Psychology and Aging Copyright 2003 by the American Psychological Association, Inc.

2003, Vol. 18, No. 3, 443– 460 0882-7974/03/$12.00 DOI: 10.1037/0882-7974.18.3.443

Aging and Dual-Task Performance: A Meta-Analysis

Paul Verhaeghen, David W. Steitz, Martin J. Sliwinski, and John Cerella
Syracuse University

The relations between dual-task effects and aging were examined through a meta-analysis of 33 studies
(with 48 independent participant groups) using latency as the dependent measure and 30 studies (with 40
independent participant groups) focusing on accuracy. Brinley plots and state traces were derived, and a
model to explicate different types of complexity (additive and multiplicative) was developed. The effects
of dual-task processing on latency were additive, and this additive cost was larger in older adults than in
younger adults and larger than predicted from general slowing. This cost was small and independent of
task complexity. The effects of dual-task processing on logit-transformed accuracy were likewise
additive, but no specific age deficit was associated with this dual-task cost.

With aging, performance on a large number of cognitive tasks
declines (see Craik & Salthouse, 2000, and Park & Schwarz, 2000,
for edited volumes of narrative reviews). Moreover, it has been
found that these performance decrements are intercorrelated (see
Verhaeghen & Salthouse, 1997, for a meta-analysis), suggesting
that decline in a single cognitive primitive may lie at the founda-
tion of those widespread age-related declines in cognitive perfor-
mance. This cognitive primitive has mostly been identified with
the concept of mental resources (see, e.g., Salthouse, 1991, 1996).
Resource explanations claim that age-related differences in (a)
mental speed, (b) mental workspace, and/or (c) mental energy are
responsible for a large part of the observed age-related changes in
more complex aspects of cognition.
To date, the role of speed has been the main focus of research
in this field (see Salthouse, 1991, 1996, for an overview). One of
the key discoveries in cognitive aging in the last quarter of the
previous century is that, within relatively broad task domains (e.g.,
lexical processing or visuospatial processing), the effects of age on
reaction-time performance can be well described by a single linear
function (see, e.g., Cerella, Poon, & Williams, 1980; Hale &
Myerson, 1996). It has been demonstrated that such a pattern of
within-domain general linear slowing can be explained using a
single underlying mechanism (Bamber, 1979; Cerella, 1994; Dunn
& Kirsner, 1988). The slope of the function relating performance
of the old to performance of the young can be considered an
estimate of the single age-related slowing factor that operates on
central processes (i.e., all processes intervening between the sen-
sory and motor stages of processing; see Cerella, 1990, for a
mathematical derivation of this claim; see Ratcliff, Spieler, &
Paul Verhaeghen, David W. Steitz, Martin J. Sliwinski, and John
Cerella, Department of Psychology, Syracuse University.
This research was supported in part by Grant AG-16201 from the
National Institute on Aging. A portion of this research was presented at the
International Congress of Psychology, Stockholm, Sweden, July 2000.
John Cerella is the author of the appendixes.
Correspondence concerning this article should be addressed to Paul
Verhaeghen, Department of Psychology, 430 Huntington Hall, Syracuse
University, Syracuse, New York 13244-2340. E-mail: pverhaeg@
psych.syr.edu
McKoon, 2000, for an alternative interpretation; and Myerson,
Adams, Hale, & Jenkins, 2003, for a defense of the original
interpretation).
The pervasiveness of the phenomenon of linear slowing effects
has led many researchers to propose it as a new null hypothesis for
cognitive aging (e.g., Cerella, 1994; Faust, Balota, Spieler, &
Ferraro, 1999; Madden, Pierce, & Allen, 1992; Perfect & Maylor,
2000). That is, it has been proposed that generalized slowing
should be considered the baseline phenomenon of cognitive aging:
The starting point of any analysis should be that a particular set of
conditions is expected to yield a linear relationship between laten-
cies of younger and older adults (with a slope larger than one and
typically a small negative intercept). This expectation has impli-
cations for the way condition effects should be measured. Typi-
cally, condition effects in latency data are expressed in additive
terms, that is, the time needed in a baseline condition is subtracted
from the time needed in a critical condition, and this index is then
used in subsequent analyses. However, given that the empirical
relationship between latencies of younger and older adults is linear
(and near multiplicative), this traditional way of looking at data
gives rise to incorrect interpretations for the simple reason that
more complex conditions take longer to process. Because complex
conditions take longer, the absolute age difference in more com-
plex conditions will, by virtue of this near-multiplicative relation-
ship, be larger than the age difference in less complex conditions
(Cerella et al., 1980). Conversely, going from simple to more
complex conditions, the condition effect will be larger in older
adults than in younger adults. Here is a numerical example. Sup-
pose that there is a baseline task that takes younger adults 500 ms
to complete and that the peripheral component is 200 ms. Assume
an age-related slowing factor of 1.5 (healthy 60 to 70-year-old
adults are typically 1.5 times, or 50%, slower than college-age
adults). One can then expect a reaction time of 200 ms ⫹ (1.5 ⫻
300 ms) ⫽ 650 ms for older adults. (Typically, peripheral process-
ing is also slowed in older adults, but by a very small factor.) Now
introduce a more complex version of the task, which increases
reaction time for the young by 500 ms, leading to a total time
of 1,000 ms. The reaction time expected for older adults is now
200 ms ⫹ (1.5 ⫻ 800 ms) ⫽ 1,400 ms. For the case at hand, an age
difference of 150 ms is recorded in the baseline condition and one
of 400 ms in the more complex condition, despite the fact that no

443
444 VERHAEGHEN, STEITZ, SLIWINSKI, AND CERELLA
additional deficit is associated with the specific processes under-
lying the increased complexity.
Research into the second aspect of mental resources, namely,
workspace, has typically focused on working memory capacity. It
has been shown that working memory capacity declines with
advancing age (for a meta-analysis, see Bopp & Verhaeghen,
2002; Verhaeghen, Marcoen, & Goossens, 1993) and that working
memory capacity mediates between age and more complex aspects
of cognition, such as episodic memory, reasoning, and spatial
ability (for a meta-analysis, see Verhaeghen & Salthouse, 1997).
Some researchers have claimed that certain aspects of working
memory functioning are partially independent of processing speed
and may yield specific age-related effects. For instance, it has been
shown that tasks requiring online processing concurrent with stor-
age and retrieval from working memory yield age differences in
speed and accuracy that are larger than those predicted from tasks
without such demands (see, e.g., Kliegl, Mayr, & Oberauer, 2000;
Mayr & Kliegl, 1993; Verhaeghen, Kliegl, & Mayr, 1997).
The third aspect of resources, mental energy, has typically been
operationalized as attentional ability. There is a rather large body
of literature on the relationship between aging and at least two
aspects of attention, namely, selective and divided attention (for
recent narrative overviews, see McDowd & Shaw, 2000, and
Rogers, 2000). Verhaeghen and De Meersman (1998a, 1998b)
conducted meta-analyses on age differences in two tasks that
measure selective attention, namely the Stroop color–word task
and negative priming. It was found that once age differences in the
baseline condition were taken into account, the observed age
deficits in Stroop interference and location negative priming dis-
appeared and that there was only a slight deficit in identity nega-
tive priming (and this might have been associated with a memory
deficit rather than with age differences in attention per se). This
makes it unlikely that age-related difficulties with selective atten-
tion are the source of latency differences in more complex cogni-
tive tasks (see, e.g., Hasher & Zacks, 1988). Rather, mental slow-
ing (whatever its cause) may be the sole source of the observed age
differences in selective attention paradigms.
What about aging and divided attention? The paradigm most
often used to measure divided attention is dual-task performance,
that is, simultaneous processing of two (and sometimes more)
sources of information. There is little doubt that mental resources
are needed to divide attention among two or more tasks (see
Pashler, 1998, for an overview of theories pertaining to dual-task
performance and its demands on the cognitive system). There is
also little doubt that there is an age-related decline in dual-task
performance in the sense that older adults are slower under dual-
task conditions than younger adults are (for reviews, see, e.g.,
Hartley & Little, 1999; McDowd & Shaw, 2000). However, under
the null hypothesis of general slowing, mental slowing is to be
expected in any kind of cognitive task. The real question then
becomes, Are older adults slower under dual-task conditions than
can be expected from the age difference observed under single-
task conditions? If the answer were affirmative, this would be
evidence for a deficit in divided attention over and beyond the
effect of normal, general age-related slowing. If the answer were
negative, then the age-related deficit in divided attention may be
merely another instance of the general slowing that pervades the
cognitive system with advancing age.
Given that dual-task performance may require the division of
mental resources among tasks, it is also reasonable to assume that
under certain circumstances, accuracy would decrease when a
secondary task gets compounded with a given single task. The
same question that we have asked about age-related slowing in
dual-task performance can be asked about accuracy and aging
under dual-task conditions: Are there age-related differences in
accuracy under the dual-task paradigm beyond those observed
under single-task conditions? (Note that the age difference need
not be in the direction of lower accuracy. It can also take the
direction of higher accuracy under dual-task conditions than under
single-task conditions, if older adults sacrifice speed to preserve
accuracy.)
In the present article, we report results from a meta-analysis on
aging and its effects on dual-task performance as measured by
reaction time and accuracy. Two types of graphical analysis were
applied to the data using hierarchical linear modeling (Sliwinski &
Hall, 1998). First, a state-trace analysis was conducted, that is, we
regressed performance in dual-task conditions on performance in
single-task performance separately for younger and older adults.
We then tested whether a single line sufficed to explain the data or
whether two different lines (one for younger adults, one for older
adults) were needed to adequately describe the relation between
dual-task performance and single-task performance. If a single line
were to fit the data, it would imply that there was no age difference
in the relation between single- and dual-task performance. If two
lines were to appear (one for young, one for older adults), the
interpretation would depend on the form the dual-task cost took
(see below). It has been shown that state-trace analysis is less
conservative than the more traditional Brinley analysis when sam-
pling variance is large in comparison to the experimental effect
(Verhaeghen & De Meersman, 1998a). Second, we conducted a
Brinley analysis (see, e.g., Brinley, 1965; Cerella et al., 1980), that
is, performance of older adults was analyzed as a function of
performance of younger adults. In this analysis, we examined
whether a single line sufficed to explain the data or whether two
different lines (one for single-task performance, one for dual-task
performance) were needed to adequately describe the relation
between the performance of younger and older adults. If two lines
were needed, this would be direct evidence for age sensitivity of
the processes involved in dual-task performance.
Theoretically, dual-task costs can take many forms. For the
latency data, we examined the two simplest cases: additive and
multiplicative costs. First, dual tasking may serve to add an addi-
tional step in the chain of processing operations needed for the
single task (e.g., a task-switching or task-preparedness stage; All-
port, Styles, & Hsieh, 1994; Monsell, 1996). In that case, the
dual-task cost would be additive, that is, it would add a fixed
number of milliseconds to the reaction time. (Note that this is the
model that is implicitly assumed in most of the literature, where
dual-task costs are calculated as the difference between latency in
the dual-task condition and latency in the single-task condition.) A
second possibility is that the effects of dual tasking may be
distributed across the central components of the single task, slow-
ing each of these components by a fixed ratio. This implies that as
the complexity of the single task (as indicated by its latency)
increases, so would the dual-task cost. In that case, dual task costs
would be multiplicative on the latency of the single task. If one
assumes that the age deficit in central processes takes the usual
 AGING AND DUAL-TASK PERFORMANCE
multiplicative form within families of processes (see, e.g., Cerella,
1990; Hale & Myerson, 1996), then the choice of model for
dual-task costs will lead to distinct predictions concerning the type
of linear functions found in both Brinley and state-trace analyses.
In Appendix A, we formalize these models and their consequences
for the dual- versus single-task traces in Brinley space and for the
young versus old trace in state-trace space. As can be seen in
Appendix A, it is only the convergence of evidence from both
types of analysis that can answer the question of what form the
dual-task cost takes and whether or not there is an age difference
in this dual-task cost.
The focus of our analysis is on the description of age differ-
ences, not on general effects of different manipulations in task or
stimulus material on dual-task processing. Some of these effects,
however, may interact with age differences and are therefore
explored here. We included five potential moderating variables in
our analysis. These emerged partially from theoretical analysis and
partially out of convenience, that is, they represent variables that
varied in a systematic way within and across the studies in our
database. The first moderating variable is whether or not the
processing requirements of the primary task are predominantly
sensorimotor (e.g., simple reaction time) or cognitive (e.g., visual
search) in nature. Sensorimotor tasks presumably place less heavy
demands on attentional resources than cognitive tasks. Addition-
ally, age differences in sensorimotor tasks are smaller than those in
cognitive tasks (Cerella & Hale, 1994). For these two reasons, age
differences in dual-task costs might be smaller in sensorimotor
tasks than in central tasks. The second moderating variable is
whether the primary task is verbal (e.g., working memory span) or
visuospatial (e.g., visual search) in nature. Age differences in
latency for verbal tasks are typically smaller than those in visuo-
spatial tasks (see, e.g., Hale & Myerson, 1996), and this distinction
might moderate the age difference in dual-task costs as well. The
third moderating variable is whether or not the input modality for
the primary task (visual or auditory) matches the input modality
for the secondary task. Within-modality dual tasking generally
yields larger costs than between-modalities dual tasking (Pashler,
1998, pp. 160 –161), and it may be possible that this difference in
difficulty also leads to an age difference in dual-task costs. The
fourth moderating variable is input modality of the primary task
(visual or auditory); the fifth is output modality of the primary task
(manual or vocal).
Method
Sample of Studies
Studies were collected by consulting the PsycINFO electronic database,
through personal contacts, and by checking references found in the articles
thus retrieved. The search was concluded in July 2001. Inclusion criteria
were: (a) The study contained experiments in which a comparison was
made between age groups, namely, younger adults (with a mean age of 30
years or younger) versus older adults (with a mean age of 60 years and
older); and (b) the study compared either latencies or accuracy or both
under dual-task conditions with the corresponding measure in single-task
performance. Table 1 presents a listing of all independent groups for the
latency analyses; Table 2 provides a listing of all independent groups for
the accuracy analyses. No study was excluded for any reason other than
those stated above.
445
State-Trace Analysis
Latency analysis. For the state-trace analysis, the mean latency data of
dual-task conditions were regressed on the mean latency data of single-task
conditions. Mean latency data were expressed in milliseconds. The as-
sumption is that a linear model adequately captures the relation between
single- and dual-task latencies; the rationale for this expectation is derived
in Appendix A.
Traditionally (see, e.g., Cerella et al., 1980; Hale & Myerson, 1996;
Verhaeghen & De Meersman, 1998a, 1998b), meta-analyses in the field of
cognitive aging have been conducted using pooled or aggregate regression
procedures. Pooled regression uses the reaction-time data from any of the
single- and dual-task experimental conditions from each study and ignores
the nesting of conditions within studies (see, e.g., Cerella, 1985, 1991).
Sliwinski and Hall (1998) described serious statistical problems that arise
using pooled regression, including ambiguous regression coefficients and
biased significance testing. In the present study, we used aggregate regres-
sion (weighting for sample size) for our first pass over the data, that is, for
identifying outlying studies. The aggregate regression approach uses only
between-studies variability and does not provide information about the
state-trace function for any single study. To take full advantage of the
nested structure of conditions within studies, we used multilevel modeling
procedures to obtain state-trace functions for each study, to obtain average
parameters (and their variances) across studies, and to compare the param-
eters from these functions in the old and young samples. This analytic
approach required the specification of a condition-level model, which repre-
sents the mean reaction time for each dual-task condition as a function of the
reaction time from a corresponding single-task condition within each study:
Dit ⫽ ␤0t ⫹ ␤1tSit ⫹ Rit, (1)
where Dit is the average dual-task reaction time from Condition i in Study
t, Sit is the average single-task reaction time from Condition i in Study t, ␤0t
is the intercept, ␤1t is the slope relating single- to dual-task reaction times
for Study t, and Rit is the residual for Condition i in Study t. The study-level
model then represents each regression parameter as a function of the
overall mean and each study’s unique effect as follows:
␤ 0t ⫽ ␤៮ 0 ⫹ U0t,
␤ 1t ⫽ ␤៮ l ⫹ U1t, (2)
where ␤៮ 0 is the average intercept across all studies, ␤៮ l is the average slope
across all studies, and U0t and Ult are the increments to intercept and slope
associated with Study t. The fixed effects, ␤៮ 0 and ␤៮ 1, provide precision
weighted estimates of the average within-study intercept and slope,
whereas the random effects, U0t and Ult, provide estimates of the within-
study regression parameter variance.
Age effects were examined in the condition level model by introducing a
dummy variable that codes for age group (age ⫽ 0 if young, age ⫽ 1 if old):
Dit ⫽ ␤0t ⫹ ␣0t(age) ⫹ ␤1tSit ⫹ ␣1t(age ⫻ Sit) ⫹ Rit. (3)
The parameter ␣0t conveys the effect of age group on the intercept of the
state-trace, and the parameter ␤lt conveys the effect of age group on the
slope. If the average ␣៮ 0 is larger than 0.0 by a statistically significant
amount, coupled with an absence of age difference in ␣៮ 1, then the differ-
ence between the state-traces of younger and older adults is situated in the
intercept only, which would be evidence of a purely additive age deficit on
the dual-task effect. If ␣៮ 1 is larger than 0.0 by a statistically significant
amount, then the slopes of the state-traces of younger and older adults are
significantly different.
Accuracy analysis. The state-trace analysis on accuracy data (i.e.,
percentage correct) was conducted analogously to the state-trace analysis
on latency data. However, no good rationale exists for expecting a linear
relation between accuracies of single and dual tasks. Additional difficulties
446 VERHAEGHEN, STEITZ, SLIWINSKI, AND CERELLA

Table 1
Latency Analysis: Sample of Studies

RT (ms), RT (ms),
Task types n single dual

Study Single Dual Young Old Young Old Young Old

Anderson (1999, Expt. 1) FR FR ⫹ VS 24 24 398 543 486 781

Anderson (1999, Expt. 2) CR CR ⫹ VS 24 24 392 494 433 625
Anderson (1999, Expt. 4) VRC VRC ⫹ VS 24 24 361 498 382 554
Baddeley, Logie, Bressi, Della Sala, & Spinnler (1986) AP AP ⫹ MT 20 28 381 418 435 479
Baron & Mattila (1989)a TM, VM TM ⫹ VM 12 12 576 744 631 857
Broadbent & Heron (1962)a STM, LTM LRC ⫹ STM, LTM 21 32 18,400 23,350 24,000 26,800
Chen et al. (1996) VRT VRT ⫹ WKG 16 16 359 376 366 391
Craik & McDowd (1987) VC VC ⫹ CRR 15 15 689 1,058 1,040 1,857
Gick, Craik, & Morris (1988) SV, WS SV ⫹ WS 18 18 4,963 5,543 5,627 6,438
Greenwood & Parasuraman (1991) LM LM ⫹ AP 10 20 544 627 548 609
Guttentag & Madden (1987) LM LM ⫹ AP 14 14 621 832 609 848
Hartley & Little (1999, Expt. 1) CI, LI CI ⫹ LI 60 60 440 496 600 709
Hartley & Little (1999, Expt. 3) CI, LI CI ⫹ LI 20 20 447 485 538 597
Hartley & Little (1999, Expt. 4) CI, LI CI ⫹ LI 20 20 421 481 697 783
Hartley & Little (1999, Expt. 5) CI, NI CI ⫹ NI 20 20 545 574 500 597
Hartley & Little (1999, Expt. 6) LI, CI LI ⫹ CI 20 20 505 549 574 652
Hartley & Little (1999, Expt. 7) LI, CI LI ⫹ CI 20 20 463 509 714 823
Hawkins, Kramer, & Capaldi (1992, Expt. 1) VD, AD VD ⫹ AD 14 14 419 581 637 951
Kramer, Larish, & Strayer (1995) VM, AA VM ⫹ AA 29 30 2,550 4,245 4,046 8,349
Li (1999, Expt. 1) DS, WS, NV NV ⫹ W, N 24 24 5,400 5,740 7,330 8,430
Li (1999, Expt. 2) DS, WS, MA MA ⫹ W, N 24 24 5,700 7,000 7,972 9,454
Light, Kennison, Prull, La Voie, & Zuellig (1996,
Expt. 2) NN NN ⫹ NA 48 48 684 747 689 758
Light, Kennison, Prull, La Voie, & Zuellig (1996,
Expt. 3) NN NN ⫹ NA 66 66 824 889 810 902
Light & Prull (1995, Expt. 1) WN WN ⫹ NA 24 24 590 673 601 687
Light & Prull (1995, Expt. 2) WN WN ⫹ NA 48 48 577 691 580 697
Light, Prull, & Kennison (2000, Expt. 2) CV, MA CV ⫹ MA 36 36 945 1,186 945 1,198
Light, Prull, & Kennison (2000, Expt. 3) CV, MA CV ⫹ MA 24 24 772 793 750 778
Lorsbach & Simpson (1988) LM LM ⫹ AP 18 18 434 589 454 648
Madden (1986, Expt. 1) VS VS ⫹ AP 20 20 649 896 653 935
Madden (1986, Expt. 3) VS VS ⫹ AP 20 20 678 1,074 689 1,078
Madden (1986, Expt. 4) VS VS ⫹ AP 16 16 636 1,020 657 1,021
Madden (1987) VS VS ⫹ AP 24 24 1,076 1,460 1,084 1,512
McDowd (1986) AP AP ⫹ MT 6 6 455 765 645 1,195
McDowd & Craik (1988, Expt. 1) VD VD ⫹ A 16 16 537 712 691 1,118
McDowd & Craik (1988, Expt. 2) VD VD ⫹ AS 36 36 522 690 680 1,044
Morris, Gick, & Craik (1988) SV SV ⫹ WN 24 24 2,825 3,625 3,675 4,388
Rogers, Bertus, & Gilbert (1994) MS, VS MS ⫹ VS 20 20 541 764 595 949
Salthouse, Fristoe, Lineweaver, & Coon (1995, Expt. 1) LRC, DS LRC ⫹ DS 40 40 1,125 1,875 1,375 2,042
Salthouse, Fristoe, Lineweaver, & Coon (1995, Expt. 2) DI, DSD, NA, LRC DI, DSD, NA, LRC 50 50 2,032 2,687 2,243 3,400
Salthouse & Somberg (1982) VS, MS, VD, TP VRT ⫹ VS, MS, VD, TP 8 8 358 332 468 543
Somberg & Salthouse (1982, Expt. 2) DF AP ⫹ DF 16 16 231 260 401 509
Tecce, Cattanach, Yrchik, Meinbresse, & Dessonville
(1982) AP AP ⫹ LRC 35 23 263 462 308 736
Tsang & Shaner (1998, nonpilot) SO, STM MT ⫹ SO, STM 15 15 1,107 2,068 1,107 2,185
Tsang & Shaner (1998, pilot) SO, STM MT ⫹ SO, STM 15 15 1,072 1,757 1,112 1,870
Tun & Wingfield (1994) VR VR ⫹ LM 18 18 527 724 627 856
Tun, Wingfield, & Stine (1991) VR VR ⫹ LR 18 18 430 464 523 573
Tun, Wingfield, Stine, & Mecsas (1992) PR PR ⫹ VR 25 25 860 996 1,068 1,475
Wright (1981, Expt. 1) DF, VRS DF ⫹ VRS 12 12 4,538 5,823 4,707 6,122

Note. RT ⫽ reaction time; Expt. ⫽ experiment. Task types: A ⫽ auditory recognition and categorization; AA ⫽ alphabet arithmetic; AD ⫽ auditory
discrimination; AP ⫽ audio probe; AS ⫽ auditory search; CI ⫽ color identification; CR ⫽ cued recall; CRR ⫽ cued recall and recognition; CV ⫽ category
verification; DF ⫽ digit-span forward; DI ⫽ digit-identity decision; DS ⫽ digit-symbol substitution; DSD ⫽ digit-symbol substitution decision; FR ⫽ free
recall; LI ⫽ letter identification; LM ⫽ letter matching; LR ⫽ letter recognition; LRC ⫽ letter recall; LTM ⫽ long-term memory; MA ⫽ mental arithmetic;
MA ⫹ W, N ⫽ mental arithmetic, extra word, extra number, and subproduct and embedded-number recall; MS ⫽ memory search; MT ⫽ manual tracking;
NA ⫽ number addition; NI ⫽ number identification; NN ⫽ nonword naming; NV ⫽ number verification; NV ⫹ W, N ⫽ number verification and extra
word, extra number or end number; PR ⫽ picture recall; SO ⫽ spatial orientation; STM ⫽ short-term memory; SV ⫽ sentence verification; TP ⫽ temporal
prediction; TM ⫽ tone matching; V ⫽ visual search and categorization; VC ⫽ visual categorization; VD ⫽ visual discrimination; VM ⫽ visual matching;
VR ⫽ verbal recall; VRC ⫽ verbal recognition; VRS ⫽ verbal reasoning; VRT ⫽ vocal reaction time; VS ⫽ visual search; WKG ⫽ working memory;
WN ⫽ word naming; WS ⫽ word span.
a
Male participants only.
 AGING AND DUAL-TASK PERFORMANCE 447

Table 2
Accuracy Analysis: Sample of Studies

Accuracy Accuracy
Task types n (%), single (%), dual

Study Single Dual Young Old Young Old Young Old

Anderson et al. (2000) CR CR ⫹ AD 12 12 79 60 67 44

Baddeley, Logie, Bressi, Della Sala, & Spinnler (1986) AP AP ⫹ MT 20 28 55 57 55 53
Chen et al. (1996) VRT VRT ⫹ WKG 16 16 93 85 83 66
Gick, Craik, & Morris (1988) SV, WS SV ⫹ WS 18 18 88 81 93 86
Guttentag & Madden (1987) LM LM ⫹ AP 14 14 99 98 98 96
Hartley & Little (1999, Expt. 2) CI CI ⫹ LI 20 20 96 91 91 80
Hawkins, Kramer, & Capaldi (1992, Expt. 1) VD, AD VD ⫹ AD 14 14 95 95 94 94
Isingrini, Vazou, & Leroy (1995) CR CR ⫹ AD 40 80 70 38 39 25
Kramer, Larish, & Strayer (1995) VM, AA VM ⫹ AA 29 30 95 90 94 85
Li (1999, Expt. 1) DS, WS, NV NV ⫹ W, N 24 24 91 92 94 91
Li (1999, Expt. 2) DS, WS, MA MA ⫹ W, N 24 24 94 89 94 89
Li, Lindenberger, Freund, & Baltes (2001) STM, WKG STM ⫹ WKG 37 40 84 70 74 48
Li, Lindenberger, Freund, & Baltes (2001) walking
accuracy WKG WKG ⫹ STM 37 40 80 65 86 65
Light & Prull (1995, Expt. 1) WN WN ⫹ NA 24 24 17 10 7 4
Light, Prull, & Kennison (2000, Expt. 1) EG, CR EG ⫹ MA, CR ⫹ MA 24 24 43 34 26 21
Light, Prull, & Kennison (2000, Expt. 2) CV, R CV, R ⫹ MA 36 36 93 86 76 56
Lindenberger, Marsiske, & Baltes (2000) recall accuracy STM, WKG STM ⫹ WKG 47 48 82 60 74 41
Lindenberger, Marsiske, & Baltes (2000) walking
accuracy STM, WKG STM ⫹ WKG 47 48 72 56 74 49
Lorsbach & Simpson (1988) LM LM ⫹ AP 18 18 91 95 91 94
Macht & Buschke (1983) FR FR ⫹ VRT 48 48 78 46 79 49
Madden (1986, Expt. 1) VS VS ⫹ AP 20 20 99 98 98 96
Madden (1986, Expt. 3) VS VS ⫹ AP 20 20 98 98 98 97
Madden (1986, Expt. 4) VS VS ⫹ AP 16 16 98 97 98 95
Madden (1987) VS VS ⫹ AP 24 24 98 97 97 96
McDowd & Craik (1988, Expt. 2) V, VS V, VS ⫹ AS 36 36 97 97 97 96
Morris, Craik, & Gick (1990, Expt. 1) FR FR ⫹ SV 16 16 96 87 76 54
Morris, Craik, & Gick (1990, Expt. 2) FR FR ⫹ SV 16 16 83 69 63 42
Morris, Gick, & Craik (1988) SV SV ⫹ WN 24 24 93 94 88 89
Park, Puglisi, & Smith (1986, Expt. 2 and 3) DI, VRL DI ⫹ VRL 38 31 84 90 61 69
Park, Puglisi, Smith, & Dudley (1987) DI, VRL DI ⫹ VRL 27 24 86 86 63 62
Park, Smith, Dudley, & Lafronza (1989, Expt. 1) DI, VR DI ⫹ VR 64 62 67 51 53 46
Perry & Wingfield (1994, Expt. 1) CR, FR LD ⫹ CR, FR 24 24 57 44 47 37
Perry & Wingfield (1994, Expt. 2) FR LD ⫹ FR 12 12 42 21 32 23
Ponds, Brouwer, & van Wolffelaar (1988) DC DC ⫹ VT 17 41 36 38 29 33
Salthouse, Fristoe, Lineweaver, & Coon (1995, Expt. 2) DI, DSD, NA, LRC DI, DSD, NA, LRC 50 50 90 80 89 82
Tsang & Shaner (1998, nonpilot) SO, STM MT ⫹ SO, STM 15 15 98 93 97 93
Tsang & Shaner (1998, pilot) SO, STM MT ⫹ SO, STM 15 15 98 96 98 96
Tun & Wingfield (1994) VR VR ⫹ LM 18 18 54 41 51 32
Tun, Wingfield, & Stine (1991) VR VR ⫹ LR 18 18 45 41 45 43
Wright (1981, Expt. 1) DF, VRS DF ⫹ VRS 12 12 99 96 86 63

Note. Expt. ⫽ experiment. Task types: AA ⫽ alphabet arithmetic; AD ⫽ auditory discrimination; AP ⫽ audio probe; AS ⫽ auditory search; CI ⫽ color
identification; CR ⫽ cued recall; CV ⫽ category verification; DC ⫽ dot counting; DF ⫽ digit-span forward; DI ⫽ digit-identity decision; DS ⫽
digit-symbol substitution; DSD ⫽ digit-symbol substitution decision; EG ⫽ exemplar generation; FR ⫽ free recall; LD ⫽ lexical decision; LI ⫽ letter
identification; LM ⫽ letter matching; LR ⫽ letter recognition; LRC ⫽ letter recall; MA ⫽ mental arithmetic; MA ⫹ W, N ⫽ mental arithmetic, extra word,
extra number, and subproduct and embedded-number recall; MT ⫽ manual tracking; NA ⫽ number addition; NV ⫽ number verification; NV ⫹ W, N ⫽
number verification and extra word, extra number, or end number; R ⫽ recognition; SO ⫽ spatial orientation; STM ⫽ short-term memory; SV ⫽ sentence
verification; V ⫽ visual search and categorization; VD ⫽ visual discrimination; VM ⫽ visual matching; VR ⫽ verbal recall; VRL ⫽ visual recall; VRS ⫽
verbal reasoning; VRT ⫽ visual reaction time; VS ⫽ visual search; VT ⫽ visual tracking; WKG ⫽ working memory; WN ⫽ word naming; WS ⫽ word
span.

were caused by accuracy scores close to 100%, which led to problems with
range restriction and to scaling issues. Therefore, a logit (or log-odds)
transformation was applied to the data (logit ⫽ ln[P/(1 ⫺ P]; with P ⫽
proportion correct). The logit transformation represents the most common
and general function for modeling proportions in linear models (McCul-
lagh & Nelder, 1989). Preliminary model fitting and inspection of the
residuals from regression analysis of the raw accuracy data indicated
evidence of heteroscedasticity, which was eliminated by applying the logit
transformation. As for the latency data, the main analysis tested whether
different equations were needed for younger and older adults after the logit
transformation had been applied.
Brinley Analysis
Latency analysis. For the Brinley analysis on latency, mean perfor-
mance of older adults was regressed on mean performance of younger
adults. Mean latency data were expressed in milliseconds. We applied the
multilayered slowing model advanced by Cerella (1990) to the data. This
448 VERHAEGHEN, STEITZ, SLIWINSKI, AND CERELLA

model assumes that aging brings about differential slowing in peripheral

processes (i.e., input and output processes) and central processes. Cerella
demonstrated that such a model yields young– old data that can be de-
scribed by a linear function:

O t ⫽ ␤ 0 ⫹ ␤ lY t, (4)

where Ot is average latency of older adults in Study t, and Yt is average

latency of younger adults in Study t. The ␤l parameter, or the slope of the
function, describes the ratio of older over younger central processing time
and thus provides an index of age-related slowing in central processing.
(The ␤0 parameter, or the intercept of the function, has no directly inter-
pretable meaning.) We adopted this equation as our model for age differ-
ences in single-task performance.
At the descriptive level, we tested whether different equations were
needed for single-task and dual-task conditions. In Appendix A, we discuss
the theoretical meaning of possible observed differences between condi-
tions. In our multilevel Brinley analysis, as in the multilevel state-trace
analysis, we distinguished a condition-level model and a study-level
model. The condition-level model represented the mean reaction time of
older adults as a linear function of the mean reaction time of younger adults
for each study, using a dummy code for investigating single- versus
dual-task effects. The study-level model represented each regression pa-
rameter as a function of the overall mean and each study’s unique effect.
Accuracy analysis. Preliminary model fitting and inspection of the
residuals from regression analysis of the raw accuracy data indicated
evidence of heteroscedasticity, which was eliminated by applying the logit
transformation, as was done for the state-trace analysis (see Verhaeghen,
Vandenbroucke, & Dierckx, 1998, for an earlier application of the logit
model in the context of detecting age-related differences in cognition). The
main analysis tested whether different equations were needed for single-
task and dual-task conditions after the logit transformation had been
applied.
Significance level of all statistical tests was set at .05.

Results
Outlier Analysis
Latency data. As a first descriptive tool, we offer a tabulation
of the complete set of aggregated latency data in Table 1; Figure
1 represents a plot of the mean latencies in state space (Figure 1A)
and Brinley space (Figure 1B). One of the possibly problematic
aspects of the present data set, as can be seen in Table 1 and
Figure 1, is the large but unbalanced diversity in reaction times.
Figure 1 shows that a large number of the data points cluster below
latencies of 2,000 ms; only a few studies have longer latencies, but
these latencies are then quite a bit larger than 2,000 ms. Explor-
atory analysis on the data of younger adults indeed shows that all
single task data points with latencies of 2,550 ms or longer are
outliers, in the sense that these latencies are situated at three or
more interquartile ranges from the mean.
During our first pass over the data, it became clear that these
long-latency points were quite influential in the estimation of the
regression parameters, especially for the Brinley analysis. For the
full data set, a linear model (WLS, weighting for sample size)
fitted the data very nicely, with R2 ⫽ .99 for single-task conditions
and .98 for dual-task conditions. For single-task conditions, the
intercept is 26 ms (SE ⫽ 48), and the slope is 1.26 (SE ⫽ 0.02);
for dual-task conditions, the intercept is 273 ms (SE ⫽ 102), and
the slope is 1.13 (SE ⫽ 0.02). Interaction analysis showed that
adding a dummy intercept term and an interaction term for dual-
task performance significantly improved fit (R2 full model ⫽ .985;
Figure 1. Scatter plot of the aggregate latency data (from Table 1). A:
Plots younger and older adults’ mean dual-task latency as a function of
corresponding mean single-task latency (state trace). B: Plots single- and
dual-task mean latency of older adults as a function of the corresponding
mean latency of younger adults (Brinley plot). RT ⫽ reaction time.
⌬R2 ⫽ .002; F change [2, 92] ⫽ 7.74. Both terms were significant,
implying that the Brinley regression lines for single-task and
dual-task performance differed in both intercept and slope. Note
that the Brinley slope for the dual-task performance in this full data
set was actually reliably smaller than the slope for single-task
performance, suggesting a smaller age deficit in central processing
under dual-task conditions than under single-task conditions.
Closer scrutiny of the data showed that this counterintuitive
result may have been due to the large influence on the regression
lines by two of the observations with long latencies: (a) the study
by Broadbent and Heron (1962) and (b) the study by Kramer,
Larish, & Strayer (1995). The former study had a very large
influence on the regression lines (Cook’s D ⫽ 8.19 and 13.94 for
the single-task and dual-task condition, respectively), probably
because of its extremely long latencies. The latter study was an
 AGING AND DUAL-TASK PERFORMANCE 449

outlier for the dual-task condition (Studentized residual ⫽ 6.20)

and was also quite influential in that condition (Cook’s D ⫽ 0.68,
which is large compared with the next rank-ordered Cook’s D,
which was 0.04). Therefore, we omitted these two studies from the
database and reestimated the regression lines. The linear model
again fitted the data very well, R2 ⫽ .98 and .98 for single-task and
dual-task conditions, respectively. For single-task conditions,
␤0 ⫽ 84 (SE ⫽ 42), and ␤l ⫽ 1.17 (SE ⫽ 0.02); for dual-task
conditions, ␤0 ⫽ 145 (SE ⫽ 49), and ␤1 ⫽ 1.17 (SE ⫽ 0.02).
Interaction analysis showed that the two regression lines were
statistically identical (R2 full model ⫽ .983; ⌬R2 ⫽ .000; F change
[2, 88] ⫽ 0.99); the resulting single regression line for both types
of conditions has an intercept of 113 ms (SE ⫽ 30) and a slope of
1.17 (SE ⫽ 0.02). This Brinley slope, however, is quite small
compared with slopes typically observed in aging research (e.g.,
Cerella, 1990). Visual inspection of Figure 1 suggests that there is
a discontinuity in the scatterplot around the 2,000-ms mark on the
abscissa: the slope of the Brinley function appears to be steeper for
tasks taking less than 2 s than for tasks taking longer than 2 s.
Therefore, in a second Brinley analysis, we examined only the data
points for which younger adults’ single-task latencies were faster
than 2 s (38 data points for single-task performance, 38 for
dual-task performance). This model produced a slope of 1.57,
which is more in line with expectations.
Given these two findings, namely, that the outliers in the data set
had very long task latencies and that the Brinley slope is more in
line with expectations for single tasks with latencies below 2,000
ms, we decided to truncate the data set and include only studies
with younger adult single-task latencies smaller than 2,000 ms
(k ⫽ 38, with k being the number of independent participant
groups).
Accuracy data. A tabulation of the complete set of accuracy
data is provided in Table 2; Figure 2 represents a plot of the
average accuracies in state space (Figure 2A) and Brinley space
(Figure 2B). Exploratory data analysis showed that no data points
could be considered as outliers. Therefore, all further analyses
were conducted on the full data set (k ⫽ 40).
Figure 2. Scatter plot of the truncated aggregate latency data (from
State-Trace Analysis Table 1, omitting all data points with younger adults’ mean single-task
latencies longer than 2,000 ms). A: Plots younger and older adults’ mean
Latency data. Figure 2A depicts the latency state space for the dual-task latency as a function of corresponding mean single-task latency
truncated data set. Multilevel modeling procedures were applied to (state trace). B: Plots single- and dual-task mean latency of older adults as
this data set. Fitting a model of the form shown in Equation 3 a function of the corresponding mean latency of younger adults (Brinley
produced the following parameter estimates: ␤៮ 0 ⫽ 109.4 plot). The regression lines are derived from the aggregate regression
(SE ⫽ 27.3), ␣៮ 0 ⫽ 50.7 (SE ⫽ 21.3), ␤៮ 1 ⫽ 0.96 (SE ⫽ 0.06), and analysis. RT ⫽ reaction time.
␣៮ 1 ⫽ 0.04 (SE ⫽ 0.03). The ␣៮ 1 parameter, which conveys the age
effect on the slope relating dual- to single-task performance, was
not significantly larger than zero. Therefore, the model was refit icantly larger than zero, indicating that the intercept was larger in
constraining ␣៮ 1 to be zero. Adding this constraint did not reduce older compared with younger adults. Thus, analysis of the fixed
the fit of the model as evidenced by a likelihood ratio test (␹2[1, effects was consistent with an additive dual-task cost and with an
N ⫽ 40] ⫽ 2.3, p ⬎ .05). The additive age-effect model yielded the additive age effect in this cost. This model accounted for over 95%
following parameter estimates: ␤៮ 0 ⫽ 92.1 (SE ⫽ 24.9), ␣៮ 0 ⫽ 77.9 of the within-study variance in mean dual-task reaction times, and
(SE ⫽ 11.5), ␤៮ 1 ⫽ 0.99 (SE ⫽ 0.06). The estimate of the ␤៮ 1 inspection of residual plots revealed no systematic pattern or
parameter (the slope of the function relating dual-task performance evidence of heteroscedasticity, suggesting the model provided an
to single-task performance) contained unity within its 95% confi- adequate fit to the data.
dence interval (0.87 to 1.11), and the intercept estimates for the We also investigated the influence of our selected moderator
younger and older adults were significantly greater than zero. This variables on the additive dual-task cost. The first moderator vari-
suggests that the dual-task cost is constant across task complexity able was whether the task relied primarily on sensorimotor or
as defined by single-task latencies. The ␣៮ 0 parameter was signif- central processes. Tasks classified as primarily sensorimotor were
450 VERHAEGHEN, STEITZ, SLIWINSKI, AND CERELLA

reaction time to visual or auditory probes, visual or auditory

discrimination, color or letter identification, tone matching and
visual matching. All other tasks were classified as primarily cen-
tral. Slopes for studies using sensorimotor (0.96, SE ⫽ .07) and
central (1.02, SE ⫽ .13) primary tasks did not differ by a statisti-
cally significant amount ( p ⫽ .68). Additive effects did not differ
between studies using sensorimotor (105.1, SE ⫽ 31.8) and studies
using central (93.3, SE ⫽ 53.3) primary tasks ( p ⫽ .80), nor did
the additive age effects differ between sensorimotor (73.4,
SE ⫽ 27.2) and central (39.8, SE ⫽ 23.3) tasks by a statistically
significant amount ( p ⫽ .14). The second moderator variable
coded whether the primary and secondary tasks were verbal or
nonverbal. The additive cost was significantly higher ( p ⫽ .03) for
verbal (271.1, SE ⫽ 65.1) compared with nonverbal (99.3,
SE ⫽ 49.6). The average slope for verbal tasks (0.75, SE ⫽ .13)
was significantly lower ( p ⫽ .05) than for nonverbal tasks (1.08,
SE ⫽ .10), although both slopes did not differ significantly
from 1.0. The verbal–nonverbal contrast did not modify any other
effect. The third moderator involved matching of input modalities.
The additive cost was higher for studies in which the input mo-
dality matched (127.7, SE ⫽ 29.6) compared with studies in which
input modalities did not match (80.1, SE ⫽ 11.5, p ⫽ .04). Input
modality, our fourth variable, did not produce any reliable effects.
Our fifth variable was vocal versus manual output. Vocal-output
tasks were associated with reduced and nonsignificant additive
dual-task costs and with reduced and nonsignificant age effects
( p ⬍ .01): ␤៮ 0 ⫽ 37.6 (SE ⫽ 35.7) versus 105.6 (SE ⫽ 26.1),
␣៮ 0 ⫽ 30.3 (SE ⫽ 21.7) versus 95.7 (SE ⫽ 12.7), respectively. The
slopes for the studies using a vocal-output mode (␤៮ 1 ⫽ 1.14,
SE ⫽ 0.14) and manual-output mode (␤៮ 1 ⫽ 0.97, SE ⫽ 0.07) for
the primary tasks did not differ by a statistically significant amount
( p ⫽ 0.28), and both slopes were not significantly different
from 1.0.
Accuracy analysis. Figure 3A shows the accuracy state space.
Accuracy scores were logit transformed prior to the multilevel
analysis. Note that this amounts to fitting a curvilinear model in the
original state space. Studies that had experimental conditions
yielding perfect performance were omitted from this analysis, as Figure 3. Scatter plot of the aggregate accuracy data (from Table 2),
the logit score for P ⫽ 1.00 is undefined. Fitting a model of the expressed as percentage correct. A: Plots younger and older adults’ mean
form shown in Equation 3 to the logit scores produced the follow- dual-task accuracy as a function of corresponding mean single-task latency
ing parameter estimates: ␤៮ 0 ⫽ ⫺0.41 (SE ⫽ 0.12), ␣៮ 0 ⫽ ⫺.03 (state trace). B: Plots single- and dual-task mean accuracy of older adults
(SE ⫽ 0.07), ␤៮ 1 ⫽ 0.95 (SE ⫽ 0.33), and ␣៮ 1 ⫽ ⫺.03 (SE ⫽ 0.03). as a function of the corresponding mean latency of younger adults (Brinley
Neither the coefficient associated with age differences in intercept plot). The regression lines are derived from the aggregate regression
nor the coefficient associated with age differences in slope was analysis on logit-transformed scores.
significant, so the model was refitted including only ␤៮ 0 and ␤៮ 1 as
parameters. This model had an intercept of ⫺0.45 (SE ⫽ 0.11) and state trace was lower for the visual compared with auditory input
a slope of 0.95 (SE ⫽ 0.03). The intercept was significantly primary-task mode, ␤៮ 1 ⫽ 0.90 (SE ⫽ 0.04), and ␤៮ 1 ⫽ 1.02
smaller than zero ( p ⬍ .01); the slope was not significantly (SE ⫽ 0.04), respectively ( p ⬍ .04). Primary-task input mode did
different from unity. This model accounted for 84.7% of the not influence the additive dual-task cost nor the effect of age ( p ⬎
within-study variance in mean dual-task accuracy. Inspection of .30 for all comparisons). None of the other candidate moderators
residual plots revealed no systematic pattern or evidence of het- influenced the dual-task costs or age effects by a statistically
eroscedasticity, suggesting the model provided an adequate fit to significant amount ( p ⬎ .20 for all comparisons).
the data. Together, then, these analyses show that (a) there was an
additive effect of dual-task performance on accuracy in logit-
Brinley Analysis
transformed space (i.e., the logit of proportion correct was reduced
by a constant amount) and (b) there was no evidence for an Latency data. Figure 2A shows the latency Brinley space for
age-specific effect on dual-task performance. the truncated data set. A multilevel regression model was fit to the
We next investigated the influence of the same set of moderator young– old response-time data to test whether the slowing func-
variables examined in the response-time analyses. The slope of the tions differed between single- and dual-task conditions. The results
 AGING AND DUAL-TASK PERFORMANCE
of the model fitting yielded the following parameter estimates:
␤៮ 0 ⫽ ⫺190.7 (SE ⫽ 48.1), ␣៮ 0 ⫽ 17.3 (SE ⫽ 29.3), ␤៮ 1 ⫽ 1.61
(SE ⫽ 0.08), ␣៮ 1 ⫽ 0.03 (SE ⫽ 0.04). The value for ␣៮ 1 was not
reliably different from zero, indicating that the average slopes for
the single- and dual-task conditions did not differ significantly.
Refitting the model with ␣៮ 1 constrained to be zero yielded the
following parameter estimates: ␤៮ 0 ⫽ ⫺199.5 (SE ⫽ 46.5),
␣៮ 0 ⫽ 35.7 (SE ⫽ 12.1), ␤៮ 1 ⫽ 1.63 (SE ⫽ 0.08). The ␣៮ 0 parameter
was significant. Thus, two lines emerged in Brinley space: a line
for single-task conditions, with an intercept of ⫺200 ms and a
slope of 1.63, and a line for dual-task performance parallel to the
line for single-task performance but elevated 36 ms above it. The
existence of two lines indicates an age-specific deficit in dual-task
performance. The model accounted for 95.6% of the within-study
variance.
We investigated the influence of the same set of moderator
variables as in the previous state-trace latency analyses. Consistent
with previous analyses (Hale & Myerson, 1996), the slope for
nonverbal tasks (1.78, SE ⫽ .12) was significantly steeper ( p ⬍
.01) than for nonverbal tasks (1.18, SE ⫽ .15). The latter slope was
not significantly larger than 1.0. However, the verbal–nonverbal
contrast did not influence the dual-task effects. The only other
moderator variable to influence age-related slowing was one that
coded for whether the input modalities for the primary and sec-
ondary tasks matched (e.g., both were visual or both were audi-
tory). Older adults exhibited less slowing on tasks in which the
input modality matched compared with studies in which input
modalities did not match (⫺133.9, SE ⫽ 33.0, p ⬍ .01). Modality
match did not interact with either dual-task cost or the Brinley
slope ( p ⬎ .10 for all).
Accuracy data. Figure 3B depicts the accuracy Brinley plot.
Fitting a full model to the logit-transformed data (omitting studies
with perfect accuracies) produced the following parameter esti-
mates: ␤៮ 0 ⫽ ⫺0.46 (SE ⫽ 0.11), ␣៮ 0 ⫽ ⫺0.03 (SE ⫽ 0.08),
␤៮ 1 ⫽ 0.94 (SE ⫽ 0.03), and ␣៮ 1 ⫽ ⫺0.03 (SE ⫽ 0.03). Given that
neither the coefficient associated with the age dummy nor the
interaction term was significant, the model was refitted including
only ␤៮ 0 and ␤1 as parameters. The intercept of this model was
equal to ⫺0.49 (SE ⫽ 0.09); the slope was equal to 0.93
(SE ⫽ 0.03). The intercept was less than zero by a statistically
significant margin, and the slope was significantly less than one.
Follow-up analyses indicated that only one moderator influ-
enced age differences in dual-task effects. Older adults exhibited
lower accuracy on tasks in which the input modality of the primary
and secondary tasks matched (␤៮ 0 ⫽ ⫺0.58, SE ⫽ 0.12) compared
with studies in which input modalities did not match (␤៮ 0 ⫽ ⫺0.25,
SE ⫽ 0.12, p ⫽ .03). Modality match did not interact, however,
with either dual-task cost or with the slope ( p ⬎ .10 for all).
Speed–Accuracy Trade-Offs
The dependent variable for dual-task performance is typically
either latency or accuracy, and participants are typically instructed
to give priority to one over the other. It is, however, possible that
the specific age-related effect of dual tasking on latency is due to
a different speed–accuracy trade-off, that is, it is possible that
compared with younger adults, older adults have a greater ten-
dency to sacrifice speed to keep accuracy high. Such age differ-
ence in strategy use alone could produce the observed dissociation.
451
To test this assumption, we plotted accuracy differences (i.e.,
single-task accuracy ⫺ dual-task accuracy) as a function of latency
differences (i.e., single-task latency ⫺ dual-task latency) for both
young and older adults (see Figure 4). Even though a formal
analysis of these data is not possible, informal inspection of the
figure suggests two conclusions. First, the data tend to cluster
along the axes of the figure, implying that the cost occurs in either
accuracy or latency, but not in both simultaneously. This is reas-
suring because it suggests that participants are complying with
instructions to change only one aspect of behavior. Second, when
data points do stray from the diagonal, they are situated in the
fourth quadrant, implying that accuracy costs lead to slower speed
and, conversely, speed costs lead to lower accuracy. If speed–
accuracy trade-offs were the typical occurrence, one might expect
the data to be situated in the third quadrant (lower accuracy
coupled with a speed-up) or first quadrant (longer latency coupled
with an increase in accuracy). Moreover, the data for older adult
groups seem to dwell into the fourth quadrant more often than
those of younger adults, strongly suggesting that, if anything, older
adults show speed–accuracy trade-off effects less often than
younger adults do.
Discussion
The main question raised in this meta-analysis is whether an
age-associated deficit in dual-task performance exists over and
beyond the effect of general age-related slowing. We examined
studies based on both latency and accuracy measures. For each set
of studies, we conducted two analyses: state-trace analysis and
Brinley analysis, using multilevel hierarchical modeling (Sliwinski
& Hall, 1998). We summarize first the findings from the latency
data and then those from the accuracy data, and we end with some
general conclusions.
With regard to latency data, the findings pointed to two conclu-
sions: (a) Dual-task costs are additive (rather than multiplicative),
and (b) these additive costs are larger in older than in younger
adults. Therefore, there is a significant age-related deficit in dual-
task performance as measured by latency costs. Two converging
Figure 4. Dual-task costs in logit-transformed accuracy as function of
dual-task costs in latency, separated by age.
452 VERHAEGHEN, STEITZ, SLIWINSKI, AND CERELLA
results support these conclusions. First, multilevel modeling of the
state-trace demonstrated that dual-task costs in both younger and
older adults are additive, that is, the requirement of having to
perform a concurrent second task adds a fixed cost to the latency
of the first task. This was evidenced by the conformation of the
state traces, which had slopes of one (viz., 0.99, with an
SE ⫽ 0.06) and intercepts greater than zero. Under the assumption
of serial processing, this result is in accordance with the hypothesis
that dual tasking has the effect of adding an additional stage to the
sequence of processes already present in the single task. The
duration of the added step, as derived from this analysis, was 92
ms in younger adults (95% confidence interval ranging from 43 ms
to 141 ms). In older adults, the added step took an extra 78 ms
(95% confidence interval ranging from 55 ms to 100 ms), for a
total duration of 170 ms. The dual-tasking stage, then, takes about
twice as long in older adults as in younger adults. This can also be
seen in the raw data: The average single-task latency in the
truncated data set was 624 ms for the young and 841 ms for the
old; the average dual-task latency in the truncated data set was 729
ms for the young and 1,057 ms for the old, respectively, defining
a cost of 105 ms for the young and of 216 ms for the old. In terms
of percentages of the single-task latency, the dual-task cost for the
young in the truncated data set was 17%, versus 26% for the old.
Second, multilevel modeling in Brinley space showed that the
Brinley line for dual-task performance was elevated above the line
for single-task performance by a constant 36 ms (95% confidence
interval ranging from 12 ms to 59 ms). This result is an unusual
one (an age deficit that takes the form of parallel lines in a Brinley
plot)—in fact, as far as we know, this is the first demonstration of
such an effect in a young– old meta-analysis. Typically, age-
related effects in Brinley plots are situated in the slope of the
young– old function and not just in the intercept. The presence of
an intercept difference in the absence of a slope difference signals
a deviation from general slowing. Like the state-trace results, this
result points to the presence of a specific age-related deficit in
dual-task performance. This effect is additive to the age-related
effect present in single-task performance.
A number of additional findings are worth pointing out. First,
single and dual tasks taking less than about 2 s to perform yielded
a larger Brinley slope than tasks taking longer to perform. The
reason appears to lie in the nature of the single tasks. As can be
seen in Table 1, the studies with slow reaction times mostly used
verbal tasks (span measures, sentence verification, and the like).
Verbal tasks typically yield smaller Brinley slopes than nonverbal
tasks (see, e.g., Hale & Myerson, 1996; Sliwinski & Hall, 1998);
in our sample of studies, the nonverbal tasks more often showed
faster reaction times. In agreement with this hypothesis, we indeed
found that in our truncated data set, age differences were larger for
visuospatial tasks (age-related slowing factor of 1.8) than for
verbal tasks (age-related slowing factor of 1.2).
Second, some of our moderator variables did indeed produce
differential dual-task costs in latency. We found an absence of
dual-task costs (and an absence of an age difference) with vocal
reaction times. A modality match in tasks produced reliably larger
dual-task costs than when tasks relied on different modalities. With
the exception of the vocal/manual reaction-time variable, however,
none of the moderator variables interacted with age, indicating that
the age effect in dual-task performance is relatively general across
types of processing; in other words, the same variables that influ-
ence dual-task costs in younger adults also influence dual-task
costs in older adults, and in similar amounts.
Turning now to the accuracy data, the main conclusions are that
(a) dual-task performance yields lower accuracy than single-task
performance and the cost is additive after logit transformation and
(b) there is no evidence for age-related differences in either single-
or dual-task performance. The first conclusion follows from state-
trace analysis. We found a decreased intercept and a slope not
different from unity, indicating that dual-task performance leads to
a fixed accuracy cost, at least after logit transformation. Thus,
dual-task costs are well described by a fixed drop in logits. Note
that a constant cost in logit-transformed space takes the shape of a
crescent below the diagonal, bounded by the origin and the
(100,100) point when transformed back to percentage correct (see
Figure 3). The second conclusion follows from both state-trace and
Brinley analyses. The state-trace analysis showed that the equa-
tions describing accuracy under dual-task conditions as a function
of single-task conditions are equivalent in younger and older
adults. Likewise, the Brinley analysis showed that a single func-
tion suffices to relate performance of older adults to that of
younger adults; no separate lines are needed to describe age-
relations in single- and dual-task performance. The visual-input
modality led to a larger decrease in accuracy than the auditory-
input modality but did not produce a reliable age difference in
dual-task cost. Additionally, none of the other moderator variables
yielded a significant age difference in dual-task cost.
Let us take stock. How is one to evaluate the age-specific effect
in latency measures of dual-task performance? This question may
be answered in several ways. First, note that for both older and
younger adults, the effect of having to perform a secondary task is
expressed as an additive cost in latency data. The simplest inter-
pretation of this outcome is to suppose that an extra stage is added
to a processing sequence. The new stage introduced in the pro-
cessing chain takes longer in older adults than in younger adults.
We also found that the added stage elevated the error rate but did
so equally in younger and older adults. The latter result is reas-
suring because Age Group ⫻ Condition interactions in latency can
be safely interpreted only in the absence of Age ⫻ Condition
interactions in accuracy (Cerella, 1990). For instance, the added
processing step might take so much time in older adults that
intermediate results decay from working memory, making dual-
task performance relatively more error prone in older adults (Salt-
house’s [1996] simultaneity mechanism of cognitive slowing). Our
findings seem to rule out this scenario. Or older adults might
sacrifice speed for accuracy in dual-task conditions, which would
lead to smaller or reversed age differences in accuracy under
dual-task conditions as compared with single-task conditions. If
this were the case, one might expect that latency dual-task costs in
older adults would be associated with a smaller decrease in accu-
racy than observed in younger adults. If anything, the opposite
seems to be the case (see Figure 4): Dual tasking seemed to be
more likely to compromise both speed and accuracy in older than
in younger adults.
In our added-stage interpretation, we assumed strict serial pro-
cessing. In this case, as elaborated in Appendix A, the appropriate
standard for evaluating the size of the age effect in speed of
dual-task performance is against the size of the age deficit present
in single-task performance. The general age deficit in single-task
performance is best defined as a ratio of the latency of central
 AGING AND DUAL-TASK PERFORMANCE
processes in older adults over the latency of central processes in
younger adults, a ratio provided directly by the slope of the Brinley
function (Cerella, 1991). Our best estimate for the general age
deficit is the slope derived from our multilevel Brinley analy-
sis, 1.63 (95% confidence interval ranging from 1.47 to 1.79). Our
best estimate for the additive cost is the intercept in the state-trace
analysis, namely, 92 ms in the young and 170 ms in the old.
Consequently, the age-related slowing ratio in dual-task cost
equals 1.85 (viz., 170/92). This slowing ratio in dual-task costs
falls outside the confidence interval for the Brinley slope. Hence,
we conclude that the age effect in dual-task cost is larger than the
general aging effect present in single-task performance.
Although the insertion of an additional processing stage is
consistent with the dual-task effect displayed in the state traces, it
is not the only possibility. Retaining the framework of serial-stage
models, Sternberg (1969) has shown that additive factors, such as
those of task difficulty and dual-task level, are also compatible
with the tenets of selective influence. Under that interpretation, the
effect of the dual-task requirement is to extend the duration of an
existing stage rather than to insert a new stage. We show in
Appendix B that the same ensemble of slowing factors continues
to account for age effects in the alternative model.
Discrete-stage models of any stripe have met with increasing
skepticism. Broadbent (1984) argued that they fail to represent the
dynamics of neural processes. A growing number of behavioral
studies (such as Kounios, Montgomery, & Smith, 1994; Meyer,
Irwin, Osman, & Kounios, 1988) have shown that information
accumulates gradually in the course of processing and is accessible
prior to the completion of a process, contrary to the all-or-none
characteristic of discrete components. These graded dynamics are
captured in the cascade model of McClelland (1979) in configu-
rations that are not incompatible with additive outcomes. In Ap-
pendix C, we develop a (schematic) cascade model of dual tasking
that successfully reproduces the state traces. We show once again
that the same ensemble of slowing factors, applied to the time
constants of the cascade, continues to provide a quantitative ac-
count of the age effects in the data.
These various models converge on a common conclusion that
dual-task performance is age sensitive, at least as far as latency is
concerned. This raises the question of whether inflated dual-task
costs might explain age differences in more complex aspects of
cognition. Now, our analysis suggests that the dual-task cost is
additive. Existing processes are not penalized in a way that would
be expressed by an elevated slope in the state trace. Rather, the age
difference is expressed by an intercept change in the state trace.
The single-task/dual-task manipulation, however, seems very close
to a manipulation advanced by Mayr and Kliegl (1993), which
does indeed drive a slope change. This manipulation is between
sequential and coordinative tasks. Sequential tasks involve pro-
cessing that proceeds in a chainlike fashion, where every interme-
diate result is immediately used in the next step; coordinative tasks
involve processing that requires intermediate storage of results
while processing is being carried out concurrently. Coordinative
processing typically leads to larger age differences than those
produced in sequential tasks (Kliegl, Mayr, & Krampe, 1994;
Kliegl et al., 2000; Mayr & Kliegl, 1993; Mayr, Kliegl, & Krampe,
1996; Verhaeghen et al., 1997), as evidenced by significant slope
differences in Brinley plots. In other words, the effect of coordi-
453
native processing appears to pervade all central processes in the
sequential tasks.
The main requirement of dual-task performance, to perform two
tasks concurrently, certainly entails coordination of performance
on the two tasks, and we have just shown that this coordinative
demand does not manifest itself in a slope difference but in an
intercept difference. Consequently, one might be tempted to con-
clude that the cause of the elevated age deficits in coordinative
data (e.g., Mayr & Kliegl, 1993) is unlikely to be the process of
coordination itself: Coordinative deficits are not an instance of
dual-tasking deficits. Further consideration, though, leads to the
following suggestion. The additive age effect in dual-task cost and
the multiplicative age effect of coordination may be two sides of
the same coin in that multiple and repeated switches between
processing streams may result in multiplicative age deficits even if
each of the switching steps by itself carries an additive deficit (see
Appendix B for a mathematical proof of this assertion). The
different age outcomes in Mayr and Kliegl’s studies (slope differ-
ences) and in the dual-task studies (intercept differences) may be
due to nothing more than the frequency of switching engendered
by the two paradigms. We should also point out, however, that this
implies that all coordinative tasks with a similar age-related slow-
ing factor should include the same proportion of task switching
interspersed with the central processing bouts. Interestingly, coor-
dinative tasks often yield very similar slowing factors indeed (see,
e.g., Verhaeghen et al., 2002). That all of these tasks are charac-
terized by similar needs for task switching is possible but quite
unlikely. Therefore, it seems that the additive age-related deficits
found here might theoretically scale up to yield multiplicative age
deficits, but this seems empirically implausible.
Some restrictions to our conclusions should be pointed out.
First, a number of potentially relevant studies could not be exam-
ined because no information on single-task performance was avail-
able. This was often the case with studies using the PRP (i.e.,
psychological refractory period) paradigm, where single-task per-
formance is imputed from performance at long stimulus onset
asynchronies. Given that these studies used different input modal-
ities for the two tasks, it is likely that exclusion of these studies led
to an overestimation of the magnitude of dual-task effects in both
younger and older adults. Second, the number of studies is rela-
tively modest. This may have resulted in suboptimal power to
detect effects in the moderator analyses. The reader may note,
however, that the technique we used is quite sensitive—we repli-
cated the analyses using traditional effect-size analysis and
weighted least-squares regression and found that effects emerging
from the aggregate regression analysis were not always detected
by these more traditional approaches. Third, a rather heteroge-
neous group of studies is lumped into one analysis. Even though
our moderator analysis suggests that age differences in the dual-
task effect do not covary with different types or complexity levels
within a particular task (with the exception of the absence of an
age effect in vocal reaction time), more and more careful experi-
mental analysis is clearly needed.
To summarize, we found that both younger and older adults
suffer from the effects of dual-task performance. Dual-task costs
on latency are additive, and older adults suffer from a larger
dual-task cost than younger adults. This dual-task cost in older
adults is larger than the effect that would be expected from general
slowing. However, because the effect is additive and independent
454 VERHAEGHEN, STEITZ, SLIWINSKI, AND CERELLA
of task complexity, there is little reason to assume that it is an
important factor in the explanation of age differences in more
complex aspects of cognition. No specific age deficit is associated
with dual-task costs on accuracy.
References
References marked with an asterisk indicate studies included in the
meta-analysis.
Allport, A., Styles, E. A., & Hsieh, S. (1994). Shifting intentional set:
Exploring the dynamic control of tasks. In C. Umilta & M. Moscovitch
(Eds.), Attention and Performance IV (pp. 421– 452). Cambridge, MA:
MIT Press.
*Anderson, N. D. (1999). The attentional demands of encoding and re-
trieval in younger and older adults: 2. Evidence from secondary task
reaction time distributions. Psychology and Aging, 14, 645– 655.
*Anderson, N. D., Iidaka, T., Cabeza, R., Kapur, S., McIntosh, A. R., &
Craik, F. I. M. (2000). The effects of divided attention on encoding- and
retrieval-related brain activity: A PET study of younger and older adults.
Journal of Cognitive Neuroscience, 12, 775–792.
Ashby, F. G. (1982). Deriving exact predictions from the Cascade model.
Psychological Review, 89, 599 – 607.
*Baddeley, A., Logie, R., Bressi, S., Della Sala, S., & Spinnler, H. (1986).
Dementia and working memory. Quarterly Journal of Experimental
Psychology, 38A, 603– 618.
Bamber, D. (1979). State-trace analysis: A method of testing simple
theories of causation. Journal of Mathematical Psychology, 19, 137–
181.
Baron, A., & Mattila, W. R. (1989). Response slowing of older adults:
Effects of time limit contingencies on single- and dual-task performance.
Psychology and Aging, 4, 66 –72.
Bopp, K. L., & Verhaeghen, P. (2002). Aging, short-term memory span,
and working memory span: A meta-analysis. Manuscript submitted for
publication.
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: Biological determinants of speed of behavior and its
changes with age (pp. 114 –149). Springfield, IL: Charles C Thomas.
Broadbent, D. E. (1984). The Maltese cross: A new simplistic model for
memory. Behavioral and Brain Sciences, 7, 55–94.
*Broadbent, D. E., & Heron, A. (1962). Effects of a subsidiary task on
performance involving immediate memory by younger and older men.
British Journal of Psychology, 53, 189 –198.
Cerella, J. (1985). Information processing rates in the elderly. Psycholog-
ical Bulletin, 98, 67– 83.
Cerella, J. (1990). Aging and information processing rate. In J. E. Birren &
K. W. Schaie (Eds.), Handbook of the psychology of aging (3rd ed., pp.
201–221). San Diego, CA: Academic Press.
Cerella, J. (1991). Age effect may be global not local: Comment on Fisk
and Rogers (1991). Journal of Experimental Psychology: General, 120,
215–223.
Cerella, J. (1994). Generalized slowing in Brinley plots. Journals of
Gerontology: Psychological Sciences, P49, 65–71.
Cerella, J., & Hale, S. (1994). The rise and fall in information-processing
rates over the life span. Acta Psychologica, 86, 109 –197.
Cerella, J., Poon, L. W., & Williams, D. H. (1980). Age and the complexity
hypothesis. In L. W. Poon (Ed.), Aging in the 1980s: Psychological
issues (pp. 332–340). Washington, DC: American Psychological
Association.
*Chen, H., Schultz, A. B., Ashton-Miller, J. A., Giordani, B., Alexander,
N. B., & Guire, K. E. (1996). Stepping over obstacles: Dividing attention
impairs performance of old more than young adults. Journal of Geron-
tology, 51A, M116 –M122.
*Craik, F. I. M., & McDowd, J. M. (1987). Age differences in recall and
recognition. Journal of Experimental Psychology: Learning, Memory,
and Cognition, 13, 474 – 479.
Craik, F. I. M., & Salthouse, T. A. (Eds.). (2000). The handbook of aging
and cognition (2nd ed.). Mahwah, NJ: Erlbaum.
Dunn, J. C., & Kirsner, K. (1988). Discovering functionally independent
mental processes: The principle of reversed association. Psychological
Review, 95, 91–101.
Faust, M. E., Balota, D. A., Spieler, D. H., & Ferraro, F. R. (1999).
Individual differences in information-processing rate and amount: Im-
plications for group differences in response latency. Psychological Bul-
letin, 125, 777–799.
*Gick, M. L., Craik, F. I. M., & Morris, R. G. (1988). Task complexity and
age differences in working memory. Memory & Cognition, 16, 353–361.
*Greenwood, P., & Parasuraman, R. (1991). Effects of aging on the speed
and attentional cost of cognitive operations. Developmental Neuropsy-
chology, 7, 421– 434.
*Guttentag, R. E., & Madden, D. J. (1987). Adult age differences in the
attentional capacity demands of letter matching. Experimental Aging
Research, 13, 93–99.
Hale, S., & Myerson, J. (1996). Experimental evidence for differential
slowing in the lexical and nonlexical domains. Aging, Neuropsychology,
and Cognition, 3, 154 –165.
*Hartley, A. A., & Little, D. M. (1999). Age-related differences and
similarities in dual task interference. Journal of Experimental Psychol-
ogy: General, 128, 416 – 449.
Hasher, L., & Zacks, R. T. (1988). Working memory, comprehension, and
aging: A review and a new view. In G. H. Bower (Ed.), The psychology
of learning and motivation (Vol. 22, pp. 193–225). San Diego, CA:
Academic Press.
*Hawkins, H. L., Kramer, A. F., & Capaldi, D. (1992). Aging, exercise,
and attention. Psychology and Aging, 7, 643– 653.
*Isingrini, M., Vazou, F., & Leroy, P. (1995). Dissociation of implicit and
explicit memory tests: Effect of age and divided attention on category
exemplar generation and cued recall. Memory & Cognition, 23, 462–
467.
Kliegl, R., Mayr, U., & Krampe, R. T. (1994). Time–accuracy functions for
determining process and person differences: An application to cognitive
aging. Cognitive Psychology, 26, 134 –164.
Kliegl, R., Mayr, U., & Oberauer, K. (2000). Resource limitations and
process dissociations in individual differences research. In U. von
Hecker, S. Dutke, & G. Sedek (Eds.), Generative mental processes and
cognitive resources: Integrative research on adaptation and control (pp.
337–366) Dordrecht, The Netherlands: Kluwer.
Kounios, J., Montgomery, E. C., & Smith, R. W. (1994). Semantic memory
and the granularity of semantic relations: Evidence from speed–accuracy
decomposition. Memory & Cognition, 22, 729 –740.
*Kramer, A. F., Larish, J. F., & Strayer, D. L. (1995). Training for
attentional control in dual task settings: A comparison of young and old
adults. Journal of Experimental Psychology: Applied, 1, 50 –76.
*Li, K. Z. H. (1999). Selection from working memory: On the relationship
between processing and storage components. Aging, Neuropsychology,
and Cognition, 6, 99 –116.
*Li, K. Z. H., Lindenberger, U., Freund, A. M., & Baltes, P. B. (2001).
Walking while memorizing: Age-related differences in compensatory
behavior. Psychological Science, 12, 230 –237.
*Light, L. L., Kennison, R., Prull, M. W., La Voie, D., & Zuellig, A.
(1996). One-trial associative priming of nonwords in young and older
adults. Psychology and Aging, 11, 417– 430.
*Light, L. L., & Prull, M. W. (1995). Aging, divided attention, and
repetition priming. Swiss Journal of Psychology, 54, 87–101.
*Light, L. L., Prull, M. W., & Kennison, R. F. (2000). Dividing attention,
aging, and priming in exemplar generation and category verification.
Memory & Cognition, 28, 856 – 872.
*Lindenberger, U., Marsiske, M., & Baltes, P. B. (2000). Memorizing
 AGING AND DUAL-TASK PERFORMANCE
while walking: Increase in dual-task costs from young adulthood to old
age. Psychology and Aging, 15, 417– 436.
*Lorsbach, T. C., & Simpson, G. B. (1988). Dual-task performance as a
function of adult age and task complexity. Psychology and Aging, 3,
210 –212.
Luce, R. D. (1986). Response times: Their role in inferring elementary
mental organization. New York: Oxford University Press.
Macht, M. L., & Buschke, H. (1983). Age differences in cognitive effort in
recall. Journal of Gerontology, 38, 695–700.
*Madden, D. J. (1986). Adult age differences in the attentional capacity
demands of visual search. Cognitive Development, 1, 335–363.
*Madden, D. J. (1987). Aging, attention, and the use of meaning during
visual search. Cognitive Development, 2, 201–216.
Madden, D. J., Pierce, T. W., & Allen, P. A. (1992). Adult age differences
in attentional allocation during memory search. Psychology and Ag-
ing, 7, 594 – 601.
Mayr, U., & Kliegl, R. (1993). Sequential and coordinative complexity:
Age-based processing limitations in figural transformations. Journal of
Experimental Psychology: Learning, Memory, and Cognition, 19, 1297–
1320.
Mayr, U., Kliegl, R., & Krampe, R. (1996). Sequential and coordinative
processing dynamics in figural transformation across the life span.
Cognition, 59, 61–90.
McClelland, J. L. (1979). On the time relations of mental processes: A
framework for analysing processes in cascade. Psychological Re-
view, 86, 287–330.
McCullagh, P., & Nelder, J. A. (1989). Generalized linear models (2nd
ed.). New York: Chapman & Hall.
*McDowd, J. M. (1986). The effects of age and extended practice on
divided attention performance. Journal of Gerontology, 41, 764 –769.
*McDowd, J. M., & Craik, F. I. M. (1988). Effects of aging and task
difficulty on divided attention performance. Journal of Experimental
Psychology: Human Perception and Performance, 14, 267–280.
McDowd, J. M., & Shaw, R. J. (2000). Attention and aging: A functional
perspective. In F. I. M. Craik & T. A. Salthouse (Eds.), The handbook of
aging and cognition (2nd ed., pp. 221–292). Mahwah, NJ: Erlbaum.
Meyer, D. E., Irwin, D. E., Osman, A. M., & Kounios, J. (1988). The
dynamics of cognition and action: Mental processes inferred from
speed–accuracy decomposition. Psychological Review, 95, 183–237.
Monsell, S. (1996). Control of mental processes. In V. Bruce (Ed.),
Unsolved mysteries of the mind: Tutorial essays in cognition (pp. 93–
148). Hove, England: Taylor & Francis.
Morris, R., Craik, F. I. M., & Gick, M.L. (1990). Age differences in
working memory tasks: The role of secondary memory and the central
executive. Quarterly Journal of Experimental Psychology: Human Ex-
perimental Psychology, 42, 67– 86.
*Morris, R. G., Gick, M. L., & Craik, F. I. M. (1988). Processing resources
and age differences in working memory. Memory & Cognition, 16,
362–366.
Myerson, J., Adams, D. R., Hale, S., & Jenkins, L. (2003). Analysis of
group differences in processing speed: Brinley plots, Q—Q plots, and
other conspiracies. Psychonomic Bulletin and Review, 10, 224 –237.
*Park, D. C., Puglisi, J. T., & Smith, A. D. (1986). Memory for pictures:
Does an age-related decline exist? Journal of Psychology and Aging, 1,
11–17.
*Park, D. C., Puglisi, J. T., Smith, A. D., & Dudley, W. N. (1987). Cue
utilization and encoding specificity in picture recognition by older
adults. Journal of Gerontology, 42, 423– 425.
Park, D. C., & Schwarz, N. (Eds.). (2000). Cognitive aging: A primer.
Philadelphia: Psychology Press.
*Park, D. C., Smith, A. D., Dudley, W. N., & Lafronza, V. N. (1989).
Effects of age and a divided attention memory task presented during
encoding and retrieval on memory. Journal of Experimental Psychology:
Learning, Memory, and Cognition, 15, 1185–1191.
455
Pashler, H. E. (1998). The psychology of attention. Cambridge, MA: MIT
Press.
Perfect, T. J., & Maylor, E. A. (2000). Rejecting the dull hypothesis: The
relation between method and theory in cognitive aging research. In T. J.
Perfect & E. A. Maylor (Eds.), Models of cognitive aging (pp. 1–18).
Oxford, England: Oxford University Press.
Perry, A. R., & Wingfield, A. (1994). Contextual encoding by young and
elderly adults as revealed by cued and free recall. Aging and Cognition,
1, 120 –139.
*Ponds, R. W. H. M., Brouwer, W. H., & van Wolffelaar, P. C. (1988).
Age differences in divided attention in a simulated driving task. Journal
of Gerontology, 43, P151–P156.
Ratcliff, R., Spieler, D., & McKoon, G. (2000). Explicitly modeling the
effects of aging on response time. Psychonomic Bulletin and Review, 7,
1–25.
Roberts, S., & Sternberg, S. (1993). The meaning of additive reaction-time
effects: Tests of three alternatives. In D. E. Meyers & S. Kornblum
(Eds.), Attention and performance XIV: Synergies in experimental psy-
chology, artificial intelligence, and cognitive neuroscience (pp. 611–
653). Cambridge, MA: MIT Press.
Rogers, W. A. (2000). Attention and aging. In D. C. Park & N. Schwarz
(Eds.), Cognitive aging: A primer (pp. 57–73). Philadelphia: Psychology
Press.
*Rogers, W. A., Bertus, E. L., & Gilbert, D. K. (1994). Dual-task assess-
ment of age differences in automatic process development. Psychology
and Aging, 9, 398 – 413.
Salthouse, T. A. (1991). Theoretical perspectives on cognitive aging.
Hillsdale, NJ: Erlbaum.
Salthouse, T. A. (1996). The processing-speed theory of adult age differ-
ences in cognition. Psychological Review, 103, 403– 428.
*Salthouse, T. A., Fristoe, N. M., Lineweaver, T. T., & Coon, V. E. (1995).
Aging of attention: Does the ability to divide decline? Memory &
Cognition, 23, 59 –71.
*Salthouse, T. A., & Somberg, B. L. (1982). Skilled performance: Effects
of adult age and experience on elementary processes. Journal of Exper-
imental Psychology: General, 111, 176 –207.
Sliwinski, M., & Hall, C. B. (1998). Constraints on general slowing: A
meta-analysis using hierarchical linear models with random coefficients.
Psychology and Aging, 13, 164 –175.
*Somberg, B. L., & Salthouse, T. A. (1982). Divided attention abilities in
young and old adults. Journal of Experimental Psychology: Human
Perception and Performance, 8, 651– 663.
Sternberg, S. (1969). The discovery of processing stages. Extensions of
Donder’s method. In W. G. Koster (Ed.), Attention and performance II
(pp. 276 –315). Amsterdam: North-Holland.
*Tecce, J. J., Cattanach, L., Yrchik, D. A., Meinbresse, D., & Dessonville,
C. L. (1982). CNV rebound and aging. Electroencephalography and
Clinical Neurophysiology, 54, 175–186.
*Tsang, P. S., & Shaner, T. L. (1998). Age, attention, expertise, and
time-sharing performance. Psychology and Aging, 13, 323–347.
*Tun, P. A., & Wingfield, A. (1994). Speech recall under heavy load
conditions: Age, predictability, and limits on dual-task interference.
Aging and Cognition, 1, 29 – 44.
*Tun, P. A., Wingfield, A., & Stine, E. A. L. (1991). Speech-processing
capacity in young and older adults: A dual-task study. Psychology and
Aging, 6, 3–9.
*Tun, P. A., Wingfield, A., Stine, E. A. L., & Mecsas, C. (1992). Rapid
speech processing and divided attention: Processing rate versus process-
ing resources as an explanation of age effects. Psychology and Aging, 7,
546 –550.
Verhaeghen, P., Cerella, J., Semenec, S. C., Leo, M. E., Bopp, K. L., &
Steitz, D. W. (2002). Cognitive efficiency modes in old age: Perfor-
mance on sequential and coordinative verbal and visuo-spatial tasks.
Psychology and Aging, 17, 558 –570.
456 VERHAEGHEN, STEITZ, SLIWINSKI, AND CERELLA
Verhaeghen, P., & De Meersman, L. (1998a). Aging and the negative
priming effect: A meta-analysis. Psychology and Aging, 13, 1–9.
Verhaeghen, P., & De Meersman, L. (1998b). Aging and the Stroop effect:
A meta-analysis. Psychology and Aging, 13, 120 –126.
Verhaeghen, P., Kliegl, R., & Mayr, U. (1997). Sequential and coordinative
complexity in time–accuracy function for mental arithmetic. Psychology
and Aging, 12, 555–564.
Verhaeghen, P., Marcoen, A., & Goossens, L. (1993). Facts and fiction
about memory aging: A quantitative integration of research findings.
Journals of Gerontology: Psychological Sciences, 48, P157–P171.
Appendix A
State-Trace and Brinley-Function Predictions
Let Sit be the response times for Task t over its several conditions i (by
convention, subscripts are assigned so that response times increase with i),
and let Ditt⬘ be the response times for the same conditions performed in the
context of the distractor task t⬘. Further, let SYit and DYitt⬘ be the values for
young participants and SOit and DOitt⬘ be the values for old participants. All
parameters represent population means, known without error. Henceforth,
we drop the subscripts t and t⬘, understanding that we are addressing a
particular pair of tasks.
Suppose that DYi ⫽ fY (SYi) and DOi ⫽ fO(SOi) for some functions fY, fO,
presumably monotonic increasing. These are the state traces. Suppose that
SOi ⫽ gS(SYi) and DOi ⫽ gD(DYi) for some functions gS, gD, presumably
monotonic increasing. These are the Brinley functions. We ask here what
constraints the form of fY, fO, gS, and gD might place on an underlying
information-processing model. Empirically, all four functions are observed
to be linear, more or less. This drastically simplifies the model space.
Case 1: Multiplicative Dual-Tasking Effect
Let us adopt a “two-compartment” representation of SYi, common in
mathematical modeling (Luce, 1986). SYi is cast as the sum of a “decision”
or “central” stage CYi and a common “residual” or “sensorimotor” stage RY,
SYi ⫽ CYi ⫹ RY. Suppose now that the dual-task influence acts to modulate
C, thus DYi ⫽ ␦Y ⫻ CYi ⫹ RY, for some ␦Y ⱖ 1; that is, the central processes
of t are slowed by a factor ␦Y when performed in the context of t⬘. The
equations for SYi and DYi can now be solved for the state trace that they
imply, DYi ⫽ ␦Y ⫻ SYi ⫹ (1 ⫺ ␦) RY. The multiplicative dual-tasking model
can be seen to impose a strong constraint on the state trace; the trace is
linear with a slope greater than unity (because ␦Y must be greater than one
to induce a dual task cost) and an intercept less than zero (because RY ⬎ 0
and [1 ⫺ ␦Y] ⬍ 0).
What about the formulations for SOi and DOi? Work of Cerella (1990),
Hale and Myerson (1996), and Kliegl et al. (1994) suggests that age effects
are expressed as multiplicative “slowing factors,” operating at the stage
level. Following these authors, we construct expressions for the old par-
ticipants, SOi ⫽ c ⫻ CYi ⫹ r ⫻ RY and DOi ⫽ d ⫻ ␦Y ⫻ c ⫻ CYi ⫹ r ⫻
RY, where r, c, d ⱖ 1. Thus, in the old participant, Component C is slowed
by the factor c, R is slowed by the factor r, and d magnifies the dual-tasking
effect ␦. Note that r, c, and d need not be distinct—we have represented the
most general case. Solving SOi and DOi for the state trace, we get the linear
equation DOi ⫽ d ⫻ ␦Y ⫻ SOi ⫹ r(1 ⫺ d ⫻ ␦Y)RY. Comparing the two
equations, the state traces show the relation slopeOld ⬎ slopeYoung (in the
presence of a dual-tasking age deficit, i.e., d ⬎ 1).
SYi, DYi, SOi, and DOi can also be solved for the Brinley functions, which
are also linear, SOi ⫽ c ⫻ SYi ⫹ (r ⫺ c)RY and DOi ⫽ d ⫻ c ⫻ DYi ⫹ (r ⫺
d ⫻ c)RY. The combined model of dual tasking in the young and aging in
the old imposes a constraint on the Brinley functions similar to that on the
Verhaeghen, P., & Salthouse, T. A. (1997). Meta-analyses of age–
cognition relations in adulthood: Estimates of linear and non-linear
age effects and structural models. Psychological Bulletin, 122, 231–
249.
Verhaeghen, P., Vandenbroucke, A., & Dierckx, V. (1998). Growing
slower and less accurate: The effects of age on time–accuracy functions
for recall from episodic memory. Experimental Aging Research, 24,
3–19.
*Wright, R. E. (1981). Aging, divided attention, and processing capacity.
Journal of Gerontology, 36, 605– 614.
state traces, slopeDual ⬎ slopeSingle. Indeed, the two pairs of functions
exhibit the invariance
slopeOld slopeDual
⫽ .
slopeYoung slopeSingle
There is also an empirical, not a mathematical, constraint on the intercepts
of the Brinley functions. Most often, one would expect that interceptSingle
⬍ 0 because the sensorimotor deficit r is typically less than the cen-
tral deficit c. Similar logic would lead one to expect interceptDual ⬍
interceptSingle.
Case 2: Additive Dual-Tasking Effect
Suppose now that the effect of dual tasking is to impose a constant
overhead V, rather than to retard C, thus DYi ⫽ VY ⫹ CYi ⫹ RY, for VY ⱖ 0.
The depiction of SYi does not change. In the old participant, VY is condi-
tioned by the multiplicative deficit d, DOi ⫽ d ⫻ VY ⫹ c ⫻ CYi ⫹ r ⫻ RY,
d ⱖ 1. Again, SOi is unchanged. This additive dual-tasking model entails
the state traces DYi ⫽ SYi ⫹ VY and DOi ⫽ SOi ⫹ d ⫻ VY. Both traces have
a slope of one and a positive intercept and are related by interceptOld ⱖ
interceptYoung. The Brinley functions are given by SOi ⫽ c ⫻ SYi ⫹ (r
⫺ c)RY and DOi ⫽ c ⫻ DYi ⫹ (r ⫺ c)RY ⫹ (d ⫺ c)VY. The two functions
display a common slope greater than one, as well as distinct intercepts.
One again expects that interceptSingle ⬍ 0, but the relation between
interceptSingle and interceptDual depends on the relative magnitude of d and
c: If d ⬎ c, then interceptDual ⬎ interceptSingle and vice versa.
Conclusions
Case 1 and 2 differ in their depiction of the dual-tasking effect, multi-
plicative in Case 1 and additive in Case 2; they do not differ in their
depiction of the aging effect, which is multiplicative throughout. The two
cases lead to distinct predictions for the state traces and the Brinley
functions. In Case 1, the state traces for young and old will be divergent,
with slopes greater than one; in Case 2, the traces will be parallel with unity
slope. The Brinley functions will be divergent in Case 1; in Case 2, they
will be parallel.
The two cases hardly exhaust the possible information-processing mod-
els, even within the narrow domain of additive stages and additive and/or
multiplicative effects. We explored two other models. Case 3 is a combi-
nation of Cases 1 and 2: Dual tasking has both a multiplicative effect ␦ ⫻
C and an additive effect V, thus, DYi ⫽ VY ⫹ ␦Y ⫻ CYi ⫹ RY. In Case 4,
the dual-tasking effect is additive but not constant; instead, the effect is
proportional to the duration of C; thus, DYi ⫽ ␦Y ⫻ CYi ⫹ CYi ⫹ RY, where
the parameter ␦Y is now conceived of as being greater than zero but perhaps
much less than one. The predictions of Cases 3 and 4 are basically the same
 AGING AND DUAL-TASK PERFORMANCE
as Case 1: Both the state traces and the Brinley functions will diverge.
Thus, divergent slopes appear to be the signature of any model in which
dual tasking engenders a multiplicative effect, regardless of the particular
architecture through which the effect is expressed.
The outcome of the meta-analysis was exceptionally clear: The data
conformed to the predictions of Case 2. What can be concluded from this?
Clearly, the multiplicative models can be rejected. It does not follow that
the additive model can be accepted, as there must be a vast number of
plausible models that predict these simple relationships. (One alternative is
developed in Appendix C.) The most that can be said is that the data are
consistent with Case 2. The tentative acceptance of Case 2 does not mean
that its parameters can be estimated. Four equations are given in the data,
and they are not necessarily orthogonal. The models, Case 2 included, have
Appendix B
Dual Tasking and Coordination Compared
Our results on dual tasking (parallel Brinley functions) and Kliegl’s
(e.g., Mayr & Kliegl, 1993) results on coordination (divergent Brinley
functions) are suggestive of the differences between Case 2 and Case 1 of
Appendix A. Here, we draw the comparison more explicitly, placing the
two paradigms in a common framework. The two-component model of
Appendix A describes the baseline task, B, in both paradigms (the single-
task conditions or the sequential-task conditions): BYi ⫽ CYi ⫹ RY and BOi
⫽ c ⫻ CYi ⫹ r ⫻ RY. The entailed Brinley function is BOi ⫽ c ⫻ BYi ⫹ (r ⫺
c)RY. Assume that the effect of the experimental manipulation (dual tasking
or coordination) is to add a quantity of “overhead” VY to the experimental
conditions E, leading to the three-component model: EYi ⫽ VY ⫹ CYi ⫹ RY
and EOi ⫽ d ⫻ VY ⫹ c ⫻ CYi ⫹ r ⫻ RY. We propose that the difference
between the two paradigms lies in the quantity VY.
There are two useful treatments of this quantity. In one, overhead in the
experimental task is taken to be a constant, VY ⫽ VE. This would reflect a
situation where the experimental manipulation adds a stage to the process-
ing stream, as seems to be the case for dual tasking. A constant overhead
leads to the Brinley function EOi ⫽ c ⫻ EYi ⫹ (r ⫺ c)RY ⫹ (d ⫺ c)VE.
Comparing this function with the function for the baseline task, the two
lines can be seen to have the same slope and different intercepts.
The second treatment of VY is more interesting. Notice that the three-
component model can be interpreted quite flexibly. V and C (and R) need
not occur in single blocks of time. Kliegl’s description of coordinative
tasks (Mayr & Kliegl, 1993) makes it appear that bouts of central process-
ing are interspersed with bouts of overhead. Repeated cycles of this sort are
not incompatible with our equations for E, provided that the sum of all the
bout durations adds up to V or C. The interesting consequence of such
cycling is that the magnitude of V is proportional to C, VY ⫽ ␦E ⫻ CY (in
(Appendix follows)
457
more than four parameters and are therefore underdetermined. Although
Case 2 cannot be fully solved, the vagaries of these particular equations do
allow some parameters to be estimated. VY is given by the observed
interceptYoung, and d by
interceptOld
.
interceptYoung
c is given by the Brinley slope and the fact that c ⬎ r can be inferred from
the sign of interceptSingle (which is observed to be negative). Finally,
the fact that d ⬎ c can be inferred from the fact that interceptDual ⬎
interceptSingle.
the simplest case). Substituting this expression for V in the equations for
EYi and EOi leads to the Brinley function
d ⫻ ␦E ⫹ c d ⫻ ␦E ⫹ c
EOi ⫽ ⫻ EYi ⫹ (r ⫺ )RY.
␦E ⫹ 1 ␦E ⫹ 1
Comparing this function with the function for the baseline task, the two
lines can be seen to differ in slope as well as in intercept.
The supposition that the baseline task is overhead free is not necessary
for the contrast we draw between dual tasking and coordination. Overhead
may be a part of the baseline task as well as the experimental task—it may
be an integral concomitant of virtually all information processing. Let us
reformulate the baseline model to include overhead. The most general case
is BYi ⫽ (VB ⫹ ␦B ⫻ CYi) ⫹ CYi ⫹ RY and BOi ⫽ d(VB ⫹ ␦B ⫻ CYi) ⫹ c ⫻
CYi ⫹ r ⫻ RY. This model incorporates both a constant overhead compo-
nent and a proportional overhead component. The entailed Brinley function
is
d ⫻ ␦B ⫹ c (␦B ⫹ 1)r ⫺ (d ⫻ ␦B ⫹ c) d⫺c
BOi ⫽ ⫻ BYi ⫹ RY ⫹ ⫻ V B.
␦B ⫹ 1 ␦B ⫹ 1 ␦B ⫹ 1
Taking this as our baseline model, the contrast between dual tasking and
coordination can still be drawn. We hypothesize that dual tasking alters
only one parameter when the BYi and BOi equations are rewritten for EYi
and EOi: Namely, VB is changed to VE (with the expectation that VE ⬎ VB).
This change in the Brinley function for the experimental task can be seen
to leave the baseline slope unaffected and to change the intercept. Coor-
dination, on the other hand, is hypothesized to alter only ␦B, which changes
to ␦E (with the expectation that ␦E ⬎ ␦B). This induces a slope difference
in the experimental Brinley function, as well as an intercept difference.
458 VERHAEGHEN, STEITZ, SLIWINSKI, AND CERELLA

Appendix C

Cascade Model Predictions

It is a common sentiment that discrete-stage models such as those
developed in Appendixes A and B are low in neural plausibility and that
actual brain processes have nonlinear dynamics. The cascade model of
McClelland (1979) has emerged as a paradigmatic alternative. Much of the
influence of this model is owed to McClelland’s demonstration that,
despite their nonlinear dynamics, cascade models can account for additive-
factors outcomes such as those reported in the state traces of our text. In
this appendix, we construct models of this stripe and go on to show that the
multiplicative slowing factors postulated in Appendices A and B continue
to apply, successfully accounting for the age effects in our data within a
cascade framework.
Let us approach the cascade models by reviewing the discrete-stage
models they supplant. Appendix B concluded with a model of dual tasking
built from three components, a residual component R or BASE, a computa-
tional component C or COMPUTE, and an executive component V or CON-
TROL. Changing our notation, let the component durations be given by ␶BASE ,
␶COMPUTE , and ␶CONTROL , so that the response time can be written RT ⫽ ␶BASE
⫹ ␶COMPUTE ⫹ ␶CONTROL. As developed, ␶BASE is a constant, ␶COMPUTEi increases
with task condition i, and ␶CONTROL takes one of two values, ␶CONTROLS , or
␶CONTROLD , depending on whether the target task is performed by itself or
together with a distractor task (with the expectation that ␶CONTROLS ⬍
␶CONTROLD ).
In this formulation, the control component operates in both single and
dual modalities, changing its duration to suit. In the terminology of Stern-
berg (1969), the effects of dual tasking are expressed through the selective
influence of a processing stage distinct from the stage that accommodates
the effects of task condition. Unlike the case of pure insertion, the stage
architecture does not change with the dual-tasking requirement, only the
stage durations. The force of Sternberg’s argument was its demonstration
that the postulates of selective influence are as consistent with a pattern of
additive factors, such as the state traces reported in our text, as are the
postulates of pure insertion.
McClelland (1979) accepted the three-stage serial architecture of Stern-
berg (1969) but questioned its discrete, all-or-nothing dynamics. Dynamics
can be captured at a schematic level by assuming that each stage is
characterized by a scalar activation level (0 ⫺ 1), accepts a scalar input (0
⫺ 1), and produces a scalar output (0 ⫺ 1) equal to its activation level. In
the discrete-stage configuration, a stage is inactive (0) until it receives an

Figure C1. Activation as a function of time for three processes linked in a cascade (dotted, dashed, and solid
curves). The reaction time (RT) of the cascade is defined by the time at which the third (solid) process crosses
an activation threshold of 2.5. The crossing point is marked by a symbol. The three functions on the left simulate
a young participant whose RT ⫽ 927 ms; the three functions on the right simulate an older subject whose
RT ⫽ 2,936 ms. The aged cascade was derived from the young cascade by applying slowing factors to the
process time constants.
 AGING AND DUAL-TASK PERFORMANCE 459

input (1). After ␶ time units, it goes active (1) and switches its output however, cannot be ignored; because of the nonlinear activation function,
from 0 to 1. The connection between this three-stage series and the RT they affect the mean. Given that A 僆 N(5,1) is normally distributed, then
equation written above is obvious. (A ⫻ E 3(t) ⫺ 2.5) will be normally distributed with a cumulative distri-
McClelland’s (1979) alternative was to suppose that activation levels are bution equal to ⌽(5E 3(t) ⫺ 2.5), where ⌽() is the cumulative normal
graduated rather than discrete. Driven by a unit step function, the activation distribution. In an important corrective to McClelland (1979), Ashby
level of a McClelland-styled stage rises from 0 to 1 on a negative expo- (1982) pointed out that this expression is not a true probability distribution
nential function of time, because some fluctuations will lodge E3 below ␪; on those trials, the
threshold will never be crossed and an RT will never be asserted. This
E1(t) ⫽ (1 ⫺ e⫺t/␶ ). happens on exactly 1 ⫺ ⌽(2.5) ⫽ .006 of the trials. Normalizing by this
In this formulation, ␶ is redefined as an exponential time constant, the
reciprocal of the exponential rate ␳. These dynamics imply that prior to its
completion, a stage or “computational process” generates partial informa-
tion. Given the serial connectivity, this information is communicated
immediately to its successor process as an input greater than zero but less
than one. In this fashion, an initial, step-function input generates a wave of
activation that sweeps through the series in what is called a cascade. Given
three linked processes with time constants ␶BASE , ␶COMPUTE , and ␶CONTROL ,
McClelland showed that the activation level or output of the final process
E3 is given by the general gamma function

E3(t) ⫽ 1 ⫺ 冉 ␳COMPUTE
⫻
␳CONTROL
␳COMPUTE ⫺ ␳BASE ␳CONTROL ⫺ ␳BASE
e ⫺␳ BASE ⫻t

␳BASE ␳CONTROL
⫹ ⫻ e ⫺␳ COMPUTE ⫻t
␳BASE ⫺ ␳COMPUTE ␳CONTROL ⫺ ␳COMPUTE

⫹
␳BASE
⫻
␳COMPUTE
␳BASE ⫺ ␳CONTROL ␳COMPUTE ⫺ ␳CONTROL
e ⫺␳ CONTROL ⫻t
冊 .

Like E1(t), E3(t) rises gradually from 0 to 1. Because 1 is approached as

a limit, the response time of the cascade cannot be defined by the instant
at which the final stage goes to 1, as in the discrete model. Instead, a
threshold ␪ must be introduced into the cascade model, and RT is defined
as the instant at which E3 reaches ␪ (thus, E 3[RT] ⫺ ␪ ⫽ 0). For
convenience, the cascade output is remapped from (0 ⫺ 1) to (0 ⫺ 5), and
by convention, ␪ is fixed at 2.5. Thus, RT is given by the root of the
expression (5E 3(t) ⫺ 2.5).
Activation functions are illustrated for two cascade models in Figure C1.
These are the two extreme cases from the set of 40 models constructed
below. The fastest cascade (three curves to the left on the time axis)
represents the performance of young participants in the easiest condition of
the single task; the slowest cascade (three curves to the right on the time
axis) represents the performance of older participants in the hardest con-
dition of the dual task. Each model consists of three cascaded processes,
and the activation function of each process is shown. The three curves all
rise from 0 and approach 5 as a limit. The curves are ordered, the first
leading the second and the second leading the third, as enforced by the
serial architecture. Also shown is the threshold level of 2.5, as a horizontal
line. The response time for the fast cascade is defined by the point at which
the third activation function crosses the threshold. It is marked by a circle
on the graph and fixes the RT at 927 ms. The slow cascade reaches
threshold considerably later, with an RT of 2,936 ms.
McClelland (1979) expanded his model to account for response-time
variances as well as means by wrapping a stochastic shell around this
deterministic core. Two sources of variability are introduced, which alter
the cascade from trial to trial. On each trial, a quantity of preactivation is
added to E3, as a constant sampled from N(0,1), where N(0,1) is the Figure C2. State traces (A) and Brinley plots (B) derived from a set of 40
standard normal distribution, and the asymptotic activation level of 5 is cascade models. Ten basic models simulated single-task conditions per-
perturbed by a constant sampled from N(0,1). Mathematically, the latter formed by young participants. The control process of the basic model was
quantity can be viewed equally well as a perturbation of the threshold lengthened to simulate 10 dual-task conditions. 10 ⫹ 10 additional models
value 2.5. Both noise sources average to zero across trials, so the mean were generated by lengthening the control process, the computation pro-
preactivation level remains zero, and the mean asymptote (or threshold) cess, and the base process by different amounts to simulate older partici-
remains 5 (or 2.5). pants. The fastest and the slowest of the 40 cascades are shown in Figure
Preactivation effects are expressed in the RT variability and, following C1. The regression lines in the two panels match those obtained in the
Roberts and Sternberg (1993), are ignored here. Asymptote fluctuations, meta-analysis (cf. Figure 2). RT ⫽ reaction time.

(Appendix continues)
460 VERHAEGHEN, STEITZ, SLIWINSKI, AND CERELLA
factor corrects for these lapses, and F(t) ⫽ ⌽(5E 3(t) ⫺ 2.5)/⌽(2.5) gives
the exact distribution of threshold crossing times for the stochastic model.
Finally, the mean RT is given by
⬁ estimated from the data. Working outward from ␶COMPUTE , factor values
RT ⫽ 冕 (1 ⫺ F(t))dt.
0
In what follows, RT refers to this integral. (For calculation purposes, ⬁ was
approximated by 10, which preserved three decimal places of accuracy.)
For the two cascades of Figure C1, the integral evaluates to RTs of 1,032
ms and 3,079 ms, values somewhat greater than the deterministic crossing
times.
So constituted, we now construct a set of cascade models that recreate
the pattern of results obtained in the meta-analysis. For this exercise,
cascade time constants were adjusted so as to more or less reproduce the
empirical state traces and Brinley functions. For this proof of concept, we
did not attempt to maximize the model fit. Model adjustments were made
by manual trial and error, with a goal of matching the regression param-
eters to one decimal place.
Least constrained in this construction were the 10 models set up to
generate task effects for young participants in the single-task condition. For
these models, ␶BASE was assigned a value of 1.12 s, ␶CONTROL was assigned
.045 s, and ␶COMPUTE ranged from 0.1 s to 1 s in 0.1-s increments. Collec-
tively, these values produced a series of RTs (i.e., SY) in the range 1 s to 2
s, which captures a decisive portion of the state trace. The dual-task
condition was emulated by increasing ␶CONTROL to 0.125 s; ␶COMPUTE and ␶BASE
were unaltered. Notice that all ␶CONTROL and ␶COMPUTE values fall below ␶BASE;
in all conditions, BASE remains the slowest, or rate-limiting, process. This
is the configuration identified by McClelland (1979) as producing purely
additive effects. This proved to be true of the task and modality effects in
our exercise. The dual RTs (DY) ranged from 1.1 to 2 s, and regressing SY
onto DY exposed a perfectly linear relationship (correlation of 1.0) with a
slope of 1.0 and an intercept of 0.09 s, matching the observed state trace.
We next introduced impairments in the young cascades in the form of
slowing factors applied to the time constants, in an attempt to recreate the
outcomes observed from older participant groups. We began by applying a
slowing factor of 1.6 to ␶COMPUTE , taken from the central slowing factor
were applied to ␶BASE and ␶CONTROL until the target regression parameters
were obtained. As finally disposed, ␶BASE received a factor of 1.4, and
␶CONTROL received a factor of 2.0. Notice that the magnitudes of the three
model parameters are rank ordered like the empirical values, slowingBASE ⬍
slowingCOMPUTE ⬍ slowingCONTROL although the absolute value of
slowingCONTROL (2.0) is somewhat greater than the empirical estimate (1.85).
We suspect that the difference in magnitude is due not to the cascade
framework but to the selective-influence framework. The latter ensures that
some of the common-path slowing bundled into the intercept of the SY ⫻
SO Brinley function reflects slowingCONTROL and not solely slowingBASE . This
contamination propagates through all the model parameters via their
interdependencies.
These assignments lead to a SO ⫻ DO slope of 1.0 and intercept of .2,
matching the state trace from older participants. Brinley functions assem-
bled from the cascade output also matched their empirical counterparts. SY
⫻ SO shows a slope of 1.6 and an intercept of ⫺.20 s; DY ⫻ DO shows a
slope of 1.6 and an intercept of ⫺.16 s. State traces and Brinley functions
from the cascade models are illustrated in Figure C2.
We mention two further observations of McClelland’s (1979). First, the
global cascade function—E3(t) in our example— does not determine the
order of the component processes. The symmetry of the terms of E3(t)
ensures that all permutations are equivalent. Second, any process that is not
rate limiting can be replaced by two child-processes whose time constants
sum to the parent process with no noticeable effect on the global activation
function. These two properties suggest that our account of coordination,
built from multiple, interleaved bouts of COMPUTE and CONTROL (Appendix
B), can be recreated in a cascade context.
Received December 19, 2000
Revision received October 10, 2002
Accepted October 15, 2002 䡲