Individual differences in the fan effect and working memory capacity q

Transcription

1 Journal of Memory and Language 51 (2004) Journal of Memory and Language Individual differences in te fan effect and working memory capacity q Micael F. Bunting a, *, Andrew R.A. Conway b, Ricard P. Heitz c a Department of Psycological Sciences, University of Missouri, 207 McAlester Hall, Columbia, MO 65211, USA b Department of Psycology, Princeton University, Green Hall, Princeton, NJ 08544, USA c Scool of Psycology, Georgia Institute of Tecnology, Atlanta, GA, USA Received 18 April 2004; revision received 22 July 2004 Available online 17 September 2004 Abstract In opposition to conceptualizing working memory (WM) in terms of a general capacity, we present four experiments tat favor te view tat individual differences in WM depend on attentional control. Hig- and low-wm participants, as assessed by te operation span task, learned unrelated sentences for wic te subject and predicate of te sentences sared concepts (fan). Sentences were learned in sets organized by subjects (Experiments 1A and 1B) or predicates (Experiments 2A and 2B). WM predicted accuracy and reaction times on a subsequent speeded verification task, but not learning. In Experiments 1A and 2A, low-wm participants ad a steeper, positively sloped fan effect for reaction times to studied items tan ig-wm participants. In Experiments 1B and 2B, fan was eliminated across but not witin memory sets, wic eliminated individual differences but not slope to te fan effect. Tese effects suggest te crux of WM is attentional control, and competition across sets causes individual differences. Ó 2004 Elsevier Inc. All rigts reserved. Keywords: Attentional control; Fan effect; General capacity; Individual differences; Inibition; Working memory q Experiments 1A and 1B of tis researc were submitted to te University of Illinois at Cicago in partial fulfillment of te requirements for te Master of Arts degree by Micael F. Bunting. We tank Bennetta Fairfax, Cristina Gabe, Antony Kamis, Dimitri Perivoliotis, and Bradley Poole for testing participants in te prescreening, and Stellan Olsson and Gary Raney for offering elpful comments on an earlier draft of te manuscript. We especially tank Stepen Tuolski for providing elp in planning and interpreting tese experiments. Micael Bunting is supported by a postdoctoral fellowsip at te University of Missouri from te Missouri Reabilitation Researc Training Program (Kristofer Hagglund, P.I.), te National Institute of Cild Healt and Human Development, and te National Institutes of Healt (Grant 2 T32 HD ). * Corresponding autor. Fax: address: buntingm@missouri.edu (M.F. Bunting). Introduction Rapidly and accurately retrieving familiar information from semantic memory sometimes requires working memory capacity. Cantor and Engle (1993, see Experiment 1 of teir report) demonstrated as muc wen tey observed tat te magnitude (or slope) of fan effects for retrieval from long-term semantic memory varied as a function of individual differences in working memory capacity. Fan effects occur wen te time required to verify a proposition encoded in long-term memory increases concomitantly wit te number of propositions collectively committed to memory (Anderson, 1983). Fan effects are te product of interference among sared concepts. Te critical result in Cantor and Engle (1993) X/$ - see front matter Ó 2004 Elsevier Inc. All rigts reserved. doi: /j.jml

2 M.F. Bunting et al. / Journal of Memory and Language 51 (2004) was tat participants wit less working memory capacity revealed more dramatic fan effects tan tose wit more working memory capacity, and te slopes of individual fan effects statistically accounted for te relationsip between working memory capacity and reading compreension, as measured by te Verbal Scolastic Aptitude Test (VSAT). Tis result as been used to promote te idea of working memory capacity as amount of activation. According to Cantor and EngleÕs (1993) teoretical framework, working memory is te activated portion of long-term memory, and working memory capacity is te total amount of activation available for cognition (a similar position was taken in Anderson & Lebiere, 1998; Engle, Cantor, & Carullo, 1992; Lovett, Reder, & Lebiere, 1999). Four experiments are reported ere; te results of wic are inconsistent wit te notion of working memory capacity as amount of activation and inconsistent wit Cantor and EngleÕs (1993) interpretation of individual differences in te fan effect. Te current experiments suggest tat individual differences in working memory capacity and te fan effect are better explained by a model in wic working memory capacity refers to te ability to resist interference rater tan a limited amount of activation. Working memory as capacity Tere is some consensus among cognitive psycologists tat working memory is a system, or set of processes, tat allows for te active maintenance of information in te face of concurrent processing and/ or distraction (Baddeley & Hitc, 1974; Baddeley & Logie, 1999; Engle, Kane, & Tuolski, 1999; Miyake & Sa, 1999). Working memory span tasks, including counting span, operation span, and reading span, were developed in accordance wit tis view of working memory (for examples of eac see Case, Kurland, & Goldberg, 1982; Daneman & Carpenter, 1980; Turner & Engle, 1989, respectively). Altoug different in surface level features, tese tasks are structurally similar, for some sort of secondary processing (be it counting sapes, solving matematical operations, or reading sentences) is imposed during encoding to detract from memory for some memoranda. For example, in te operation span task used in te present set of experiments and described more torougly in te Prescreening, participants verify te veracity of simple matematical equations wile attempting to remember strings of unrelated words for later recall. Eac trial consists of a matematical operation followed by a memoranda (e.g., Is (9 * 2) 2 = 15? Road), and recall is cued after sets of 2 5 operationword strings. Working memory span is te number of words correctly recalled. Assessed as suc, working memory capacity reliably and strongly predicts more complex cognitive beaviors suc as reasoning, problem solving, and reading compreension (Engle, 2001). Working memory span was once commonly described as te capacity to sare resources among competing goals (Cantor & Engle, 1993; Daneman & Carpenter, 1980; Just & Carpenter, 1992; Turner & Engle, 1989). Daneman and Carpenter (1980), in te tradition of Baddeley and HitcÕs (1974) working memory model, proposed tat te processing and storage components of teir reading span task made competing demands for a sared resource, and tey ypotesized tat more efficient processing permits more room for storage. Oters, including Turner and Engle (1989), took a similar capacity view of working memory, but tey differed from Daneman and Carpenter in te extent to wic tey conceived of tis resource as a domainspecific. Turner and Engle made a convincing argument for te domain-generality of working memory wen tey demonstrated tat operation span, wic does not involve reading, accounted for as muc variance in multiple measures of reading ability as reading span. Reading span and operation span account for te same variance in reading compreension, and tey load on te same factor in factor-analytic studies (Conway, Cowan, Bunting, Terriault, & Minkoff, 2002; Engle, Kane et al., 1999; Kane et al., 2004). Terefore, te processing component need not be reading in order to predict reading compreension. According to te general capacity model of working memory capacity, working memory is a domain-free memory capacity (Engle et al., 1992). Te dual components of operation span, as in oter span tasks, require active maintenance (i.e., remembering words for later recall) in te face of concurrent processing (i.e., solving matematical operations). Working memory as attentional control Te general capacity model was closely associated wit a size metapor for memory capacity and empasized te measurement of memory ability, but tis model as not eld up to empirical investigation (e.g., Conway & Engle, 1994; for a review see Engle, 2001). Engle and colleagues, wo were largely responsible for proposing and advancing te general capacity model, ave found tat pure memorial processes fail to account for te covariation between working memory span and iger-order cognition. Tey ave instead argued tat working memory is equivalent to sort-term memory plus a general, controlled-attention ability (Engle, 2001; Engle, Tuolski, Lauglin, & Conway, 1999; Kane & Engle, 2003). It is tis attentional control component tat is responsible for te predictive utility of working memory span tasks. According to Engle and colleaguesõ controlled attention teory of working memory, working memory capacity refers to te attentional ability to maintain activation for goal-relevant information and ignore goal-irrelevant or distracting information via control mecanisms suc as goal maintenance and inibition.

3 606 M.F. Bunting et al. / Journal of Memory and Language 51 (2004) In support of tis view, Engle and colleagues ave produced a line of correlation-based studies in wic were found positive and predictive relationsips between working memory capacity and measures of attentional control. Tis researc includes Conway, Cowan, and BuntingÕs (2001) demonstration tat individual differences in working memory capacity are predictive of sadowing errors in a dicotic listening ( cocktail party ) paradigm. Kane and Engle (2003) demonstrated a similar relationsip but wit te Stroop paradigm. Te Stroop effect, a quintessential marker of te failure of attentional control, emerges in a color naming task wen te to-be-named color appears in te form of a color word and wen te ink color and color name do not matc (e.g., te word blue written in te color red wen te goal is to name te color red). Kane and Engle (2003) sowed tat tose wit less working memory capacity are more likely to make Stroop errors tan tose wit more working memory capacity, but only wen te Stroop task was composed of mostly congruent trials (e.g., te word blue sown in te color blue). Wen te task was so biased, tose wit less working memory capacity could rely on te easier prepotent response of reading te word sown. Adopting tis strategy, owever, is detrimental to performance on te occasional incongruent trial in wic te to-be-named color and color name do not matc. Kane and Engle suggested tat individual differences in te Stroop effect are evidence of te ig spanõs ability to remain focused on te task goal, wic is someting tat takes attentional control. Haser and colleagues also implicate attentional control, specifically te ability to suppress proactive interference, in teir view of working memory capacity (Haser & Zacks, 1988; Haser, Zacks, & May, 1999). Lustig, May, and Haser (2001; see also May, Haser, & Kane, 1999) sowed tat te reading span task is susceptible to te effects of proactive interference (PI). Lustig, May et al. were motivated by te ypotesis tat PI builds over many consecutive sets in most any memory span task but especially in te reading span task. Terefore, tey manipulated set order, presenting sets eiter in ascending (smallest to largest) or descending order (largest to smallest), and ypotesized tat te descendingorder condition would pose less difficulty even for tose prone to interference effects, suc as older adults in teir study. Arguably, te effect of PI sould be minimized if te largest (and most difficult) sets are presented first, before te effects of PI ave mounted. Not only did tis manipulation improve span scores, but te relationsip between span and reading compreension was attenuated. From tis perspective, te crux of working memory span tasks is interference, and te ability te control interference from prior episodes is te source of individual differences (Haser, Tonev, Lustig, & Zacks, 2003; Lustig, Haser, & Tonev, 2001; Lustig, May et al., 2001; May et al., 1999). Te relevant distinction for us is not between tese two teories of attentional control but ow te concept of attentional control is an important sift from a focus on memory capacity and te size metapor for memory. For tat reason, we will use te term attentional control to refer collectively to EngleÕs teory and HaserÕs teory, wic ave more commonalities tan differences and wic would make identical predictions in our studies. We use te term capacity view to refer to te opposing view of working memory capacity tat is focused on quantifying memory capacity. Working memory and te fan effect Te sift from memorial processes or memory ability to attentional control in contemporary accounts of working memory brougt us to reevaluate Cantor and EngleÕs (1993) experiment wit te fan paradigm and teir primary explanation of teir effect. In Cantor and Engle (1993) and in te basic fan paradigm (Anderson, 1983), participants first learned many facts about various people in various locations (e.g., te artist is in te ouse, te plumber is in te park, te teacer is in te boat). In a subsequent speeded recognition task, tey verified te verity of some statements; some were learned target facts and some were unstudied foils (e.g., te teacer is in te train). Cantor and EngleÕs materials were also used in Experiment 1 of tis report and are listed in Table 1. In Cantor and Engle, an interference effect was manifest in reaction time to te queried facts; te more facts tat were associated wit a person or a location, te longer it took participants to verify just one fact about tat person or location. In te basic fan paradigm, a fan effect on retrieval can appen for accuracy measures, too. Cantor and Engle (1993) found tat te slope of te fan effect for reaction times was greater for tose wit low working memory capacity tan tose wit ig working memory capacity. (Working memory span was assessed wit te operation span task, and te low and ig groups were tose in te lower and upper quartiles of te distribution for operation span, respectively.) Cantor and Engle (1993) advocated for te general capacity view of working memory and attributed individual differences in te fan effect to differences in te activation capacity of working memory. Tey argued tat working memory is equivalent to te activated portion of semantic long-term memory and tat working memory capacity is equivalent to source activation. Tis view is consistent wit oter capacity-limited models of spreading activation (e.g., Anderson, 1983). To account for te fan effect, Anderson (1983) suggested tat learned facts are stored in networks of semantic-associates. Wen a test sentence is presented in a fan paradigm like tat used by Cantor and Engle (1993), te memory

4 M.F. Bunting et al. / Journal of Memory and Language 51 (2004) Table 1 Studied and non-studied Person and Location pairs in Experiments 1A (d) 1B() Location-terms Person-terms Lawyer Artist Plumber Fireman Teacer Doctor (1) Studied Person and Location pairs Boat d d Park d d Curc d d Bank d d Diner Hotel Museum Airport House d d Store d d Zoo d d Train d d Stadium Jail Teater Camper (2) Non-studied Person and Location pairs Boat d d Park d d Curc d d Bank d d Diner Hotel Museum Airport House d d Store d d Zoo d d Train d d Stadium Jail Teater Camper Note. Propositions were of te form, Te person is in te location. nodes corresponding to te subject and predicate of te sentence are activated, and tis activation spreads evenly to related nodes. For example, if given te sentence, a plumber is in te park, activation sould spread to oter locations associated wit te person (e.g., plumber) and to oter people associated wit te location (e.g., park). Participants correctly verified facts in te recognition test wen activation spreading from one node intersected wit activation spreading from anoter node. Because activation spreads equally and evenly to all related nodes, te number of oter associative links limits te speed at wic activation spreads. Greater source activation, owever, speeds te spread of activation. Hence, arguing from Cantor and EngleÕs vantage, one could attribute faster reaction times and smaller fan effects for tose wit more working memory capacity to more source activation, or capacity. An attentional control view of working memory, consistent wit EngleÕs controlled attention view and HaserÕs inibition view, can also account for Cantor and EngleÕs (1993) effect and an additional prediction tat te general capacity view cannot. Individual differences in te fan effect may be due to te ability to avoid making spurious connections (cf. Lustig, Haser et al., 2001, 2001). Tose wit low working memory capacity may keep irrelevant propositions unnecessarily active, tereby overloading working memory. Te inibition view was in its infancy at te time te Cantor and Engle paper was publised, but Cantor and Engle offered tis as an untested ypotesis and callenge to teir own interpretation: If low-span subjects are more vulnerable to interference tan are ig-span subjects, te fan task could reflect te difference. It remains to be seen weter te differential fan effect observed between ig-

5 608 M.F. Bunting et al. / Journal of Memory and Language 51 (2004) and low-span subjects in te present studies would be found wen te fact-retrieval task involved minimal interference (p. 1111). Two kinds of interference are apparent wen one considers carefully ow participants learned and were later tested on te fan propositions in Cantor and Engle (1993). Cantor and Engle employed a blocked design for te learning procedure. Participants were exposed to te propositions in sets organized by te person-term of te sentences, and te duration of te study time was proportionate to te set size. Referring to te stimuli in Table 1, one can see tat te tree propositions for plumber (te plumber is in te park... curc... bank) were studied togeter and were terefore episodically unique. Altoug tere was interference wit te person-term plumber as a result of te fact tat it was associated wit multiple locations, tat interference was limited witin a single learning episode. Te same was not true for interference from multiple instances of te location-terms. Te location-terms were repeated across blocks witin te learning procedure. For example, te location-term, park, was sared wit two person-terms, plumber and teacer. Tis particular form of cross-episode interference sould be especially difficult for tose wit low working memory capacity. An attentional control view of working memory would predict tat interference from irrelevant episodes is detrimental to low spans but not ig spans (Lustig, May et al., 2001). Our goal was to test te interference ypotesis wit te manipulation of response competition between Experiments 1A and 1B. We replicated Cantor and EngleÕs (1993) design in Experiment 1A and sougt one critical effect: a fan size-by-span interaction, wic would indicate tat te slope of te fan effect for te low-span group was greater tan te slope for te ig-span group. (Note. Cantor and Engle reported te fan effect as a function of person-fan, rater tan location-fan or total-fan. Because te purpose of Experiment 1A was to replicate Cantor and Engle, te same approac was taken ere for all experiments). In Experiment 1B, we eliminated interference from across episodes but not interference witin episodes. Location-fan was eld at one (i.e., interference or overlap among locations in te propositions was eliminated). Person-fan was greater tan one to cause a fan effect. According to a capacity-limited view of working memory capacity, individual differences in te fan effect sould remain because activation is still divided as person-fan increases and igand low-span participants differ in te total amount of activation. In contrast and according to attentional control views of working memory capacity, te elimination of overlap among locations (Experiment 1B) sould reduce response competition from irrelevant learning episodes and, terefore, eliminate or attenuate individual differences in te fan effect. A main effect of fan was ypotesized in Experiments 1A and 1B. Participant screening Participants were screened for working memory capacity on te operation-word span task. Turner and Engle (1989) originally reported tis score, and it consistently correlates wit performance on tests of cognitive ability (e.g., VSAT and RavenÕs Progressive Matrices; see Cantor & Engle, 1993; Conway et al., 2002; Daneman & Merikle, 1996; Engle, Kane et al., 1999). Estimates of internal consistency reliability, suc as coefficient as and split-alf correlations, are typically in te range of (Conway et al., 2002; Engle, Kane et al., 1999; Kane et al., 2004). Operation span scores are also stable over time (e.g., over a period of tree monts in Klein & Fiss, 1999). Participants were tested individually and solved algebraic equations wile trying to remember series of unrelated words. A pool of 66 matematical operations (stated in te form of a question) was randomly paired wit an equal number of memoranda (ig-frequency, one-syllable, and semantically unrelated words; e.g., Is (8 * 3) 3 = 21? CALF). Te stimuli were previously reported in LaPointe and Engle (1990). Eac operation began wit multiplication or division of two integers, and a tird integer was eiter added to or subtracted from te result. A solution, wic was correct alf of te time, was provided for eac equation. Participants read eac operation aloud, said yes or no to verify its veracity, and read a word. Te experimenter initiated presentation of eac operation-word string by key press and advanced to te stimuli immediately after te participant read te word. Tree question marks followed eac set, wic ad 2 6 memoranda, and cued participants to recall and write te words in serial order on a response seet. Te response seet contained an equal number of blanks for eac trial. Participants were neiter told ow many words to expect on an upcoming trial, nor were tey reminded ow many words occurred on a previous trial. However, guessing was encouraged, and recall was not timed. Tree series of eac set size (for a total of 15 sets) were presented in a random order tat was te same for all participants. Tree additional series of two items eac were practice. Eac participantõs working memory span was te sum of te correctly recalled words for trials tat were recalled in te correct order and witout error (maximum score was 60). Accuracy at solving te matematical operations was monitored but was not taken into account in te calculation of working memory span. However, participants below 85% matematical accuracy were excluded. A large number of participants were prescreened on te operation span task for use in tese experiments and oter researc in our laboratory. Only tose in te upper or lower quartiles were eligible for tese experiments. Experiments 1A and 1B were conducted in one semester, and quartile boundaries were

6 M.F. Bunting et al. / Journal of Memory and Language 51 (2004) determined by operation span scores from 340 participants. Quartiles for Experiments 2A and 2B were determined by operation span scores from 400 participants. In bot cases, te cutoff score for te lower quartile was 10, and upper quartile scores began at 19. Participants were undergraduates from te subject pool of te Psycology Department at te University of Illinois at Cicago and received course credit in excange for participating in te screening or any of te reported experiments. Participants ad Englis as a first language, normal or corrected-to-normal vision, and normal use of te dominant and to make rapid keyboard and written responses. Experiment 1A: Interference from location-terms, a replication of Cantor and Engle (1993) Metod Participants Forty-two participants (23 ig-span and 19 lowspan; M(SD) span scores = (5.14) and 6.79 (2.12), respectively) were selected from a pool of 340 participants prescreened on operation span. Eigt additional participants (2 ig-span and 6 low-span) were excluded for not acieving 63% accuracy on te recognition component of te fan task. Tere were 8, 24, and 32 verification trials per level of fan, respectively, so participants could miss 3 responses at Fan 1, 9 responses at Fan 3, and 12 responses at Fan 4. Materials and procedure Te materials and procedure for te fan task (acquisition and recognition) came from Cantor and Engle (1993). Te tasks were computerized and te experimental session lasted min. In tis and eac subsequent experiment, te time between te Prescreening and te fan task was different for eac participant, but most participants completed bot tasks witin a week and all participants completed bot tasks witin a six week period. Acquisition and retention task. Participants learned 16 sentences of te form, te person is in te location. See Table 1 for te materials. Eac location-term was associated wit two person-terms, and eac person-term was associated wit one, tree, or four location-terms. Hence, person-fan was 1, 3, or 4, and location-fan was 2. Sentences were grouped by te person term (e.g., all sentences for plumber were presented togeter). Te algoritm n(10s) + 10s, were n equals te number of sentences to be displayed, determined te presentation duration per group. Oral recall was tested after eac set was presented once (presentation of a person-term wit a question mark cued recall). A set of propositions was scored as correct if all of te appropriate sentences were recalled in any order and witout te generation of any additional incorrect sentences. Te experimenter monitored recall and indicated by key press weter a set was recalled correctly or not. Tis memorization-recall procedure was repeated until a set was correctly recalled on tree consecutive cycles. Wen tis criterion was met, eac was presented once more. Te number of memorization-test cycles per fan set was recorded for data analysis. Recognition task. Te speeded recognition test ad four blocks of 32 trials eac (16 studied and 16 non-studied). Non-studied (foil) sentences were generated by switcing person-terms between sets of studied sentences so tat tey still reflected te same fan size. Participants verified te veracity of eac sentence by key press (keys 1 and 3 represented yes and no, respectively). Reaction time and errors were recorded for data analysis. Results Data were analyzed from te acquisition and recognition tasks. Te dependent measure for te acquisition task was te number of memorization-recall cycles required to reac criterion. Te dependent measures for te recognition task were reaction time and accuracy. Te independent measures for bot were working memory capacity (i.e., span, ig, and low) and propositional fan size (fan: 1, 3, and 4). Trial type (studied and nonstudied sentences) was an additional independent measure in te recognition task. Acquisition and retention task Te low-span group did not require more memorization-test cycles tan te ig-span group to meet te acquisition criterion. Te mean number of memorization-recall cycles (see Table 2) for eac level of span was submitted to a fan span mixed-design analysis of variance (ANOVA). Tere was a significant main effect of fan, F(2,80) = 12.42, MSE = 0.25, p <.001. Ortogonal pairwise comparisons sowed tat fewer memorization-recall cycles were required on average to reac criterion at Fan 1 (M = 3.12) tan at Fan 3 (M = 3.51) or Fan 4 (M = 3.63). Te difference between Fan 3 and Fan 4 was not significant. (Note: alpa was adjusted downwards by means of te Bonferroni correction for multiple analyses). Tere was neiter a main effect of span nor an interaction between fan and span, F(1,40) < 1.0, MSE = 0.55, ns, and F(2,80) = 1.51, MSE = 0.25, ns, respectively. Recognition task Reaction time. In tis and all experiments, reaction times are for correct responses, and reaction times on incorrect response were omitted from data analysis.

7 610 M.F. Bunting et al. / Journal of Memory and Language 51 (2004) Table 2 Means (and SD) for te number of memorization-test cycles as a function of fan size and span group for all experiments Span group Fan size Experiment 1A Hig-span 3.13 (.43) 3.37 (.51) 3.63 (.61) Low-span 3.11 (.21) 3.68 (.84) 3.63 (.76) M 3.12 (.35) 3.51 (.69) 3.63 (.67) Experiment 1B Hig-span 3.11 (.26) 3.41 (.80) 3.61 (.87) Low-span 3.27 (.59) 3.79 (.72) 3.89 (.53) M 3.19 (.46) 3.60 (.77) 3.75 (.73) Experiment 2A Hig-span 3.05 (.15) 3.27 (.48) 3.48 (.66) Low-span 3.12 (.27) 3.38 (.52) 3.38 (.57) M 3.08 (.22) 3.33 (.50) 3.43 (.61) Experiment 2B Hig-span 3.05 (.15) 3.36 (.58) 3.52 (.61) Low-span 3.22 (.33) 3.39 (.45) 3.74 (.64) M 3.13 (.27) 3.38 (.51) 3.63 (.63) Note. Perfect recall would result in tree memorization-test cycles. Reaction times greater tan 3.0 SDs from eac participantõs mean were replaced by te value at 3.0 SD. Tis adjustment affected less tan 1.0% of scores in tis and eac remaining experiment. We tested te tree-way interaction among fan, span, and trial type, but based on Cantor and Engle (1993),we did not expect tat te fan span interaction would vary by trial type. Terefore, mean reaction times (in milliseconds; see Table 3) were submitted to a fan span trial type mixed-design ANOVA. As predicted and as found in Cantor and Engle (1993), tere was an interaction between fan and span, F(2,80) = 3.30, MSE = 146,983, p <.04. Based on KeppelÕs (1991) formula for repeated measures, te effect size (x 2 ) was.09, wic is a medium Table 3 Means (and SD) for correct response times and error rates to studied and non-studied items on te recognition test of Experiments 1A and 1B for ig- and low-span participants Probe type Response times (in ms) Mean errors Fan 1 Fan 3 Fan 4 Fan 1 Fan 3 Fan 4 Experiment 1A Hig span Studied 1437 (268) 1791 (389) 1809 (375).04 (.21).74 (1.01) 1.57 (1.41) Non-studied 1737 (447) 2011 (460) 1998 (372).39 (.50) 1.22 (1.35) 1.17 (1.15) M 1587 (328) 1901 (406) 1903 (359).22 (.29).98 (.94) 1.37 (.91) Low span Studied 1545 (355) 2144 (671) 2216 (734).11 (.46) 1.63 (1.64) 3.16 (2.22) Non-studied 1985 (401) 2503 (511) 2509 (576).63 (.90) 2.63 (1.98) 3.16 (2.79) M 1765 (329) 2324 (560) 2362 (633).37 (.52) 2.13 (1.41) 3.16 (1.71) Experiment 1B Hig span Studied 1407 (509) 1761 (515) 1814 (430).32 (.72).91 (.97) 1.36 (1.22) Non-studied 1719 (455) 1896 (456) 1901 (439).45 (.67) 1.32 (2.01).91 (1.38) M 1563 (454) 1828 (468) 1857 (396).39 (.49) 1.11 (1.05) 1.14 (.88) Low span Studied 1379 (358) 1819 (476) 1938 (363).09 (.29) 1.00 (1.11) 1.41 (1.71) Non-studied 1758 (463) 1923 (371) 1902 (330).73 (.88) 1.82 (2.34) 1.32 (1.73) M 1568 (383) 1871 (409) 1920 (305).41 (.53) 1.41 (1.53) 1.36 (1.57)

8 M.F. Bunting et al. / Journal of Memory and Language 51 (2004) to large effect (Coen, 1977). A follow-up test to tis interaction confirmed tat te slope of te fan effect for reaction time was greater for low-spans (M = 211, SE = 40) tan ig-spans (M = 113, SE = 24), t(df = 40) = 2.17, p < Tis comparison of slopes for ig- and low-spans is in keeping wit our primary ypotesis tat ig- and low-spans would differ in te magnitude (or slope) of teir fan effects. Cantor and Engle sowed tat te fan effect slope accounted for te relationsip between working memory and verbal abilities (as assessed by te verbal scolastics aptitude test, VSAT) and argued tat te magnitude of te fan effect appears to reflect te same mecanism as te working memory spans. Te pattern of te fan span interaction was similar for bot targets and foils, and tus te test of te full tree-way interaction was not significant, F(2,80) < 1.0, MSE = 42,796, ns. In te absence of any indication tat te results are different for targets and foils, we did not analyze tese items separately. Additional analyses are reported in Appendix A. Error rate. Mean errors during verification (see Table 3) were submitted to an analysis like tat for te reaction time data. A significant interaction between fan and span indicated tat te slope of te fan effect of errors was greater for low- tan ig-span participants, F(2,80) = 8.50, MSE = 1.67, p <.001. Te effect size (x 2 ) was.26, a large effect (Coen, 1977). A follow-up test to tis interaction confirmed tat te slope of te fan effect for errors was greater for low-spans (M = 1.39, SE =.11) tan ig-spans (M =.38, SE =.06), t(df = 40) = 4.52, p <.001. We did not analyze targets and foils separately because te pattern of te fan effect was similar for bot, as indicated by te fact tat te tree-way interaction of fan span trial type was not statistically significant, F(2,80) < 1.0, MSE = 1.92, ns. Additional analyses are reported in Appendix A. Discussion Te low- and ig-span groups acquired te materials in an equal amount of time, but important individual differences emerged in te recognition task. Te lowand ig-span groups revealed fan effects for reaction times and errors wit te increasing number of propositions associated wit te person-terms of te sentences, but te slopes of te fan effects were larger for te low-span group. Tese results are similar to tose of Cantor and Engle (1993, see Experiment 1 of teir 1 By slope, we mean te slope of te linear regression line troug te data points at eac fan level, wic takes into account te fact tat te distance between Fan 1 and Fan 3 is greater tan te distance between Fan 3 and Fan 4. report). Cantor and Engle reported tat individual differences in working memory capacity predicted te fan effect for reaction times, but tey did not find additional differences in te fan effect for errors. It sould be noted tat te studied propositions in Cantor and EngleÕs (1993) experiment were atypical in one respect. In teir experiment and te current Experiment 1A, te propositions at Fan 1 and Fan 3 were nested subsets of te Fan 4 propositions. For example, as seen in Table 1, te locations for lawyer and plumber togeter completely overlap wit te locations for teacer. In oter versions of te fan paradigm (e.g., Anderson, 1983), tere is overlap in te items from different fans, but tese overlaps do not result in complete subset relationsips. Suc subset relationsips could affect te way in wic Fan 4 propositions were represented in memory. In particular, any probe involving a Fan 4 person could be answered by reference to te corresponding Fan 1 and 3 items (e.g., using te above example, any probe involving teacer could be answered by retrieving te locations associated wit lawyer and plumber instead of retrieving te teacer locations directly). Tis implies tat te participants need not store four locations wit te Fan 4 person-terms at all; rater, tey need only store te two related person terms (and rely on te location-fans from tere). Tus, te representation tat participants may be constructing essentially turns te Fan 4 items into Fan 2 items. Te materials in te remaining experiments were constructed to avoid suc subset relationsips. Te results of Experiment 1A are consistent wit te view of working memory capacity as amount of activation, but susceptibility to response competition from te overlap among fan materials, specifically location-terms, may account for te individual differences observed ere. In Experiment 1A, location-fan was eld constant at two wile person-fan varied from one, tree, and four. In Experiment 1B, eac person-term was associated wit a unique location (i.e., location-fan = 1). Two alternative outcomes were foreseen. First, according to a capacity-limited perspective (Cantor & Engle, 1993), te span effects sould remain because personfan as not canged. Tat is, one would predict tat more working memory capacity sould permit faster spreading activation to te multiple location-terms associated wit te person-terms at Fan 3 and Fan 4. An alternative outcome is tat eliminating tis specific form of interference (overlapping location-terms) will at least attenuate, if not eliminate, any effect of span. An attentional control view of working memory would make tis prediction based on researc suggesting tat low spans are prone to interference from irrelevant episodes and spurious connections (Kane & Engle, 2003; Lustig, May et al., 2001). Reports from te participants also lead us to expect te second outcome. Many participants reported tat

9 612 M.F. Bunting et al. / Journal of Memory and Language 51 (2004) tey used mnemonics to elp tem learn te fan propositions. For example, some reported trying to remember just te first letter of eac location associated wit a person-term. So, plumber PCB migt ave elped a participant to remember tat te plumber was in te park, curc, and bank. Based on attentional control teories of working memory capacity, we would not expect igand low-span participants to differ in integration and mnemonic strategies suc as tis (cf. Engle et al., 1992). But, we would expect tem to differ in teir susceptibility to interference from competing responses tat do not fit teir mnemonic cunk. Experiment 1B: No interference from location-terms Metod Participants Forty-four new participants (22 ig-span and 22 low-span; M(SD) span scores = (6.15) and 6.77 (2.56), respectively) were selected from te same pool of participants wo completed operation span for Experiment 1A. Two additional participants (1 igspan and 1 low-span) were excluded from analyses for not acieving 63% accuracy on te recognition task. Materials and procedure Te materials and procedure were uncanged from Experiment 1A, wit te exception tat tere was no overlap among location-terms (location-fan = 1). Studied and non-studied person- and location-terms are sown in Table 1. Results Te independent and dependent variables were te same as tose for Experiment 1A. Acquisition and retention task As in Experiment 1A, te low-span group did not require more memorization-test cycles tan te ig-span group to meet te acquisition criterion. Te mean number of memorization-recall cycles (see Table 2) for eac level of span was submitted to a fan span mixed-design ANOVA. Tere was a fan effect, F(2,84) = 17.19, MSE =.213, p <.001. Ortogonal pairwise comparisons sowed tat Fan 1 (M = 3.19) propositions were acquired faster tan Fan 3 (M = 3.60) or Fan 4 (M = 3.75) propositions. Te difference between Fan 3 and Fan 4 was not statistically significant. (Note: alpa was adjusted downwards by means of te Bonferroni correction for multiple analyses). Tere was neiter a main effect of span nor an interaction between fan and span, F(1,42) = 2.79, MSE =.88, ns, and F(2,84) < 1.0, MSE =.22, ns, respectively. Recognition Task Reaction time. Mean reaction times for correct responses (see Table 3) were submitted to a fan span trial type mixed-design ANOVA. Individual differences in working memory capacity were unrelated to te slope of te fan effect for reaction times. Neiter te test of te full treeway interaction nor te interaction between fan and span were significant, F(2, 84) = 1.02, MSE = 49,226, ns, and F(2,84) < 1.0, MSE = 66,794, ns, respectively. Te effect size for te fan span interaction (wic was te critical effect in Experiment 1A) was.01, and te power (by carts from Pearson & Hartley, 1972) was.88 to detect an effect size (x 2 ) of.09, or te effect size in Experiment 1A. Terefore, we cannot attribute te failure to detect te critical effect to a lack of power. However, tere was a fan effect for reaction times, and te slope of te fan effect was significantly greater tan zero, F(2,84) = 40.96, MSE = 66,794, p <.001. Additional analyses are reported in Appendix B. Error rate. Mean errors (see Table 3) were submitted to a fan span trial type mixed-design ANOVA. As for te reaction time analysis, individual differences in working memory capacity were unrelated to te fan effect slope for errors. Neiter te test of te full tree-way interaction nor te interaction between fan and span were significant, F(2,84) < 1.0, MSE = 1.49, ns, and F(2,84) < 1.0, MSE = 1.83, ns. Te effect size for te fan span interaction was.01, and te power (by carts from Pearson & Hartley, 1972) was greater tan.98 to detect an effect size (x 2 ) of.26, or te effect size in Experiment 1A. As in te case of te analysis of reaction times, we cannot attribute te failure to detect te critical effect to a lack of power. Tere was a significant, positively sloped fan effect for errors, F(2,84) = 11.81, MSE = 1.83, p <.001. Additional analyses are reported in Appendix B. Discussion Te outcome of tis experiment was consistent wit our predictions about working memory capacity and attentional control. Te ig- and low-span groups revealed nearly identical fan effects for reaction times and errors wit increasing person-fan, but te absence of location-based interference eliminated te relationsip between individual differences in working memory and fan. Tat is, te low-span group performed like te ig-span group wen te materials did not contain overlapping location-terms. Te fact tat tere is still a main effect of fan sould come as no surprise. Radvansky, Spieler, and Zacks (1993, see also Radvanksy, 1999) demonstrated tat fan effects can be attenuated and even eliminated wen te materials can be organized into real world situations, or situation models. In Radvansky et al. (1993, Experiment 1), participants did not sow fan effects for

10 M.F. Bunting et al. / Journal of Memory and Language 51 (2004) reaction times or errors wen multiple inanimate objects were represented in a single location (i.e., a single situation model), but tey sowed fan effects wen multiple objects were in multiple locations (i.e., multiple situation models). However, te same was not true wen Radvansky et al. used person-terms in place of inanimate objects (see teir Experiment 3). In tis case, Radvansky et al. sowed tat facts about people being in locations do not readily give rise to person-based or location-based organization, and in fact participants did not spontaneously use situation models to represent people in general locations. Our current experiment is not a direct replication of Radvansky et al. (1993, Experiment 3), but our person- and location-terms are similar, as are our results. Participants in our study were asked to represent a single person in multiple locations. Tis not only does not conform to a real-world situation or an easily conceivable situation model, but our materials were not suggestive of one type of situation over anoter. Power analyses offered reassurance tat we ad sufficient power to detect effects of te same magnitude as te critical effects in Experiment 1A. Noneteless, our conclusions for Experiment 1B are contingent upon furter analysis of te between-experiments effect of interference. Wat we really need to know is weter te differences in slope between ig- and low-span participants canges depending on te interference condition. If an attentional control account of working memory is correct, ten we sould expect to see a difference in te performance for low-span, but not ig-span participants between experiments. Experiments 1A and 1B: Between-experiments analyses Te critical effect in Experiment 1A te fan span interaction was te same for targets and foils, and terefore we collapsed across te trial type variable in te between-experiments analysis of te recognition data. Te dependent measures are reaction time and accuracy in recognition. Te independent measures are working memory span and location-based interference, wic was present in Experiment 1A but not in Experiment 1B. Because we did not observe any relevant differences in learning rates for ig- and low-span participants, furter analysis of te acquisition data were not conducted. Recognition task Reaction time Reaction times for correct responses were submitted to a fan interference span mixed-design ANOVA. Te test of te full tree-way interaction was not significant, F(2,164) = 1.56, MSE = 52,955, ns. However, even in te absence of tat effect, te trends in te data are meaningful. In Fig. 1, ig-spans clearly did not differ between experiments and ad nearly identical reaction times in Experiments 1A and 1B. Low-spans performed muc like ig-spans wen we controlled for location-based interference in Experiment 1B, but reaction times for te low-spans in Experiment 1A are clearly different from te rest. A significant interference span interaction provides statistical support for tese claims as well, F(1,82) = 3.70, MSE = 435,869, p =.05. Tere is a simple effect of interference for low-spans but not ig-spans. Collapsing across fan, low-spans were significantly faster in Experiment 1B witout location-based interference (M = 1786 ms, SE = 72) tan in Experiment 1A (M = 2150 ms, SE = 105), F(1,39) = 8.53, MSE = 158,051, p <.006. Te mean reaction times for ig-spans were statistically equivalent in Experiments 1A (M = 1797 ms, SE = 66) and 1B (M = 1750 ms, SE = 88), F(1, 43) < 1.0, MSE = 133,716, ns. Additional analyses are reported in Appendix C. Error rate Mean errors (see Fig. 1) were submitted to a fan interference span mixed-design ANOVA, and te test of te full tree-way interaction was significant, F(2, 164) = 3.14, MSE =.875, p <.046. For low-spans, te fan effect slope was significantly greater in Experiment 1A, te interference experiment (M =.92, SE =.11), tan in te absence of location-based interference in Experiment 1B (M =.34, SE =.11), F(1,39) = 14.53, MSE =.24, p <.001. Te fan effect slopes for ig-spans were statistically equivalent in Experiments 1A (M =.38, SE =.06) and 1B (M =.27, SE =.08), F(1, 43) = 1.31, MSE =.11, ns. Additional analyses are reported in Appendix C. Discussion and interim conclusion Te preceding analyses of errors and reaction times provide collaborating evidence tat te difference in slopes for ig- and low-span participants depends on te presence of location-based interference. Altoug we probably lacked te statistical power to detect te interaction among fan interference span for reaction times, we argue tat a trend toward te desired effect is apparent in Fig. 1 and supported by oter significant effects. Te significant, between-subjects interaction between interference and span sows tat low-spans responded faster in te absence of location-based interference (Experiment 1B) tan in its presence (Experiment 1A). Critical, too, is te fact tat ig-spans performed equivalently between experiments and like te low-spans in Experiment 1B. Experiments 1A and 1B togeter support te ypotesis tat te ability to control attention in te service of preventing attention to irrelevant episodes, not te amount of LTM activation, accounts for individual

11 614 M.F. Bunting et al. / Journal of Memory and Language 51 (2004) Fig. 1. Between-experiments effects for Experiments 1A and 1B. Mean verification times (in ms) and mean errors for ig- and lowspan participants as a function of fan size (1, 3, and 4) and te presence (Experiment 1A) or absence (Experiment 1B) of interference from overlapping location-terms. Fan is collapsed across trial type (studied and non-studied items). Bars represent standard error of te mean (RT, reaction time). differences in te fan effect. All forms of interference were not eliminated in Experiment 1B, but interference from competing learning episodes was removed. Tis result is consistent wit control accounts of working memory. Lustig, May et al. (2001) argued tat susceptibility to interference from irrelevant episodes is te crux for wy te reading span works te way it does. Te results of Experiment 1B seem consistent wit tat interpretation. Te implications of tis finding for teories of working memory are discussed furter in te General discussion. Experiment 2A: Interference from person-terms Location-fan was manipulated between Experiments 1A and 1B (but was eld constant witin eac experiment) wile person-fan varied from one, tree, and four. We next tested if a similar pattern of results would attain from te manipulation of te presence or absence of overlapping person-terms. Experiment 2A was similar to Experiment 1A wit respect to te fact tat participants experienced interference bot witin a learning episode and across episodes. Tere were significant canges as well. First, te design of Experiment 2A was adjusted to prevent subset relationsips. As discussed in Experiment 1A, Cantor and EngleÕs (1993) smaller fan propositions (Fan 1 and Fan 3) were completely nested witin te largest set (Fan 4). Tis is atypical and may ave inadvertently affected ow te Fan 4 items were represented. Te stimuli for Experiment 2A did not contain complete subset relationsips. To accommodate tis, fan size was canged from 1, 3, and 4 in Experiments 1A and 1B to 1, 2, and 3 in Experiments 2A and 2B. Tere was a second significant cange wit respect to ow te stimuli were grouped during learning. In Experiments 1A and 1B, te propositions were presented for study according to te personterms of te sentences, so tat all of te propositions for a person-term were presented togeter. In Experiments 2A and 2B, te propositions were organized for acquisition according to te location-terms. Location fan varied from 1 to 3, and all of te person-terms associated wit a single location were studied togeter. Te person-terms, but not location-terms, overlapped across sets in Experiment 2A. Te critical difference between Experiments 2A and 2B was te absence of overlapping person-terms in Experiment 2B (person-terms were associated wit just one location). Metod Participants Forty-tree new participants (22 ig-span and 21 low-span; M(SD) span scores = (3.20) and 6.90 (2.11), respectively) were selected from a pool of 400 participants wo completed operation span. Five additional participants (2 ig-span and 3 low-span) were excluded from analyses for not maintaining 63% accuracy on te recognition task.

12 M.F. Bunting et al. / Journal of Memory and Language 51 (2004) Materials and procedure Te experimental session lasted min. Te tasks were computerized, and participants were tested individually. Acquisition and retention task. Participants learned 12 sentences of te form, te person is in te location (see Table 4 for te materials list). Eac person was associated wit two locations, and eac location was associated wit one, two, or tree people. Hence, person-fan was 1, 2, or 3, and location-fan was 2. Sentences were grouped by te location term (e.g., all sentences for train were presented togeter). Te procedure was oterwise uncanged from Experiment 1A. Recognition task. Te speeded recognition test ad four blocks of 24 randomly-ordered trials eac (12 studied and 12 non-studied). Tere was one added criterion for te construction of te fan propositions and nonstudied foils. In order to avoid subset relationsips, te Fan 1 and Fan 2 items were not completely nested witin Fan 3, as was te case in te current Experiment 1A and Experiment 1 of Cantor and Engle (1993). In Experiments 1A and 1B, people were switced between sets of studied sentences so tat tey still reflected te same fan size. In tis experiment, te non-studied items do not reflect te same fan size. Te procedure was oterwise uncanged from Experiment 1A. Results Data were analyzed from te acquisition and recognition tasks. Te dependent measure for te acquisition task was te number of memorization-recall cycles required to meet criterion. Te dependent measures for te recognition task were reaction time and accuracy. Te independent measures for bot were working memory capacity (i.e., span; ig and low) and propositional fan size (fan: 1, 2, and 3). Trial type (studied and nonstudied sentences) was an additional independent measure in te recognition task. Acquisition and retention task As in Experiments 1A and 1B, te low-span group did not require more memorization-recall cycles tan te ig-span group to meet te acquisition criterion. Table 4 Studied and non-studied Person and Location pairs in Experiments 2A (d) 2B() Person-terms Location-terms Boat Park Curc Zoo Store Train (1) Studied Person and Location pairs Lawyer d d Artist d d Plumber d d Teacer d d Fireman d d Coac Carpenter Doctor d d Journalist Cook Realtor Musician (2) Non-studied Person and Location pairs Lawyer d d Artist d d Plumber d d Teacer d d Fireman d d Coac Carpenter Doctor d d Journalist Cook Realtor Musician Note. Propositions were of te form, Te person is in te location.