Memory &. Cognition 1992.20(4),4t3-423 Working-memory mediation of adult age differences in integrative reasoning TTMOTHY A. SALTHOUSE Georgia Institute of Technolagy,Atlontq Georgia Three research methods were used to inveetigate the hypotheeized mediational influence of working memory on age-related differencee in integrative ieasoning. Reeulte from all three pro cedureswere consietentwith the-hypothesisbecause(1) statietical control ofan index ofworking memory attenuated the age differences in reasoning accuracy, (2) young adults were more accurate than older adults in a measure reflecting the preservation oflinformation during procegsing, and (3) yorrng adults performing the task with a concunent memory load exhibited i qualitative pattern of performance similar to that of older adults performing thetask without a concurr€nt memory load. Although there has recently been considerableinterest in working memory as a possible mediator of adult age differencesin cognition (for reviews, see Light, l99l; Salthouse, 1990; Stine & Wingfield, 1990), rhe amounr of evidence directly relevant to the mediational role of working memory is still quite limited. The research reported in this article was therefore designedto use three investigative procedures to examine the hypothesis that working memory functions as a mediator of the agerelated differences in at least some cognitive tasks. The existence of negative relations between age and measuresof fluid or processaspectsofcognition has been well documentedon the basisof researchdating from the early decadesof this century (see Salthouse, l99lb, for a review). Despite a considerableamount of research,no completely satisfactory explanation of these age-cognition relations is yet available. However, in the last few years there has been growing interest in the role of working memory as a proximal (i.e., assessedat the time of testing) mediator of age differences in at least some measuresof cognitive functioning. One of the reasonsfor the enthusiasmabout explanations based on working memory is that the construct of working memory seemsmore amenable to operationalization than do other relatively general constructs such as attentional capacity. That is, if one conceptualizesworking memory as involving the simultaneous storage and processing of information, then it presumably can be measuredby tasks in which there are both storageand processingrequirements.For example, the two tasks used to operationalize working memory in the current project, computation spanand reading This research was supported by NIA Grant R37 AG06826. I would like to thank J. Jackson and A. Kersten for assistance in testing subjects, and the members of the Cognitive Aging Group at Georgia Tech (R. Babcock, J. Earles, C. Hereog, A. Mattina, W. Rogen, B. Rypma, and A. Smith) for helpful comments on an earlier draft of this manuscript. Corrcspondence should be addrcssed to T. A. Salthouse, School of Psycholog, C.rcorgiakstitrte of Technologr, Adanra, CA 308324170. span,involved subjects'retentionof digit or word information while concurrently carrying out processingconcerned with solving arithmetic problems or answering sentence-comprehensionquestions. The focal cognitive task in this project was integrative reasoning involving abstract verbal materid. A typical problem is illustrated in the top panel of Figure l. Prob. lems can vary in the number of premisesdescribing relations between two terms, but dl evenrually end with a question asking what will happento one term if a specified changeis introduced in another term. An advantage of this particular task is that problem diffielty (as rcflected by decision accuracy) can be systematically altered by rnanipulation of the number of premises presented prior to the question, presumably because demands on working memory are increasedby the additional requirementsfor the storage and coordination of information associated with more premises. The tlree investigative proceduresused in this project were all designed to examine the hypothesis that agerelated differences in the integrative reasoning task are at least partially attributable to age-associatedreductions in some aspectof working memory. The statistical control procedure tests the prediction that the magnitude of the age differences should be greatly reduced if people of different ages are statistically equated with respect to their working-memory ability. Becauseworking npmory involves the preservation (storage) of information during processing, the experimental analysis procedure examines the implication that older adults should be less accurate than young adults at maintaining relevant information while engagedin the performance of the task. And finally, if a low level of working nrcmory can be consideredanalogous to a concurrent memory load, then the panern of performance exhibited by older adults should be qualitatively similar to that found in young adults who perform the task while also remembering other information. This expectation is examined in the simulation procedure. 413 Copyright 1992 PsychonomicSociety, Inc. 414 SALTHOUSE relatively narow and closely linked to particular qpes of information and specific kinds of processingoperauons. R and S do the OPPOSITE The Salthouse (l99la) studies included two distinct Q and R do the SAME working-memory measures, but the group-administration whatwill happento S? format may have created a spurious associationbetween lf Q INCREASES, the working-memory and reasoningmeasuresbecauseof factorsunrelatedto working memory (e.g., time restrictions necessitatedby group testing). An attemptwas made (1 Relevant, 1stposition) to minimize the possibility of an artifactual relation beG and H do the SAME tween the working-memory and reasoning measuresin the present study by administering the tasks by meansof F and G do the OPPOSITE what will happento H? computers. This allowed research participants to spend lf G DECREASES, as much time as desired in each phase of the workingmemory tasks and to respond at their own pace in the reasoning task. (RecognitionProbe, 1st position) If working memory is important for cognitive tasksbeD and E do the SAME cause it allows information to be preserved while other information is being processed,then one should expect E and F do the OPPOSITE *** D and E do the SAME*** to find better preservation of information during the performance of a cogritive task among people wtnse workingmemory systems are highly effective than among people Figure l. Illustration of sample problems in the integrativewhose working-memory systemsare less effective. This reasoning task. prediction can be tested by using versions of the experimental analysisprocedure to evaluatethe availability Statistical control procedures have been used in two of information presentedearlier in an ongoing cognitive earlier research projects with this same integrative- task. Information availability can be examined with a vareasoningtask. Salthouse,Mitchell, Skovronek, and Bab- riety of different methods, but the critical point is that cock (1989) administered the reasoning task and the the evidence for age differences in working memory is computation-spanworking-memory task to 120 adultsbe- derived while the subject is actively engagedin the pertween 20 and 80 years of age. Multiple-regression equa- formance of a cognitive task and not from a separatetask tions predicting reasoningaccuracyrevealedthat age was deliberately designedto assessworking memory. In the associatedwith an R' of .278 when consideredalone. but terminology of Salthouse(1990), measuresobtainedwith the R' value for age was reduced to only .l 19 after par- the experimental analysis procedure are within<ontext tialing the variance associatedwith the working-memory measuies, whereas those typically used in the statistical measure.Two additional studies,eachinvolving over 220 control procedure are out-of-context measures. Previousresearch(e.g., Salthouse,[rgg, Palmon, & adults between 20 and 80 years of age who were adMitchell, 190; Salthouseet al., 1989)with the current ministered the integrative-reasoningtask and two workingmemory tasks, were reported by Salthouse(l99la). The integrative-reasoningtask examined information availabilworking-memory tasks in these studies were group- ity in terms of the contrast between two types of trials. administered versions of the computation-span and Trials in which all premises are relevant to the decision (illustrated in the top panel of Figure l) require both memlistening-spantasks (seeSalthouse& Babcock, 1991, for further description). Resultssimilar to those of Salthouse ory and integration of information. That is, becausethe et al. (1989) were obtained in both studies.To illustrate, question concerns terms originally mentioned in differvalues of R2 associatedwith age in the prediction of ent premises, information from several premises must be reasoningaccuracy were .121 and .104 before statistical integrated and coordinated to determine the answer. In control of a compositemeasureof working memory, but contrast, no information integration is required in trials were only .015 and .036 after such control. The results in which only one premise is relevant becausethe quesof these earlier studies, therefore, suggestthat a fairly tion concerns terms mentioned in a single premise. An large proportion of the age-related differences in the example is illustrated in the middle panel of Figure 1, in integrative-reasoning task may be mediated by reductions which the question refers to two terms originally related in working memory. In terms of percentages,the vari- in the first premise. Becausetrials in which all premises anceassociatedwith agewas reducedby 57 .2%,87 .6%, are relevant require both memory and integration whereas those with a single relevant premise do not require acrossand 65.4To, respectively, across the three studies. Although the statistical control procedure can be very premise integration, it has been assumedthat compariinformative. it does have several limitations. For exam- sons of accuracy in t}te two types of trials would be inple, the fact that only a single measureof working mem- formative about the relative importance of memory and ory was used in the Salthouseet al. (1989) study raises integration processesin the performance of this particuthe possibility that the assessmentof working memory wits lar task. (All Relevant) WORKINGMEMORYAND REASONING In both earlier srudies(Salthouseet al., 1989, lg0), decision accuracydecreasedas more premiseswere presented, and the magnitude of accuracy decline for trials with one relevant premise was nearly the same as that for trials with all relevant premises. Furthermore, the same generalpatternwas evident in adultsofall ages.Theseresults were interpreted as suggesting that a critical factor in the variation in accuracy as a function of number of premises and as a function of adult age rnay be the preservation of information, and not its integration or coordination. That is, the principal determinantof the variations in reasoning accuracy associatedwith additional premises or with increasedage appearsto be the ability to maintain previously presentedinformation rather than the ability to integrate information that is available in memory. As with almost any experimental technique, however, objections can be raised with respect to how results of the manipulations are best interpreted. One reservation about the contrast between one-relevantand all-relevant trials is that even performanceon one-relevanttrials may not be a pure measure of information availability, but might also reflect the ability to transform the format of the information from the terms "opposite" and ..same" used in the premises to the terms "increase" and ..decrease" used in the question. Rather than representing the ability to maintain information without integration, therefore, accuracy of decisions in one-relevant trials could reflect the ability to maintain information and to convert it from one representationalformat to another. This concern can be addressedby attempting to obtain a more direct assessmentof information availability in the form of recognition probes for previously presented premise information. An example of a probe-recognition trial used in the current experiment is illustrated in the bottom panel of Figure l. Notice that the probe is presented in the same format as the premises, and the subject merely has to decide whether it is the same as. or different from, one of the premises presentedearlier in that trial. Supplementing the contrast of performance in one-relevant and all-relevant trials with these recognition probes should therefore allow a more definitive evaluation of the hypothesis that older adults perform pmrly on integrative-reasoning tasks becausethey are less likely to preserve relevant information than are young adults. Becauserecognition probes provide a more direct assessment of the status of information in memory, a finding of age differences in measures of probe-recognition accuracy would provide additional support for the inference that a failure to maintain critical information during processing is an important determinant of age differences in integrative-reasoning performance. The principal assumptions underlying the simulation procedure are that the effects of reduced working-mernory functioning can be mimicked by various experimental manipulations, and hence that the pattern of performance exhibited by young subjects performing the reasoning task with one of those manipulations should be qualitatively similar to the pattern exhibited by older subjects perform- 4I5 ing thetaskundernormalconditions.BaddclcyandHitch (1974;Hitch & Baddeley,1976)havepropoaedrharrequiring subjectsto maintaina concurrentmemoryload reduceseffective working-memoryfunctioning because someaspectsof working memory havcto bc devotedto preservingthetask-irrelevant memory-loadinformation. If this assumptionis corr€c.t,and if a major faclor contributing to the agedifferencesin reasoningperforrnarrce is a diminishedworking nrcmoryamongolder adultsrelative to young adulS, then the patternsof perforrnancc for older adultsshouldbe expectedto be very similar to thosefor young adults performing the task with a concurrent memory load. Bottr groups shorld have lower levels of performancethan young adults without I concurrrentnpmory l@d, ard thc panernof pcrfornrala differerrcesbetweenyoung adults and old aduls should resemblethat berweenyoungadultswith and without the concurrentmemoryload. The major limitation of the simulationapproachis that directquantitativecomparisons aregercrally rn nEaningful when there is no basisfor speci$ing the correspondencebetweena particular numberof yearsof agc differenceanda given amountof concurrentmemoryload. Qualitativecomparisonsevaluatedwith rcspoct!o a common refererrcegroup are also complicatcdbecauscimpairmentsin performancesan be producedin a variety of different ways. Confidenccthat the simulationis accuratethustendsto increasein proportionto the nnmber and patternof correspondences that can be establishcd. Fortunately,the reasoningtaskinva*igatcdin this pmject allowsserreralpossibleconmstsbetweendrepcrfmrrance of older adultsandthat of youngadultswith a corrcurrent memoryloadby comparingeachagainstthc pcrformance of youngadultswith rcspectto ( I ) ttle effectof additional premiseson decisionaccuracy,(2) trc pofiernof utrry ofone-relevanttrials relativeto all-relevantfials, (3) thc patt€rnof accuracyof orrc-relevanttrials asa fuirctionof the serialpositionof the relevantpremise,and(4) thepattern of accuracyof recognition-probetrials as a function of the serial position of the relevantpremise. To summarize,if an important factor contributing to adult age differencesin integrativereasoningis an agerelatedreductionin working npmory, thenorp wouldexpect (l) substantialattenuationofthe agedifferenccsin reasoningafter statisticalcontrol of measuresof working memory, (2) significantly lower accuracyfor older adultscomparedwith youngadultsin measuresof information availability obtafuredwhile subjecSare engaged in theperformanceof thetask,and(3) a qualitativelysimilar patternof performancefor older adultsard for young adultsperforning thetaskwith a concurrentrrcrnory led. The study describedbelow was designcdto test eachof thesepredictions. In additionto the primary focus on working npmory, there was also a secondaryintercst in the role of percepual-comparisons@ as a potential mediatorof the agedifferencesin both reasoningand working memory. Previoussodies have revealedthat statisticalcontrol of 416 SALTHOUSE an indexofperceptuals@ greatlyattenuated theagedifferencesin measuresof reasoningand working memory (Salthouse, l9la; Salthouse & Babcock,l99l). However, the samepaper-and-pencil measuresof perceptualcomparisonspeedwere usedin all of the earlier studies, andconsequentlythe generalizabilityof this result is not yet known. Two computer-administeredmeasuresof perceptual-comparison speedbasedon the Digit Symbol SubstitutionTest (Salthouse,1992)were thereforeemployedto examinethe influenceof perceptualspeedon the relationsbetweenageand performanceon both the reasoningand working-memorytasks. METHOD SubJccts Characteristics of the three samplesof adults who participated in this project are summarized in Table l. Older adults were recruited from ncwspaperadvertiscmentsand referrals from other participantsand were paid a nominal fee for their participation. Both samplesof young adults were college studentswho received extra credit in psychology courses for their participation. It is apparent in Table I that the three groups were sirnilar with respectto averageyearsof edrcation ard self-ratedtrealth,ard thus any age-related differences in the performance rneasuresare unlikely to be confounded with differences in at least these characteristics. hocedure All participants performed four tasks, in the sameorder, before the reasoningtask. The first task, digit-symbol, was a computeradministeredversion of the WAIS-R Digit Symbol SubstitutionTest (Wechsler, l98l). Displays in this task consistedof a code table associatingdigits and symbolsand a singletest stimulusconsisting of a digit-symbol pair. The code table remained constant across trials, but the identity ofthe digits and symbols in the test stirnulus varied from trial to trial. The task was to press the / (slash) key on the keyboard as rapidly as possible when the digit and symbol in the test stimulus matchedaccording to the code table and to press tlrcZkey on the keyboard when they did not match. The digit-digit task was very similar but had no symbols. The code table was therefore uninformative in this task becausethe decisions were to be made on the basis of the physical identity of the pair of digits. In both tasks,a practiceset of 18 trials was followed by an experimenal set of 90 trials. Becauseaccuracy was greater than 95% in both tasks for all three groups, the median time per responsein the experimental trials served as the dependent measure. The two working-memory tasks required subjects to remember information while also carrying out specified processing. The tobe-rememberedinformation in the reading-span task was the last word in the sentenoe,and the processingconsistcdof selectingwhich of three alternatives was the corr@t answer to a simple question about the sentence.The to-be-rememberedinformation in the computation spantask consistedofthe secod digit in an arithnreticproblem, and the processingconsistedof selecting which of three alternatives was the correct answer to the problem. Materials in both tasks were identical to thosc described in Salthouseand Babcock (l9l). Each successiveitem (scntenceor arithmaic problem) could be viewed ai long as desired, and the subjectswere allowed as much time as neededto type their recall responses.A trial consisted of the presentationof the designatednumber of senteircesor arithmetic problems followed by an anempt to recall the last items from all of the presentedsentencesor arithmetic problems in the order in which they were pres€nted. Spanswere determined by the longest sequenccin which the target items were recalled correcdy and the processing was performed without mistakes on at least two of the three trials. A practice phase with trials containing sequencesof up to three items preceded the experimental phase in each task. The reasoning task was first described, then administered in a practice block of seventrials. Two experimental blocks of63 trials each were then presented.All participants received the sametrials in the samesequence.Forty-five ofthe trials were normal reasoning trials, and l8 were recognition-probetrials. Across the two blocks there were an equal number of increase (or satne) and decrease(or diferenr) trials for eachof the five combinationsof number and type ofpremises(i.e., l-1, 2-1,2-2,3-1, and 3-3, where Tsble I Mesns gnd Standard Deviationsfor Variables in the Three Samples young Young-Memory Load ord N 3 Proportionfemales 30 .53 0 .50 SD Age Health Education 20.l 1.5 t4.2 DS time DD time CSpan RSpan Reas. acc. Probe acc l.l3 0.53 4.4 2.9 79.1 7'7.3 1.0 0.7 l.l 30 .30 SD 68.2 t.7 t4.5 0.14 l.9l 0.05 0.77 2.1 2.2 1.5 1.9 ll.l 59.9 r5.2 6.4 SD 5.'l 0.9 2.2 20.3 1.6 1 4 I. 1.5 0.6 1.3 0.36 0.27 t.7 1.6 9.2 12.2 l.l9 0.53 4.7 3.3 63.6 68.6 0.19 0.(X 1.8 1.8 l1.8 15.6 Note-Health is self-rating on a scale where I = excellent and 5 = poor. Education is self-reported years of formal education completed. DS time is time in seconds in digit-symbol task. DD time is time in seconds in digit-digit task. CSpan is the span in the computation-spantask. RSpan is the span in the reading-span!ask. Reas. acc. is percentagecorrect in the normal-reasoning trials. Probe acc. is percentagecorrect in the probe-recognition trials. WORKINGMEMORYAND REASONING the first digit refersto the total numberof premisesand the second digit to the number of relevantpremises).All types of trials, including the recognition-probetrials, were rurdomly intermixed within the experinrntal block, but all participantsreceivcdthe sanre random s€quence. Premiseswere presentedsequentiallyfor 4 seceach. The ques_ . tion remainedvisible until the subject'sresponse(Z for incriase, / [slash] for decrease),but the instructionsemphasizedthat the responsesshould be made as rapidly as was consistentwith maximum accuracy. Recognition-probetrials were identical to the reasoningtrials except that the recognition probe replaced the question requiring an increase or decreasedecision. One half of the probe trials had the samepremiseasone ofthose pres€ntedearlier in the trial, and one h4f !"d a premise that differed from an earlier prcmise by a reversal of the original relation (e.g., a premise with a sanrerelation was changed to one with an opposite relation). Decisions that the premise was identical to onc pr€sent€dearlicr were communicated by pressingthe / (slash)key, and decisionsthat the premts€was different were communicatedby pressingthe Z key. Trials for the subjectsin the concurrent-memory-load condition were preoded by the presenadon of a sct of five randomly sclected digits. The subjectsviewed the digits for 4 sec, after which thc first display of the reasoningtask was prescnted.Immediatelyfollowing_thedecisionresponseto the reasoningqucstionor recognition probe, the subjectwas askedto recall the five digits in rhe order in which they were presentedby typing them on the computerkeyboard. Performance in the reasoning tasks was only analyzed for trials.in which the subjectswere correct on the digit recall. This includedl0l trials for the averagesubject,with a raige acrosssub. jects of 4E to 125. REST.JLTS Statistical Control Analyscs Multiple-regression analyseswereconductedto examine the statisticalcontrol predictionsofattenuatedagerelationson the measureof reasoningaccuracyaftei partialing the varianceassociatedwith working memory. Becausemeasuresof informationavailability, in the form of probe-recognition accuracy,andof perceptualspeed, from thedigit-symbolanddigit-digit tasl$, wer€obtained from eachparticipant,thesemeasureswere also entered into theregrcssionanalyses.Compositenreasures of working memoryand perceptualspeedwere createdby averagingthe e scoresfrom the measuresin the computation spanand readingspantasksto createa working-memory compositeand averagingthe e scoresfrom the mqNures in the digit-symbl and digit-digit rasks to create a perceptual-speed composite.(Correlationsbetweenthe measuresin the total sampleafter partialingout agewere .38 betrveen the two working-memorymeasures and .56 betweenthetwo perceptual-speed measures. Theremaining correlations, againafterpartialingage,were -.0g betweencomputationspanand digit-symbol, -.07 be_ tweencomputation spananddigit-digit, -.16 between readingspananddigit-symbol,and -.O$ betweenreading spanand digit-digit.) Resultsof the regressionanalysei basedon the completesamplesof youngandold aOutts, andof youngadultswith andwithouta Concurrent mem_ ory load,are summarized in the top panelof Table2.t 417 Entriesin the first columnof Table2 referto the cumulativeR2afterthevariablein thatrow andtheimnrcdiately precedingrows had beenenteredinto the regression equation.Valuesin the secondcolumnindicatethe incrementin Rr whenthc variablein that row is added to the regressionequation.The valuesin thethird column are F ratiosevaluatingthe statisticalsignificanccof the initial R'or theincrementin Rr associated with theadded variable.Finally, the entriesin the fourth columnindicatethe perc€ntage by which thegroupdifferenccswere attenuated by satisticalconrol of theprecedingvariables. For example,thediffererrcebetweenth€ initial R! for age of .479andthe incrementin Rr for ageafter controlling the influenceof workingmemoryof .248is .23I , whicfi correspondsto an att€nuationof 4t.2%. Becausesomesubjectshad an averageaccuracyvery close to chance,the analyseswere repeatedaftcr omitting subjectswith lessthan80%correctresponscs on the reasoningtrials with only onepremise. This is an arbitrary criterion, but it doesensure.that the levelof understanding and degreeof motivationwas sufficientlyhigh thar the task could be performedwith moderateaccuracyin the simplestcondition.Thesedata, bascdon 26 young adults, 16older adults,and 12 youngadultswith a concurrentmemoryload, are displayedin the bottompanel of Table2. Three points should be noted about the resultsof the analysesinvolvingthe groupsof youngadultswithouta concurrentmemoryloadandold adults.The first is that therclativeannuntof afienuation (i.e.,48.2% in thcconplet€sarnpleand32.4% in the restrictedsample)of the age-associated varianceafter control of working memory is similar to that foundin the previousstudiesusingthe statisticalcontrolprocedure(e.g., Salthous€ et al., 1989; Salthouse, l99la). The secondpoint is that the reduction in age-related varianceassociatedwith control of the probe-accuracy rneasureis approximatelythe sanreasthatassociated with theworking-memory (i.e., 49.3%vs. 48.2% in measure the completesample)and that the additional reduction when both measuresare controlled simultaneouslyis rathersmall(i.e., to 61.8%).This suggcsts that,ar least with respectto theage-related differerrceson reasoningaccuracy, the two variablesare not in&pendentbut instesd have a certain amountof straredor common variance. Finally, it can be seenthat statistical control of the perceptual-speed measure,either by itself or in combination with other variables,resultedin an attenuationof the age-relatedvarianceby 80% togO%. The apparent implication is that a large proportion of the age-related differences in this reasoningtask is mediatedby agerelatedreductionsin speedofprocessing,as indexedby the perceptual-speed measures. Resultsfrom the contrastsof young adults with and without a concurrentmemory load are presentedin Table 2 primarily for purposesof comparisonwith the age contrasts. Notice that unlike the differences between 4I8 SALTHOUSE Table 2 Results of Hierarchical RegressionAnalyses on Average ReasoningAccuracy Young vs. Old Young vs. Young-Memory Percent Attenuated Increment in R N Load Percent Attenuated Increment in R' All Subjects 2'l.4* Group .479 53.28* .321 Working memory Group .283 .531 34.41* 30.12* 48.2 .009 -346 551 0.77 29.41* Probe accuracy Group .396 .639 .243 62.50* 38.2't* 49.3 .461 .617 156 68.56* 23.191 51.4 Working memory Probe accuracy Group .283 .463 .&6 .180 .1 8 3 44.74* 28.35* 28.92* 61.8 .009 .461 .622 .l6l I.31 67.01* 23.87* 49.8 Perceptual speed Group .46 .547 .081 58.55r 1 0 . 2*1 8 3 .I .100 .370 .210 9.08* 24.38* 15.9 Percepoal speed Probe accuracy Group .46 .605 .67 .1 3 9 .062 78.22* 23.36* 10.43* 87.1 .100 .480 .626 .380 .t46 15.03* 56.77* 21.86* 54.5 Percepoal speed Working memory Group .46 .529 .582 88.9 .100 .r 0 l .382 .001 .281 9.09* 0.06 25.q* t2.5 Perceptual speed Working memory kobe accuracy Group .M .529 .620 .67| 89.4 .100 .l0l .480 .629 .001 .379 .t49 t4.87* 0.10 56.l3* 22.03* 53.6 Group .426 29.72* .t47 Working memory Group .tu .r22 .472 .288 13.62* 2t.261 12.4 .2' to .148 6.03 7. 2 8 Probe accuracy Group . t66 . 5l 7 .351 I3.38* 28.451 17.6 .312 .48 .136 20.33* 8.88* '1 .5 Working memory Probe accuracy Group .184 .254 .531 .070 .277 14.94* 5.67 22.44* 35.0 .122 .34 .482 .222 .138 8.25* 14.94* 9.31* 6.1 Perceptual speed Group .451 .490 .039 34.52* 3.03 .o94 36.1 Perceptual speed Probe Accuracy Croup .451 .525 -5& .o'74 .039 36.49t 4.36 2.M .2n .rt'l 5.33 4.30 90.8 .361 .M .244 .105 7.6* 16.02* 6.86 28.6 94.r .l17 .2N .305 .083 .105 5.89 4.t9 5.28 93.0 .tl7 .2ffi .3U .494 .063 .053 .063 .091 .051 62.3t4 8.43* 7.10 '17.93* 10.54* 15.28* 8.64* .452 -5.0 Subjects Satisfying Accuracy Criterion Percepoal speed Working memory Group .451 .505 .530 Perceptual speed Working memory Probe accuracy Group .451 .505 .547 .5't7 .054 .425 .054 .u2 .030 39.32* 6.45 3.43 39.46* 4.'t2 3 . 7| 2.62 6.35 .lt7 90.8 .083 .184 .l l0 7.95* 5.59 12.33r 7.16 -0.l I I 25.2 *P < .ol. young and old adults, the performance differences associated with the memory-load condition were not substantially attenuatedby statisticalcontrol of the workingmemory or perceptual-speedvariables. This is to be expected becausethe two young adult groups were very similar in these out-of-context measures(cf. Table l). Analyses of Information Availability Accuracy in the reasoningtask as a function of the number of presentedpremises is displayed in Figure 2. The top panelcontainsthe datafrom all subjects,andthe bottom panelpresentsresultsfrom the subsetof subjectsin eachgroupwith accuracyof at least80% on trials with only one premise. Notice that accuracydecreaseswith morepremisesandthat the absolutelevel of performance is very similar for trials with one relevantpremiseand for trials in which all premiseswere relevant.The analysisof variance(ANOVA) results,summariz€din Table 3, confirm that the main effects of group and number of premiseswere siSnificantbut that the relevanceeffect WORKINGMEMORYAND REASONING eral waysin whichaverageaccuracycouldhavedecreas€d with the pres€ntation of additionalpremises.For example, accuracymight haveremainedinvariantfor either the first or the last premisein the trial but declinedon otherpremises,or it mighthavedeclinedneadyuniformly acrossall premisepositions.The results,displayedin Figure 3, indicatethat the secondcharacterization is more accurate.That is, for all threegroups,accuracywaslower with more premisesregardlessof the serial positionof the relevantpremise.Furthermore,youngadultswithout a concurrentmemoryload were alwaysmore accurate nt thanweresubjectsin the othertwo groups.Resultsof the o ANOVAs with agegoung, old) andpremiscgpe (i.e., o l-1,2-1,2-2,3-1,3-2,arf,3-3,wherethefirstandsec0.4 ond digitsrefer to numberof premisesand positionof Young Old Young- Mom. Load relevantpremise,respectively)'asfactorsaresummarized c A lRl e l 6 / . + + + o in Table4. Of particularinterestis theabsence l Rslev. ----;--of anage x premisetype inrcractionin both the analysisbasedon all of the dataand in the analysisbasedon the datafrom n the restrictedsample.Virtually identicalrcsultswereevi0.9 dentin theanalysescontrastingthetwo youngadultgroups distinguished by the presence or absence of a concurrent 0.8 memoryload. Figure 3 also contains probe-recognition accuracy o.7 ploaed as a function of serial position of the probcd prcmise.Thegeneralpatterncloeelyrcsemblesthatof the one-relevantreasoningtrials exceptwhen the trials containednvo premises,wherethe older adultsandttrcyoung adults with a concunent memory load were somewhat more accuratein thc recogrition trials than in thc reasoning trials. Results of the age grorrp x premise type ANOVAs similar to thosecondrrctedon the accurrcy of ' t 2 3 one-relevantreasoningtrials are summarizedin Table 5. Numberof Premises The major point to be emphasizedfrom thesc&u is that the agediffererce in recognitionaccuracywaseliminated by restricting the sampleto subjectswith accuracyof at Figure 2. Accuracy in trials with all premisesrelevlnt to the do' least 80% in the reasoningtrials with a single premise. cision or with only onc premisc rrlevant to the decision es a function of the number of prcsented premises. The top penel prcsents To verify that the elimination of the agedifferenceswas data from all subjects, and the bottom poml showsdala from sub' not simply attributableto the reducedpower associated jects with at least t0% accuracy on trids with a single prcmise. with a smallersamplesize,an analysisof covariancewith accuracyin one-premisereasoningtrials asthe covariate wasalso conductcd.The agegroup main effect wasalso (i.e., one relevant vs. all relevant) was not significant. not significantin this analysis[F(1,57) = 0.85, MS, = primary patterns The difference betweenthe in the top 131.061,thusconfirmingthe resultsfrom the restrictcd andbonompanelsofFigure 2 is that in the bottompanel sampleanalysis. the differencesbetweenthe youngandold groupsappear to increaseasadditionalpremisesarc presented.This pat- Qualitative Sinituity of Age end tern wasverified in the ANOVA asthe interactionof age Concurrent-Memory-Loed Contrests group X numberof presentedpremiseswas significant Inspectionof Figures2 and 3 and Tables3, 4, and 5 in the data from the restrictedsample.With this single reveals that the performanceof all three groups was exception,the resultsfrom the analysesinvolvingthecon- qualitatively very similar. This is evidentin the declines trastbetweenyung adultswith ad witlrut theconcurrcnt- in accuracywith additionalp,remises, in the neadyequivamemory-loadrequirementweresimilar, particularly with lent levels of accuracyfor one-relevantand all-relevant respectto t}teabsence ofinteractionsinvolvingthegroup trials, andin the serial-positionfunctionsfor one-rclevant factor. reasoningtrials and for probe-recognitiontrials. StatistiAccuracywas also analyzedon one-relevanttrials as cal supportfor this similarity is evidentin the generalab. a functionof the serialpositionof the relevantpremise. senceof inleractionsinvolving the group variable in the Theseanalvseswere of interestbecausethere were sev- ANOVA resultssulrunarizedin Tables3, 4, and 5. L I I 419 ,l 420 SALTHOUSE Young Young - Mem. Load otd 0.9 o.7 o o t o 0.5 0.4 () Reasoning Recognition c 9 o E I (L r o.s 0.8 o.5 0.4 Premise Position Figure 3. Accuracy in reasoning trials with one relevant premise and in recognition-probe trials asa frmction of the serial pmition of the rehvant or probed premise.The top panel prcsents data from all subjects, and the bottom panel shows data from subjects with rt le{st ffi% sccuracy on trisls with a single prcmisc. DISCUSSION The results of this study suggestthat there are at leasl two factors involved in the age differences in integrativereasoning performance. One factor is fairly broad, in that it affects performance in all conditions of the task and may reflect lack of understanding, low motivation, or some type of sensory or motor deficit. It is assumedthat the influence of this factor can be minimized by eliminating subjects with low levels of performance in the simplest trials involving only one premise. The magnitude of the age differences when the sample was restricted in this manner was somewhat smaller than that in the complete sample, but there were only two caseswhere the pattern ofresults appeareddifferent. One difference occurred in the analysessunmarized in Table 3, where the interaction Table 3 Results of Analyses of Variance on ReasoningAccuracy Young vs. Old Young vs. Young-Memory Load df MS" dl MS, All Subjects 47.50* 658.24 Group I,58 Numberof premises Group x number 2,rt6 " 69.87* 2.63 205.45 I ,58 2,1t6 " Relevance(onevs. all) Group x relevance 1,58 ' 0.60 |.ffi 1 3 5l.9 r,s8 " Number x relevance Group x numb€r x relevance 2,t16 " 0.72 0.50 rw.62 2,116 " Group SubjectsSatisfyingAccuracyCriterion I,N 28.29* 363.41 t40.00* 9.69* 125.33 r,37 2,74 " |,N " r.74 1.83 133.0'1 t,3'7 ' 2,80 0.65 97.Q 2,'74 " Numberof premises Group x number 2,80 " Relevance(one vs. all) Group x relevance Number x relevance Group x numb€r X relevance *p < .ol. ' 0.u 2'7.68* &.14* 0.37 2.93 1.43 1.43 0.36 8| 5.04 t84.98 6.16 107.33* 3.5r 0.70 o.g 2.U 3.20 422.86 133.41 t22.94 125.88 I 14.90 104.59 WORKINGMEMORY AND REASONING 421 Teble4 !fg!Sg!4g!ry5 "lrgryrygg! leasoglg 1gllgy withone.Retevanr hemise _ _ I."q!q vt- ry __ df --!---g-L v.qC '1.-v9rl1C-l49mory Load df F MS, All Subjects Group l,58 33.93* 765.62 1,58 2 9 . 8 |t 9,61.98 Premisetype Group x premisetype 5,290 13.851 1.23 3t4.24 5,290 t4.52' 0.46 357.24 Group 1,ltt) 16.39* 590.68 t.3'l Premisetype croup x premisetype 5,200 t9.20* t.22 292.80 5,185 Subjects Satisfying Accuracy Criterion 'P < .ol. - 8.86i 635.42 15.371 0.92 334.63 Teble 5 rlllnancc on kobe.Recognirion Accurry Io{$fl41dry= _ _Yqlg r. qd df F Group I,58 Premisetype Group x premisetype 5,290 Group l,,to 5,2m Mi, All Subjects 9.40* 1139.93 22-t2* 0.66 339.29 L-flg "'. Iguls __4L_ - J__ {ggg.y l4.a MS, 1,58 4.77 t422.O7 5,290 22.t6r t.1l 357.45 Subjects Satis[ing Accuracy Criterion Premisetype Group x premiserype 18.62' o.37 985.06 |.37 288.88 5,185 0.A2 t7:15i l.l4 t214.24 31 0 . 5 3 tP < .ol. of age x numberof premiseswas significantin the restrictedsamplebut not in the completesample.This pat_ tern, illustratedin the bonompanelof Figure2, is ionsistentwith earlier findingsof larger agedifferenceswith a greaternumberof premises(Salthouseet al., 1999, 1990).Inclusionof datafrom subjectswho performedat or nearchancelevelsin the taskapparentlyobscuredthis interactionin the analysesbasedon data from the complete sample. The seconddifference in the results of the complete sampleand the restrictedsamplewas in the analysisof probe-recognitionaccuracy,summarizedin Tabli 5. In this case-,the agedifferenceswere no longer significant when subjectsperforming poorly in the simplestreasoning problemswereeliminatedfrom theanalyses.This pat_ tern, which is supportedby the absenceof significantige differencesin the analysisof covariance,suggeststhit, at leastwhen there is someassurancethat evlryone un_ denands thetaskandis apparentlymotivatedto perform, young and old adults are equivalent in the aUitity to preserveuntransformedinformation during processing. A similar result in a cube-comparison task was reportA by Salthouse ard Skovronek(19y2\,whereit wasconcluded that agedifferencesin working memoryare pronounced only whenthe stimulusinformationhasto be manipulated or transformedin somefashion.This interpretation may also apply in the presentstudy becausetransformation (from same or opposite to increase or decrease) is required with the reasoningquestions,but not with the recognition probes. The tendency for young adults to have nearly the same accuracy in the one-relevant reasoning trials and the probe-recognitiontrials (i.e., 79.4% vs. 77.97o) but for older adults to be less accurate in the reasoningtrials (i.e., 66.7Vovs. 73.6Vo)is also consistent with this interpretation.3 No single set of results is definitive, but the combined results from three quite different procedures appear fairly convincing in suggesting that working memory is a major factor contributing to adult age differences in integrative reasoning. That is, working memory soemsto be implicated becauseofthe substantial attenuation ofthe age differences after statistical control of an index of working memory, the presence of age differences in at least some measures of information availability obtained during the performance of the task, and the qualitatively similar pattern ofperformance differences between young adults and old adults and between young adults performing under normal and under concurrent-memory-load conditions. Each of these procedures has limitationi, but becausethey are not the same limitations, confidence in one's inferencesis enhancedwhen the results from each procedure converge on the same interpretation. It therefore seems reasonableto conclude that one ofthe causesofadult age differences in certain cognitive tasks is a limited ability 422 SALTHOUSE to preserve information during processing, which can be viewed as a consequenceof an impairment in working memory. The fact that the premises in the reasoning task were presentedsequentially, and for a duration ofonly 4 sec, raises the possibility that there was a greater influence of working-memory or perceptual-spe€dfactors in this task than in more traditional reasoning tasks. This interpretation is plausible, but it should be noted that similar patterns ofage differences, and attenuationsofthose differences after statistical control of an index of working memory, were observedin previous studiesin which the premiseseither were presentedsequentially under selfpacedconditions (Salthouseet al., 1989)or were all presented simultaneously in a paper-and-pencil format (Salthouse, l99la). It therefore seemsunlikely that the results of the current study have limited generalizability because of the specific method used to present the stimulus materials. If it is the case that age-relatedreductions in working memory play an important role in the age differences in this, and perhapsother, cognitive tasks, the question immediately arisesas to what is responsiblefor age-related differences in working memory. Results from other research suggestthat the speedofexecuting relatively simple operations probably contributes to the age differences in working memory. As an example, Salthouseand Babcock (l9l) found that a large proportion of the age differences in two measuresof working memory were reduced after statistical control of measuresof perceptual speed, and these findings were replicated and extended in the studiesreported by Salthouse(l99la). The present study was not primarily designed to investigate this issue, but some of the results are obviously relevant. For example, an analysis conducted with the compositeworking-memory measureas the criterion revealed that the attenuationofthe age-relatedeffects on that measure '15%, was nearly from an R2 of .251 for age when considered alone to a value of .063 after control of perceptual speed. Furthermore, the results summarized in Table 2 reveal that statistical control ofthe perceptual-speed measures reduced the age differences in reasoning accuracy by an amount larger than that produced by control of working memory and that there was little additional attenuation of the age differences by also controlling working memory. In view ofthe apparent importance ofperceptual speed in mediating the relations between age and performance in both reasoningand working-memory tasks, it is desirable to consider exactly what is measuredby the tests of perceptual speed. Of particular concern is the possibility that becausethe digit-symbol substiotion test involves nine digit-symbol pairs, it might representmemory factors as much as perceptual-speedfactors. The issue of what is responsible for age-related differences in the digit-symbol test was recently investigatedby Salthouse (19y2).The following results, either basedon original data or cited from previously published studies reviewed in that article, led to the conclusion that memory factors were relatively unimportant determinants of the age-related differencesin digit-symbol performance:(l) age differences are still evident when all participants have learned the digit-symbol associationsto a criterion ofperfect recall; (2) agedifferences either remain constant or increasewith additional opporrunities to learn the digit-symbol pairs; (3) young and old adultshave similar serial-positionfunctions when responsetimes to individual items are arnlyznd according to the position of the digit-symbol pair in the code table, suggestingthat the code table was searched in the same manner by both age groups; (a) the age differencesremain constant in relative terms as the number of digit-symbol pairs, and hence the presumed memory demands, is varied; (5) young and old adults devote the same proportion of their responsetime to merely copying the symbols; and (6) adding working memory to the prediction equation after perceptual speedresulted in little further attenuationof the age differencesin digit-symbol performance. Results supporting an interpretation that the age differences were largely determined by the speed at which elementary operations could be executed were (l) high correlations betweendigit-symbol performance and performance on other measuresof perceptual speed and (2) little unique age-relatedvariance in digit-symbol performance after statistical control of variance in otler measuresof perceptual speed. The casefor the current composite measureof perceptual speedas a reflection of speedfactors more than memory factors is even stronger than that for the digit-symbol measure alone because of the inclusion of the digit-digit measurein the perceptual-speedcomposite. It is difficult to imagine how memory factors could have contributed to performance on the digit-digit task in which sameldifferenrdecisionswere made about pairs of simultaneously presented digits. Of course, other influences may be operating, but it seemsreasonableto suggeston the basis of the preceding argumentsthat the perceptualspeedcomposite used in this study probably does reflect some fairly basic aspectof the speedat which certain kinds of information can be processed. One possible interpretation of the relation between speed and working memory is that working memory has a dynamic quality, perhapssomewhatanalogousto someone trying to juggle several objects simultaneously. That is, just as the number of items that can be successfullyjuggled depends on the rate at which they can be caught and tossed, so might the limits on the number of distinct ideas that can be kept active (or mentally juggled) in working memory be set by the rate at which information can be processed. From this perspective, therefore, working memory might be interpreted as the set of items currently active in consciousness,and age differences in working memory might hypothesized to originate because increased age is associatedwith a reduction either in the ability to activate new information or in the ability to rnaintain the activation of old information. The hypothesisthat age-relateddifferences in working memory might be mediated by reductions in the speed of executing relevant operationsobviously needsto be con- WORKINGMEMORYAND REASONING firmed with additionalevidencefrom convergingproce_ dures.The causeof age differencesin procelsing-speed must also eventuallybe explained.Nevirtheless, results from this and other recent studiesseemto suggesta plau_ sible, and testable, interpretation of the caris"esof adult age.differences in cognition that merits further investi_ gatlon. R.EFERENCES BroosLpy, A. D., a Hrrcs, G. 1..(19.74). Working memory. In c B::"II_H Thep sychotosy of learni ng and mit ii,ai ion qv ot. l, B, pp. 647-fi67), New york: Academic press. Hncx, G. J., c Brooruy, A D. (lg6). Verbal reasoning and working memory. Quanerly Joumal of Expenmenut psycholosyi 2t, ffi342i. _ Lrcxr, L. L. (1991). Memory and aging: foui typoit"."s in search of data. Annual Review of psychotosi, 42, 33i_3j6. ^ sALTHousE,T. A. ( 1990).Working memory as a processing resource in cognitive aging. Developnenral Review, 10, l0l-124. _ S^LrHousE, T. A. ( l99la). Mediarionof adulr age differences in cog_ nition by reductions in working rnemory and spefo ofprocessing. ps!chological Science, 2, l1g-193. Srlrnousr, T. A. (l99lb). Theoreticalperspectives on cognttrve ag_ rng. Hillsdale, NJ: Erlbaum. Sr_r-rnousr,T. A. (1992). W!3t do adult age differences in rhe Digit Symbol Substitution Test reflect? Joumal of Gerontology: psyciological Sciences,47. l2l-129. SelrHousr, T. A., a Bracocx, R. L. (1991). Decomposing adult age differences in working memory. Developmental isychotogy, 27, 763-776. S r r - r x o u s s , T . A . , L p c , c ,S . , p r l r r l o x , R . , e M r r c x e r _ 1 , D. R. D. (1990). Memory factors in age-relateddifferences in simple reasoning. Psychologyand Aging,5,9-15. 423 SeLrsouse, T. A., MtrcHELr-,D. R. D., SxovRoNEx, 8., a Brx__ax, R. L. ( 1989). Effects of adult age and working memory on reason_ ing and spatial abilities. Journalif Experimentil psychology: Izarn_ ing, Memory, & Cognition, lS, 507-516. ,|gg2). Ser-rHousr. T. A., a Srovnoxrx, E. Within_conrexrassessment of age differences in working memory. Joumal of Gerowology: Psychologica!Sciences,42. il0_120. Srxr, E. A. L., a WrNcnpm, A. (1990).How much do working nEm_ ory deficits contribute to age differences in discoursc memory? Eu_ tournol of Cognitive psychology. 2, 2gg_3i0y'.. __ropean WECHsLER,D. (1981). The lVectsler Aduli hrelligence katc_Revised. New York: Psychological Corporation. NOTES l. Interactions involving the group variable with each predictor vari, able were also examined to determine whcther the innuclce of the predictor varied across groups. None of these inre.".tlons *a" significant ar the designated(.01) significancelevcl. 2. thar there are six possiblc premisc typcs becausca single rel.Note evant prcmls€ can appear in the first position when on€, two, or thr@ prcmises are presentcd, in the socond position when two or three pr€m$€s are presented, and in the third position only when ihree premrscs are presented. 3: x trial type (one-relevanr reasoning, recognrrron tn^l-g:_qloup probe) ANovA on the data from the restricted samples'revealed thar the age group.x trial type interaction failed to reachthe cnt€non significance level [F(l,zlo) : 4.08, MS. = E5.78, p = .051. (Manuscript received May 2g, l99l; revision accepted for publication lanuary 7, lWl.)