Serial-positioneffectsforitemsandrelationsinshort-termmemoryTimJones1andKlausOberauer1,21DepartmentofPsychology,UniversityofBristol,Bristol,UK2DepartmentofPsychology-CognitivePsychologyUnit,UniversityofZurich,SwitzerlandTwoexperiments usedimmediateprobedrecall of words toinvestigateserial-positioneffects. Itemmemorywas testedthroughprobingwithasemanticcategory. Relationmemorywas testedthroughprobing with the words spatial location of presentation. Input order and output order weredeconfoundedbypresentingandprobingitemsindifferent orders. Primacyandrecencyeffectsoverinputpositionwerefoundforbothitemmemoryandrelationmemory. Bothitemand relationmemorydeclined over output position. The finding of a U-shaped input position function for item memory rulesout anexplanationpurelyinterms of positional confusions (e.g., edgeeffects). Either theseserial-positioneffectsarisefromvariationsintheintrinsicmemorystrengthoftheitems, ortheyarisefromvariationsinthestrengthofitempositionbindings,togetherwithretrievalbyscanning.Keywords: Serial-position curves; Short-termmemory; Primacy effects; Recency effects; Computationalmodelling.For over a hundred years, researchers haveinvestigated peoples performance on simplememorytaskstobetterunderstandthecognitiveprocesses involved in remembering and forgetting(e.g., Ebbinghaus, 1885/1964; Nipher, 1876, 1878).Onesuchtaskistopresentpeoplewithalistofitems and then ask themto recall the itemsimmediatelyafterpresentation. Theresultstendto showa U-shaped serial-position curve, withbetter memoryfor items presentedearlyinthelist (primacyeffect)andlateinthelist (recencyeffect)thanforthoseinthemiddle.Although the bow-shaped serial-position curveis one of the oldest and best established factsabout short-term memory, no comprehensivetheoryhasyetemergedexplainingit. Thereare,however, anumber ofproposedmechanismsthatcould cause primacy and recency effects. The goalof our present work is to contribute furtherevidence for narrowing down the set of candidatemechanisms. Inparticular, weaimtodistinguishserial-position effects on two different tests ofmemory: itemmemory and relation memory.Itemmemoryreferstorememberingwhichitemshave been presented on a list, irrespective of theirpositions. Relationmemoryadditionallyrequiresmemoryfor relations betweenitems (e.g., theirorder in the list), or memory for relations of itemstosomeaspectoftheircontext(e.g., whichitemhas been presented in which location). Thedistinctionbetweenitemandrelationmemoryistheoreticallyinterestingbecauseit ispossibletomaintain items purely through the temporarilysustained activation of their representations,Addresscorrespondenceto: TimJones, DepartmentofPsychology, UniversityofBristol, 12aPrioryRoad, Bristol, BS81TU,UK.E-mail:Timothy.Jones@bristol.ac.ukork.oberauer@psychologie.uzh.chThis research was supported by a grant from the Economic and Social Research Council (ESRC), Grant RES-062-23-1199. KlausOberauerisnowattheDepartmentofPsychologyCognitivePsychologyUnit,UniversityofZurich,Switzerland.Memory,2013Vol.21,No.3,347365,http://dx.doi.org/10.1080/09658211.2012.726629#2013Taylor&Franciswhilst relation memory arguably requires thebinding of each itemto a context or positionmarker (e.g., Brown, Neath, &Chater, 2007;Brown, Preece, & Hulme, 2000; Burgess & Hitch,1999, 2006; Farrell, 2006, 2012; Henson, 1998).Therefore,disentanglingeffectsofserialpositiononitemmemoryandonrelationmemorypro-vides information on whether serial-positionaffects the memory strength of items (e.g.,throughsustainedactivation) or thestrengthofitemcontextbindings,orboth.InputorderandoutputorderMany common methodologies for investigatingshort-termmemoryconfoundtheorderinwhichitems are encoded (input order) with the order inwhich they are retrieved (output order). Forexample, in forward recall, output order is alwaysthesameasinputorder, andtheresultingcurvetends to showlarge primacy and only a littlerecency, whereas in backward recall, output orderis always the reverse of input order, and theresultingcurvetypicallyshowslargerecencyandonly a little primacy (e.g., Li &Lewandowsky,1993, 1995; Madigan, 1971). This demonstratestheimportanceofoutputorderinshapingserial-position curves. Effects of input order and outputorder needtobe separatedbecause they mostlikely reflect different mechanisms. Effects ofinput order are likely to arise fromprocessesduringencoding, whereaseffectsofoutputordermust arise fromprocesses duringretrieval. Ex-perimentsinvestigatingtheeffectsofinputorderand output order separately (Cowan, Saults,Elliott, &Moreno, 2002; Oberauer, 2003)revealthat they have different effects on memoryperformance. In this article, we focus on theeffects of input serial positionfor tworeasons:First, there is a richer set of theoretical ideasabout thesources of input-positioneffects (seeTable1)thanforthoseofoutput-positioneffects(for a review of hypothetical sources of both kindsof effects, see Oberauer, 2003). Second, input-positioneffectsaremoregeneral becausemem-ory for lists always involves input order, butoutput order comes intoplay only whenmorethanoneitemistested.Several methods canbe used to control foreffects of output order. One is to test for only oneitempertrial (e.g., Davelaar, Goshen-Gottstein,Ashkenazi, Haarman, & Usher, 2005; Hay,Smyth, Hitch, & Horton, 2007; McElree &Dosher, 1989; Monsell, 1978). Another methodis to test all of the items from a trial, one at a time,TABLE1Potential loci ofprimacyandrecency,andthepredictionsofthegenericmodel foritemandrelationmemoryTypeofmemory Locusofserial-positioneffects Retrievalmechanism Serial-positioncurveItemmemory Itemactivation Directcueingwithcategory UItempositionbindings FlatPositionaldistinctiveness FlatItemactivation Scanninginputpositions UItempositionbindings UPositionaldistinctiveness Inverted-URelationmemory Itemactivation Directcueingwithspatiallocation UItempositionbindings FlatPositionaldistinctiveness FlatItemlocationbindings UaPositionlocationbindings FlatSpatialdistinctiveness FlatItemactivation ScanninginputpositionsbUItempositionbindings UPositionaldistinctiveness UItemlocationbindings FlatPositionlocationbindings USpatialdistinctiveness UNote. Boldfont indicatesviablecombinationsof mechanismsaccordingtoourdata. Positionreferstoserial input position.Locationreferstospatiallocation.UU-shapedserial-positioncurvewithprimacyandrecency.Flat flatserial-positioncurvewithnoprimacyandnorecency.Inverted-Uupsidedownserial-positioncurve,withbettermemoryformiddlepositions.aOnly if bindings are stronger at extreme input positions, not extreme spatial locations.bInall cases the scanning retrievalmechanismshiftstheserial-positioncurvesalittletowardsprimacy.348 JONESANDOBERAUERinadifferentorderfromtheirorderofpresenta-tion, suchthatinputorderandoutputorderareuncorrelated (Cowan et al., 2002; Oberauer,2003). Inthe present experiments we usedthesecond method, for reasons that are explainedbelow.Serial-positioneffectsintestsof itemmemoryandrelationmemoryPrevious experiments testing input serial-positioneffectssuggestthatitemmemoryischaracterisedby a pronounced recency advantage, with little orno primacy effect, whereas relation memorydemonstratesamorebow-shapedcurve. Amea-sureofitemmemorycanbeobtainedbytestsofitemrecognitionthataskwhetheraprobestimu-lus has been presented in the memory list,regardlessofitsrelationtootherlistitemsortoa particular position in the list. When a single itemis probedshortlyafter encodingamemorylist,serial-position curves typically show extendedrecency and only a small, one-item primacy effect(e.g., Hayetal., 2007; McElree&Dosher, 1989;Monsell, 1978). Whenitemrecognitionistestedbyaseriesofprobesratherthanasingleprobe,suchthat test order is unconfoundedwithpre-sentationorder, therecencyeffect is still largerthantheprimacyeffect(Oberauer,2003).Itemmemory can also be tested by probedrecall usingcues that arealreadyrelatedtotheitems throughapermanent associationinlong-termmemory, rather thanthroughatemporaryassociation formed on the current trial. Forinstance, Davelaar et al. (2005) presentedalistofwordsthatallbelongedtodifferentcategoriesand then probed for a particular item by present-ingitscategoryasacue.Thisisthemethodusedtotest itemmemory inthe current study. Thesemantic links between items and categoriesareestablishedinlong-termmemory, sothatthecorrect response can be determined fromthelong-termsemanticassociationbetweencategoryanditem, togetherwithmemoryforwhichitemshavebeenpresentedonthecurrentlist.Memoryfortherelationsofthecurrentlistitemstoeachother, ortotheirlistpositions, areirrelevantforperformance. Davelaar et al. (2005)foundclearrecencyandnoprimacywhenprobingforsingleitems froma list by category name at typicalpresentation rates (800 ms per item); primacyemerged only at very rapid rates (300 ms per itemor faster). These findings suggest that, at leastwith conventional presentation rates, tests of itemmemory result ina strong recency effect, withlittleornoprimacyeffect.Relationmemorycanbetestedwithasingleprobeindicatingtherequesteditemsserialposi-tion (e.g., Farrell & Lelie` vre, 2009; Nairne, Ceo, &Reysen, 2007), orthespatial locationassociatedwiththerequesteditem(e.g., Avons, Wright, &Pammer,1994).Ineithercase, theresulttendstobeabow-shapedserial-positioncurve,withmorepronouncedprimacythanforitemmemory.This comparisonof findings intheliteraturesuggeststhattherecencyeffectreflectsprimarilyan advantage in item memory (activationstrength)forthemorerecentlypresenteditems,whereastheprimacyeffectreflectsanadvantageinrelationmemory(itemcontext bindings) forthe initially presentedlist items. However, thiscomparison is compromised by two challenges,whichweaddressnext.How to separate and compare itemmemoryandrelationmemoryAn effort to separately assess serial-positioneffects on itemmemory and relation memoryfaces twochallenges. Thefirst challengeis thattests of item memory and tests of relation memoryoften differ not only in whether they require itemmemory or relationmemory, but alsoinotherregards, whichconfoundtheircomparison. Suchpotential confounding variables include the methodoftesting(recall vs. recognition), thenumberofitemstested(asingleitemortheentirelist),andpeoples expectancy (i.e., knowing in advancewhether item memory or relation memory will betested). Thesevariables arelikelytoaffect theshapeoftheserial-positioncurve(e.g.,Bhatarah,Ward, &Tan, 2008; Duncan&Murdock, 2000;Oberauer, 2003).The second challenge arises fromthe task-purity problem: We cannot be certain that tests ofrelation memory actually draw on relationalrepresentations (i.e., bindings), and we cannotbecertainthattestsofitemmemorydonotrelyonrelationalrepresentations. Forinstance, atestof relation memory that has been much studied intheshort-termmemoryliteratureisserial recall,requiring memory for items in their order ofpresentation. Order memory is explained in manymodelsofserialrecall bybindingsbetweeneachSERIALPOSITIONFORITEMSANDRELATIONS 349itemandapositionmarker (e.g., Brownet al.,2000; Burgess&Hitch, 1999, 2006; Farrell, 2006,2012; Henson, 1998). However, forward serialrecall can also be accomplished without suchbindings, by establishing an activation gradientacrosstheitemrepresentations,suchthatsucces-sivelistitemsareactivatedtoadecreasinglevel(Grossberg&Stone,1986;Page&Norris,1998).Conversely, testsofitemmemorymightnotrelyexclusivelyonthememorystrengthofitems.Forexample,onewaytomeasureitemmemoryistorequest that participants recall items in theirappropriatepositions, but thenscoreanitemascorrect even if it is produced in the wrongposition, aslongasitwasonthelist(e.g., Bjork&Healy, 1974; Fuchs, 1969). Withthis method,recall is almost certainlydependent onmemoryforitempositionbindings.In the present study we address these chal-lenges as follows: Wetestedbothitemmemoryandrelationmemory by probedrecall. Partici-pants encoded a list of words presented in randomorderacrossavertical arrayof spatial locations.Each word belonged to a different semanticcategory. Relation memory was tested by probingevery word from the list with its spatial location inanewrandomorder(Oberauer,2003).Recallinganitembyits spatial locationrequires memoryfor the relation between each word and its spatiallocation. Itemmemory was tested by probingevery word from the list with its semanticcategory in a new random order (Davelaaret al.,2005). Thistestrequiresmemory forwhichwordswerepresentedinthecurrentlist,becauseovertheentireexperimentseveralwordsbelong-ing to each category are presented, but it requiresnomemoryfor anyrelations of thecurrent listwordstootherwordsortotheircurrentcontext.The relationbetweena category probe anditsassociatedlistwordispartofsemanticlong-termmemory and therefore does not have to be held inshort-termorworkingmemory.This experimental paradigmequates tests ofitemandofrelationmemoryfortestingmethod(i.e., probedrecall) andnumber of items tested(i.e., the entire list). To control test expectancy, inExperiment 1 we included a postcued condition inwhich the kind of probe (spatial probes orcategoryprobes)wasrevealedonlyafterencod-ing(Duncan&Murdock, 2000). Byrandomisinginputorderandoutputorderseparatelyoneachtrial, wedeconfoundtheeffectsof input andofoutputorder(Oberauer, 2003); inthisarticleweconcentrateontheeffectofinputserialposition.Presentationof itemsinrandomorderacrossthespatiallocations also ensures thatinput orderis uncorrelated with spatial location. For thisreasonpeople cannot easily rely ona primacygradient overinput ordertorecoverwhichitemwasinwhichspatiallocation. Therefore, ourtestof relation memory arguably tests temporarybindings betweenwords andtheir spatial loca-tions. Conversely, our test of itemmemory isarguably a fairly pure test of item memory,becausepeoplesscoredependsonlyonwhetheror not they remember which member of the probecategorywas presentedonthecurrent list. It isnotnecessarytorecoverthebindingsbetweenawordanditsinputpositionorspatiallocationtoachieveitemrecall.Theseefforts toaddress thetask-purity pro-blemnotwithstanding, wecannotbecertainthatrecall of the itemin a given spatial locationactuallyreliesonitemlocationbindings, orthatrecall of theitembelongingtoagivencategorydoes not rely on any itemcontext bindings.Therefore, wedonot relyonour twokinds oftestaspuretestsofitemmemoryandofrelationmemory, respectively. Rather, we consider arange of possible retrieval processes throughwhichpeoplecouldperformthesetwotestsandimplement these processes in computational mod-els. Themodelscombinethesealternativeretrie-valprocesseswithvariousmechanismsthathavebeenproposedasexplanationsforserial-positioneffects at encoding. Weturntotheseproposedmechanismsnext.Proposedcausesofserial-positioneffectsatencodingThere are several proposed causes of primacy andrecencyeffectsoverinputpositioninimmediatememory.Thecurrentstudyaimstoconstrainthebroadlocus where serial-positioneffects occur,rather than specify their exact mechanistic causes.For example, serial-position effects might beestablishedover thestrengthof itemrepresenta-tions; overthestrengthofbindingsbetweenitemand context representations (i.e., bindings be-tweenitems andinput-positionmarkers, or be-tween items and spatial locations); or through thedistinctiveness of context representations. Afac-torial combination of these possibilities withvarious retrieval mechanisms leads to a largenumber of potential sources of serial-position350 JONESANDOBERAUEReffects (seeTable1). Webelievethat rejectinganyimplausiblecombinations is avaluablestepinnarrowing down the potential mechanismsinvolved.Here,thepotentialcausesandretrievalprocesses considered are outlined. Following pre-sentationof theempirical results, wereport theoutcomeof modellingall possiblecombinationsof loci of serial-position effects with differentretrievalmechanismsandruleoutcertaincombi-nationsofmechanismsasimplausible.Strengthofitemrepresentations. Onepotentialcause of serial-position effects is the strength (e.g.,activation)ofitemrepresentations.Itemstrengthcouldvarythroughtheamountofattentionpaidtoeachitem(e.g.,Farrell&Lewandowsky,2002;Page&Norris, 1998). Ithasoftenbeenassumedthatmoreattentionispaidtoearlyitemsinalist(e.g., Brownet al., 2000), possiblyduetotheirnovelty (e.g., Farrell & Lewandowsky, 2002),creating a primacy gradient in itemstrength,whichshouldtranslate intoaprimacyeffect inbothtests of itemmemoryandtests of relationmemory. Arecency gradient in itemstrengthcouldarisefromretroactiveinterferenceduringencoding.Thisoccurswhentheencodingoflateritems interferes withthemaintenanceof earlieritems, perhapsbecauselateritemssteal featuresfrom earlier items (e.g., Nairne, 1990; Oberauer &Kliegl, 2006). A recency gradient in item strengthshouldgeneratearecencyeffectintestsofitemandofrelationmemory.Strengthofitemcontextbindings.Thestrengthofitemcontextbindingsmayalsobe affectedbyattention during encoding. The oscillator-basedassociativerecall model (OSCAR; Brownet al.,2000) assumes a gradual decline over inputposition in the learning rate for associationsbetweenitemsandpositionmarkers. Thiswouldlead to a primacy gradient if and only if items areretrievedthroughassociations withtheir input-positionmarkers. The strengthof itemcontextbindingscouldalsoformarecencygradient(i.e.,strongerbindingstowardstheendofthelist), ifretroactive interference impairs itemcontextbindings (Oberauer & Kliegl, 2006). In thepresentparadigm, agradientofbindingstrengthcanapplyeither tobindings betweenitems andmarkers of their input position, or tobindingsbetweenitemsandtheirspatiallocations.Contextual distinctiveness. Intheliteratureonimmediate serial recall, primacy and recencyeffectsareoftenexplainedthroughthedifferentdistinctiveness of their contexts. These distinctive-nessmechanismsassumethatitemsareretrievedbycueingwiththeirassociatedcontext markers.For instance, if items are cued by positionmarkers, the positional distinctiveness of thesemarkerswillaffectthelikelihoodofordererrorsat eachinput position. Initssimplest form, thiskindof explanationappealstoedgeeffects: Thefirst and last items in a list have only oneneighbouring itemalong the input position di-mension, whereasitemsinthemiddleofthelisthave two neighbours. This means that the first andlast items have less chance of having theirpositionconfusedwithneighbouringitems (e.g.,Brownet al., 2007; Brownet al., 2000; Henson,Norris, Page, &Baddeley, 1996; Lee &Estes,1977).Amore sophisticated version of positionaldistinctiveness is implemented in Hensons(1998) startendmodel. The positional contextis formed by a combination of two activationgradients: a start gradient, with activationdecreasing exponentially fromthe start of thelist, and an end gradient, with activationdecreasing exponentially fromthe end of thelist. Recall isinitiatedbyreinstatingthecontextfor the itemat a certain position. Due to theexponentialshapeofthegradients,thecontextismost distinctivefromothercontextsat thestartandtheendof thelist. Therefore, thereis lesschance of confusion between these items andothers with neighbouring contexts, leading tobetter recall for start and end items (primacyand recency effects). Henson (1998) suggestedthat startendcontext markers donot have todenoteinput position, but couldalsodemarcateother relevant contexts (e.g., spatial locationsdemonstrating spatial distinctiveness), if theseare more suitable for the task. As our test ofrelation memory involves probing with spatiallocationcues, weconsidertheeffectsof distinc-tiveness across boththeinput positionandthespatiallocationinourcomputationalmodels.Afurthervariant of positional distinctivenesshas beenproposedbyBotvinickandWatanabe(2007). In their model of serial recall, representa-tions of numerical rank serve as position markers.Basedonsingle-cell recordingsof rank-sensitiveneuronsintheparietalcortex,theyassumedthatthe distinctiveness of rank representations de-creased with increasing numerical rank. Thistranslatesintoaprimacygradientondistinctive-nessofpositionmarkersforinputposition.SERIALPOSITIONFORITEMSANDRELATIONS 351Retrieval mechanisms. We consider two me-chanismsbywhichitemsareretrievedintestsofitemmemory. The first is whenthe temporaryactivation of items above a threshold leveldiscriminates them from less activated representa-tions ofother items (e.g., Farrell & Lewandowsky,2002; Page &Norris, 1998). In this case, onlyvariabilityinitemactivationlevels couldcauseserial-positioneffects. Alternatively, itemmem-orycanbeestablishedby bindingallitemsofthecurrent trial to a context representation thatdistinguishes the current trial fromprecedingtrials(e.g., Brownetal., 2007; Burgess&Hitch,2006; Henson, 1998; Lewandowsky &Farrell,2008). The context representation is then usedtocuetheassociateditems, andthestrengthofbinding to items can also contribute to serial-positioneffects.The most obvious mechanism for rememberingthe relation between an itemand its spatiallocationover the short termis by maintainingdirect bindingsbetween item representationsandrepresentations of corresponding locations. In thiscase, thestrengthofitemlocationbindings, andthespatialdistinctivenessoflocationrepresenta-tions, would determine recall. However, it ispossiblethatbothitemsandlocationsareboundtopositionmarkerscodingtheirinputorder,andpeople scan through the input-position markers toprobe for both items and their locations (e.g.,Cowan, Saults, &Morey, 2006). Inthiscase, anydirectbindingbetweenspatiallocationsanditemrepresentationswouldbeirrelevantforretrieval.Serial-position curves would be determinedinstead by the strength of bindings betweeninput-positionmarkers anditems, andbetweeninput-positionmarkers andspatial locations, aswell as thedistinctiveness of theinput-positionmarkers.ThepresentexperimentsOur study aims to constrain the possible mechan-isms that contribute to serial-position effectswhen encoding a list of items into short-termmemory. To this end, we provide a controlledempiricalseparationoftestsofitemandrelationmemoryandcomparetheresults tothepredic-tionsof computational modelsof variouspoten-tial memory mechanisms. We used Oberauers(2003)randomprobedrecall methodfordecon-founding input and output order. The methodusedfortestsofitemandrelationmemorywereidentical until retrieval. Categorycuesprovideatest of item memory, whereas spatial location cuestestrelationmemory.We report two experiments: Experiment 1aimed to investigate the effects of probe type(i.e., probing itemmemory withcategory cues,andprobingrelationmemorywithlocationcues)onprimacyandrecencyeffectsoverinput posi-tion. In addition, the impact of strategic encodingwas investigatedbyincludingamanipulationofpre- versus postcueing. To foreshadow, we ob-tainedU-shapedserial-positioncurveswithbothprimacyandrecencyfortestsofitemmemoryaswell as tests of relationmemory, different fromDavelaar et al. (2005). Experiment 2examinedtheimpact of proactiveinterferenceonprimacyand recency over input position to test onepossible cause for the difference between ourserial-positioncurvesforitemmemoryandthoseofDavelaaretal.EXPERIMENT1:PROBETYPEANDSTRATEGICENCODINGThe goal of Experiment 1 was toseparate thecontributions of itemand relation memory toprimacyandrecencyeffects over input positiononmemoryforlists. Weusedcategoryprobestotest itemmemoryandspatial-locationprobestotest relation memory. To isolate serial-positioneffects of input position fromthe confoundingeffectsofoutputorder,wetesteditemsinanew,randomorder that was uncorrelatedwithinputorder.Toexplorepossibleeffectsofstrategicencod-ing based on task demands (e.g., Duncan &Murdock, 2000), participants learnt whethertheywouldbetestedbyacategoryprobeor aspatialprobeeitherbeforelistpresentation(pre-cueing)orafterlistpresentation(postcueing).Weexpectedthat itemmemory wouldshowmostly recency, because this is what previousinvestigations havefoundusingcategoryprobes(Davelaar et al., 2005) or item recognition(Monsell, 1978; Oberauer, 2003). For relationmemory (positional probes) we expected primacyandrecencyeffectsoverinput positionof aboutequal size, or evenmoreprimacythanrecency,based on previous findings fromserial recall,probed recall, and relational recognition (e.g.,Cowan et al., 2002; Farrell &Lelie` vre, 2009;Healy, 1974; Oberauer, 2003). As notedinthe352 JONESANDOBERAUERintroduction, the comparison of item memory andrelation memory across these experiments iscompromisedbyvariousconfounds. Thepresentstudyprovides adirect, within-experiment com-parisonof atest of itemmemoryandatest ofrelationmemoryavoidingtheseconfounds.MethodParticipants. Sixteen participants, 15 womenand1manbetweentheagesof17and35years,took part in Experiment 1. They were all psychol-ogyundergraduates at theUniversityof BristolwhowerenativespeakersofEnglishandpartici-patedtoreceivecoursecredit.Design. Each trial included seven Englishwords fromseven different sets (described intheMaterials sectionbelow), presentedvisuallyinsevendifferentboxesarrangedverticallyonacomputer screen. All seven words were thenprobed for recall using either itemprobes orspatial position probes. The probe type wasindicated either before the trial began (precueing)orafterthewordshadbeenpresented(postcue-ing). The combination of pre/post cueing andprobetyperesultedinfour within-subjects trialtypes. The spatial position of each word wasdeterminedat randomoneachtrial. Theorderinwhichwordswereprobedforrecall(i.e., theiroutputorder)wasalsovariedrandomlyoneachtrial, withtheconstraint that everycombinationofinputpositionandoutputposition(e.g., wordat input position3 probedfor recall at outputposition 2) occurred four times for each trial type.Foreach trial type thisrequired196 (774)words to be recalled, using 28 trials, resulting in atotal of 112 trials. The type of cueing wasmanipulatedbetweensessions. Participants tookpart in two sessions with 56 trials in each. Half oftheparticipantstookpartintheprecuedtrialsintheirfirstsession, andtheotherhalfcarriedoutthe postcued trials in their first session. Within theprecued trials for each session, half of theparticipants did the 28 category-probed trials first,whilst the other half didthe 28 spatial-probedtrials first. Inthepostcuedsession, probetypesoccurredinrandomorder. Categorieswerecho-sen randomly for each trial, with no categoryrepeatedonconsecutivetrials. Wordswerecho-sen at randomfrom10 possibilities for eachcategory. The first session was preceded byeight practicetrials, andthesecondsessionwaspreceded by four practice trials, all of which wereexcludedfromthedataanalysis.Materials. Wordsof oneortwosyllableswereselectedfromthecategorynorms of VanOver-shelde, Rawson, &Dunlosky(2004). Wecreated17 sets of 10 words belonging to 17 categories; anattempt was madetominimisethevariationofwordfrequencywithineachcategoryusingfre-quencies fromthe CELEXdatabase (Baayen,Piepenbrock, &vanRijn, 1995). Eachtrial wasconstructed by selecting seven sets at random,avoiding repetitions of sets across consecutivetrials, and then selecting from each set one of the10 words at random. Before the experimentbegan, theparticipant was presentedwithalistof all 170words tobeusedintheexperiment,grouped into the 17 categories, to clarify thecategorymembershipofthewords.Procedure. Eachtrial beganwithacuetothetrial type (CATEGORIES, QUESTIONMARKS, or START) displayed in the middleofacomputerscreenusingblacktextonawhitebackground. Participants werecuedwithcate-gories forthecategory-probedtrials, questionmarks forthespatial-probedtrials, andstartfor the postcued trials. After 2 s, this cuedisappeared, and500ms later sevenblackrec-tangles appeared, one above the other, left of thecentreof thescreen. Followinga500-ms pause,the seven words were presented, one at a time, indifferent boxes ina randomorder. Eachwordremainedvisible for 800 ms; the offset of onewordcoincidedwiththeonsetofthenextword.After the last word disappeared, and a further 700ms pause, thefirst probeappeared. For spatialpositionprobes,thisinvolvedthepresentationofa red question mark in one of the boxes, in whichcasetheparticipantwasexpectedtosayoutloudthe word that appeared in that box for theexperimenter to write down. For category probes,a categoryname would appearin red tothe rightoftheboxes,andtheparticipantwasexpectedtosayoutloudthewordinthememorylistbelong-ingtothatcategory. Participantswereinstructedtosaydontknow iftheycouldntrememberaword. Once the word had been recalled, theparticipant pressedthespacebar, andafter 200msthenext probeappeared, andsoonuntil allseven words had been recalled. After recalling thefinal word ofthetrial and pressingthespace bar,there was a 3.5 s pause with a blank white screen,andthenthenexttrialbeganwiththedisplayofthetrialtypecue.SERIALPOSITIONFORITEMSANDRELATIONS 353Before each block of precued trials, an on-screenmessage madeclear howthe next28 trialswouldbeprobed:withcategoriesorquestionmarks. Participants were tested for two 60-minute sessions consisting of four blocks each,andtheywereencouragedtohaveabreakaftereachblock.ResultsAnanalysis of variance(ANOVA) was carriedout with four within-subjects factors: pre/postcueing, probe type, input position, and outputposition, and proportion of answers correct in theappropriatepositionas thedependent variable.The results up to quadratic contrasts are pre-sentedinTable2.Figure1shows that therewereroughlysym-metrical primacyandrecencyeffects over inputposition. Thelackof aninteractionwithprobetype indicates that the shape of the input positioncurvewassimilarforspatialprobesandcategoryprobes. Serial position did interact with thetypeof cueing, however: Precueing resulted inanelevatedprimacyeffect (seeFigure1). Thisobservation is reflected in the significant maineffect of cueing and the interaction of cueing withthe linear contrast of input position. The lack of athree-way interaction between cueing, probe type,and input position suggests that the effect ofprecueingwasthesameforbothprobetypes.TABLE2ResultsofANOVAforExperiment1,testingeffectsofpre/postcueing,probetype,inputposition,andoutputpositionVariables Contrast F MSE p g2pPre/Post 5.48* 0.281 .033 .268Probetype 142.52*** 0.557 B.001 .905Input L 0.158 1.025 .697 .01Q 41.113*** 0.19 B.001 .733Output L 147.889*** 0.16 B.001 .908Q 72.573*** 0.041 B.001 .829Pre/PostInput LL 6.834* 0.188 .02 .313LQ 1.65 0.08 .218 .099ProbeTypeInput LL 3.415 0.072 .084 .185LQ 2.306 0.134 .15 .133Pre/PostOutput LL 0.038 0.102 .847 .003LQ 0.202 0.073 .66 .013ProbeTypeOutput LL 0.694 0.11 .418 .044LQ 0.538 0.059 .474 .035OutputInput LL 81.757*** 0.025 B.001 .845LQ 10.383** 0.08 .006 .409QL 17.914*** 0.05 .001 .544QQ 5.516* 0.035 .033 .269Pre/PostProbeTypeInput LLL 1.148 0.078 .301 .071LLQ 1.889 0.052 .189 .112Pre/PostProbeTypeOutput LLL 0.208 0.031 .665 .014LLQ 0.306 0.033 .588 .02Pre/PostOutput Input LLL 0.114 0.03 .74 .008LLQ 1.242 0.041 .283 .076LQL 3.861 0.035 .068 .205LQQ 3.465 0.035 .082 .188ProbeTypeOutput Input LLL 0.051 0.077 .825 .003LLQ 15.602*** 0.045 .001 .51LQL 0.002 0.057 .962 0LQQ 4.55* 0.049 .05 .233Pre/PostProbeTypeOutputInput LLLL 1.037 0.068 .325 .065LLLQ 0.173 0.06 .684 .011LLQL 4.627* 0.034 .048 .236LLQQ 0.913 0.05 .354 .057Note. Thedegrees of freedominall cases are(1, 15). ANOVAanalysis of variance. g2p partialetasquared; L linearcontrast; Qquadratic contrast; LQinteractionof linear contrast andquadratic contrast; Pres. rate presentationrate;Input inputposition;Output outputposition.*pB.05,**pB.01,***pB.001.354 JONESANDOBERAUERThere was a clear decline in accuracy overoutput position, and an interaction between inputposition and output position, with later inputpositionsbeingmorestronglyaffectedbyoutputinterference.Themaineffectofprobetypereflectedmuchbetterrecallwithcategoryprobesthanpositionalprobes. Therewerealsosomesignificant higherorder interactions that are difficult to interpret, soarenotdiscussedhere.For a more direct comparisonwithpreviousstudiesthatprobedonlyoneitempertrial (e.g.,Davelaar et al., 2005), the effects of type ofcueing, probe type, and input position wereinvestigated for only the first output position.Thesignificant contrasts fromthis ANOVAarereportedinTable 3. The significant linear andquadraticcontrastsoverinputpositionreflectaninput position curve with both primacy andrecency; the recency effect was more pronouncedthan the primacy effect (see Figure 2). Thequadratic contrast of input position interactedwithtypeof cueingandwithprobetype. Theseinteractionsreflectthefactthatthemiddleitemssuffered most from postcueing (compared toprecueing) and from using spatial location probes(comparedtocategoryprobes).DiscussionThere were clear serial-position effects over inputposition for both item and relation memory, eventhoughmemoryof input positionisnot directlyrequired to complete either task. A comparison ofserial-position curves over input position for itemandrelationmemory showednodifferences inprimacy and recency when averaging over alloutputpositions,butslightlygreaterprimacyandrecency for relation memory than item memory atthe first output position. The first output positionresulted in best recall, and itemmemory wasgenerally superior to relation memory, so that theserial-positioncurveforitemmemorywhenonlythe first-probeditem was considered was closetoceiling*this alone could have flattened thisserial-positioncurverelativetothat forrelation1 2 3 4 5 6 700.10.20.30.40.50.60.70.80.91Input PositionProportion CorrectPre-Cued CategoriesPost-Cued CategoriesPre-Cued PositionsPost-Cued PositionsFigure 1. Accuracy as a function of probe type, pre/postcue-ing, andinput positioninExperiment 1 averagedover alloutput positions. Error bars represent within-subjects standarderrors(e.g.,Bakeman&McArthur,1996).TABLE3ResultsofANOVAforExperiment1,testingeffectsofpre/postcueing,probetype,andinputpositionatOutputPosition1Variables Contrast F MSE p g2pPre/post 4.03 0.097 .063 .212Probetype 137.627*** 0.085 B.001 .902Inputpos. L 10.257** 0.16 .006 .406Q 93.082*** 0.04 B.001 .861Pre/Post ProbeType 0.352 0.067 .562 .023Pre/Post Input LL 2.119 0.061 .166 .124LQ 12.954** 0.038 .003 .463ProbeTypeInput LL 3.064 0.042 .1 .17LQ 14.887** 0.057 .002 .498Pre/Post ProbeTypeInput LLL 2.288 0.064 .151 .132LLQ 0.133 0.039 .72 .009Note. Thedegrees of freedominall cases are(1, 15). ANOVAanalysis of variance. g2p partialetasquared; L linearcontrast; Qquadratic contrast; LQinteraction of linear contrast and quadratic contrast; Pre/post pre/postcueing; Input inputposition.*pB.05,**pB.01,***pB.001.SERIALPOSITIONFORITEMSANDRELATIONS 355memory. Therefore, theconservativeconclusionat this point is that serial-position curves for itemmemoryandfor relationmemorywereroughlyparallel.The presence of a quadratic contrast over inputpositionwithout alinearcontrast indicatessym-metrical primacyandrecencyforbothitemandrelation memory. Previous experiments havefound predominantly recency for tests of itemmemory (e.g., Hay et al., 2007; McElree & Dosher,1989; Monsell, 1978; Oberauer, 2003), andDavelaar et al. (2005) found only recency with noprimacy. Twomethodological features might beheldresponsiblefor thedifferencebetweenourresultsandthoseofpreviousstudies. Inpreviousstudies participants had advance knowledge of thekindof probe, andtests of itemmemoryofteninvolved the output of a single item per trial (e.g.,Davelaar et al., 2005; Hay et al., 2007; McElree &Dosher, 1989; Monsell, 1978).Thefirstexplanationcanberuledoutbyourfindingthat theprimacyeffect was, if anything,even stronger in the precued condition than in thepostcuedcondition. Thesecondexplanationap-pears more promising. When we limited theanalysis tothefirst output positionfor adirectcomparisontoexperiments testingonlyasingleitem, we obtained more pronounced recency thanin the overall analysis. A small primacy effect wasstill foundfor bothitem-memory andrelation-memorytests. Thismatchesmanyprevioustestsof item memory (e.g., Hay et al., 2007; McElree &Dosher, 1989), although not Davelaar et al.(2005), whosemethodologywas most similar toours and resulted in no measurable primacyeffect. The difference between our results andthose of Davelaar et al. is explored further inExperiment2.Therewasageneral declineinaccuracywithoutput position(Cowanet al., 2002; Oberauer,2003), confirming a role of output interferenceinitemand relation memory. The interactionbetweeninput positionandoutput positionsug-gests that output interference had different effectsdepending on the input position. Like Cowanetal. (2002), wefoundthat output interferenceimpacted onthe recency portionof the serial-position curve more strongly than the primacyportion. This explains why the serial-positioncurve at the first output position, which isunaffected by output interference, shows lessprimacy and more recency than the serial-positioncurveaveragedoveralloutputpositions.Toconclude, weobtainedessentiallyparallelserial-position curves over input position for testsof itemmemory andof relation memory. Thisresultgoesagainstourexpectation,gleanedfromtheliterature, thatitemmemoryischaracterisedmore by recency, whereas relation memory ischaracterised by serial-position curves that aresymmetricallyU-shaped, or evenhavestrongerprimacythanrecency. Our results therefore donot support the speculation that the recency effectintestsrequiringbothitemandrelationmemoryisprimarilydrivenbyitemmemory, whereastheprimacy effect is primarily driven by relationmemory. Weexplorethetheoreticalimplicationsof our findings more thoroughly through asystematicexplorationof computational models,describedafterExperiment2.EXPERIMENT2:PROACTIVEINTERFERENCEDavelaar et al. (2005) suggestedthat long-termmemorymaycontributetotheprimacyeffectintests of short-term memory, perhaps becauseearlier items have a greater opportunity forrehearsal (e.g., Tan &Ward, 2008)and are there-fore more likely to be stored into long-termmemory (e.g., Rundus, 1971). In their Experiment2, Davelaaretal. usedasmall pool ofpotential1 2 3 4 5 6 700.10.20.30.40.50.60.70.80.91Input PositionProportion Correct Pre-Cued CategoriesPost-Cued CategoriesPre-Cued PositionsPost-Cued PositionsFigure 2. Accuracy as a function of probe type, pre/postcue-ing, and input position for Experiment 1 at only the rstoutput position. Error bars represent within-subjects standarderrors(e.g.,Bakeman&McArthur,1996).356 JONESANDOBERAUERwordsforeachcategoryandrepeatedcategoriesfromtrial to trial to create a high level ofproactiveinterference(whereinterferencefromwords on previous trials confuses retrieval for thecurrent trial). They suggested that proactiveinterferenceimpairsrecall fromlong-termmem-ory,butnottheshort-termstore,therebyprovid-ingaclearer pictureof what is goingonintheshort-term store. In contrast, in our Experiment 1we sampled categories froma larger set andalsosampledwordswithineachcategoryfromalargerset. Therewasthereforelessrepetitionofcategories andof words across trials, thus argu-ablyreducingtheamount of proactiveinterfer-ence. With the additional assumption thatprimacy items are encoded more strongly intolong-termmemorybecausetheyhavemoretimeto be encoded, proactive interference couldexplain why we obtained more primacy thanDavelaaretal.: Memoryfortheprimacypartofthelist couldhavebeenimpairedbythehigherlevelofproactiveinterferenceinDavelaaretal.Experiment 2 used a similar method to Experi-ment 1, except that the same categories were usedforeachblockof fourconsecutivetrialssothatproactiveinterferencecanbuildupwithineachblock. Thenumber of items toberememberedwas reducedto5tofurther minimiseanycon-tribution fromlong-termmemory. Probe typewaspostcuedinthisexperimentfortworeasons:First, postcueingrulesoutanyeffectofthekindof probe (category probe or spatial probe) onprocesses during encoding, ensuring that theserial-position effects are not confounded withstrategic adjustments to the expected kind ofprobing. Second, Experiment 1 showedthat, ifanything, the precued method produces evenmoreprimacythanthepostcuedmethod. If weobtain primacy in tests of itemmemory withpostcueing, we can be confident that this primacyisindependentofwhetherornotpeopleknowinadvancehowtheywillbeprobed.MethodParticipants.Thirtyparticipants,15womenand15 men between the ages of 17 and 35 years, tookpart in Experiment 2. They were all nativespeakers of Englishandparticipatedtoreceiveeithercoursecreditor10cash.Design. EachtrialincludedfiveEnglishwordsfromfivedifferent categories, presentedvisuallyinfivedifferentboxesonacomputerscreen.Allfive words were then probed for recall usingeithercategoryprobesorspatialpositionprobes.Therewere80trials,groupedinto20sets,with4trials in each set. The four trials in each particularsetusedwordsfromthesamefivecategories, sothatproactiveinterferencecouldbuildupwithinthe set. No category was repeated on consecutivesets, so that proactive interference should bereduced between sets. Accordingly, we subdividedthetrials ineachset intoonefirst trial, twomiddletrials, andonelasttrialperset. Thismeantthattherewere20first trials, 40mid-dletrials, and20lasttrialsintotal. Proactiveinterferenceshouldbeminimal onthefirst trialandmaximalonthelasttrialineachset.Output order was variedrandomly, withtheconstraint that for the first andlast trials,everycombinationof input positionandoutputposition (e.g., word at input position 3 probed forrecall at output position 2) occurred two times foreach probe type, whereas for middle trials,everycombinationof input positionandoutputpositionoccurredfourtimesforeachprobetype.This required 50 (552) words to be re-called, using10trials perprobetypeforfirstand last trials, and 100 (554) words,using20trialsperprobetypeformiddle trials.The spatial position of each word was determinedatrandomoneachtrial.The20setsof trialswereconstructedsothateverycombinationof probetypefor thefirsttrialandprobetypeforthelasttrial(22 4combinations) occurred5 times. Twomiddletrials were allocated randomly to each set, so thatthe probe type of the last trial in each set couldnot bepredicted. After thetrial structures andtheir order within each set had been fixed,categoriesandwordswereselectedforthem.Sets of four trials werepresentedinrandomorder in four test blocks of five sets (20 trials) each.Thetest blockswereprecededbyeight practicetrials, which were excluded from the data analysis.Materials and procedure. The materials werethesameas thoseinExperiment 1except thatonly four words were used per category toincreaseproactiveinterferencethroughincreas-ingthelikelihoodofchoosingthesamewordondifferent trialswithinaset. ThetimingintervalsandprocedurewereidenticaltothoseofExperi-ment1. Participantsweretestedfora90-minutesessionconsistingof four blocks, andtheywereencouragedtohaveabreakaftereachblock.SERIALPOSITIONFORITEMSANDRELATIONS 357ResultsAfactorial analysis of variance(ANOVA) wascarriedoutwithfourwithin-subjectsfactors:trialposition(firstversuslastinasetoffour), probetype, input position and output position, andproportionofanswerscorrectintheappropriatepositionas the dependent variable. The resultsfromthisANOVAuptoquadraticcontrastsarepresentedinTable4.Figure3shows theinput serial-positioncurvewithextendedprimacyandmodest recencyandalsothemaineffect of trial position(first versuslast inaset), whichdidnot interact withinputposition. Thesedatashowthattherewasasmalleffectofproactiveinterference(andreleasefromproactiveinterferencebetweensets)thataffectedalllistpositionsapproximatelyequally.Trialposi-tiondidnot interact withprobetype, F2.076,MSE0.14, p.16. Nevertheless, we testedtheeffect of trial positionforeachprobetypesepa-rately.Theeffect wassignificantonlyforcategoryprobes, F13.785, MSE0.085, p.001, butnotforspatialprobes,F0.55,MSE0.189,p.464.There was aninteractionof probe type andinputposition, withgreaterprimacyandrecencyfor positional probes thancategoryprobes. Theexperiment also replicated the main effects ofprobe type (better recall for category probes) andthelineardeclineofrecalloveroutputposition.DiscussionThere was an effect of proactive interference,butno evidence that it selectively suppressedTABLE4ResultsofANOVAforExperiment2,testingeffectsoftrial,probetype,inputposition,andoutputpositionVariables Contrast F MSE p g2pTrial 7.339* 0.135 .011 .202Probetype 218.484*** 0.39 B.001 .883Input L 19.072*** 0.364 B.001 .397Q 64.504*** 0.05 B.001 .69Output L 154.502*** 0.109 B.001 .842Q 36.927*** 0.082 B.001 .56Trial ProbeType 2.076 0.14 .16 .067ProbeTypeInput LL 0 0.118 1 0LQ 22.164*** 0.043 B.001 .433ProbeTypeOutput LL 6.136* 0.118 .019 .175LQ 11.919** 0.059 .002 .291OutputInput LL 4.032 0.101 .054 .122LQ 2.201 0.097 .149 .071QL 18.376*** 0.083 B.001 .388QQ 1.239 0.108 .275 .041Trial ProbeTypeInput LLL 0.451 0.062 .507 .015LLQ 1.217 0.047 .279 .04Trial ProbeTypeOutput LLL 1.295 0.062 .264 .043LLQ 1.99 0.044 .169 .064Trial Output Input LLL 1.528 0.056 .226 .05LLQ 0.205 0.057 .654 .007LQL 1.579 0.073 .219 .052LQQ 0.008 0.089 .931 0ProbeTypeOutput Input LLL 4.945* 0.042 .034 .146LLQ 4.722* 0.073 .038 .14LQL 1.924 0.097 .176 .062LQQ 0.704 0.117 .408 .024Trial ProbeTypeOutputInput LLLL 1.164 0.048 .29 .039LLLQ 0.521 0.046 .476 .018LLQL 0.464 0.087 .501 .016LLQQ 1.887 0.11 .18 .061Note. Thedegrees of freedominall cases are(1, 29). ANOVAanalysis of variance. g2p partialetasquared; L linearcontrast; Qquadraticcontrast; LQinteractionoflinearcontrastandquadraticcontrast; Input inputposition; Output outputposition.*pB.05,**pB.01,***pB.001.358 JONESANDOBERAUERperformanceintheprimacyportionoftheinputpositioncurve. Thepresent datashowthat theprimacy effect over input position is robustagainstamanipulationofproactiveinterference,forbothrelationmemoryanditemmemory.This result rules out one more potentialexplanation for why we obtained a primacy effectin our test of itemmemory whereas Davelaaretal. (2005) found none, using essentially thesame procedure. We could notthink of any otherdifferencebetweentheexperiment of Davelaaretal.andoursthatmightbeheldresponsibleforthe different results. One might speculate that thefact that onlyasingleitemwas tested, or thatthere was a time limit onrecall (1.5 s) intheexperiment of Davelaar et al., contributed toeliminating primacy, but there is no reason tobelieve that these features of the experiment haveany impact on the primacy effect. We also need toconsider the possibility that the experiment ofDavelaaret al. simplyhadinsufficient powertodetecttheprimacyeffect. Asouranalysisofthefirst-probediteminExperiment1 hasshown,theprimacyeffect forthefirst-recalleditemisrela-tivelysmall, sothat it couldbemissedduetoalackofpower(Davelaaretal.,2005,hadonly30trialsintherelevantpresentation-ratecondition,compared to our 56 in Experiment 1). We cannotconclusively resolve the remaining discrepancybetweenourresultsandthoseofDavelaaretal.Fortunately, this discrepancy is of little conse-quence for the question of what causes serial-position effects in list memory. As we explainbelow, our investigation of various potentialsources of serial-positioneffects does not hingeontherebeingbothprimacyandrecencyeffectsonitem memory:Aslongasatleastone oftheseeffectsisobtained,ourconclusionshold.Toconclude, weobtainedconsistent evidenceacross two experiments that there is both primacyandrecencyforitemmemoryaswell asrelationmemory. Thisresult wasalsoobtainedwithtwofurther experiments of ours, not reportedhere,usingthesameparadigm. Ourresultisinagree-ment withpreviousstudiestestingitemmemorythrough different methods (e.g., Cowan et al.,2002; Hayetal., 2007; McElree&Dosher, 1989;Monsell, 1978; Oberauer, 2003). Therefore, weare confident that there are both primacy andrecency effects over input position in tests of itemmemory as well as relation memory. We nowconsider possible explanations of these effects,undertakingasystematicexplorationofpotentialmechanisms and retrieval processes through com-putationalmodelling.MODELLINGCAUSESOFSERIAL-POSITIONEFFECTSWeconstructedagenericcomputational frame-worktocomparethedifferent potential loci ofserial-positioneffects (for theMatlabcode, seeFigure S1, which is available via the supple-mentary tab on the articles online page athttp://dx.doi.org/10.1080/09658211.2012.726629).We did not aimto quantitatively fit the data.Rather, weinvestigatedwhether thequalitativepatternof our data*inparticular theU-shapedserial-position curve over input for both item andrelationmemory*canbereproducedbyimple-mentingserial-positioneffectsatdifferentlociinthemodel (i.e., strengthofitemrepresentations,strength of bindings between items and inputpositions, strength ofbindingsbetween itemsandspatial locations, distinctiveness of input positions,and distinctiveness of spatial locations). Anycombinations of loci and retrieval modes thatcannot generate these U-shaped serial-positioncurvesforbothkindsof probeareruledout bythepresentdata.1 2 3 4 500.10.20.30.40.50.60.70.80.91Input PositionProportion Correct Low-PI CategoriesHigh-PI CategoriesLow-PI PositionsHigh-PI PositionsFigure3. Accuracyas afunctionof probetype, proactiveinterference(lowvs. high), andinputpositioninExperiment2, averagedover all output positions. Error bars representwithin-subjectsstandarderrors(e.g., Bakeman&McArthur,1996).SERIALPOSITIONFORITEMSANDRELATIONS 359All hypothetical causes of serial-positioneffectsexploredinoursimulationswereassumedto vary across serial position according to aU-shapedfunction. For instance, simulations in-vestigatingwhether variationinitemactivationcontributes to serial-position effects assumed thatitems at thebeginningandtheendweremorestrongly activated than those in the middle of thelist. Thiswasintendedtogiveeverymechanismthebest chanceof producingaU-shapedserial-position curve, so that any mechanisms that couldnot,evenunderthesegenerousconditions,couldbediscounted. Weevaluatedall simulations forwhether theyproducedU-shapedserial-positioncurvesoverinputposition.Weusedagenericarchitectureinwhichserialorder is representedby associating itemrepre-sentationstocontextmarkers; thisassumptionissharedby the most successful models of serialrecall (Lewandowsky & Farrell, 2008). Within thisarchitecture, weconsideredsixsourcesofserial-positioneffectsinconjunctionwithfourretrievalmechanisms(seeTable1). Theyareinstantiatedinvariouscombinationsinpublishedmodels.Serial-positioneffectsinitemmemoryFigure 4 demonstrates the major structures in-volvedinourmodelofmemoryforitems. Thereare item representations, denoted by letters;position markers representing ordinal positionsintheinputsequence, denotedbynumbers; andbindingsbetweenthetwo. Thepositionmarkerscan overlap each other (the shaded areas onFigure4). Thereforeactivationof PositionMar-ker 1 will also partially activate Position Marker 2totheextent that theyoverlap. Theamount ofactivation that feeds from each position marker tothe item presented in that position depends on thestrength of binding between the position anditem. The relative amount of activation that feedsto neighbouring items depends on positionaldistinctiveness, asmodelledbytheproportionofoverlapbetweenneighbouringpositionmarkers.Duringretrieval, thecategorycueusedintheexperimentwasassumedtoprovideextraactiva-tiontotheappropriatelist item, but not toanyotherlistitems(seeFigure4). Activationhadtosurpass a thresholdfor retrieval tooccur. Thisthreshold was set higher than the activationprovided by the category cue alone, reflectingtheaimtorecall anitemthat is bothfromthecorrectcategoryandfromthecurrentlist. Gaus-siannoiseinthemodel contributedtowhetheritemactivation exceeded the threshold or not.Theaddednoisecouldoccasionallyliftanextra-list itemmatching the currently cued categoryabove threshold, thereby producing an error (i.e.,anextralist intrusion), but thesewerenot expli-citlymodelled.Tworetrieval mechanismswereconsideredinthemodel. Thefirst makesnouseof theitempositionbindings.Itaddstheactivationfromthecuedcategory(whichispositiveonlyfortheonematchingitem)totheintrinsicactivationsoftheitems themselves (i.e., the activation they re-ceivedatencoding)andrecallstheitemwiththehighestresultingactivation,provideditsurpassesthe threshold. The second mechanism implementsretrievalbyscanningthroughthelist. Itusesthepositionmarkersonebyoneinforwardordertocueitems. At eachstep, activationconferredtotheitemsfromthecurrentpositioncueisaddedtotheintrinsicitemactivationandtheactivationfromthecategorycue. Whenoneormoreitemsexceedthethreshold, theitemwiththehighestresultingactivationisretrieved,andthescanningprocedure finishes. If no item exceeds the thresh-old,thenextpositionmarkerisactivated,andsoon until the end of the list. An omission isreturned if no items are recalled after all positionmarkers have been used as cues (i.e., scanning hasreachedtheendofthelist). Wecombinedthesetworetrieval processeswiththreepossibleloca-tions of serial-positioneffects: memorystrength(i.e., activation) of items, strength of bindingsFigure 4. Representationsinvolved inthemodelofmemoryfor items: Overlapping position markers denoted by numbers,item representations denoted by letters, and bindings betweenthetwo.AcategorycueisdepictedgivingextraactivationtoItemb.360 JONESANDOBERAUERbetweenitemsandpositions, anddistinctivenessofpositions.Thefindingsfromsimulationswiththemodelwereclear:primacyandrecencygradientsontheintrinsic activation of item representationsproducedaU-shapedserial-positioncurvewitheitherretrievalmechanism. Primacyandrecencyoverthestrengthof itempositionbindingspro-ducedaU-shapedserial-positioncurveonlywithretrieval by scanning. Variation in positionaldistinctiveness (i.e., higher overlap of mediumthanofendpositions)onlymatteredwithretrie-val by scanning, and it could not produce aU-shaped curve. In fact, higher positional distinc-tivenessattheendsthaninthemiddleofthelistledtobetterrecall of itemsat medial positions.This is becauseoverlapinthepositionmarkersgives items extra chances of being recalled (whencuedwithnearbypositionmarkers). Intests ofitemmemory, it does not matter if anitemisrecalled in response to the wrong positionmarker as cue, it onlymatters that theitemisrecalled at all. Medial position markers havemoreoverlapswithothermarkersthanthoseatthe extreme ends, and therefore medial itemshave more opportunities of being activated abovethreshold, leading to an inverted-U serial-positioncurve.Serial-positioneffectsinrelationmemoryFigure 5 shows the structures involved in themodel of relationmemory. It isanextensionofthemodel foritemmemorythat includesrepre-sentationsof spatial locations. Theshadedareasrepresent overlap in memory for nearby locations.In addition to their bindings to input-positionmarkers, items werealsoboundtospatial loca-tions. Moreover, spatial locationswereboundtoinput-position markers, such that the spatiallocation in which the first-presented itemap-pearedwas boundtothefirst positionmarker,andsoon. InFigure5, thesebindings areonlyshown for the first serial position to avoid clutter,but thereweresuchbindings for all serial posi-tions. As in the experiments described above,spatiallocationandinputpositionwereuncorre-lated across trials, and therefore we assignedspatial locationstoinput positionsat randominthe simulations. In addition to the three sources ofserial-positioneffectsdiscussedinthecontextofitemmemory (i.e., intrinsic activationof items,bindingsbetweenitemsandinput positions, anddistinctivenessofinputpositions), serial-positioneffects inrelationmemory couldbe causedbyvariation in the strength of itemlocation bindings,thestrengthofbindingsbetweenspatial locationsandinput positions, andby variationinspatialdistinctiveness (i.e., by varying how much adjacentlocations overlap in memory, as for the inputpositionsabove).We explored two retrieval mechanisms. Thefirstistousetheprobedspatiallocationasacuetodirectlyretrievetheitemlinkedtoitthroughitemlocationbindings. Thisworkedinthesameway as cueing with input position above. The cuedspatial locationpartially activatedany overlap-ping locations, and activation was forwarded fromlocations to items according to the strength of theitemlocationbindings. Aftertheadditionoftheintrinsic strength of itemrepresentations andGaussiannoise, the most activateditemwouldberetrievedaslongasitsactivationexceededathreshold.Thesecondretrieval procedureusesscanningthroughthelist inorderofpresentation. Peoplecouldscanthroughtheinput-positionmarkersinforwardorder, simultaneouslyrecallingthespa-tial locationandtheitemthat isboundtoeachpositionmarkeruntil theyfindthelocationthatmatchesthegivenlocationprobe(Cowanetal.,2006). The item retrieved at the same time wouldthen be reported. Such an indirect retrievalprocess is conceivable if retrieval in order ofpresentationis mucheasier thanretrieval of anitemby its direct binding to the spatial cue.Thereissomeevidencethatretrieval inforwardorderiseasierthanretrieval inanyotherorderFigure 5. Representationsinvolvedinthe modelofmemoryfor relations: an extension of the model for items addingspatial locations, denoted by S numbers, spatialitembindings, and input positionspatial bindings. Only oneexample of spatialitem and positionspatial bindings isshowntoavoidclutter, whereinthemodel therewouldbefourofeach.SERIALPOSITIONFORITEMSANDRELATIONS 361(Lange, Cerella, &Verhaeghen, 2011; Lange,Verhaeghen, & Cerella, 2010), rendering retrievalby scanning plausible. Scanning involved steppingthroughtheinputpositionmarkersoneatatimeand retrieving the most activated spatial location.Activation was conferred to spatial locations fromthe position markers through the bindings be-tween them; the degree of activation for eachlocation depended on the overlap between neigh-bouring input positions and the strength ofbindings between input positions and spatiallocations. Once a location was retrieved thatmatchedtheprobedlocation, thecurrent input-position marker was used to cue for its associateditemasdescribedintheprevioussection. Ifanyitems activationexceededthethresholdat thisstep, the itemwith the highest activation wasselected as output, otherwise an omission wasreturned.Again, primacy and recency in the intrinsicactivationofitemrepresentationsproducedaU-shapedserial-positioncurvewitheitherretrievalmechanism. With direct cueing using a spatiallocation, onlythestrengthof bindings betweenitems and spatial locations and the distinctivenessofspatiallocationscouldhaveapotentialimpacton the serial-position curve over input position. Infact, spatial distinctiveness hadnoeffect ontheserial-position curve over input order becausespatial locations were uncorrelated with inputpositions. Strength of bindings between itemsand spatial locations affected primacy and re-cencyonlyifthebindingswerestrongeratmoreextremeinput positions; it didnot matterif thebindings werestronger at moreextremespatiallocations.Retrieval byscanningthroughinputpositionstendedtoproduceaslightlyasymmetricalserial-positioncurve, weighted towards primacy. Thiswas due tothe forward-scanning nature of themechanism. For example, assumethat apartici-pant is cued with the fourth spatial location,whichin thistrialcorrespondedtothefinal inputposition. Due to noise in the retrieval process, thefourth spatiallocation mightbe retrievedalreadyby an earlier input position marker. Here thescanning stops, and an attempt is made to retrievethe itemat that inputposition. The mechanism isblindtothefactthatastrongeractivationofthefourth spatial location would probably haveoccurredhadit beencuedwiththefinal inputposition. Earlierinput positionssufferlessfromthis problem as they have a greater opportunity tocuetheirmoststronglyrelatedspatiallocation.With scanning as the retrieval mechanism,primacyandrecencyinthestrengthof bindingsbetween items and spatial locations, or thestrengthofbindings betweenitems and positions,producedaU-shapedserial-positioncurve. Anyoverlap(orreduceddistinctiveness)intheinputposition markers also produced a U-shaped curve.This is because recalling items in the correctpositionsisimportantforrelationmemory.Med-ial positions have more overlaps, leading to ahigher chance of recalling the wrong spatiallocation, or the wrong item, either of whichresults in a mismatch of probed location andrecalleditem.Table1summarisestheresultsofoursimula-tions. The combinations of retrieval processes andpresumed mechanisms of serial-position effectsthatareviableinlightofourdataareprintedinbold. Thetableshowsthat ourfindingsruleoutabouthalfofthetheoreticalpossibilities.GENERALDISCUSSIONThe goal of the present experiments was toanalyseserial-positioneffectsinshort-termrecallseparately for itemmemory and for relationmemory. The experimental design allowed adirect comparisonof testsof item memory (usingcategoryprobes) withtests that unambiguouslyrequired relation memory (binding of items totheirspatialpositionatpresentation)inawithin-subjects comparisonthat avoids confounds withother variables, and using a procedure thatseparates input fromoutput order. Based onpreviousliteratureusingcategoryprobes(Dave-laar et al., 2005) or itemrecognition(Monsell,1978; Oberauer, 2003), weexpecteditemmem-ory toshowmostly recency andlittle primacy.Contrary to this expectation, our results withboth probe types showed primacy effects thatwereatleastasstrongascorrespondingrecencyeffects.Serial-positioneffectsoverinputpositionOur design provided a separation of serial-position effects along the dimensions of inputandoutputposition.Thisallowsustoinvestigatetheeffectsoverinputpositioncontrolledfortheimpact of output position. The separation of itemmemoryandrelationmemoryhelpsustofurther362 JONESANDOBERAUERnarrow down the set of candidate mechanisms forserial-positioneffects.Asdemonstratedbyoursimulations, primacyandrecencyinitemmemorycannot becausedbyamechanisminvolvingpositional confusions/distinctivenessalone. Suchamechanismpredictsaninverted-Ushapedserial-positioncurve, con-traryto theU-shaped curve obtainedin thedata.Distinctivenessisacrucial componentofseveralcurrentmemorymodels*forexample, thescale-independent memory, perception, and learningmodel (SIMPLE; Brownet al., 2007), thestartend model (Henson, 1998), and the model ofBotvinick and Watanabe (2007). To account for aU-shapedserial-positioncurveinitemmemory,distinctiveness models are required to makeadditional assumptions. For SIMPLE(Brownetal., 2007) athresholdfor distinctiveness is pro-posed. Items whose positions are more confusablewith each other (e.g., medial items in a list) mightfall belowthe threshold and therefore not berecalledat all.This seems tous animplausibleassumption because it implies an inefficient use ofmemory. Itwouldbemorehelpful toattempttoremember these items, but potentially confusethemandrecall themintheincorrect positions.As position does not matter for item memory, thiswouldleadtomore chances tocorrectly recallmedial items, resulting in an inverse-Userial-positioncurve,asshowninourmodel.Inthestartendmodel, inadditiontodiffer-ential distinctiveness, strength is introducedintothe calculationleading tothe selectionofan item for recall. Extreme position markers havegreater strength than more medial markers. Itemsthat fall below a threshold of strength areomitted. This strength parameter in the startend model is akin to strength of bindings betweenitemsandinputpositionsinourmodel. Withoutit, the startend model would be unable toproduce a U-shaped serial-position curve foritemmemory.Explanations of primacy andrecency effectsover input positions basedonvariations inthestrength of bindings between items and inputposition markers face an obvious difficulty:Neither successful item recall nor successful recallof itemlocation relations requires informationaboutanitemsinputposition.Ifweassumethatretrieval isaccomplishedthroughdirectbindingsbetweentherelevant cues (i.e., thecategoryorthe spatial location) and the target item, then ourtest of itemmemorywouldbesensitiveonlytodifferencesintheitemsintrinsicactivation; ourtest of relation memory would in addition besensitivetovariationsinthestrengthofbindingsbetweenitems andspatiallocations. The strengthof bindings between items and their input positionmarkers can play a role only if we assume a moreindirect retrieval process based on scanningthroughthelistinitsorderofpresentation.Our experiments showlargelyparallel serial-position curves for tests of itemmemory andorder memory. Parsimony therefore stronglysuggests that thesamemechanisms arerespon-siblefor serial-positioneffects inbothkinds oftests. Table 1 shows that only two models arecompatible with the U-shaped serial-positioncurveforbothitemandrelationmemory.Oneisa model with serial-position effects on the in-trinsicactivationof items. Theotherisamodelwith serial-position effects on the strength ofbindingsbetweenitemsandtheirinputpositions,combinedwithretrievalbyscanning.Theimplications of our modellingresults donot dependonthe shape of the serial-positioncurve; inparticulartheydonotdependontherebeing both primacy and recency effects in thedata. We implemented symmetric U-shaped serial-positiongradientsondifferentparametersinthemodels(becausethisistheshapeweobservedinthe experiments), and as a consequence weobtainedeither U-shapedserial-positioncurves,or serial-positioncurves that didnot reflect theimplementedserial-positiongradientsatall (i.e.,they were either flat or had an inverted U shape).Therefore, our simulations show that serial-positiongradients oncertainmodel parameters(inparticular, itemactivationandbindings be-tweenitems andinput positions) translate intocorrespondingserial-positioneffectsinthedata,whereas serial-positiongradients onother para-metersdonot. Thisresult isindependent of theshapeoftheserial-positiongradients. Therefore,itdoesnotmatterwhetherthereisbothprimacyand recency in item memory and relationmemory*as long as there is either primacy orrecency(or both) initemmemoryaswell asinrelationmemory,ourconclusionshold.To conclude, our data place constraints onexplanationsofserial-positioneffectsinimmedi-atememory. TheU-shapedserial-positioncurveover input position for both itemand relationmemory cannot be explained purely by positionaldistinctiveness. Either these serial-position effectsarisefromvariationsintheintrinsicstrength(oractivation) of the items, or they arise fromvariations in the strength of bindings betweenSERIALPOSITIONFORITEMSANDRELATIONS 363items and input positions. In the latter case,memory must rely on retrieval by scanningthrough the list in the order of presentation,rather than on direct bindings between itemsandthegivenretrievalcues.Manuscriptreceived1March2012Manuscriptaccepted29August2012Firstpublishedonline24September2012REFERENCESAvons, S. E., Wright, K. L., &Pammer, K. (1994).Theword-lengtheffect inprobedandserial recall.TheQuarterlyJournal of Experimental PsychologySectionA,47,207231.Baayen, R. H., Piepenbrock, R., & van Rijn, H. (1995).The CELEX lexical database (Release 2) [CDROM]. Philadelphia, PA: University of Pennsylvania,LinguisticDataConsortium.Bakeman, R., & McArthur, D. (1996). Picturingrepeated measures: Comments on Loftus, Morrison,andothers. BehaviorResearchMethods, Instruments,&Computers,28,584589.Bhatarah, P., Ward, G., &Tan, L. (2008). Examiningtherelationshipbetweenfreerecall andimmediateserial recall: The serial nature of recall and the effectof testexpectancy.Memory& Cognition,36,2034.Bjork, E. L., &Healy, A. F. (1974). Shorttermorderanditemretention. JournalofVerbalLearningandVerbalBehaviour,13,8097.Botvinick, M., &Watanabe, T. (2007). Fromnumer-ositytoordinal rank: Again-eldmodel of serialorder representationincortical working memory.JournalofNeuroscience,27,86368642.Brown, G. D. A., Neath, I., &Chater, N. (2007). Atemporal ratio model of memory. PsychologicalReview,114,539576.Brown, G. D. A., Preece, T., &Hulme, C. (2000).Oscillator-based memoryfor serialorder. Psycholo-gicalReview,107,127181.Burgess, N., &Hitch, G. J. (1999). Memoryforserialorder: Anetworkmodel of thephonological loopanditstiming.PsychologicalReview,106,551581.Burgess, N.,&Hitch, G. J. (2006). Arevisedmodelofshort-term memory and long-term learning of verbalsequences. Journal of Memory andLanguage, 55,627652.Cowan, N., Saults, J. S., Elliott, E. M., & Moreno, M. V.(2002). Deconfounding serial recall. Journal ofMemoryandLanguage,46,153177.Cowan, N., Saults, J. S., & Morey, C. C. (2006).Development of working memory for verbalspatialassociations. Journal ofMemoryandLanguage, 55,274289.Davelaar, E. J., Goshen-Gottstein, Y., Ashkenazi, A.,Haarman, H.J.,&Usher, M.(2005).Thedemise ofshort-term memory revisited: Empirical and compu-tational investigationsofrecencyeffects. Psycholo-gicalReview,112,342.Duncan, M., &Murdock, B. (2000). Recognitionandrecall with precuing and postcuing. Journal ofMemoryandLanguage,42,301313.Ebbinghaus, H. (1964). Memory: Acontribution toexperimental psychology (H. A. Ruger &C. A.Bussenius, Trans.). NewYork, NY: DoverPublica-tions.(Originalworkpublished1885)Farrell, S. (2006). Mixed-list phonological similarityeffectsindelayedserial recall. Journal of MemoryandLanguage,55,587600.Farrell, S. (2012). Temporal clusteringandsequencinginshort-termmemoryandepisodic memory. Psy-chologicalReview,119,223271.Farrell, S., & Lelie` vre, A. (2009). Is scanning in probedorder recall articulatory? Quarterly Journal ofExperimentalPsychology,62,18431858.Farrell,S.,&Lewandowsky,S.(2002).Anendogenousdistributed model of ordering in serial recall.PsychonomicBulletinandReview,9,5979.Fuchs, A. H. (1969). Recall for order andcontent ofserial wordlists inshort-termmemory. Journal ofExperimentalPsychology,82,1421.Grossberg, S., &Stone, G. (1986). Neuronal dynamicsofattentionswitchingandtemporalorderinforma-tioninshort termmemory. Memory&Cognition,14,451468.Hay,D.C.,Smyth,M. M.,Hitch,G.J.,&Norton,N.J.(2007). Serial positioneffects inshort-termvisualmemory: A SIMPLE explanation? Memory &Cognition,35,176190.Healy, A. F. (1974). Separating item from orderinformation in short-term memory. Journal of VerbalLearningandVerbalBehaviour,13,644655.Henson, R. N. A. (1998). Short-term memory for serialorder: Thestartendmodel. CognitivePsychology,36,73137.Henson, R. N. A., Norris, D. G., Page, M. P. A., &Baddeley, A. D. (1996). Unchainedmemory: Errorpatterns rule out chaining models of immediateserial recall. Quarterly Journal of ExperimentalPsychology,49A,80115.Lange, E. B., Cerella, J., & Verhaeghen, P. (2011). Easeof access to list items in short-term memory dependsontheorder of therecognitionprobes. Journal ofExperimental Psychology: Learning, Memory &Cognition,37,608620.Lange, E. B., Verhaeghen, P., &Cerella, J. (2010).Dualrepresentation of item positions in verbalshort-term memory: Evidence for two accessmodes. EuropeanJournal of CognitivePsychology,22,463479.Lee, C. L., &Estes, W. K. (1977). Orderandpositioninprimary memory for letter strings. JournalofVerbal Learning and Verbal Behaviour, 16,395418.Lewandowsky, S., & Farrell, S. (2008). Short-termmemory: Newdata and a model. Psychology ofLearningandMotivation,49,148.Li, S.-C., &Lewandowsky, S. (1993). Intralist distrac-torsandrecall direction: Constraintsonmodelsofmemory for serial order. Journal of ExperimentalPsychology: Learning, Memory &Cognition, 19,895918.364 JONESANDOBERAUERLi, S.-C., &Lewandowsky, S. (1995). Forward andbackwardrecall:Differentretrievalprocesses.Jour-nalofExperimentalPsychology:Learning,Memory&Cognition,21,837847.Madigan, S. A. (1971). Modality and recall orderinteractions in short-term memory for serialorder. Journal of Experimental Psychology, 87,294296.McElree, B., &Dosher, B. A. (1989). Serial positionand set size in short-term-memory*The time courseofrecognition.JournalofExperimentalPsychology:General,118,346373.Monsell, S. (1978). Recency, immediate recognitionmemory, andreactiontime. Cognitive Psychology,10,465501.Nairne, J. S. (1990). Afeature model of immediatememory.Memory&Cognition,18,251269.Nairne,J.S.,Ceo,D.A.,&Reysen,M.B.(2007).Themnemoniceffectsofrecallonimmediateretention.Memory&Cognition,35,191199.Nipher, F. E. (1876). Probabilityof errorinwritingaseriesofnumbers. AmericanJournalofScience, 12,7980.Nipher, F. E. (1878). Onthedistributionof errors innumberswrittenfrommemory. Transactionsof theAcademyofScienceofSt.Louis,3,ccxccxi.Oberauer, K. (2003). Understanding serial positioncurvesinshort-termrecognitionandrecall. JournalofMemoryandLanguage,49,469483.Oberauer,K., &Kliegl, R. (2006). Aformalmodelofcapacity limits in working memory. Journal ofMemoryandLanguage,55,601626.Page, M. P. A., & Norris, D. (1998). The primacy model:Anewmodel ofimmediateserial recall. Psycholo-gicalReview,105,761781.Rundus, D. (1971). Analysisof rehearsal processesinfreerecall. JournalofExperimentalPsychology, 89,6377.Tan, L., &Ward, G. (2008). Rehearsal inimmediateserial recall. Psychonomic Bulletin &Review, 15,535542.VanOvershelde, J. P., Rawson, K. A., &Dunlosky, J.(2004).Categorynorms:Anupdatedandexpandedversionof theBattigandMontague(1969)norms.JournalofMemoryandLanguage,50,289335.SERIALPOSITIONFORITEMSANDRELATIONS 365Copyright of Memory is the property of Psychology Press (UK) and its content may not be copied or emailed tomultiple sites or posted to a listserv without the copyright holder's express written permission. However, usersmay print, download, or email articles for individual use.