP varies from that of the P3b with which it appears to overlap (see Friedman, 1990b; Paller et al., 1987; and Johnson, in press for similar *The label, "P3b," refers to a positive potential elicited typically by task-relevant, infrequently occurring events, with a peak latency between 300 and 1000 ms post-stimulus (depending upon the complexity of the task). The scalp distribution of the P3b is usually, but not always, characterized by a parietally-focused amplitude maximum. P3b's scalp distribution can also be modulated by task requirements (see Johnson, 1993 for a complete discussion). Infrequently occurring, task-irrelevant, unusual or novel events, elict a "novelty P3," with a more frontaUy-oriented scalp distribution than the P3b (see discussion in the Assessments of Frontal Lobe Function in the Elderly section of this chapter). In addition, another P3 component, the "P3a," has been observed in the waveforms elicited by very infrequently occurring deviant events, when those events are "unattended," as well as when they are attended (see N. Squires et al., 1975). The P3a occurs with a peak latency of around 280 ms and is recorded with a Cz-maximum scalp distribution in a midline series of electrode sites.
352
D. Friedman and M. Fabiani
arguments). In fact, as can be seen in the figure, the memory effect (indicated by the shaded areas) has a frontally-oriented scalp distribution, whereas the P3b displays a parietally-maximal scalp distribution. In their initial study, Pallet et al. (1987a) labeled this difference "Din" (for _difference in subsequent memory) and, consistent with an interpretation that this activity reflected at least some aspects of encoding, showed that Dm was larger for semantic compared to orthographic orienting conditions (see also Pallet, 1990; Friedman, 1990b). The data of Friedman (1990b) and Pallet (1990) have been interpreted as suggesting that Dm reflects elaboration, an effortful process conducive to the later retrieval of events.
A
Subsequently R e c o g n i z e d Subsequently Unrecognized
.....
B
r - - - O.--
6"
Subsequently Recognized Subsequently Unrecognized
0
>4
Fz
"','/-/,,-. . . .
-~'-,: ,v,' l,t
"0 ~3 0
-e 0
Q'O'
E
