The ties that bind what is known to the recall ofwhat is new

Transcription

1 Psychoomic Bulleti & Review 2000, 7 (4),604-6/7 The ties that bid what is ow to the recall ofwhat is ew DOUGLAS L. NELSONad NANZHANG Uiversity ofsouth Florida, Tampa, Florida Cued recall success varies with what people ow ad with what they do durig a episode. This paper focuses o prior owledge ad disetagles the relative effects of 10 features of words ad their relatioships o cued recall. Results are reported for correlatioal ad multiple regressio aalyses of data obtaied from free associatio orms ad from 29 experimets. The 10 features were oly wealy correlated with each other i the orms ad, with otable exceptios, i the experimets. The regressio aalysis idicated that forward cue-to-target stregth explaied the most variace, followed by bacward target-to-cue stregth. Target coectivity ad set size explaied the ext most variace, alog with mediatedcue-to-targetstregth. Fially, frequecy, cocreteess, sharedassociatestregth, ad cue set size also cotributed sigificatly to recall. Tae together, idices of prior word owledge explai 49OJ6 of the recall variace. Theoretically drive equatios that use free associatio to predict cued recall were also evaluated. Each equatio was desiged to codese multiple idices of word itercoectivity ito a sigle predictor. Few memory researchers would argue with the propositio that recall ivolves a iteractio betwee what is ow ad what is ew. Early wor o free recall (Deese, 1965; Jeis & Russell, 1952) ad cued recall (Bahric, 1970) showed that memory is facilitated i each ofthese tass by preexistig coectios amog the words. More recet wor idicates that such coectios ca facilitate implicit as well as explicit recall (e.g., Nelso, Schreiber, & Holley, 1992), ca ihibit as well as facilitate performace (M. C. Aderso & Spellma, 1995), ad ca produce false memories (McEvoy, Nelso, & Komatsu, 1999; Roediger & McDermott, 1993). Studies cocerig the iteractio betwee ow ad ew iformatio ofte ivolve a two-step research process. Free associatio data are used as a yardstic for assessig what is ow, ad recet study i a experimetally defied episode is used to determie what is ew. Whe preexistig owledge exerts a ifluece durig a curret episode, researchers draw coclusios about the role ofprior experiece i memory. Of course, ow iformatio comes i may forms, ad oe purpose ofthis paper is to separate the relative effects ofa umber ofdifferet word features ow to affect cued recall. The secod purpose is to develop ad evaluate measuremet algorithms desiged to reduce multiple word-to-word coectios to sigle-value predictors of cued recall. This research was supported by Grat MHl6360 from the Natioal Istitute of Metal Health to D.L.N. The authors tha Taao Komatsu for her help i orgaizig the data ad Cathy McEvoy for her commets o a earlier draft. Correspodece cocerig this article should be addressed to D. L. Nelso, Departmet of Psychology, Uiversity of South Florida, Tampa, FL ( elso@lua.cas.usf.edu). I oe exemplar of the two-step approach, subjects study idividually preseted words or targets ad are the preseted with other words as retrieval cues for promptig target recall (see Nelso, McKiey, Gee, & Jaczura, 1998, ad Nelso, Schreiber, & McEvoy, 1992, for reviews). The test cues are called extralist cues because they are uavailable durig the study trial ad caot be directly ecoded alog with their to-be-recalled targets. This feature is importat because subjects are forced to rely o preexistig coectios durig retrieval, maig it a ideal tas for ivestigatig the iteractio betwee ow ad ew iformatio. The mutable ature of this tas represets aother useful feature. I additio to maipulatig the features ofthe studied targets ad cues, the ecodig ad retrieval coditios ca be varied idepedetly. Durig either or both phases, subjects ca be implicitly or explicitly orieted, they ca be focused o differet levels ofstimulus iformatio, cotext ca be varied, ad i short, a host of experimetal variatios ow to affect cued recall ca be varied. Give that the study trial ormally taes about 1 mi ad the test trial aother 2 4 mi, the extralist cuig tas is ai to a sematic primig tas with logerdelays ad a strogerepisodic compoet. Results obtaied i the extralist cuig paradigm idicate that preexistig features cotribute substatially to cued recall by ifluecig what is activated about to-beremembered target words durig study ad by ifluecig the effectiveess ofcues preseted as retrieval aids durig testig (Nelso et al., 1998; Nelso & Schreiber, 1992; Nelso, Schreiber, & McEvoy, 1992; Nelso & Xu, 1995). Recall is more liely whe targets have fewer preexistig associates, whe more ofthese associates are coected to each other ad bac to their targets, ad whe they are high i cocreteess ad low i prited frequecy. Test Copyright 2000 Psychoomic Society, Ic. 604

2 PREDICTING CUED RECALL 605 cues are more effective whe they are coected to fewer associates, oe of which is the target. Recall is also more liely whe test cues are more strogly related to their targets (the forward coectio), whe targets are more strogly related to their cues (the bacward coectio), whe cues ad targets share mediated coectios (e.g., apple-tree-orage), ad eve whe they share associates (e.g., both apple ad orage produce fruit as a associate). Such results are robust ad occur uder a variety ofecodig ad retrieval coditios, but what is more importat, they idicate that what subjects have leared about words outside the episode cotributes substatially to recall withi the episode. This wor is similar to other wor o the role of preexistig iformatio i that it relies o assessig prior owledge-i this case, lexical owledge-ad the attempts to maipulate that owledge i the cotext of a cotrolled episodic experiece. The fact that word attributes, such as umberofassociates (set size), ifluece recall validates the free associatio yardstic as a useful idex ofwhat people ow whe they have bee exposed to similar cultures. If the yardstic was ivalid or ureliable, there would be o reaso to expect that the features idexed by this method would have systematic effects (Nelso, McEvoy, & Deis, 2000). For example, iffree associatio could ot be used to estimate reliably how may strog associates are lied to a give word ad if such associative iformatio was ot geerally shared, this attribute should fail to have cosistet effects o recall. The umber ofassociates lied to a give word ad other word features idexed by free associatio couldbe regarded as just aother collectio of radom variables that cotribute to error. Despite the apparet success of this wor o prior owledge, there is a importat drawbac. Such wor relies o idexig features i free associatio i order to predict memory performace i recall ad recogitio. Although the features are maipulated uder carefully cotrolled coditios, they may ot be idepedet. As far as coclusios about preexistig represetatios are cocered, the research is as correlatioal as it is maipulatioal, ad it is legitimate to as whether the ostesible effects produced by ay give feature are direct or are produced through a associatio with a correlated attribute. This problem represets the classic dilemma of worig with materials variables, ad researchers ca solve the problem i oe ofthree ways. First, the potetial role of prior owledge ca be treated as radom oise ad the igored i research o the effects ofvariables that ca be directly maipulated durig the experimetal episode. Oe problem with this approach is that evolvig experimetal paradigms ad theories essetially treat subjects as bla tablets ad igore the potetial ifluece ofprior iformatio altogether. Secod, prior owledge ca be istilled uder carefully cotrolled coditios, ad its ifluece o performace ca be directly studied. This approach represets the strogest traditio i experimetal sciece, but it raises legitimate questios about ecological validity. Fially, i the measure-the-maipulate traditio, prior owledge ca be assessed i oe tas, ad this iformatio ca be used to predict ad costrai performace i aother. All ofour wor o cued recall has bee ofthis type, but as was oted, such wor suffers from the possibility ofcorrelatedattributes that evetually must be disetagled. This paper has two specific aims. The first ad mai aim is to address the problem of correlated features directly (see Rubi & Friedly, 1986). We examied correlatios betwee features idexed by free associatio i a ormative database ivolvig 58,000+word pairs (Nelso, McEvoy, & Schreiber, 1999) ad applied both correlatio ad multiple regressio procedures to a experimetal database for extralist cued recall ivolvig 2,000+ pairs ad approximately 1,000 subjects. The ratioale drivig these aalyses assumed that low correlatios amog word features i each database would idicate that they ca be regarded as fuctioally idepedet. Alteratively, the presece of atural correlatios i the ormative database would suggest depedece amog the features ad would forecast a potetial for cofoudigs i the experimetal database. Natural relatioships amog the features ofwords ca be reduced or elimiated durig the list-buildig process by holdig these features costat across coditios. For example, whe iterest is focused o ivestigatig the ifluece of forward stregth betwee the test cue ad the target, pairs are ot selected at radom, but with a view toward represetig this iformatio as a variable. Other features ow to affect recall are cotrolled (e.g., word frequecy). Attempts to maipulate some features while cotrollig others is a complex process i which may costraits are applied, ad although such costraits may elimiate some cofoudigs, they may itroduce others as a result of the questioable assumptio that what is ot cotrolled cacels out as radom oise. The use of multiple lists i ay sigle experimet should al1eviate such cofoudigs, but this procedure provides o guaratees, particularly agaist atural relatioships. What is more, i log-termresearchprojects, features left to vary radomly early i the research oly come to be cotrolled later i the research (e.g., Nelso, Schreiber, & Xu, 1999). Hece, we thought it would be importat to apply multiple regressio procedures to the results of experimets o extralist cued recall coducted uder similar coditios i the last 10years i our laboratory. Such procedures ca be used to determie whether a particular feature had a effect o probability of cued recall that is above ad beyod ay effects producedby its correlatio with other features. The secod aim ofthis paper was to evaluate measuremet algorithms desiged to reduce multiple idices of word-to-word coectio stregth to sigle predictors. Research o cued recall idicates that may differet types of preexistig coectios bid words together to ifluece success. Some ofthese coectios cocer the associative orgaizatio of related words activated by a studied target word, ad others cocer the various types ofcoectios that li this target with its test cue. Rather

3 606 NELSON AND ZHANG Set size Feature Frequecy Cocreteess Coectivity Resoace Forward stregth Bacward stregth Shared associate stregth Mediated stregth Table 1 Defiitios of Features Defiitio Number of associates of a word produced by two or more subjects i a discrete free associatio tas. Prited frequecy of occurrece from Kucera ad Fracis (1967). Ratigs of word cocreteess o a 1-7 scale. Mea umber of coectios amog the associates of a target. Probability that a associate of the target produces it i free associatio. Probability that a test cue produces its target i free associatio. Probability that a target produces its test cue i free associatio. Stregth of a test cue ad target whe coected to a third word. Calculated by multiplyig the probability that the test cue produces Word X by the probability that the target produces Word X ad summig the results whe more tha a sigle shared associate is ivolved. Stregth of a test cue ad target whe coected through a third word. Calculated by multiplyig the probability that the test cue produces Word X by the probability that Word X produces the target ad summig the results whe more tha a sigle mediator is ivolved. tha simply beig a fuctio of activatig a sigle li betwee cue ad target, recall is best described as the meshigofthe associative etwors activatedby eachof these etities (Nelso et ai., 1998). I the extralist cuig tas, recall appears to be the result ofa itersectio of the test cue ad its activated associates with the target ad its activated associates. However, whe attemptig to cotrol or to maipulate a preexistig relatioship betwee related words, ivestigators traditioally resort to free associatio orms ad use oly a sigle idicator for predictig coectio stregth. Typically, forward cueto-target stregth is used as the oly metric (e.g., Bahric, 1970; Nelso & McEvoy, 1979), but bacward target-tocue stregth has also bee used (Humphreys & Galbraith, 1975; Nelso et ai., 1998) because it carries implicatios for the ecodig specificity priciple (Tulvig & Thomso, 1973). These metrics have prove their usefuless i predictig memory performace, but their uidimesioal ature places limits o their effectiveess, give that we ow that the coectios relatig ay two words are liely to be much more complicated. A idex that icorporates multiple sources ofword-to-word coectio stregth will ease list costructio ad reduce cofoudig i ay tas i which word-to-word relatioships are critical. Twosummary idices were developed, ad their effectiveess was evaluated i multiple regressio aalyses by determiig how well they predicted the probability ofcorrect extralist cued recall. METHOD Procedure I what we will call the stadard procedure for the extralist cuig tas, a sigle study trial o a list of idividually exposed target words was followed by a sigle self-paced test trial. Durig study, targets were preseted visually at 3 sec per word uder itetioal learig istructios idicatig that memory for the words would be tested. No advace iformatio about the ature ofthe test was provided. Durig the test trial, a sigle cue for each target was preseted to prompt the recall of its target, ad testig followed immediately after a short istructioal period. All the subjects were give explicit test istructios idicatig that they should use each test cue to help them recall oe of the studied target words. Fiftyeight percet ofthe items were studied ad tested uder this stadard coditio; 23% were give itetioal learig istructios, but the items were preseted at faster or at slower rates; ad 19% were studied or tested uder differet istructios (cocreteessratig ad vowel-amig study istructios, or must-guess ad must-ot-guess test istructios). Materials The 2,272 cue-target pairs were tae from 29 differet experimets ru i our laboratory from These pairs were selected from a large free associatio database that was started i the 1970s that ow icludes free associatio resposes to 5,000+ words ad their resposes for a total of more tha 72,000 pairs. These orms ad the procedures for collectig them ca be obtaied o the Web (Nelso, McEvoy, & Schreiber, 1999). Table I provides the defiitios for the features ofcues, targets, ad cue-targetpairs that were defied by usig the orms ad that were evaluated i these experimets, icludig set size, forward stregth, ad so forth. Table 2 presets the sources ofthe data, as well as a idicatio of what variables were maipulated i each experimet. The published studies are idicated by the last ames of the authors, the dates of publicatio, ad the specific experimets ivolved so that these refereces ca be examied for additioal details-for example, the specific pairs used i each experimet, the umberof subjects, ad so o. The upublished studies are beig prepared for publicatio ad do ot differ i procedures or i item selectio procedures. Although target set size was maipulated more tha ay other variable, o sigle variable was maipulated i all the experimets. All the experimets had 24 pairs i each list, ad at least two differet lists were used i each oe. A average of10.61 subjects (SD = 3.19, rage = 8-24) studied ad were tested o each pair, with probability of correct recall computed by dividig the umber of subjects who correctly recalled the target, relative to the total tested o that target. Table 3 summarizes the experimetal database i a way that closely parallels wor with the idividual experimets. The purpose ofthis table is to covey the mai descriptive statistics for each feature aalyzed i the multiple regressio aalyses to follow ad to show that, whe comparable criteria are used for defiig features, they have similar effects i the database ad the idividual experimets. The first colum idicates the selected features that were chose because they had bee maipulated i prior experimets ad

4 PREDICTING CUED RECALL 607 Table 2 Source of Materials Source of Materials ad Data What Feature Was Varied Nelso, McEvoy,& Schreiber (1990), Experimet 2 TSS Nelso & Schreiber (1992), Experimets I & 2 TSS, TCON Nelso, Schreiber, & McEvoy (1992), Experimets I & 3 FS, QSS, TSS Nelso, Schreiber, & McEvoy(1992), Experimet 4 QSS, ISS Nelso, Beett, Gee, Schreiber, & McKiey (1993), Experimets 1, la, 3-5 TSS, TC Nelso, Beett, Gee, Schreiber, & McKiey (1993), Experimet 2 TSS, TC, SAS Nelso, McEvoy, Jaczura, & Xu (1993), Experimets 1-3,5 TSS Nelso & Xu (1995), Experimets I & 2 TSS, TFREQ Nelso, Beett, & Leibert (1997), Experimet I FS,SAS-MS, TSS Nelso, Beett, & Leibert (1997), Experimet 2 MS-SAS Nelso, McKiey, Gee, & Jaczura (1998), Experimet I TC, TR Nelso, McKiey, Gee, & Jaczura (1998), Experimet 2 TSS, TC, SAS Nelso, McKiey, Gee, & Jaczura (1998), Experimet 3 FS, BS Upublished TSS, TR, TC Upublished FS-BS-SAS-MS, TSS, TC Upublished FS, TC Upublished TC, TFREQ Upublished TSS, Remote TSS Note--QSS, cue set size; TSS, target set size; TC, target coectivity; TR, target resoace; FS, forward stregth; BS, bacward stregth; SAS, shared associate stregth; MS, mediated stregth; TCON, target cocreteess; TFREQ, target frequecy; Remote TSS, set size of the associates of the target. had bee show to affect cued recall. The secod colum reports the average value ofthe specified feature for the 2,272 items used i the 29 experimets (actually, because of missig cocreteess ratigs, the sample sizes rage from 2,171 to 2,272). Stadard deviatios are reported ext to the meas throughout the table. Hece, for cue set size, the data idicate that the 2,272 test cues had a average of associates (SD = 5.20). The ext four colums report the stadard criteria used for defiig the variables i the idividual experimets ad the effects of these variables o the probability ofcorrect cued recall. These criteria were always defied as less tha, greatertha, or equal to some fixed value determied by examiig the frequecy distributio for the feature ad the selectig criteria at the extremes. For example, test cues ad targets with small ad large sets have bee defied as those with 8 or fewer associates at the low ed ad those with 18 or more associates at the high ed, respectively, as determied by free associatio orms. The third colum reports the criteria used for each feature. The fourth colum idicates the mea ad stadard deviatio achieved for the feature at each criterio, ad the fifth colum idicates the item sample size that meets this criterio. Fially, the sixth colum reports the probability of cued recall achieved at each criterio. To cotiue the cue set size example, cues with smaller sets teded to produce higher levels ofrecall (.63) tha did those with larger sets (.49). That all cotrasts were sigificat was idicated by t tests o the recall probabilities. Variables that affect cued recall i subjectbased aalyses for data collected i idividual experimets have the same effects i item aalyses based o data pooled over subjects. Rather tha describig each effect i detail, i order to coserve space, readers are advised to sca across the rows describig each feature to gai a sese ofthe ature ofeach maipulatio, the umber ofitems i the database that meet the criteria, ad the effect of Table 3 Features, Their Descriptive Characteristics, ad Their Ifluece o Probability of Correct Recall Value Maipulatio Recall Feature Average SD Criterio M SD Sample Size P SD Cue set size :s ': Target set size :s ': Target frequecy 99 \77 :s \ ': Target cocreteess :s ': , Target coectivity :S ': Target resoace :s ': Forward stregth :S ': Bacward stregth :S , ': Shared associate stregth :S \ ':.05, Mediated stregth :S \.001 1, ':

5 608 NELSON AND ZHANG Table 4 Correlatios Amog the Attributes ofassociative Networs i the Normative Database ad i the Experimetal Database Attributes QSS TSS TFREQ TCON TC TR FS BS SAS MS QSS TSS TFREQ TCON TC TR FS BS SAS MS Note-Values above ad below the diagoal are from the ormative ad the experimetal databases, respectively. QSS, cue set size; TSS, target set size; TFREQ, [log(.5 + target frequecy)]; TCON, target cocreteess; TC, target coectivity; TR, target resoace; FS, forward stregth; BS, bacward stregth; SAS, shared associate stregth; MS, mediated stregth. the feature o recall. We draw attetio to two separate treds. First, as idexed by cocreteess ratigs for test cues ad targets, the word sample is biased toward cocrete words, as opposed to more abstract oes. Cocreteess was ivestigated early o i the project, ad as a result of this wor, this variable was routiely computed ad was ot allowed to be cofouded with ay ofthe variables that were beig maipulated. Secod, although the same criteria were used to defie low ad high levels offorward ad bacward stregth, the latter varied over a wider rage, particularly because high bac stregth items tededto have stroger coectios tha did high forward stregth items. Bac stregth, as will be discussed below, was ot ivestigated util later i the project, ad ufortuately, it was ot always carefullycotrolled, despite evidece (Humphreys & Galbraith, 1975) ad theory (Tulvig & Thomso, 1973) idicatig that it could be importat. RESULTS Correlatios Amog the Features Table 4 presets the correlatios betwee the features i the ormative database (above the diagoal) ad those i the experimetal database (below the diagoal) that will be evaluated i the mai regressio aalysis. Except for cocreteess, which was based o a sample of51,753, the ormative values show i Table 4 were based o 58,143 pairs of words cosistig ofthe 5,019 ormed words (cues) ad their resposes (targets). For this aalysis, oly targets havig their critical associates ormed were icluded, so this sample represets 81% ofthe pairs i the free associatio orms (Nelso, McEvoy, & Schreiber, 1999). Geerally speaig, the correlatios amog the features i the ormative database were low, suggestig that they were largely idepedet, but because of the large sample size, ay correlatio greater tha.01 was sigificat after Fisher's r-to-z trasformatio. Despite the low correlatios, several patters emerged, ad they were similar for cues ad for targets ormed idepedetly as cues. For example, test cues with larger associative sets teded to be more wealy related to their resposes tha were those with smaller sets. The correlatios ofcue set size with forward, bacward, shared associate, ad mediated stregth were small ad egative. Similarly, with the exceptio offorward stregth, the correlatios oftarget set size with stregth measures teded to be small ad egative. Although words with smaller sets ted to have more coectios from their associates bac to the target (resoace) ad fewer coectios amog their associates (coectivity), the correlatios are relatively small (r = -.21 ad.25, respectively). Table 4 also shows that wea atural relatioships emerged amog some ofthe idices of stregth. Positive correlatios were obtaied for forward-bacward stregth (r =.29), for forwardmediated stregth (r =.17), ad for mediated-shared associate stregth (r =.40). The magitudeofthe forwardbacward relatioship idicates that, although associative symmetry is relatively commo i atural laguage, asymmetry is the more typical patter whe stregth is idexed through free associatio. I additio, the correlatio amog the two idirect idices ofstregth suggests that may pairs are related both because the two words produce the same associate ad also because they share a two-step mediator. The lower half of Table 4 presets the correlatios amog the attributes for the pairs used i various experimets for which all 10 idices were available. I geeral, these values appear to be relatively low ad ofmagitudes approximately similarto those i the ormative database. The correlatio betwee the 45 correlatios associated with the two databases was r =.84, which was sigificat (z = 7.94). Ideally, ifperfect cotrol had bee attaied i the experimetal database, the 45 correlatios amog the attributes would have bee zero i every case, ad the overall correlatio betwee the two databases would have bee lower. This was ufortuately ot the case, ad several troublesome discrepacies emerged. The oe raisig the most problems idicated that target set size had bee uitetioally ad partially cofouded with bacward target-to-cue stregth ad, to a lesser extet, with shared associate ad mediated stregth. Targets with smaller sets teded to be more strogly coected to their test cues. I additio, their resoace teded to be higher, ad the

6 PREDICTING CUED RECALL 609 TableS Results of the Multiple Regressio Aalysis Evaluatig Probability of Correct Cued Recall as a Fuctio of Each Attribute Attribute Stadard Error Stadard Coefficiet t Value Itercept Cue set size Target frequecy Target cocreteess Target set size Target coectivity Target resoace Forward stregth Bacward stregth Shared associate stregth Mediated stregth coectivity oftheirassociates teded to be slightly lower. These partial cofoudigs emerged because the earlier experimets igored bacward stregth ad idirect stregth, as well as coectivity ad resoace stregth, i costructig the lists. By way ofcotrast, the effects of forward stregth were ow early o i the research (e.g., Bahric, 1970; Nelso & McEvoy, 1979), ad this factor was always cotrolled or maipulated so it was ever cofoudedwith target set size. Evetually, the ifluece ofthe other features i the associative cuig tas were recogized or discovered, ad as they were, tighter cotrol was achieved. I what follows, we preset the results of multiple regressio aalyses i a effort to disetagle the separate effects ofthese variables. Mai Regressio Aalysis Table 5 presets the results ofthe multiple regressio aalysis ofthe 10 attributes used for predictig probability of correct cued recall. The assumptios uderlyig this regressio were evaluated by examiig the residuals (Myers & Well, 1995, pp ). Residual probabilities of correct recall were plotted agaist fitted probabilities of recall for each attribute, ad each plot showed the residuals to be evely ad liearly related to fitted probability. Frequecy distributios of the residuals closely approximated ormal distributios, ad the results ofthe Durbi-Watso (D) statistic, as well as the serial autocorrelatios, provided o evidece for lac of idepedece. The D statistics for the 10 attributes averaged 1.49, with the largest deviatio from this value beig 1.52, ad the serial autocorrelatios averaged.26, with the largest deviatio beig.24. Hece, the aalysis ofthe residuals suggested that assumptios ofliearity, homogeiety ofvariace, ormality, ad idepedece were reasoable for each attribute etered ito the equatio. The results ofthe mai multiple regressio aalysis idicated that the 10 features were sigificatly related to recall [R =.58; F(lO,2120) = , MS res =.045]. Table 5 presets the stadard errors, the stadardized coefficiets, ad the t values for the itercept ad each feature. The stadardized regressio coefficiets show i the middle colum oftable 5 are calculatedas if all ofthe idepedet variables had meas of zero ad stadard deviatiosofoe. Such coefficiets ca be compared regardless ofthe differeces i the scales of the variables ivolved, ad they ca be used to determie which attributes i the regressio equatio cotributed most to predictig cued recall. I the preset database, forward stregth (.37) cotributed the most, followed by bacward stregth (.21) ad target coectivity (.19) ad the by target set size (-.15) ad mediated stregth (.15). Frequecy (-.11), shared associate stregth (.08), ad cocreteess (.09) cotributed about equally to recall, ad cue set size (-.04) also produced a small but sigificat effect. Target resoace (.03) had a small positive effect that was ot reliable. Other tha reducig the overall correlatio (R =.56) ad reducig the differeces betwee the two measures ofidirect stregth, deletig cocreteess ad frequecy had little ifluece o the results ofthe mai regressio. The mai results were also uaffected by restrictig the aalysis to ous (1,285 cue-target pairs), or were they iflueced i ay sigificat way whe additioal cue characteristics were added to the 10 characteristics used i the mai aalysis (i.e., cue frequecy, resoace, ad coectivity). Neither cue frequecy or cue resoace had sigificat effects. Although cue coectivity had a sigificat positive effect, this variable has ever bee ivestigated i a maipulatioal desig, so the specific fidigs were omitted i order to coserve space. Furthermore, deletig data obtaied uder vowelamig istructios had little effect o the mai regressio aalysis, but we ote that these istructiosproducediterestig cotrasts i effect sizes whe separate regressio aalyses were coducted for each istructioal coditio. Cocreteess ratig istructios ( = 186 targets) produced results that were early idetical to those of stadard itetioal learig istructios ( = 1,872 targets). I cotrast to both ofthese istructioal sets, vowelamig istructios ( = 138 targets) produced smaller beta coefficiets for target set size, coectivity, bacward stregth, ad shared associate stregth ad larger betas for forward stregth ad mediated stregth. I short, the poor ecodig associated with vowel amig appeared to decrease reliace o iformatio ecodedabout the target ad to icrease reliace o iformatio provided directly by the test cue. Fially, we ote that a uforced stepwise regressio (F-to-eter = 2.41 ad F-toremove = 2.39) of the same data produced results that

7 610 NELSON AND ZHANG were similar to those ofthe multiple regressio aalysis i that oly resoace was left out ofthe model. Each feature added sigificatly to explaied variace after prior features were accouted for. They emerged i steps i the followig order: forward stregth, target set size, target coectivity, bacward stregth, target frequecy, mediated stregth, target cocreteess, shared associate stregth, ad cue set size. Thus, target set size cotributed sigificatly to recall probability after the effects offorward stregth were accouted for, ad so forth. The regressio results cofirmed what previous idividual experimets had show whe these variables were maipulated i the extralist cuedrecall tas (e.g., Nelso & McEvoy, 2000; Nelso & Schreiber, 1992; Nelso & Xu, 1995). More importat, they show that each feature affects recall i the expected directio whe the ifluece of the remaiig features has bee cotrolled statistically by multiple regressio. Cue ad target set size have sigificat egative effects o cued recall (Nelso, Schreiber, & McEvoy, 1992). Targets are more liely to be recalled whe they are cocrete (Nelso & Schreiber, 1992) ad whe they occur less frequetly (Nelso & McEvoy, 2000; Nelso & Xu, 1995). The results also idicate that some features cotribute more to correct recall tha do others. Direct forward coectios betwee the test cue ad the target have the largest effect o cued recall, followed by bacward coectios ad the by idirect coectios (Nelso, Beett, & Leibert, 1997). I additio, coectios amog the associates ofthe target (Nelso, Beett, Gee, Schreiber, & McKiey, 1993) cotribute more tha resoat coectios (Nelso et al., 1998) ad have about the same ifluece as target set size effects (Nelso, Schreiber, & McEvoy, 1992). PROPORTION OF VARIANCE EXPLAINED The purpose ofthis sectio is to determie how much of the variace i extralist cued recall ca be explaied by idices of word owledge tae from free associatio. The results of a regressio aalysis that icluded eight free associatioidices idicatedthat a liear combiatio ofthese measures was correlated.56 with probability of cued recall, producig a adjusted R2 of.32 (target cocreteess ad frequecy were omitted i this aalysis). Thirty-two percet ofthe total variace i the extralist cued recall tas ca be predicted from free associatio measures tae before a experimet. This estimate, however, is also affected by the variace associated with each tas, so it is specific to this data set (Myers & Well, 1995, pp , ). A more accurate estimate taes the reliability of each tas ito accout, which the sets the ceilig for the correlatio. The reliability ofeach tas was assessed by usig test-retest procedures. I the discrete free associatio tas, two separate reliability studies have bee coducted by reormig the same words o differet samples ofsubjects. Nelso ad Schreiber (1992) reormed 155 items, ad Nelso et al. (2000) reormed 120; each ivestigatio produced the same average reliability of.89. To estimate the reliability ofthe extralist cued recall tas, idetical cue-target pairs leared uder similar ecodig coditios i differet experimets were culled from the preset study. The probability of correct cued recall was calculated for the first ad secod occurrecesofthe pair, with idividual pairs radomly assiged to first ad secod positios havig averages of ad subjects per pair. This procedure amassed 225/2,272 pairs ad was possible because some ofthe pairs were used i more tha oe experimet. Small differeces i ecodig coditios were allowedfor example, the same words presetedat 3- or 4-sec presetatio rates could be paired, as could cocreteessratig ad itetioal learig istructios, ad i most cases, the same pair was draw from a differet list cotext. If aythig, allowig slight differeces was expected to mae the estimate of reliability more coservative, ad give these costraits, reliability was estimated to be.73. Give the observed correlatio betwee free associatio ad cued recall of.56 i the multiple regressio aalysis ad respective estimates ofreliability of.89 ad.73, the true correlatio betwee free associatio ad extralist cued recall is Measures tae from free associatio predict approximately 49% of the variace of extralist cued recall whe the reliabilities of the two tass have bee tae ito accout. Hece, early oe halfof the variace i cued recall ca be attributed to prior owledge, as idexed by free associatio. MEASUREMENT ALGORITHMS The results of the mai regressio aalysis idicated that, although they produced differet effect sizes, 10 variables related to preexistig lexical features iflueced the lielihood ofextralist cued recall. Kowledge ofthis iformatio will be useful to ayoe worig o cuig or primig tass ivolvig related words, but such owledge also meas that maipulatig oe of these variables requires cotrollig the others. Give the availability of a large ormative database (Nelso, McEvoy, & Schreiber, 1999), this tas is as possible as it is dreadful to implemet. A meas for reducig some of these variables to sigle predictorswould be useful, ad it also would be helpful to have ey idices expressed i the same form. For example, i the idividual experimets, the idices ofresoace ad coectivity were based o coutig lis amog the items maig up the targets' associative sets. Coutig the lis proved to be a efficiet way to capture these variables, but such couts effectively assig a coectio stregth of uity to each li, whe i fact the lis vary i stregth, just as with ay word-to-word associatio. Expressig these variables i terms ofa stregth idex would tae such variatio ito accout ad would put the idices ofresoace ad coectivity more o a par with idices ofdirect ad idirect stregth. However, ulie sigle-coectio measures such as forward stregth, resoace ad coectivity are liely to ivolve may coectios, ad some theoretically drive meas ofcombiigthem ito a sigle idex

8 PREDICTING CUED RECALL 611 is required. I this paper, we used a equatio associated with PIER 2 to begi this process.? PIER 2 purports to explai how preexistig owledge about word relatioships affects cued recall ad offers a pricipled meas for data reductio (Nelso et al., 1998). Geerally speaig, the process aspects of the model attribute cued recall to the computatio ofthe itersectio of the test cue ad its associates with the target ad its associates. The formal aspects ofthe model rely o a set of three equatios that estimate et cue-target stregth. The first equatio was desiged to capture the ifluece ofresoace ad coectivity, the secod icorporates the effects of direct ad idirect coectio stregths, ad the third addresses the egative effects that set size has o cued recall. These equatios costitute a theory about implicitly activated associates ad their role i icremetig the activatio level ofa studied word i log-term worig memory (LTWM). With the additio of other processig assumptios ad parameters, these equatios ca be implemeted as a process model that is evaluated through simulatio procedures (e.g., Raaijmaers & Shiffri, 1981). I additio ad more germae to the preset aims, these equatios ca be implemeted as measuremet algorithms desiged to reduce multiple sources of stregth to sigle values that ca be used to predict performace i cued recall ad other tass. Such a implemetatio represets a ext step i the traditio of cotrollig or maipulatig forward ad bacward stregth i order to predict cued recall (Bahric, 1970; Nelso et ai., 1998; Tulvig & Thomso, 1973) ad primig (e.g., Kaha, Neely, & Forsythe, 1999). It is a ext step because, istead of usig oly a sigle source of stregth, the model ca be used to icorporate the combied effects of two or more variables i a sigle idex. I this paper, we focus o Equatio I ad oly briefly describe the results of icorporatig the results of this equatio ito Equatio 2. Equatio 1 The ratioale uderlyig Equatio I assumes that comprehesio processes associated with readig the target durig study automatically activate a implicit represetatio ofthe target ad its associates i LTWM. The stregth of this represetatio icreases with activatio returig to the target from its associates (resoace), as augmeted by coectios amog the associates (coectivity). The equatio ca be most easily uderstood as attributig resoace ad coectivity effects to activatio that spreads bac to the target from its associates (1. R. Aderso & Pirolli, 1984). More formally ad i its most elemetary form, the stregth of the target as a implicit represetatio, SeT;), is S(0)=ai+~:.a}ri}' (1) } where a i is the activatio ofthe implicit represetatio of the target i as a result ofits presetatio, ad a} is the ac- tivatio ofassociate j occurrig as a result ofpresetig the target. Furthermore, 'I} = Wi} +2,a'}, ad it represets the resoace comig bac to target i from associate j as modified by associate-to-associate coectios withi the set for all Wi} > 0, ad for j ::1=, ::1= i. wi} is the preexistig li comig bac to target i from associatej. Note that the wi} > 0 restrictio idicates that the ifluece ofcoectivity o a give associate is zero uless that associate is lied bac to the target. Equatio 1 ca be implemeted as a measuremet algorithm because it depeds o word-to-word coectios amog the target's associates, ad the mea preexistig stregth ofeach ofthese lis ca be estimated by usig free associatio orms (see Nelso et ai., 2000, for the theory uderlyig these estimates). A complete example calculatio of this use of Equatio I is provided i Appedix A. The first tas i evaluatig the usefuless of such calculatios was to determie how strogly they were related to the coectio cout idices that they would be replacig. For example, if Equatio I effectively captures the idices ofresoace ad coectivity, it should be positively correlated with each idex. Ufortuately, this was ot the case. Although it was positively correlated with resoace (r =.62, = 2,131), it was egatively related to coectivity (r = -.21). Words with more coectios amog their associates teded to have weaer resoat coectios (r = -.19). What is more, droppig the Wi} > 0 restrictio had virtually o effect o the outcome. Hece, eve whe the resoace coectio assumptio has bee relaxed, Equatio 1predicts that coectivity will have a egative effect o recall, a predictio that stads i cotrast to the results of the mai regressio aalysis. Equatio la Give the goal ofdevelopig a stregth idex that effectively icorporated both resoace ad coectivity, we sought a alterative.formulatio that could serve the same purpose as Equatio 1 i PIER 2 but that was more strogly related to coectivity. I oe variatio that was far more simple to calculate tha Equatio 1, the computatio ofpreexistig target stregth was implemeted by addig its omial stregth to the stregths ofits associates ad its resoat ad associate-to-associate coectios. More formally, the stregth ofthe target is idexed as S(0)=[S(T,T)+ ~S(Ai,T)] (la) +! [S(T,A})+ ~S(Ai,A})). where S(T, 1') represets the stregth ofthe target's represetatio as a result of its presetatio (omially set

9 612 NELSON AND ZHANG to 1.00), S(A i, T) is associate-to-target or resoace stregth, S(T, A) is the stregth ofthe target-to-associate lis, ad S(Ai' A) is the stregth collected from associateto-associate coectios, or what we call coectivity. I the secod part ofthe equatio, the summed stregth acquired by each associate i the set is determied by the sum ofthe iputs from the target ad from other associates i the set. A computatioal example is show i Appedix B, usig estimates of coectio stregths based o the free associatio orms, as was doe for Equatio I i Appedix A. What is iterestig about Equatio IA is that coectivity amog the associates adds directly to target stregth ad exerts a ifluece idepedet of resoace. Equatio IA was positively correlated ( = 2,131) with the coectio cout idices it was meat to replace (resoace, r =.21, ad coectivity, r =.58). It was also more strogly related to coectivity tha to resoace, a result that was i accord with the mai regressio aalysis. These correlatios idicated that Equatio 1A captured the idices of resoace ad coectivity better tha did Equatio I. Whe Equatio la was substituted for the two coectio cout idices i the mai regressio aalysis while leavig all ofthe other features i the regressio (e.g., frequecy, set size, ad so o), the overall predictability was little chaged (R =.56; F = , MS e =.05). The beta coefficiet for Equatio la was.16 which was sigificatly related to cued recall, ad the beta coefficiets for the other features were ot appreciably altered, as compared with the mai regressio aalysis. Fially, addig the square ofthis equatio to this regressio aalysis icreased R by oly.006, which suggests that the relatioship betwee Equatio 1A ad probability ofrecall is adequately represeted by a liear fuctio. Equatio 2 The results ofequatio IA were folded ito the computatio ofpier 2's Equatio 2 for each word pairig i the experimetal database. Equatio 2 computes et cue-target stregth ad is defied as S(Qj,T;) = 'LSjSi +'LSjSi' (2) where Qj is the test cue, T, is the target, ad are the associates that joi the cue ad the target through coverget ad mediated lis (Nelso et ai., 1998). Whe implemeted as a measuremet algorithm, Equatio 2 computes the itersectio betwee the test cue ad its associates ad the target ad its associates by cross-multiplyig word-toword coectio stregths ivolvig the target, the test cue, ad the associates that idirectly li them together. A example calculatio is show i the lower part ofappedix B. The results were the etered ito a regressio aalysis i place ofthe idices for target resoace ad coectivity ad the four separate idices ofcue-target stregth. Set size, cocreteess, ad frequecy were icluded, as i the mai aalysis. This regressio was sigificat [R =.55; F(5,2125) = , MS res =.05], with Equatio 2 accoutig for the majority of the variace (stadardized beta =.46, SE =.006). The stadardized betas were similar i magitude to those obtaied i the mai regressio aalysis ofthe 10 variables ad were the followig: cue set size, -.02 (SE =.00 I), target set size, -.27 (SE=.001), cocreteess,.09 (SE =.004), ad frequecy, -.09 (SE =.007). Each beta was sigificat, except for cue set size. Addig the square ofequatio 2 to this regressio sigificatly icreased R to.54. Although this.02 icrease i R was statistically sigificat, it was umerically smail suggestig that the liearity assumptio remais reasoable for the rages ofequatio 2 foud i the preset database. Nevertheless, a cautio is i order. Forward cue-totarget stregth was the best predictor of recall for this database, but 80% ofthe pairs had stregths ofless tha.26. Higher values were purposely avoided i order to avoid ceilig effects o probability of recall; had we allowed the full rage ofstregth for each type ofcoectio, we have o doubt that Equatio 2 would have show a stroger curviliear relatioship with recall. Uder such coditios, addig two or more high-value coectio stregths would o loger be equivalet to addig two or more low-value coectio stregths, because of differeces i scale. Uder such coditios, Equatio 2 would begi to fail as a predictor. Hece, Equatio 2 represets a effective summary idex ofsix separate features, because it successfully explais a substatial portio ofthe variace ofextralist cued recall, but its usefuless as a predictor is limited to relatively weaer pairigs. Oly uder such coditios is the liearity assumptio liely to hold. DISCUSSION The extralist cuig tas ivolves a mix of episodic learig ad cuig based o preexistig iformatio. Ulie paired-associate tass, the test cues are uavailable durig the study episode, ad subjects must rely o what they ow about the relatioship betwee cues ad their targets to achieve specified retrieval goals. I may ways, the tas mimics everyday cue-depedet memory performace ad provides a ideal method for studyig the itersectio betwee ow ad ew iformatio. However, although the coditios uder which the ew iformatio is acquired ca be maipulated ad otherwise carefully cotrolled, what is supposedly ow about the cue-target relatioship ca oly be iferred from measuremets tae from other sources. Examples of these sources iclude cocreteess ratigs (Paivio, Yuille, & Madiga, 1968), frequecy couts (Kucera & Fracis, 1967), ad free associatio (Bahric, 1970; Nelso et ai., 1998). I these ad other examples, measuremets are tae from some idividuals i order to ifer what other idividuals are liely to ow. The success ofthis eterprise is ormally determied by whether a particular feature idetified through the measuremet procedure has a effect o tas performace. I the extralist cuig tas, may features oflexical owledge have bee idetified ad measured through the use offree associatio ad have bee show

10 PREDICTING CUED RECALL 613 to affect recall i this tas (e.g., Nelso et al., 1998; Nelso, Schreiber, & McEvoy, 1992). I this sese, this approach to the ivestigatio of the itersectio of ow ad ew iformatio has bee successful. Nevertheless, a higher criterio for success requires that the idividual features so idetified exert their ifluece above ad beyod their relatioships with other features (Rubi & Friedly, 1986). For example, the effect of the size of a word's associative set o recall has bee ow for may years (e.g., Nelso & McEvoy, 1979). Extesive empirical wor has show that this effect is idepedet ofcocreteess (Nelso & Schreiber, 1992), frequecy (Nelso & Xu, 1995), coectivity (Nelso et ai., 1993), ad word ambiguity (Gee, 1997). Despite all ofthis wor, set size has ot bee ivestigated with respect to other word features (e.g., bacward stregth), but what is more importat, it has ot bee ivestigated i the cotext ofsimultaeous variatios i all ofthe lexical features ow to affect cued recall. A experimet that crossed 10 or more features simultaeously would be as impractical to do as it would be impossible to uderstad. There are clear limits to idividual experimets whe it comes to ivestigatig the simultaeous effects ofa costellatio ofvariables ow to affect performace. I the preset article, this dilemma was circumveted through the use of correlatioal ad multiple regressio aalyses o a database comprisig thousads ofcue-target pairigs tae from a large umber ofidividual experimets. The correlatioal aalyses for the 10 features represetig the mai focus ofthe article idicated that, for the most part, the correlatios amog the features were low ad reflected the magitude ofthe relatioships observed i the much larger ormative database. Geeral1y speaig, the results ofcorrelatioal aalyses usig the ormative database suggest that the 10 features are mostly idepedet. I the experimetal database, forward stregth was ever cofouded with target set size as a result ofdirect attempts to cotrol this relatioship. Mostly by accidet, forward stregth was ever cofouded with bac stregth, eve though a modest correlatio exists i the ormative database (r =.29). The oe importat exceptio to this patter oflow correlatios ivolved target set size ad bac stregth. These variables are wealy correlated i the ormative database (r = -.15) but more strogly correlated i the experimetal database (r = -.39), because bac stregth was left ucotrol1ed i most ofthese experimets. Target set size ad bac stregth were partially cofouded i the database, ad although hardly a ew idea, it is useful to be remided, at least occasioally, that maipulatig oe feature while lettig aother vary radomly provides o guaratee agaist cofoudig them, eve whe they are poorly related i atural laguage. Despite the partial cofoudig of target set size ad bac stregth ad other wea relatioships, the results of the multiple regressio aalyses idicate that each ofthe features maipulated i these experimets cotribute to the probability of extralist cued recall. Multiple regressio adjusts for correlatios amog variables, so it is possible to estimate the relative cotributios of each feature to recall by usig stadard beta coefficiets. These coefficiets idicated that forward stregth has the largest effect o extralist cued recall, followed by bacward stregth, coectivity, set size, idirect measures ofstregth, cocreteess ad frequecy, ad lastly, by target resoace. As was expected, each feature has positive effects o recall, except cue ad target set size ad target frequecy, which have egative effects o recall. These results were largely uaffected by various restrictios placed o the database, such as limitig the items to ous. What is most importat, these results cofirm what idividual experimets have show. Features that affect recall i idividual experimets maipulatig two or three features at a time have the same effects o recall i multiple regressio ivolvig simultaeous variatio i all ofthe features. The fidigs are robust i that they geeralize over differet subjects with similar cultural bacgrouds ad over differet items with similar associative characteristics. Fially, whe tas reliabilities are tae ito accout, 49% ofthe variace i extralist cued recall appears to be determied by preexistig owledge about the test cue ad target ad about the associative lis that coect them. Both the idividual experimets ad the multiple regressio aalyses idetify a collectio of word features that ifluece extralist cued recall, ad it would be practically ad theoretically useful to capture this ifluece i a algorithm that combies at least some ofthese features ito sigle predictors. Early theoretical wor with this tas suggestedthat both forward cue-to-targetstregth ad bacward target-to-cue stregth would be importat. The preset fidigs recapitulate a portio ofthe developmetal history of research o extralist cued recall, i which propoets of geeratio-recogitio models stressed the importace of forward stregth (Bahric, 1970), whereas propoets ofecodig specificity stressed the importace of bac stregth (Tulvig & Thomso, 1973). The ecodig specificity priciple stated that the test cue had to be ecoded while the target was beig ecoded i order for it to.be a successful cue. Each group was correct, as the preset fidigs idicate that these features represet the best overall idividual predictorsof recall success. Ofcourse, either approach ca explai as much variace as PIER 2's Equatio 2 implemeted as a measuremet tool, because it icorporates six differet sources ofstregth. Two ofthese sources, resoace ad coectivity, cocer target activatio stregth, ad four ofthese sources cocer direct ad idirect coectios liig the cue ad the target. This implemetatio depeds o usig free associatio orms to estimateword-toword coectio stregths ad o the theory uderlyig the model for determiighow to combie these stregths ito a sigle idex. PIER 2's Equatio I was used as a meas for expressig the coectio cout idices for resoace ad coectivity as a idex of target stregth. However, this equatio failed to capture coectivity i a way that reflected the ifluece of this variable o recall. The equatio assumes

11 614 NELSON AND ZHANG that the activatio level of the target icreases with the stregth ofthe resoat coectios comig to it from its associates, as augmeted by coectivity amog the associates themselves. Coectivity comes ito play idirectly by boostig resoace returig to the target. The feasibility ofequatio 1 as a measuremet idex was evaluated by calculatig it for each target i the database, usig idices of word-to-word coectio stregths provided by free associatio. Our expectatio was that this equatio would have to be positively correlated with coectio cout idices of resoace ad coectivity i order for it to be useful for predictig the effects of these variables o recall. However, although Equatio 1 was positively correlated with resoace, it was egatively correlated with coectivity ad predicted that this variable would have a egative rather tha a positive effect o recall. I light ofthis problem, Equatio 1 was redesiged to avoid the depedet relatioship betwee coectivity ad resoace. I Equatio la, the sum ofthe target-toassociate stregths was added to the sum ofthe resoat stregths ad the sum of the stregths stemmig from associate-to-associate lis. The result was used as a idex ofthe activatio stregthofthe target ad its associates. Theoretically, this equatio ca be coceptualized as a parallel activatio process, with all the coectios i a associative etwor cotributig additively to stregtheig the target ad its associates. I comparig the two equatios, oe ey differece is that Equatio la allows associate-to-associate coectivity to cotribute directly to target activatio stregth, whereas Equatio 1 permits oly a idirect cotributio through resoat coectios. This differece tured out to be critical, because Equatio la was positively correlated with resoace ad with coectivity ad was more strogly related to coectivity. Whe this equatio was substituted for the separate coectio cout idices of resoace ad coectivity i the regressio aalysis, the results idicated that it sigificatly predicted probabilityofrecall. Whe the results ofequatio 1A were fed ito PIER 2's Equatio 2, a sigle predictor was derived from six separate idices ivolvig pairwise coectios (resoace, coectivity, forward stregth, bacward stregth, mediated stregth, ad shared associate stregth). This ew idex of cue-target stregth predicted probability of cued recall about as effectivelyas whe the idices were separately etered ito the regressio equatio. At least for the extralist cued recall tas, a sigle value produced by a theoretically based algorithm ca tae the place of six separate measures. I selectigword pairs for experimets, this rule is liely to ease list costructio, as well as to reduce cofoudig. However, two caveats apply to usig this equatio for predictigword relatedess. First, its use should be cofied to wea cue-target pairs i order to maitai liearityofscale. Secod, the rule mayormay ot wor for other tass ivolvig pairs ofrelated words. The value of the equatio has bee derived solely from data o extralist cued recall, ad its geeralizatio ca be determied oly whe we ow whether performace i other tass depeds o computig the same sources of coectio stregth. For istace, this implemetatio ofequatio 2 should ot effectively predict target recovery uder implicit coditios of learig ad testig. Uder implicit coditios, subjects are ecouraged to produce the first word that comes to mid that is related to the test cue, ad it is uliely that they will be searchig for a specific word recetly see i the testig cotext. Uder this relaxed retrieval criterio, target recovery is more proe to be the result ofprimigthe target as a associate ofthe test cue, as opposed to computig a cue-target itersectio, ad therefore, it is more liely to be iflueced by forward ad mediated lis tha by bacward ad shared associate lis. Similarly, it is clear that lexical decisio is a fuctio of the forward ad bacward stregths of the prime-target pairs (e.g., Kaha et al., 1999), but we do ot ow whether idirect coectios ad coectivity will affect decisio time, or whether forward ad bacward coectios will have additive effects. I these cases, a less complex rule or the idices ofthe idividual features may fuctio better tha a algorithm that icludes all the sources of coectio stregth. For the preset, oe mai advatage ofequatio 2 is that it alerts researchers to the differet types of coectios that li words together, while providig a structured algorithm for explorig their ifluece i a variety of differet cogitive tass ivolvig related words. By way offial commet, we ote that implemetig PIER 2 as a measuremet model follows i a log traditio ofusig free associatio data to estimate the stregth ofpreexistig coectios amog word pairs i order to forecast the effectiveess of various types of primig cues. PIER 2 implemeted as a measuremet tool follows i this traditio ad is differet oly i that it has bee desiged to express multiple sources ofcoectio stregth as a sigle value, as opposed to usig oly a sigle source of stregth while igorig others. Oe difficulty iheret i this approach is that there is more tha oe way to icorporate the ifluece ofmultiple sources ofiformatio, ad as the cotrast betwee Equatios 1 ad la maes clear, theories are required i order to mae reasoed decisios about how to combie these sources. The preset fidigs idicate that a equatio that idexes coectivity idepedet ofresoace will be more liely to capture the effects ofthis word feature tha will oe that assumes that it must be expressed through resoat coectios. This coclusio, i tur, idicates that implemetig PIER 2 as a measuremet algorithm carries implicatios for developig it further as a process model desiged to explai the effects ofthe same variables. We are ow more wary ofthe assumptio that activatio must spread bac to the target i order to heighte its activatio ad lea more toward oe that allows coectivity amog the target's associates to cotribute idepedetly. At the least, the preset results suggest that each coceptualizatio of target activatio stregth eeds to be modeled ad evaluated o its merits.

12 PREDICTING CUED RECALL 615 REFERENCES ANDERSON, 1. R., & PIROLLI, P.L. (1984). Spread of activatio. Joural ofexperimetal Psychology: Learig. Memory. & Cogitio, 10, ANDERSON, M. C; & SPELLMAN, B. A. (1995). O the status of ihibitory mechaisms i cogitio: Memory retrieval as a model case. Psychological Review, 102, BAHRICK, H. P.(1970). Two-phase model for prompted recall. Psychological Review, 77, DEESE, J. (1965). The structure ofassociatios i laguage ad thought. Baltimore: Johs Hopis Uiversity Press. GEE, N. R. (1997). Implicit memory ad word ambiguity. Joural of Memory & Laguage, 36, HUMPHREYS, M. S., & GALBRAITH, R. C. (1975). Forwardad bacward associatios i cued recall: Predictios from the ecodig specificity priciple. Joural ofexperimetal Psychology: Huma Learig & Memory, 1, JENKINS, J. 1., & RUSSELL, W. A. (1952). Associative clusterig i recall. Joural ofabormal & Social Psychology, 47, I. KAHAN, T. A., NEELY, 1. H., & FORSYTHE, W. J. (1999). Dissociated bacwardprimig effects i lexicaldecisio ad prouciatio tass. Psychoomic Bulleti & Review, 6, KUCERA, H., & FRANCIS, W. N. (1967). Computatioal aalysis of preset-day America Eglish. Providece, RI: Brow Uiversity Press. Mclivov, C. L., NELSON, D. L., & KOMATSU, T.(1999). What is the coectiobetweetrue ad false memories? The differetialroles of iteritemassociatiosi recall ad recogitio.joural ofexperimetal Psychology: Learig. Memory, & Cogitio, 25, MYERS, J. L., & WELL, A. D. (1995). Research desig ad statistical aalysis. Hillsdale, NJ: Erlbaum. NELSON, D. L., BENNETT, D. J., GEE, N. R., SCHREIBER, T. A., & Mc KINNEY, V. (1993). Implicit memory: Effects of etwor size ad itercoectivity o cued recall. Joural ofexperimetal Psychology: Learig, Memory, & Cogitio, 19, NELSON, D. L., BENNETT, D. 1., & LEIBERT, T. W. (1997). Oe step is ot eough: Maig better use of associatio orms to predict cued recall. Memory & Cogitio, 25, NELSON. D. L.. & Mclivov, C. L. (1979). Ecodigcotext ad set size. Joural ofexperimetal Psychology: Huma Learig & Memory, 5, NELSON. D. L.. & McEvov, C. L. (2000). What is this thig called frequecy? Memory & Cogitio, 28, NELSON. D. L., McEvoy, C. L.. & DENNIS. S. (2000). What is free associatio ad what does it measure? Memory & Cogitio, 28, NELSON. D. L., McEvoy. C. L., & SCHREIBER. T. A. (1990). Ecodig cotext ad retrieval coditios as determiats of the effects of at- ural category size. Joural ofexperimetal Psychology: Learig. Memory. & Cogitio, 16, NELSON, D. L.. & Mclivov. C. L., & SCHREIBER, T. A. (1999). The Uiversity ofsouth Florida word associatio. rhyme ad wordfragmet orms [WWW documet). Available: FreeAssociatio/ NELSON, D. L., McKINNEY, V. M., GEE, N. R., & JANCZURA, G. A. (1998). Iterpretig the ifluece of implicitly activated memories o recall ad recogitio. Psychological Review, 105, NELSON, D. L., & SCHREIBER, T. A. (1992). Word cocreteess ad word structure as idepedet determiats of recall. Joural of Memory & Laguage, 31, NELSON, D. L., SCHREIBER, T. A., & HOLLEY, P. E. (1992). The retrieval of cotrolled ad automatic aspects of meaig o direct ad idirect tests. Memory & Cogitio, 20, NELSON, D.L., SCHREIBER, T.A.,& McEvoy,C. L. (1992).Processig implicit ad explicit represetatios. Psychological Review, 99, NELSON, D. L., SCHREIBER, T. A., & Xu, J. (1999). Cue set size effects: Samplig activated associates or cross-target iterferece? Memory & Cogitio, 27, NELSON, D. L., & XU,1. (1995). Effects of implicit memory o explicit recall: Set size ad word frequecy effects. Psychological Research, 57, PAIVIO, A., YUILLE, 1. C; & MADIGAN, S. (1968). Cocreteess, imagery, ad meaigfuless values for 925 ous. Joural ofexperimetal Psychology Moograph Supplemet, 76 (I, PI. 2), RAAIJMAKERS, 1. G., & SHIFFRIN, R. M. (1981). Search of associative memory. Psychological Review, 88, ROEDIGER, H. L., III, & McDERMOTT, K. B. (1993). Implicit memory i ormal huma subjects. I F. Boller & 1. Grafma (Eds.), Hadboo ofeuropsychology (Vol. 8, pp ). Amsterdam: Elsevier. RUBIN, D. C.; & FRIENDLY, M. (1986). Predictig which words get recalled: Measures of free recall, availability, goodess, emotioality, ad prouciability for 925 ous. Memory & Cogitio, 14, TuLVING, E.,& THOMSON, D. M.(1973).Ecodigspecificityadretrieval processesi episodicmemory. Psychological Review, SO NOTES I. The true correlatio betwee free associatio ad cued recall was computed as r Tx Ty = i / '?rx ~rvy.56 = /.89' For people iterested i checig the preset models or i creatig ew models, the cued recall database used for this paper is available o dis o request ad ca be foud o the Web: ( -elso/cuedrecalldatabase). APPENDIX A A Example Calculatio of Equatio 1 for the Word DINNER I. X associatio matrix for the word DINNER, showig its associates ad their itercoectio stregths as idexed by free associatio orms. These values are used to calculate Equatio I. DINNER Dier Supper Eat Luch Food Meal Dier Supper, Eat Luch Food Meal Calculatio ofthe et activatio stregth ofthe target word DINNER, usig Equatio I, as defied by