Memory & Cognition1979,7(2),95·112

Component-levels theory of the effects of spacingof repetitions onrecall and recognition

ARTHUR M. GLENBERGUniversity oj Wisconsin, Madison, Wisconsin 53706

Spacing repetitions generally facilitates memory for the repeated events. This articledescribes a theory of spacing effects that uses the same principles to account for bothfacilitatory and inhibitory effects of spacing in a number of memory paradigms. Increasingthe spacing between repetitions is assumed to result in the storage of greater amounts ofinformation of three types or levels: contextual, structural (associative), and descriptive. Con­textual information is encoded automatically, while the encoding of the structural anddescriptive information depends on control processes utilized. Remembering involves accessingthe stored information using retrieval cues containing information on any level that matchesthe stored information. The ultimate effectiveness of the spacing is controlled by this matehingbetween the retrieval cues and the stored information. Previous experiments demonstrating theoperation of these principles on the structural and descriptive levels are reviewed. Three newexperiments are reported that illustrate interactions between stored information and retrievalcues based on contextual information.

The effect of arepetition has always been of greatimportance to all theories of learning and memory.That performance improves with repetitions is one ofthe basic phenomena to be explained. Nonetheless,there is not an adequate explanation of when, how, orwhy repetitions are effective. No single-process theory,nor many multiprocess theories, can account for all,or even most, of the empirical findings. The goal of thisarticle is to present a theoretical framework thatintegrates a number of different repetition-effectphenomena and that makes a variety of new predictions.

Specifica11y, the theory is designed to account fordistributed practice and spacing effects found withverbal stimuli in a majority of verbal learning andmemory paradigms. The distributed practice effect refersto enhanced memory performance on repeated iternswhose presentations are distributed (either through timeor through time and other presentations), as opposed toperformance on items whose presentations are massed orcontiguous. In addition, in many situations the amountof distribution is positively correlated with performance.As the spacing, or lag, between the presentationsincreases, performance genera11y shows a concomitant,negatively accelerated improvement (Melton, 1970).Distributed practice effects, in one sense, have beendifficult to explain because of their ubiquity across bothtasks and information processing strategies (Underwood,1969; Underwood, Kapelak, & Malmi, 1976). On theother hand, recent research (Glenberg, 1976, 1977) has

This research wassupported in part by NIMHGrant MH26643.I arn indebted to Cherryl Bryant, Anita Creek, and Pam Srnithfor their assistance in collecting and analyzing the data. Requestsfor reprints should be sent to Arthur Glenberg, Departrnent ofPsychology, University of Wisconsin, Madison, Wisconsin 53706.

pointed to ways in which the effect can be manipulated.Clearly, an explanation of distributed practice effectswill have to deal with a variety of episodic memorysituations, as weil as the conditions that determinethe effectiveness of arepetition, and hence, representsa good base from which to pursue repetition effects ingeneral.

Many commonly accepted assumptions about theworkings of memory are used in developing the theory;the relation between this theory and more generaltheories (e.g., Anderson & Bower, 1973; Craik &Lockhart, 1972; Kintsch, 1974) is intended to beobvious. Within these assumptions, distributed practiceeffects are explained using an amalgamation of variousencoding variability approaches to spacing effects (e.g.,Bower, 1972; Estes, 1955; Melton, 1970). Encodingvariability positions propose that a stimulus undergoeschanges in encoding from presentation to presentation,and that the extent of change is positively correlatedwith the spacing of the presentations. These changes inencoding add information to the episodic representationof the stimulus's presentations and underlies theimprovement in performance.

As Hintzman (1974), Maskarinec and Thompson(1976), and others have noted, what it is that is assumedto be variably encoded is dependent on specificforumlations of encoding variability. It has beenproposed that the variability resides in the semanticinterpretation of events (Bower, 1972; Madigan, 1969),the context in which the events are encoded (Medin,Note 1), and the subjective organization in which theevents are embedded (Melton, 1970). The presenttheory combined these notions by proposing thatspacing effects can be due to variable encoding of anyor a11 of three types of informational components.

Copyright © 1979 Psychonomic Society, Inc. 95 0090-502X/79/020095·18$2.05/0

96 GLENBERG

The theory adds to past formulations by specifying theconditions under which a given type of informationalcomponent will be variably encoded, and the conditionsunder which the variability will influence memoryperformance. In addition, the theory proposes that thedifferent components share two abstract principles ofoperation that determine the effects of spacing, oneconcerning the processes of encoding and storage ofinformation, and the other concerning the processesat retrieval. First, arepetition is potentially effectiveto the degree that the second presentation allows forthe storage of information distinct from that stored atthe first presentation. Second, the realization of thispotential is controlled by the conditions at the time ofthe memory test, the retrieval environment.

The remainder of this paper is divided into fourseetions. The first seetion describes the theoreticalframework in detail. The second seetion provides aselective review of the literature on spacing effectsrelevant to testing the theory. The third seetion consistsof three new experiments, motivated by the theory,that expand the boundary conditions of the spacingeffect, as well as providing additional tests of parts ofthe theory. The last section evaluates the theory inregard to other explanations of distributed practiceeffects and in regard to its own completeness.

COMPONENT-LEVELS THEORY

The name "component levels" is derived from tworelated aspects of the theory. The theory specifies threetypes of components that represent three types ofinformation assumed to be important in most verballearning and memory tasks. Additionally, thesecomponents may be roughly assigned to positions withina hierarchy (level) by considering their "generality": thenumber of different traces in which they are included.

Encoding and Storage ProcessesUpon the presentation of a verbal stimulus, it is

assumed that pereeptual processes encode physicalfeatures of the stimulus. These features are then usedto access the long-term permanent representation ofthe stimulus in semantic memory. This semanticmemory representation is assumed to be composed ofpropositions giving the various features (e.g., its sound,spelling pattern, grammatical components, meaning,associations) of the item (Kintsch, 1974).

A multicomponent episodic trace is created torepresent the occurrence of the stimulus in the experi­ment. The components of the episodic trace are alsopropositional, some of which are drawn from semanticmemory. The specific components included in the traceare jointly determined by the stimulus being processed,the nature of the processing task, and the subject'sstrategies, as well as the context in which the stimulusis presented.

The components can be classified as one of threetypes, depending on the information encoded.Contextual components are automatically encodedupon the presentation of a to-be-rernembered (TBR)item, and they represent the context in which the itemis presented. Structural components encode the relation­ship discovered or imposed among sets of TBR itemsas a result of active processing. Descriptive componentsencode the specific item. The type of descriptivecomponents encoded depends on the cognitive processesbeing used in the task (e.g., Craik & Lockhart, 1972).

Components on the three levels differ as to theprobability that they are included in traces repre­senting different items, that is, the generality of thecomponents. A component that is included in manytraces is said to be general. Since the same contextualcomponents are included in all traces encoded in aspecific context, they are, typ ically, general cornpo­nents. The same structural components are includedonly in traces subjects have successfully interassociated.Hence, these components are usually less general thanthe contextual components. Descriptive componentsmay be unique to single traces. Processing demands ofa specific task may, however, change the generality ofa component. For example, if a great number ofmemorized items are associated to the same cue (e.g., asuperordinate category), then the structural componentrepresenting the association will be very general.

Contextual components. These components representthe context in which the event is encoded and includesuch information as the characteristics of the physicalenvironment, the time, and the learner's cognitive andaffective stage (Anderson & Bower, 1974; Kintsch,1974). The implicit assumption is that the events in amemory experiment do not occur in a cognitive vacuum,but are incorporated into the individual's stream ofconsciousness. For example, the trace representing thepresentation of a word in a free recalllist encodes notonly the fact that a specific word was presented, butthat it occurred in a given room, during an experiment,in a certain list, while performing a given cognitiveoperation.

Contextual information is given special status in twoways. First, it is assumed that components representingthe context are encoded automatically with theperception of the individual TBR events. Second, aswill be described below, it is assumed that contextualinformation can influence the encoding of othercomponents included in the trace.

Arepetition is potentially effective to the extent thatdifferent information is stored at the two presentationsof a repeated event. The conditions under which newinformation is stored at the second presentation dependon the component. Since contextual information isautomatically encoded, new contextual information isstored simply as a function of change in the context.Having a subject learn in two different rooms (environ-

COMPONENT·LEVELS THEORY OF SPACING EFFECTS 97

mental contexts) is sufficient for the encoding of newenvironmental information (S. M. Smith, Glenberg, &Bjork, 1978). Other, less drastic changes in the contextcan result in the storage of new contextual componentson the second presentation; for example, changes inthe cognitive and affective states of the subject willalso induce the storage of new components.

Different aspects of the context typically havedifferent rates of change. For example, changes inthe set of items being processed occur relativelyquickly, and in effect, continuously. We can thinkof these as local contextual components; that is, thesecontextual features directly influence processing only ina constrained, local, temporal region. Some aspects ofthe context change more slowly over the course of anexperimental session (e.g., the subject may be gettinghungry over the course of the experiment). Finally,some aspects of the context, the global features, aretypically unchanging or changing very slowly and are,hence, very general. Experiments 1-3 below demonstratehow these different aspects of the context can beexperimentally separated, and how they affectperformance.

The amount of change in the context, and hence theamount of change in the contextual components storedin the trace of a repeated item, is directly related to thetime or lag between the presentations of the repeateditem. This positive correlation between the number ofdifferent components stored in the trace and the repe­tition lag will be referred to as "differential storage."

Structural components. At the next level are thecomponents representing the structure that the subjectimposes on the individual events: which items areassociated, grouped, categorized, or chunked together.These structural components are not stored automatic­ally, but depend on the control processes engaged bythe subject. For example, Bjork (1975), Glenberg andBradley (in press), and others have demonstrated thatwhen subjects engage in rote repetitive rehearsal,associations among TBR items are not formed. On theother hand, processes requiring the generation ofrelations do successfully structure the items.

The specific structural components inc1uded in thetrace are determined by the local context. Cognitiveaspects of the context-most importantly, the otheritems currently being processed-will affect the structureimposed on a set of items. That is, associations willbe made only between items currently being processed(Anderson, 1972; Kintsch, 1974).

Whether or not new structural components areformed at the repetitions of stimuli depends on thecontext and on the encoding and storage processesbeing used by the subject. The learner must be engagingin some sort of structural analysis (e.g., relating TBRstimuli within a story, interactive imagery, etc.) ifnew structural components are to be added to the traceat the repetition. In addition, there must be a change in

the structure assigned to the stimuli, a change inducedby the context. For example, the stimulus may berepeated while the subject is processing a set of itemsdifferent from those on previous presentations, allowingfor the storage of new components in the trace of therepeated stimulus. Since changes in the context inducechanges in the structural components, the structuralcomponents will themselves be characterized bydifferential storage. That is, if the learner is engagingin some sort of structural processing, then the number ofdifferent structural components stored in the trace willincrease with the repetition lag.

Descriptive components. In the case of a TBR word,these components include information as to itsorthography, articulation, meaning, and so on. Whereasthe contextual and structural components are createdanew at the presentation of stimuli (they encode theevent as different from other presentations of the samestimulus), the descriptive components are copied fromthe semantic memory representation of the stimulus.The specific components copied into the episodic tracedepend on the processing in which the subject is engagedand the context. Control processes affect the depth(Craik & Lockhart, 1972) to which the word is encoded,that is, the kinds of descriptive components includedin the trace. Additionally, given that a physical stimulushas many different senses (Reder, Anderson, & Bjork,1974), the local context affects the selection ofcomponents copied into the newly formed episodictrace. The process by which this selection takes place isassumed to be as folIows: The pool of words currentlybeing processed primes related propositions in memory.The components copied into the new episodic trace willinclude the primed propositions at the task-determineddepth of analysis. Since these primed components arerelated to the traces being processed, they facilitatethe formation of structural (relational) components andgive memory its associative character.

Descriptive components are, in general, the mostunique components included in the trace. It is possible,however, to manipulate the encoding of descriptivecomponents or to design stimuli that all include thesame components. For example, requiring subjectsto process the pronunciation and sound of a set ofrhyming stimuli would result in the inclusion ofsimilar descriptive components in all the traces. Thesecomponents would then be general.

Variability in the specific descriptive componentsthat are encoded at the two presentations of a repeatedstimulus is a function of both control processes and thelocal context. The subject may be induced to encodedifferent aspects of the stimulus on its two occurrencesby changing the orienting task between the occurrences.Or, more commonly, changes in the local context mayeffect a change in the primed components includedin the trace (Bower, 1972).

So far, the theory has described the differential

98 GLENBERG

storage of three classes of representational eomponentsthat ean oeeur upon the repetition of a stimulus in amemory task. The traee of a stimulus repeated after asubstantial lag will inelude more eontextual, struetural,and deseriptive eomponents than the traee of an itemrepeated after a shorter lag. Nonetheless, the theory doesnot always predict a spaeing effeet as the result ofdifferential storage. The amount of differential storageplaees an upper bound on the ultimate effeetiveness ofthe repetition for later remembering. The extent and thedireetion of the eorrelation between differential storageand performance on a memory task is eontrolled by theretrieval environment.

Retrieval ProeessesRemembering a specifie stimulus requires retrieval

of the episodie traee representing the oeeurrenee of thestimulus in the experiment. Aeeess to episodie traees isprovided by retrieval eues. In eomponent-levels theory,a retrieval eue eonsists of eomponents similar to thoseincluded in the episodie traees. The eomponents in theeue provide temporary, limited-capacity parallelaetivation of identieal eomponents included in theepisodic traees (Lockhart, Craik, & Jaeoby, 1976). Thedegree of aetivation of a eomponent in an individualtraee is inversely related to the number of traees inwhieh the eomponent is included (its generality). The

. degree to whieh the traee is aetivated is a monotoniefunction of the summed aetivation of its individualeomponents. Therefore, (1) traee aetivation and retrievalare funetions of the number of eomponents shared bythe eue and the traee, (2) trace aetivation deereases asthe generality of the eomponents in the traee or the eueincreases, and (3) in general, traee aetivation inereaseswith the number of eomponents included in the traee.

This retrieval seheme follows the eneoding specificityprinciple (Tulving, 1976) in that the eue is only effeetiveto the extent it shares eomponents with the episodietraees. In addition, a type of eue-overload hypothesis(Stein, 1978; Watkins & Watkins, 1975) is implieated,in that a eue loses its effeetiveness as the number oftraees with whieh it shares eomponents inereases.

The speeifie eomponents included in the eue aredependent on the nature of the memory test, that is,the information available at the time of the test (Flexser& Tulving, 1978). Typiea11y, this information includesinstruetions, the eontext at the test, and any explicitretrieval cues provided. Consider first a reeognitiontest where the subjeet is provided with physieal eopiesof the original stimuli and distraetors and is asked todiseriminate between the old and new stimuli. Thesubjeet begins the test by eneoding a stimulus, therebyprodueing a eue eonsisting of eomponents representingthe eontext of the test and deseriptive eomponents ofthe stimulus. These eomponents then aetivate identiealeomponents in the episodie traces.

If testing oeeurs after a short retention interval,

then the loeal eontextual eomponents in the eue willaetivate traees eontaining eomponents representingthe same eontext that are included in traees of reeentevents. With longer retention intervals, the eontextualeomponents play a deereasing role in traee aeeess fortwo reasons. First, after a long retention interval, theloeal eontextual eomponents in the eue will noteorrespond to the loeal eontextual eomponents in theepisodie trace, so the eomponents in the eue will notaetivate any eomponents in the traee. Seeond, the globaleontextual eomponents shared by the eue and theepisodie traee are included in many traees (theeomponents are general or overloaded), so that eaehtraee will be only weakly aetivated.

The predominant souree of traee aetivation on areeognition test will be the deseriptive eomponents.The eue will contain a plethora of unique deseriptiveeomponents. Generally, if the eue is physieally identiealto a previously presented word, the traee of the euedword will be strongly aetivated and result in reeognition.On the other hand, if the eneoding of the item at inputwas impoverished (i.e., eontained few eomponents dueto inattention , fast presentation, shallow proeessing,ete.) or the eneoding of deseriptive eomponents at thetest is radiea11y different from the eneoding at input(e.g., two different meanings of a homograph), the euewill not retrieve the traee.

Retrieval in other situations is eontro11ed by thesame processes but, at times, with eomponents atdifferent levels. Consider retrieval in free reea11. Shortlyafter list input, the loeal eontextual eomponents onthe test will aetivate the traees including identiealeomponents, those proeessed near the end of the list.Thus a reeeney effeet is produeed in immediate freereeall, but not in a delayed free reea11 where the loealeontextual eomponents in the eue do not eorrespondto the eomponents in the traees (Tulving, 1968). Theglobal eontextual eomponents will heighten theaetivation of all TBR items. This eomponent ean be usedas a gate to eliminate extraexperimental eonfusions.Beeause these eomponents are included in a11 TBRtraees, however, the eue is relatively ineffeetive.

Traees reeovered with the eontextual eue eanthemselves beeome sources of retrieval cues. If strueturaleomponents (representing eategorieal or subjeetiveorganizations) are included in the reeovered traees, thenthese eomponents ean also serve as effeetive cues. Sineeonly a few traees are typically included in a subjeetivestrueture (e.g., Mandler, 1972), these traees will a11 behighly aetivated by shared struetural eomponents. Ofcourse, the effeetiveness of reeovered items as a soureeof struetural eomponents will depend on the subjeet'sprevious proeessing, that is, whether or not the itemswere struetured during input. Fina11y, the set ofdeseriptive eomponents included in a reeovered traeeean be used to aetivate traees having identieal deseriptiveeomponents. Additiona11y, the deseriptive eomponents

COMPONENT-LEVELS THEORY OF SPACING EFFECTS 99

are used to retrieve the lexical representation of theitem in order to generate an overt response.

When components from multiple levels are includedin the cue, trace access is predominately controlled bythe most specific components, since the more generalcomponents will tend to be overloaded. Nonetheless,once a trace has been accessed, general components inthe trace may be used to control performance (e.g.,contextual components may be used for judgments ofpresentation frequency).

Distributed Practice EffectsIncreasing the repetition lag will, typically, lead to

differential storage of the components on one or morelevels, and differential storage will, in general, facilitateremembering. The more different components includedin the trace, the more likely that the trace will sharecomponents with the cue and, hence, be activated andretrieved. The ubiquity of the lag effect is explainedby the built-in redundancy of most experiments;differential storage may be eliminated on some levels,but the remaining levels can still produce a distributedpractice effect.

The facilitation in remembering due to increasingthe repetition lag may not, however, always be found.Two situations arise in which differential storage doesnot manifest itself in performance as a monotonieallyincreasing lag effeet. Both situations result from traceaceessibility being conditional upon sharing eomponentswith cues used at the time of remembering. First,differential storage may occur at one level and retrievalprompted with cues containing less general components.Since access to the trace is controlled by the mostunique components in the trace, the differential storagewill not greatly influence performance. This predictionis tested in Experiment 1. During input, differentialstorage of contextual components is left uneontrolled.Differential storage is eliminated, however, in thestructural and descriptive eomponents. Free recall,using the putative contextual cue, results in a distributedpractice effect. When descriptive components areincluded in the cue (cued recall), the distributed practieeeffect is eliminated.

The second situation is more complicated. On anylevel, manipulating the relationship between the cue andthe information stored in the trace may produceperformance, as a function of the repetition lag, that ismonotonically decreasing, inverted U shaped , or mono­tonically increasing. Figure 1 illustrates these predictionsconsidering only one level.' On the left are timesduring which an event may be presented and repeated.Corresponding to each time are the components thatare encoded at that time. The gradual change in theeomponents available for encoding is due to changesin the context, although contextual changes in anexperiment are eertainly not as orderly as shown inthe figure.

COM- REPETITION STRUcTURE

TIME PONENTS LAG 1 2

;1 ql':~;j PRESENTATION TI"ES i7+tB t6+t8 i:5+t8 tl+t8

t2 hijK COMPONENTS IN TRACE mnopq I","op~ klmnopC\ ghijnopq

t3 ijkl CUE I4 44

t4 jklm NUtvtBER OF(mnop)

t5 klmnCUE 2 2ACTIVATEO(Inqs)

t6 Imno COMPONENTSCUE 3

4'5/" 4'61" 4-7/" 4·8/nt7 mnop ("'andorn)

t8 nor q

Figure l. On the left are times at which an event may berepeated and the components (on one level) stored at the presen­tations. On the right are components included in traces formedafter repetitions at four lags, the structure of three retrievalcues, and the number of components in the traces activated bythe cues.

Consider an event presented at times t s and t 8 .

The repetition results in a trace containing componentsk-q, as indicated on the right. Also shown are thecomponents in traces corresponding to events repeatedat t8 after four different lags. The increase in thenumber of components included in the traees illustratesdifferential storage as a function of repetition lag.

The figure also details trace activation provided bythree cues. For illustrative purposes, these cues consistof components on only one level. These cues differ intheir bias toward including eomponents stored at t 8. Cuebias is defmed as the tendeney to include in the cuecomponents stored at one presentation of the TBRitem, and to exclude from the cue eomponents storedat other presentations of the TBR item. The componentsin a strongly biased eue are sampled from a probabilitydistribution, having small variability, centered on theeomponents eneoded at the presentation toward whichthe cue is biased. As the bias decreases, the probabilitydistribution becomes more variable. An unbiasedeue eontains components drawn from a reetangularprobability distribution, so that components availableat all presentations are equally likely to be included inthe cue.

Cue bias can be controlled through direet experi­mental manipulation (Glenberg, 1977; Experiment 3 ofthis article), or indirectly by the retention interval.When the second presentation of a repeated event isfollowed by a short retention interval, the local contextat the test will be similar to the local context at thesecond presentation. The contextual componentsincluded in the retrieval eue will, therefore, be biasedtoward those eneoded at the second presentation. Inaddition, on a recognition test, to the extent that theloeal context influences the encoding of descriptivecomponents, the cue will contain descriptive compo­nents biased toward those encoded at the secondpresentation (Glenberg, 1976).

The first eue in Figure 1 is an example of one that ishighly biased toward the components encoded at theseeond presentation. The number of components shared

100 GLENBERG

by the cue and the trace is relatively large, leading tomuch activation and relatively high performance. Note,however, that the amount of activation is a decreasingfunction of the repetition lag. A "reverse" distributedpractice effect is generated with this cue.

The second cue is also an example of bias towardthe components presented at the second presentation,but the bias is not as great as in the first cue. Traceactivation, based on the second cue, is a nonmonotonie,inverted U-shaped function of the repetition lag.

The third cue is not biased toward either presenta­tion. The numbers entered into the table are theexpected number of activated components assumingthat four different components (out of a possible n)are randomly chosen for inclusion in the cue. Only inthis last case is performance a monotonie function ofthe repetition lag. In most experiments the retrievalcues are unbiased, resulting in the third cue type and thecommon, monotonically increasing lag effect (seeGlenberg, 1976, for a mathematieal derivation of thesethree types of spacing functions).

Nonmonotonie spacing functions have been weIlestablished in the paired associate paradigm (Atkinson& Shiffrin, 1968; Glenberg, 1976; Peterson, Wampler,Kirkpatrick, & Saltzman, 1963; Young, 1971).Component-levels theory provides a description of theconditions that control the shape of the spacing functionin this paradigm as weIl as in others.

TESTS OFTHE THEORY

The theory is in need of evidence supporting thenotion of different levels of components and cues, andthat spacing effects may occur due to differential storageon any or all of the levels. Although component-levelstheory makes use of many previous theories, not allof the tests of these theories can serve as tests ofcomponent-levels theory. The major problem is thatsimply increasing the lag between presentations of arepeated item may induce differential storage on anumber of different levels. Therefore, demonstratinga lag effect does not isolate a level. Likewise, eliminatingdifferential storage on one level does not necessarilyimply the absence of a lag effect if differential storageis not eliminated from other levels relevant to theretrieval cues.

Two approaches may be used to isolate spacingeffects on a single level. The first is to simply eliminatethe effects of differential storage on all but the targetlevel. This can be accomplished either by insuring thatthe number of new components on the nonrelevantlevels added at the second presentation is small and doesnot vary across lag conditions (encoding constancy),or that the number of new components added at thesecond presentation is always very great and doesnot vary across lag conditions (maximum encodingvariability). Unfortunately, encoding constancy may be

unattainable. The local context is always changing, sothat differential storage of these components will occur.The changing local context mayaIso influence theencoding of descriptive and structural componentseven when the same semantieally biasing contexts areused at each presentation.

A second approach to isolating spacing effects to agiven level is to eliminate the influence of componentson general levels by presenting subjects retrieval cues ofless general components. As discussed before, traceaccess is predominately controlled by the most specificcomponents in the cue.

Once a single level has been isolated, tests ofcomponent-levels theory's explanation of the lag effectmay be made. The two major types of tests involvemanipulations at either input or retrieval. At the time ofstorage, the experimenter may enhance the lag effectby promoting differential storage or eliminate theeffect by eliminating differential storage either throughencoding constancy or maximum encoding variability.At the time of retrieval, the relationship of the cue tostored information may be manipulated. As describedabove (see Figure 1), this manipulation can change theshape of the lag function from monotonieally increasing,to inverted U shaped, to monotonically decreasing.

Descriptive ComponentsIn general, increasing the lag between presentations

will increase the number of descriptive componentsincluded in the trace and lead to an increase inrecognition performance. A number of investigators(Earhard & Landry, 1976; Glenberg, 1976; Hintzman,1969, using recognition time as the dependent variable;Kintsch, 1966; Wiehawut, Note 2) have observed thiscorrelation. In addition, Earhard and Landry, as weIlas Wiehawut, found that inducing maximum encodingvariability of descriptive components tended toeliminate the lag effect. As predicted by the theory,recognition performance was high and constant acrossspacing conditions. Both articles also reported, however,that recognition performance increased with lag in thoseconditions where encoding constancy was attempted.This finding may reflect continuous changes in the localcontext that prevent encoding constancy.

The predictions illustrated in Figure 1 were tested atthe level of descriptive components by Glenberg (1976).A continuous recognition memory paradigm was used.On each occurrence of an item, the subject was toindieate whether the item was previously presented(old) or new. The lag between the first two presentationswas factorially combined with the retention intervalbetween the second and third presentations. After ashort retention interval, the local context at the thirdpresentation will be similar to the context at the secondpresentation. Consequently, both the contextual anddescriptive components in the cue will be biased as inCue Types 1 and 2 in Figure 1. After a long retention

COMPONENT-LEVELS THEORY OF SPACING EFFECTS 101

interval, the eontext at the third presentation will beasymptotieally related to the eontext at the first andseeond presentations. The eomponents inc1uded in theeue will be unbiased, as in Cue Type 3 in Figure 1. Aspredieted, the probability of eorreet1y reeognizing anold item was an inverted U-shaped funetion of lagwhen tested after a short retention interval, and amonotonically inereasing funetion of lag when testedafter a long retention interval.

Distributed praetiee effeets in the paired assoeiateparadigm may be aseribed to the deseriptive eomponentsor struetural eomponents. The ambiguity arises wheneonsidering what a subjeet learns when a paired associateis studied. If learning is eoneeptualized as "hooking-up"a response to a stimulus by ereating a strueturaleomponent, then differential storage of strueturaleomponents may oeeur. On the other hand, Greeno(1970) has had mueh sueeess explaining paired associatephenomena with a cognitive theory of learning. The firststage of learning is eoneeptualized as the storage of atraee eneoding a unitary relation between the stimulusand the response. The seeond stage involves learning toeneode the stimulus so that it retrieves the relation.Distributed praetiee effeets may, therefore, be due todifferential storage of deseriptive (stimulus) eomponentsduring the seeond stage of learning. Component-levelstheory remains neutral on this point. For eonvenienee,Greeno's (1970) interpretation of paired assoeiatelearning will be aeeepted, and distributed praetice effeetswill be analyzed at the level of deseriptive eomponents.Inereasing the lag between presentations of the pair,whieh is eorrelated with ehanges in the loeal eontext,offers the opportunity for differential storage of(stimulus) deseriptive eomponents in the traee. On thetest, the subjeet eneodes the stimulus member of thepair as a eue to the relational traee. Inereasing therepetition lag should, in general, inerease the number ofdifferent eomponents in the traee, and therefore inereasethe number of eomponents shared with the eue. Oneethe traee has been retrieved, the response must bedeeoded and output. Note that the proeesses thatproduee distributed praetiee effects in the pairedassociate paradigm are essentially the same as those inthe recognition memory paradigms.

In agreement with the assumption that distributedpractice effects in the paired associate and recognitionmemory paradigms are predominantly based on descrip­tive eomponents and cues, lag effects involving pairedassociates are essentially the same as for reeognitionmemory experiments and follow similar time courses.First, inereasing the lag between presentations ofpaired associates typically produees increases in cuedrecall. Second, the shape of the lag effect can bemanipulated by varying the retention interval betweenthe second presentation and the test. At very shortretention intervals, the lag funetion is deereasing(Peterson, Hillner, & Saltzman, 1962). With slightly

longer retention intervals, the lag function is an invertedU-shaped function (Peterson et al., 1963). Finally, withlong retention intervals, the funetion is monotonicallyinereasing (Glenberg, 1976).

Structural ComponentsWhen an item on a free recall list is repeated, it may

beeome associated with different items on its twopresentations. This differential storage of strueturaleomponents inereases with ehanges in the eontext(especially the other words being proeessed), and heneewith the lag. Thus, as the repetition lag inereases, thenumber of different associations involving the repeatedword inereases, and the prob ability of reeall inereases.This prediction of an inerease in free reeall with therepetition lag has been eonfirmed in many investigations(Glenberg, 1977; Madigan, 1969; Melton, 1970). Ofcourse, some of the effeet of lag seen in free reeall maybe due to differential storage of eontextual eomponentsand deseriptive components, in addition to the strueturalcomponents.

Some attempts have been made to isolate the effeetsof struetural eomponents in free recall. Glenberg (1977)eompared reeall of subjeets who were eneouraged toform associations among list items and subjects whowere prevented from doing so. In the former condition,the lag effect is influenced by differential storage ofstruetural, descriptive, and contextual components.In the latter eondition, the spacing effect should beattenuated due to the elimination of differential storageof struetural eomponents. As predicted, the subjectseneouraged to form associations produced a strong lageffeet in reeall. The other group showed an increase inreeall from Lag 2 to Lag 5, and then no improvement(out to Lag 17). Similar findings were reported byD'Agostino and DeRemer (1973, Experiment 1, freereeall, and Experiment 2, image-same eondition).Apparently, eliminating differential storage of structuraleomponents reduees the spaeing effeet from moderateto long lags, but does not change the effect at shorterlags.

DeRemer and D'Agostino (1974) limited theprocessing time on the seeond presentation of repeatedevents. This manipulation should also reduce differentialstorage of structural components. Again, spacing effectswere eliminated from moderate to long lags. In general,these results may be interpreted as showing that thespacing effeet in free recall is predorninantly due todifferential storage of structural components. Differ­ential storage of contextual and descriptive componentsseems to have a small effeet at the short lags, but not atmoderate or long lags.

Nonmonotonie lag effects based on structuralcomponents may be produeed by manipulating therelationship between the struetural components in theretrieval cues and the stored information in a manneranalogous to that using deseriptive components.

102 GLENBERG

G1enberg (1977, Experiments 1 and 2) repeated wordsin a free reca11 list at lags of 2, 5, and 17. Each repeatedword was assigned two once-presented words that couldbe used as recall cues. These possible cues werepresented between the two presentations of the repeatedword, one cue adjacent to each presentation. Whenreca11 was cued, each repeated word was cued with oneof the two possible cue words.

It was assumed that while studying the list, thesubjects imposed a structure on the individual wordsthat was at least in part determined by contiguity ofthe words (Wallace, 1969). Therefore, the structuresinc1uding repeated words will tend to inc1ude thepossible cue words. As the lag between the presentationsof repeated words increases, the number of differentstructural components inc1uded in the trace of therepeated word will increase; that is, there will bedifferential storage. When recall is cued with unbiasedcontextual cues (free recall), the differential storageof contextual and structural cues will produce themono tonic, increasing lag function. When recall is cued,the cue words provide a source of structural componentsbiased toward one presentation-the presentation adja­cent to the once-presented cue word-and the spacingfunction should be nonmonotonic. As predicted, freerecall resulted in the monotonie lag effect, while cuedrecall, using the biased cue word, resulted in an invertedU-shaped lag function.

Contextual ComponentsIndirect evidence for the operation of contextual

components in producing a spacing effect is providedby experiments reported by Maskarinec and Thompson(1976, Experiment 2) and Shaughnessy (1976). Bothused an incidental learning procedure with orientingtasks that forced subjects to process individual wordsand provided no incentive to form associations betweenthe words. According to component-Ievels theory,this manipulation should eliminate differential storageof structural components. (Shaughnessy discusses amechanism whereby associations between words maybe formed as a by-product of his orienting task, buthis analysis does not apply to the Maskarinec andThompson task.) Within the list some items were givenmassed presentations (MPs) and other items were givendistributed presentations (DPs). Half of the subjectsused the same orienting task on a11 presentations, whilehalf of the subjects used a different task at eachpresentation. After performing the orienting task, thesubjects were unexpectedly asked for free recall.

Maskarinec and Thompson (1976) reasoned that thetraces formed in the different-task condition encodemany descriptive components regardless of therepetition lag. That is, in order to perform the different­condition orienting tasks, the subject must encode thesame descriptive components when the presentationsare massed as when they are distributed. Therefore,

Maskarinec and Thompson predicted (from a simpleencoding variability point of view) that the massed­distributed difference would be eliminated in thedifferent-task condition. Shaughnessy (1976) deriveda similar prediction from a different theoreticalperspective.

If performance is tested with a recognition test,component-levels theory would also predict nodifference between MP and DP. That is, the efficacy ofthe descriptive components used as retrieval cues inrecognition is equated across presentation conditions bythe orienting tasks. The theory makes a differentprediction, however, for free recall. Presumably, theorienting tasks eliminated storage of structuralcomponents, but the tasks should not have eliminateddifferential storage of automatically encoded contextualcomponents. Free recall, in which contextual compo­nents are the only available retrieval cues, shouldtherefore result in a spacing effect. As predicted, therecall after DP was superior to the recall after MP inboth the Maskarinec and Thompson (1976) and theShaughnessy (1976) experiments.

The memory trace of an event also inc1udescontextual components representing the physicalenvironment in which the event is perceived (Abemethy,1940; S. Smith & Guthrie, 1924). A type of differentialstorage of these environmental context components isillustrated in an experiment by S. M. Smith et a1. (1978).The subjects were presented for study the same list ofwords in two sessions separated by 3 h. Half of thesubjects studied the list in the same room in bothsessions, the other subjects studied the list in twodifferent rooms. Reca11 was attempted in a third session,3 h later, in a room new to a11 of the subjects. Thosesubjects who studied the list in two different roomsrecalled an average of 40% more words than the subjectswho studied the list of words twice in the same context.This finding can be explained by assuming that thecomponents in the traces that represent the globalenvironmental context underwent little change for thewords studied twice in the same context. For thesubjects that studied in two roorns, however, the tracescontained components representing both environments.At testing, in a new room, the contextual eue did notinc1ude many components stored at either the first orthe second presentation. Nonetheless, the total numberof components shared by the cue and the traces wasgreater when the words were studied in two contextsthan in one context, producing the enhanced recal1.

The environmental context cue should be open tothe same manipulations as the other cues. For instance,as the difference between the two study contextsincreases, recall should show a concomitant increasewhen testing is in a eontext randomly related to theinput contexts (e.g., Cue 3 in Figure 1). When reca11 istested in a eontext that is similar to one of the inputcontexts (e.g., Cue 1 or 2 in Figure 1), recall should

COMPONENT-LEVELS THEORY OF SPACING EFFECTS 103

rather than as individual words, (2) the subjects wouldhave no trouble finding a relationship between the twowords and storing that relationship, and (3) the twopairs in the varied condition would be viewed as relatedpairs (or a triple ), rather than two unrelated pairs. Ifthese assumptions are met, then we can represent theimportant components of the memory trace, after thetwo presentations, as folIows:

The "D" represents descriptive components, the "S"represents the structural relation between the stimulusand response, and the "C" represents the contextualcomponents. For the traces presented in the variedconditions, the subscripts indicate that the stimulusdescriptive components come from two differentstimuli. For the DP traces the primes indicate variabilityin components (descriptive and contextual) due tolocal contextual changes. As can be seen, there isdifferential storage of contextual components that ishighly correlated with the MP-DPvariable, but relativelyindependent of the varied-constant variable.

Next, consider the descriptive components. Since twostimulus terms are always presented in the variedconditions, there are more different stimulus descriptivecomponents stored than in the constant conditions.Within the varied conditions there is no difference inthe descriptive components between the MP- andDP-varied pairs. Within the constant conditions theMP-DP variable does produce differential storage ofdescriptive components. Local contextual changesincrease the number of different descriptive componentsin the DP-constant pairs relative to the MP-constantpairs.

Component-levels theory predicts a differentinteraction between the varied-constant and MP-DPvariables on the cued and free recall tests. On the cuedrecall test the subject is given a stimulus to use as a cueto retrieve the pair. Since descriptive components areavailable, the contextual components will not greatlyinfluence performance. In the constant conditions,the encoded stimulus (the retrieval cue) is more likelyto share descriptive components with pairs in the DPcondition than in the MP condition. Therefore, recallof the DP pairs should be superior to recall of the MPpairs. In the varied conditions the subjects are givenboth stimuli, one at a time, with which to retrievethe response. The likelihood that these cues sharecomponents with the traces is equivalent for the MP andDP pairs in the varied conditions, but greater than forthe pairs in the constant conditions. Therefore, in thevaried conditions cued recall is predicted to beindependent of the MP-DP variable and greater than inthe constant condition.

In free recall the retrieval cue consists of the context

not increase monotonically with the difference betweenthe contexts. The latter prediction is tested andconfirmed in Experiment 3.

In summary, this review has shown that component­levels theory can account for many of the standardfindings in experiments on the spacing of repetitions,as well as integrating results across paradigms. Alsoillustrated are various methods that can be used to testthe theory. Unfortunately, most of the experimentsreviewed were not specifically designed to test thetheory, and therefore, some of the analyses are post hoc.The three experiments to be reported here were designedto provide new tests of the theory's explanation ofdistributed practice effects, especially in regard tocontextual components. Experiment 1 provides evidencefor distributed practice effects due to descriptivecomponents and local contextual components. Theresults of the experiment also provide support for theassumption that cuing with unique componentseliminates effects due to differential storage of generalcomponents. Experiment 2 examines the interactionbetween structural components and contextual compo­nents, while providing direct evidence for cue bias.In addition, the experi-nent provides a plausible accountfor the very long lag effects observed in free recall.Experiment 3 does not include distributed practice asan independent variable, but involves a more directmanipulation of differential storage of componentsrepresenting the environmental context. The experimentreplicates and extends S. M. Smith et al. (1978). Theresults demonstrate that differential storage of theenvironmental contextual components can lead toimproved memory performance, but only in conjunc­tion with retrieval cues that take advantage of thisdifferential storage.

EXPERIMENT 1

This experiment was designed to examine interactionsbetween descriptive components and local contextualcomponents. The subjects in the experiment were askedto study a list of paired associates. The list was com­posed of normatively related pairs (e.g., blade-knife).All pairs (except for primacy and recency buffers)were repeated within the list. Half of the pairs weregiven MP, and the others were given DP. At the secondpresentation, half of the pairs were repeated exactly(constant condition), while for the remaining pairs, theright-hand member was the same, but the left-handmember was changed, for example, spoon-knife (variedcondition). After the input stage of the experiment,half of the subjects were given a free recall test, andthe other half of the subjects were provided with theleft-hand members of the pairs as cues for the right-handmembers.

Normatively related ward pairs were chosen to insurethat (1) the subjects would study the pairs as pairs

MP-constant DS + CDP-constant DD'S + CC'

MP-varied D1DzS +CDP-varied D1DzS +cc'

104 GLENBERG

9r-----------~---,

Figure 2. Mean cued and free recall of response terms as afunction of type of repetitions and the spacing between therepetitions.

The subjects given free recall were asked to write down asmany of the words as they could remember, either singly or inpairs. The subjects given the cued recall tests were handed thebooklet and instructed to try to write down a response wordnext to each stimulus word. They were encouraged to write aword on every line, but they were not forced to guess. Recallwas self-paced for both groups.

VARIED..CONSTANT

j---~

DISTRIBUTED

CUED RECALL

VARIED

--------~---Ir------

...

CONSTANT_...

---------------...-

FREE RECALL

MASSED

2

7

3

.8

ow.J

:i. .6lrlCl:

55~Cl.04Cl:Cl.

Results and DiscussionThe probability of a Type I error is set at .05 for all

statistical analyses.The data of main interest are presented in Figure 2.

The interaction between type of recall, varied-constant,and distribution was statistically significant. To facilitatethe interpretation of the results, two identical analysesof variance were performed, one for cued recall andone for free recall. Each analysis was an 8 by 2 by 2(counterbalancing by varied-constant by distribution)analysis, with the last two factors being within subjects.For cued recall, the effects of varied-constant repe­tition [F(l ,24) =10.69, MSe =1.35], distribution[F(1 ,24) = 5.81, MSe = .84], and their interaction[F(1 ,24) =5.44, MSe =.63] were all significant. Forfree recall, the effect of varied-constant [F(1 ,24) = 6.29,MSe = 1.61] and distribution [F(1,24) = 5.67, MSe =1.41] were significant. The F ratio for the interactionwas only .18.

Clearly, all of the predictions of component-levelstheory were upheld. For cued recall, the descriptivecomponents in the cue appear to provide access to thetraces. In the constant conditions, this cue along withdifferential storage of descriptive components pro ducesthe MP-DP difference. In the varied conditions, maxi­mum encoding variability of descriptive components isinduced for both MP and DP pairs. Therefore, cuedrecall is high and unaffected by the MP-DP variable.In free recall, retrieval is controlled by the contextual

MethodSubjects. A total of 64 subjects was drawn from intro­

ductory psychology courses taught at the University ofWisconsin-Madison. Participation in experiments is one way offulfilling a course requirement. They were randomly assigned tothe free and cued recall conditions.

Materials and Design. Sixty-one tripies of words were drawnfrom various association norms. What was to become theresponse term in the present experiment was the stimulus wordin the norms. One of the other words was a high-frequencyresponse, the other was a low-frequency response. Thirteen ofthe tripies were used for practice and buffer pairs; the other 48tripies were used in the main experimental conditions.

The basic list structure consisted of five once-presentedprimacy buffer pairs, followed by the presentation of 24 pairs,each of which was repeated, and ending with five once-presentedrecency buffer pairs. Half of the repeated pairs were in theconstant condition; that is, both members of the pair wererepeated. For the other pairs, the stimulus term was changedwhile the response term remained the same on the secondpresentation (varied condition). Orthogonal to this manipu­lation, half of the pairs were given MP and half of the pairs weregiven DP. The distribution lag varied from two to six interveningpairs. The presentations (excluding the buffers) were arrangedin blocks of eight. Within a block, one pair of each of the fourtypes was presented twice.

Counterbalancing over subjects insured that (l) each pairwas used equally often in each of the four main conditions,(2) in the varied condition each of the two stimuli associatedwith a given response was used equally often on the first andsecond presentations of the pair, and (3) in the constantcondition each stimulus was used equally often.

The subjects who received cued recall were given a bookletcontaining 54 pages. On each page was a single stimulus term anda blank space for the response paired with that stimulus. Thefirst six stimuli consisted of three stimuli from each of thebuffers. The remaining 48 pages consisted of two cued recalltests for each response term. One of the tests was in the first24 pages; the other was in the last 24 pages, For those pairsgiven varied presentations, both of the presented stimuli wereused as cues once. The order of the presentation of sets of 24recall trials was counterbalanced over subjects.

Procedure. The subjects were run individually. The pairs wereprojected, one at a time, using a Kodak Carousel projector.The total study time for each pair (including the time neededto change slides) was 4 sec. The subjects were instructed thatthey were to see a long list of pairs of words. They were to tryto memorize the pairs as pairs, not as single words. In addition,they were told that their memories for the pairings would betested, although they were not given any specific details of thetesting procedure. The subjects viewed three practice pairsbefore beginning the main list.

at the time of the test. This cue is more likely to sharecomponents with the pairs given DP than MP, regardlessof the varied-constant variable. DP items should berecalled better than MP items in both varied andconstant conditions. In addition, the pairs given variedpresentations will be recalled better than the pairs giventhe constant presentations. If the subject falls to storea relation between the stimulus and response of a pairgiven varied presentation on the first presentation, thesecond presentation offers an essentially independentopportunity for learning. Unless there is a sufficientchange in context (and hence a chance to encode a newrelationship between the stimulus and response), thiswill not occur for the pairs given the constant repetition.

COMPONENT-LEVELSTHEORY OF SPACING EFFECTS 105

components avai1able on the test. Since differentialstorage of contextual components is independent ofthe constant-varied manipulation, the MP-DP effect isfound in both conditions.

In the introduction, predictions were made assumingthat the effect of differential storage of general compo­nents could be eliminated by providing retrieval cuescomposed of unique components on a different level.This assumption, if valid, provides a convenient methodfor testing the theory. The differences between theresults on the cued and free recall tests provide somevalidation for the assumption. Differential storage ofgeneral contextual components underlies the MP-DPeffect in free recall where the retrieval cue is composedpredominately of unique descriptive components,differential storage of contextual components does notaffect recall.

Predictions similar to those made above have beendisconfirmed in other studies. Shaughnessy, Zimmerman,and Underwood (1974), for example, reported anexperiment similar in design to Experiment I, but withquite different results. There are a number of differencesbetween the experiments, of which the most importantfor cornponent-levels theory is that unrelated wordswere used to construct the pairs. This difference mayhave led the subjects in the varied conditions to storecompletely different traces for each presentation,nullifying many of the predictions.

EXPERIMENT 2

This experiment was designed to examine the effectsof differential storage of contextual components andtheir interaction with structural components. Asdiscussed before, the context is composed of somerelatively quickly changing aspects (e.g., the wordscurrently being processed) and some components thatchange less quickly (e.g., the current list being processed,the subject's mood). The major concern of Experiment 2is with the latter, more global components of thecontext. The basic idea is to isolate the effect ofdifferential storage of global contextual componentsby insuring constant (with respect to lag) encodingvariability of descriptive components and structuralcomponents. On the recall test the relationship betweenthe contextual cue and the components in the trace ismanipulated as in Figure 1 to produce monotonicallyincreasing and decreasing lag functions.

The design involved the presentation of 10 or 11lists of words for free recall. Within each list, wordswere repeated at lags of 0, 2, 8, and 20. Some of thelists were recalled immediately after their presentations,others were not recalled until a final recall test. Thelists not recalled immediately will be referred to asthe criticallists. Embedded within the criticallists wererepeated words whose second presentations were inList 10. The seria1 position of the critical lists is, there-

fore, inversely related to the between-lists repetitionlag. The words repeated between the lists all havemaximum or elose to maximum encoding variability ofdescriptive and structural information (based on changesin the local context), regardless of the between-listslag. Only the global contextual components can beaffected by differential storage.

After presentation of List 10 (which contained therepetitions of the words repeated between lists), halfof the subjects were given an immediate final free recallfor all of the lists. The global context during this testwill be similar to the global context during the presenta­tion of the recent lists. The contextua1 componentsin the cue, therefore, will share many components withthe traces formed during presentations in recent lists(short between-lists lags) and fewer components withtraces whose first presentations were in the earlier lists(long between-lists lags). The contextual cue is biasedlike Cue 1 or 2 in Figure l, hence recall of the wordsrepeated between lists is predicted to be monotonicallydecreasing or an inverted U-shaped function of thebetween-lists lag. Based on the same reasoning, thetheory predicts a general list-recency effect. Thecontextual components in the cue are biased toward(inelude components from) the last few lists. Itemspresented only on the last lists shou1d be recalled betterthan those presented only in the first lists (Bjork &Whitten, 1974).

For the other subjects, List 10 was not immediate1yrecalled, but List 11 was presented and it was recalled.These subjects received a delayed fmal recall of all thelists after an interval of 1.5 to 2.5 h. At this time thebias in the global contextual components of the cue isgreatly reduced. As with Cue 3 in Figure 1, recall of thewords repeated between lists is predicted to be a directfunction of differential storage of the global contextualcomponents and to increase monotonically with thebetween-lists lag. Additionally, since the cue is unbiased,the list-recency effect should be eliminated. Itemspresented only on the last lists should be recalled aboutas well as items presented only on the first lists.

MethodMaterials and Design. Four different list structures were used.

One of the structures was used for List 10, which containedthe second presentations of the words repeated between lists,one was used to construct the four critical lists, and the othertwo structures were used to construct the remaining six lists.All of the structures contained 54 serial positions, of which thefirst 5 and last 6 were filled with once-presented words to makeup the primacy and recency buffers, respectively. Except for theList 10 structure, within the body of the structures were placesfor four exemplars of words repeated at each lag of 0, 2, 8, and20. In addition, there were spaces for 11 words presented oncein the list. The differences between the list structures were inthe assignment of the various conditions to serial positions.In each of the four critical lists, 4 of the 11 once-presentedwords were given a second presentation in List 10.

Within the body of List 10 were two exernplars of wordsrepeated within the list at each of Lags 0, 2, 8, and 20, 11 words

106 GLENBERG

208WITHIN-L1ST LAG

/",./~-----------------'"

"'....

IMMEDIATE RECALLNON-CRITICAL LISTS

o 2o'-;!;-_;!;- + """

----~-----------------~...----10 y"'

35r-~=c:~=-:--,-,-----------:IO

.30

o~

~25

Results and DiscussionThree different analyses of recall of the repeated

words were conducted. The dependent variables were(I) immediate recall from the noncritical lists inPositions 1, 2, 4, 5, and 7, (2) final recall from thepreviously unrecalled (critical) lists in Positions 3, 6, 8,and 9, and (3) final reeall of the words repeated betweenlists. Each analysis of variance, except for the last,included the faetors delay of final recall, list order,counterbalancing within a list, list serial position , andwithin-lists lag. The last two factors were within-subjectsmanipulations. The analysis of the recall of the wordsrepeated between lists did not include the within-listslag factor.

Consider first the immediate recall from thenoncritical lists. The important means from thisanalysis are presented in the upper panel of Figure 3.The effect of within-lists lags was statistically significant[F(3,240) =46.65, MSe =.62] . There was also a maineffect for list serial position [F(4,320) =3.01, MSe=.78) ;recall declined from about 30% on the first list tobetween 25% and 27% on the others. There was no maineffect for delay of final reeall, although this factor didinteract with the within-lists lag [F(2,240) = 3.59,MSe = .62). For unknown reasons the subjects in thedelayed final recall condition did not show an increasein recall from Lag 8 to Lag 20 on the immediate recalltests. Other significant effects include a main effect forwithin-lists counterbalancing and higher order inter­actions involving the counterbalancing variables. Theseinteraetions appear to be due to unsystematic variationsin recall and do not affect the main conclusions.

The final recall of the words repeated within thecritical lists offers a number of tests of component­levels theory. Onee the eontextual cue provides aeeessto a trace, the structural components in the trace may

Figure 3. Mean recall of words repeated within the lists.Solid lines represent recall from subjects who received theimmediate final recall. Dotted lines represent recall from subjectswho received the delayed final recall.

given a single presentation, and the second presentations ofthe 16 words that had received their first presentations on thefour criticallists.

Lists presented in Positions I, 2, 4, 5, and 7 were recalledimmediately after their presen tations. The critical lists werepresented in Positions 3, 6, 8, and 9. The between-lists repe­tition lags were, therefore, 0, 1, 3, and 6, corresponding towords whose first presentations were in Lists 9, 8, 6, and 3,respectively.

Four different list orders (assignments of specific lists to listserial positions) were used. Across the different list orders, eachof the critical lists occurred equally often in Serial Positions 3,6, 8, and 9. This procedure served to counterbalance bothspecific words to serial positions and the words repeatedbetween lists to between-lists lags. The serial positions of thefive noncritical lists were randomized in each list order underthe constraint that a given list could not occur more than oncein a given serial position. Within each list order, the wordswithin a list were counterbalanced to insure that a repeatedword appeared equally often in each of the between- and within­lists lag conditions. List 10 (and for the subjects who receivedthe delayed final recall, List 11) was the same for all subjects.

In all, 418 common, single-syllable, four- and five-letternouns were used as TBR words. The words were assigned to thelists and conditions randomly except for the constraint thateach list have approximately the same proportion of four- andfive-letter words.

Procedure. All subjects were run individually. They wereinstructed that the experiment was designed to measure theeffect of recalling one list upon the recall of the next list.Consequently, they would have to try to memorize each listof words, although only some would be recalled and the restwould be followed by arithmetic problems. The subjects werealso informed that some words would be repeated. They werenot informed of the total number of lists, which lists would berecalled immediately, or that a final recall would be requested.

The words were - projected one at a time with a KodakCarousel slide projector controlled by external timing apparatus.The total study time per word, incIuding the time betweenslides, was 3 sec. After the presentation of each list, the subjectswere given one of two instructions, either to recall the wordsor to do arithmetic problems. When instructed to recalI, theywere given 1 min to free recall the words. After presentationof the critical lists, the subjects were instructed to do arithmeticproblems. Simple multiplication problems were presented, oneevery 2 sec, for a total of 1 min. The subjects were required towrite down the answers to the problems.

The subjects receiving the immediate final recall received anadditional instruction immediately following the presen tationof List 10. They were told at that time to try to recall all ofthe words from all of the lists. In addition, they were told thatthey would receive a bonus of 1 cent for every word they couldrecall.

The subjects receiving the delayed final recall performedarithmetic following List 10. They then studied and recalledList 11. These subjects returned to the laboratory between 1.5and 2.5 h later. In the second session, these subjects were askedto recall all of the words from all of the lists. They were alsopaid the bonus of 1 cent for each word recalled.

Subjects, A total of 120 volunteers was drawn from theUniversity of Wisconsin introductory psychology classes.Participation in experiments is one way of fulfiHing a courserequirement. The first 60 subjects were assigned to theimmediate [mal recall condition. The next 60 subjects wereassigned to the delayed final recall condition. The two halvesof the experiment were separated by the semester break and,therefore, the two sets of subjects were drawn from differentcIasses. Although the subjects were sampled at different times ofthe year, the two halves of the experiment were conductedduring the same relative parts of the acadernic semesters.

COMPONENT-LEVELS THEORY OF SPACING EFFECTS 107

be used as an associative cue to other traces. Sincestructural components within each list have undergonedifferential storage, they produce a within-lists lageffect. This effect will be independent of the list serialposition since the stored relationships represented bythe structural cues are not influenced by any globalcontext bias at the test. For the final recall of wordsrepeated in the critical lists, the main effect of within­lists lag was significant (see Figure 3) [F(3,240) = 2.88,MSe =.41]. Within-lists lag did not interact with eitherlist serial position or delay of final recall (both Fs < 1).

Final recall of words repeated within the criticallists, averaged over within-lists lag, is presented inTable 1. The list-recency effect on the immediate finalrecall illustrates the effect of biased global contextualcomponents in the cue. The flat list serial positionfunction in the delayed final recall indicates an unbiasedglobal contextual cue. The interaction was significant[F(3,240) = 4.21, MSe = .59]. This confirrnation ofchanges in the bias of the global contextual componentsis a necessary condition for the predictions, based onthese changes, of recall of the words repeated betweenthe lists.

Analysis of recall of words repeated between thelists (Figure 4) reveals main effects for delay of finalrecall [F(l,80) = 18.58, MSe = 1.21] and within-listscounterbalancing [F(4,80) = 3.69, MSe = 1.21]. Thebetween-lists lag (list serial position) did not contribute,by itself, any significant variability (F < 1). Theimportant prediction is that, given the changes in thebias of the global contextual components indicated inTable 1, between-lists lag will interact with delay of finalrecall. Recall on the delayed final recall is predicted tobe a monotonically increasing function of between-listslag. Recall on the immediate final recall is predicted tobe either a monotonically decreasing function ofbetween-lists lag (Cue Type I, Figure 1) or an invertedU-shaped function of the between-lists lag. To discrimi­nate between the latter two possibilities, the interactionmean square is broken into orthogonal contrasts. Onlythe linear component of the interaction of between­lists lag and delay of final recall was significant[F(l,240) = 6.13, MSe = .55]. This resu1t indicates thatrecall on the immediate final recall is best describedby Cue 1 in Figure 1, while recall on the delayed finalrecall is best described by Cue 3.

lt should be noted that the results depicted inFigure 4 are not a simple consequence of the list serialposition variable or retention interval. Suppose, forexample, that the decrease in the immediate final recallfunction is due to forgetting of the information stored

Table 1Proportions Recalled of Words Repeated Within the CriticaI Lists

- - --- --- ------,

Figure 4. Mean recall of words repeated between lists onthe fmal recall. The solid line represents immediate final recall;the dotted line represents delayed final recall.

at the first presentation, which has a retention intervalpositively correlated with the between-lists lag. Thissupposition cannot explain the increase in the delayedrecall function. Neither can the increase in the delayedrecall function be explained as a list-primacy effect;the first presentation of the longest between-lists lagcondition is in List 3. Additionally, there is virtually noprimacy effect indicated in Table 1. The best explana­tion of these results is that provided by component­levels theory. Differential storage of global contextualcomponents and retrieval using global contextualcomponents can be manipulated in ways directlyanalogous to the storage and retrieval of descriptive(Glenberg, 1976) and structural (Glenberg, 1977)components.

An interesting implication of these resu1ts is that,under the right conditions, there may be no limit to theimprovement in performance that can be obtained byincreasing the repetition lag. After short lags differentialstorage of descriptive components may have peaked,but differential storage of structural components is stillcontinuing. After moderate lags, perhaps on the order of20 to 40 intervening items in a free recall list, structuralcomponents have reached their limit of differentialstorage, but components of the global context may havenot reached their peak differential storage. Hence, thetheory can explain the effects of very long lags thathave been found in free recall, and it predicts long lageffects in other situations where differential storageof contextual components could facilitate retrieval.

List Serial PositionTime of Final

Recall

ImmediateDelayed

3

.11

.13

6

.13

.11

8

.13

.12

9

.17

.11

EXPERIMENT 3

Like Experiment 2, this experiment was designed totest predictions of the theory concerning the differentialstorage and retrieval of global contextual components.

108 GLENBERG

The repetition lag, however, was not manipulated.Differential storage was direetly eontrolled by ehangesin the stimulus eonditions. Subjeets viewed a single40-word list in two experimental sessions separated bya 3·h intervaL The sessions were held in either the samephysical environment or different environments.Presumably, the former eondition results in the storageof similar global eontextual eomponents at bothpresentations, while the latter eondition results intraees eontaining different eomponents representingboth environments. Reeall of the list was requestedin a third session held 3 h after the seeond. The thirdsession was eondueted either in one of the old environ­ments or in a new neutral environment dissimilar to theinput environments.

The environment provides a souree of globaleontextual eomponents that are used to retrieve thestored information (S. M. Smith et a1., 1978). Consider,for subjeets who reeall in the neutral environment,how those eomponents interaet with the stored globaleontextual information. The total pool of eomponentsshared by the neutral environment and both inputenvironments is greater than the pool of eomponentsshared by the neutral environment and either inputenvironment alone. Therefore, when the contextualeomponents provided by the neutral eontext are theonly retrieval cues, words studied in two differentinput environments will be more aetivated and betterreealled than the words studied twiee in a single inputenvironment.

When reeall is eondueted in one of the original inputenvironments, the global eontextual componentsprovided by the environment are biased; for the subjeetswho studied in two different environments, thecontextual components match those stored at onepresentation (the one in the test environment), but notthose stored at the other presentation. The situationis analogous to that represented by Cue 1 in Figure 1and the immediate final reeall of Experiment 2. Reeallin the biased test eontext will not favor the subjeetswho study the list twice in different environments.The operation of the theory is too inexplicit, however,and our measurement of environmental similarity istoo erude to confidently predict that subjeets whostudy the list twiee in the biased test environment willrecall more than the subjeets who study the list twieebut only onee in the test environment.

In summary, the theory predicts an interaetionbetween the input environment (same or different)and the test environment (biased or neutral). It isalso likely that reeall in the neutral environment willbe inferior to recall in the biased environment. That is,the total number of components availab1e in the neutralenvironment and included in the traces may be lessthan the number of components availab1e in the biased(old) environment and shared by the traces.

MethodSubjects. A total of 64 subjects was drawn from the student

population at the University ofWisconsin-Madison. The partici-

pants were paid $3 for their time or could use the experienceas one way of fulfilling an introductory psychology courserequirement.

Materials and Design. Three environmental contexts wereused. Room A was located on the fifth floor of the BrogdenPsychology Building. It was decorated to look like a laboratory.The room contained a computer and assorted peripheralequipment. The subject sat in a soundproof booth in the room.The booth was approximately 6 x 6 ft and contained achair,desk, computer terminal, and various small pieces of equipmentdirectly in front of the subject. In this context the stimuli werepresented auditorily via a tape recorder.

Room B was in the basement of the same building andappeared more like an office than a laboratory. The floor wascarpeted, the walls were decorated with posters, books andplants were in the room, and the lights were partially concealedbehind a brightly colored print sheet. The room also containedchairs, tables, and cabinets. Stimuli were presented visually,on slides, in this context.

Room N, the neutral environment, was located on thesecond floor of the building. This room had bare walls andfloor and contained only combination desk-chairs.

The four major conditions were formed by the factorialcombination of two levels of input context (sarne or different)and the test context (biased or neutral). A total of 16 subjectswas assigned to each major condition. In the same inputcontext/biased test context condition, the order of the contextsin which the three sessions were conducted was AAA for eightsubjects and BBB for eight subjects. For the subjects in thesame input context/neutral test context condition, half receivedthe AAN order and half the BBN order. In the different inputcontext/biased test context condition, four subjects wereassigned to each of the context orders ABA, ABB, BAA, andBAß. The context orders for the different input context/neutraltest context condition were ABN and BAN, which were usedequally often.

All subjects studied the same list composed of 40 common,single-syllable four-letter nouns. The first liveand last five wordswere treated as primacy and recency buffers, respectively, andwere eliminated from the data analysis.

Procedure. All subjects were met at a waiting room on thethird floor of the building, and then they were escorted to thefirst room. The subjects were instructed that they were to studythe list of words for a test. The words were presented one at atime at a 3-sec rate, either visually (Room B) or auditorily(Room A). Immediately following this presentation, the subjectswere instructed to rate each word in terms of its affectivevalueon a continuum from good to bad. The same list was againpresented in the same modality at a IO-sec rate, during whichthe subjects rated the words. The subjects were dismissed afterrating the words. The affect rating was used to insure adequatememory for the words and to give the session a senseof closureso that subjects wouId not actively rehearse the words betweensessions. They were not informed that they would be requiredto recall the words at a later session.

Three hours after the first session, the subjects were met atthe third-floor waiting room and escorted to either the same ora different context. The same procedure (two presentations ofthe list) wasfollowed.

Three hours after the second session, the subjects wereagain met on the third floor and escorted to either a biased(A or B) or neutral (N) context. They were given 10 min tofree recall the words.

Results and DiscussionThe means of the four major conditions are presented

in Table 2. Statistica1 analysis consisted of a 2 (inputeontext) by 2 (test eontext) by 2 (Room A or Room Bfirst) analysis of variance. Recall by the subjects whosefirst sessions were in Room A (.67) was superior torecall by the subjects whose first sessions were in

COMPONENT-LEVELS THEORY OF SPACING EFFECTS 109

GENERAL DISCUSSION

Table 2Proportions RecaUed of Repeated Lists

Comparison to Other Explanationsof Spacing Effects

Component-levels theory as currently formulated

Room B (.54) [F(1 ,56) == 28.20, MSe == 8.94]. Thiscounterbalancing effect did not interact with any othervariable and will not be considered further . The maineffect of test context did not reach standard levels ofsignificance [F(1 ,56) == 2.94, p < .09], although themeans were ordered as predicted.

The critical result is the Input Context by Test Contextinteraction. This effect was significant [F(1,5 6) == 4.55].Reca11 after studying in two different environmentswas superior to recall after studying in a single environ­ment, when tested in a new environment. Nonetheless,variability in the study environments does not affectreca11 when tested in one of the original study environ­ments. In terms of component-levels theory, differentialstorage of the global contextual components facilitatesreca11 when retrieval is based on unbiased contextualcomponents. On the other hand, when retrieval is cuedwith biased global contextual components, the effectsof differential storage are modified as predicted.

One objection to the foregoing surnmary is that theeffects attributed to the global context may be due todifferential storage of components on many levels.That is, studying in two different environments mayhave led to greater differential storage of contextual,structural, and descriptive components than studyingtwice in the same environment. This objection , however,will not explain the interaction. Assurne for a momentthat more structural components were stored in thedifferent input context conditions than the same inputcontext conditions. This advantage would account forthe difference in recall when tested in the neutralcontext. The objection predicts the same difference,however, when recall is requested in a biased context.That is, the variability in the structural componentsshould manifest itself in an identical manner in bothtest contexts.

In conclusion, this experiment, along with Experi­ment 2, has demonstrated that manipulations at thelevel of the global context have effects analogous to theeffects of similar manipulations of the local context(Experiment 1), structural components (Glenberg,1977), and descriptive components (Glenberg, 1976).Differential storage at any level sets an upper boundon performance. The components in the retrieval cuecontroI the level of performance within this bound.

provides a qualitative explanation for the effects ofspacing on reca11 and recognition of repeated events.In addition, the theory generates a number of testablepredictions and suggests methods for testing the theory,as weIl as allowing for expansion in many directions.This explanatory power has been purchased, however,at the expense of simplicity. Predietions depend onconsidering a number of sources of information duringinput processing, the mode of processing, changes onboth of these dimensions during the repetition intervaland at successive presentations, sources of informationat testing, retrieval processes, and how a11 of thesefactors interact. An important question is whether thistheoretical expense buys any explanatory advances oversimpler accounts.

Consider an encoding variability theory based onlyupon the descriptive components. This sort of theorycould certainly account for a11 of the phenomena thatcomponent-levels theory can predict on that one level,including those found in the recognition and pairedassociate paradigms. The single-level theory can also bestretched to include the relatively straightforwardfindings in other paradigms. It falls, however, whenconsidering interactions between levels of componentsand interactions on levels other than the descriptivecomponents. For example, Glenberg (1977) found thatthe monotonie lag effect in free reca11 is modified bythe retrieval environment. The monotonie functioncould be described as reflecting differential storage ofdescriptive components, making the traces easier toretrieve, recognize, or decode. It is difficult to see,however, why, in such a model, providirtg input wordsthat were adjacent to the repeated words would modifythe shape of the lag function. To account for such afmding, the model must represent interitem relations.Similarly, the single-level encodirtg variability modelcould predict that cuing reca11 with the stimuli in thevaried condition of Experiment 1 would elirninate thespacing effect. The reinstatement of the effect underfree (contextually cued) reca11 would, however, pose amystery for the model. The single-level structural modelwould fare no better in accounting for these data.

A single-component model based on the contextualcomponents does better than the other single-levelmodels (see Hintzman, 1976). Part of the superiority isdue to contextual components being themselves alignedon a dimension (from local to global) that encompassesa range of cues varying in generality. Nonetheless,the contextual dimension alone fails to explain between­levels irtteractions. For example, in Experiment I, thedistributed practice effect demonstrated in free recal1can be taken as evidence for the differential storageof contextual components. The single-componentcontextual model could then explain the interactionbetween varied-constant presentation and distributionin cued recall in one of two ways. First, the modelcould propose that the retrieval cues used in cued recallobscure the differential storage of the contextualcomponents. This proposal, however, cannot account

Input Context

.64 .62

.54 .63

Same Different

BiasedNeutral

TextContext

110 GLENBERG

for the distributed practice effect in the constant presen­tation, cued recall data. Altematively, the model couldpropose that the varied presentation induces maximumencoding variability. This notion accounts for the inter­action in cued recall, but it incorrectly predicts the sameinteraction in free recall. That is, if varied presentationsinduce maximum encoding variability of contextual ele­ments that eliminates the distributed practice effect incued recall, then the varied presentations should alsoeliminate the distributed practice effect in free recall.

Component-levels theory also compares favorablyto the many "deficient processing" explanations(Hintzman, 1976) of distributed practice effects. As adass, these explanations propose that performancefollowing massed practice is depressed due to incompleteor inefficient processing compared to distributedpractice conditions. Hintzman (1976) further classifiesthe theories by the locus of the deficient processing(on the second presentation or between the first andsecond presentations) and whether the processingdeficiencies are due to voluntary or involuntaryprocesses. None of these models considers retrievalprocesses. Therefore, they cannot predict the sorts ofinteractions seen in Experiment 1. For example, sincethe distributed practice effect is eliminated by variedpresentations (in cued recall), we can assurne that variedpresentations eliminate the deficient processing. There­fore, these theories must incorrectly predict the absenceof a distributed practice effect in free recall.

A third dass of theories appears to be having somesuccess at least in predicting a new type of spacingeffect. These explanations depend on an analysis ofthe retrieval requirements during input, specifically,retrieval of the first presentation information at the timeof the second presentation (i.e., study-phase retrieval).Thios and D' Agostino (1976) report that when retrievalof first occurrence information (the sentence context ofan object phrase) at the time of the second presentationis encouraged, then a recall test produces a monotonie,increasing spacing effect. When the first presentationsentence is provided at the second presentation, thespacing effect is attenuated or eliminated. Whitten andBjork (1977) report a similar phenomenon.

Thios and D'Agostino (1976) propose that the study­phase retrieval may modify the retrievability of thefirst presentation information when similar retrievalinformation is used at later tests. After a short spacinginterval, study-phase retrieval of first occurrenceinformation may be accomplished by a relatively easy,temporally directed backward scan. As the spacinginterval increases, this type of retrieval operationbecomes less effective, and study-phase retrieval beginsto depend on a reconstructive process that attempts tomatch the encoding of the currently presented itemwith previous encodings. If it is assumed that laterretrieval typically depends on this reconstructiveretrieval, and that the reconstruction retrieval on thetest is facilitated by study-phase reconstructive retrieval,then the monotonic lag effect is predieted.

This retrieval-based explanation is potentially capableof predicting many of the interactions between spacingand the retrieval environment when it is assumed thatthere is positive transfer between either type of study­phase retrieval and the corresponding test retrieval.For example, when testing recognition memory aftera short retention interval, retrieval will be based on thebackward-scan mechanism. Previous study-phaseretrievals at short spacing intervals may facilitate thetest retrieval, while previous study-phase retrievals atlong spacings (where constructive retrieval is used)may not. Therefore, after a short retention interval,MPs are predicted to be recognized better than DPs.Recognition memory tested after a long retentioninterval may be based on constructive retrieval. Theconstructive retrieval practiee during distributed studywill facilitate retrieval at the delayed test, yielding themonotonie, increasing spacing effect.

This retrieval-based theory can only predict beneficialeffects of the repetition lag when the retrieval processat the test is similar to that used during study. Onlythen will the retrieval practice benefit retrieval at thetest. It is difficult, however, to reconcile this positionwith spacing effects in the typical free recall paradigmwhere the retrieval environment on the test is unlikethe putative cue at the second presentation, the nominalstimulus itself.

Problems with Component-Levels TheoryThe theory is vague and ill-specified on a number

of points, partly because of our general ignoranceabout the workings of memory, and partly because thetheory makes similar predietions regardless of the exactspecification. One such area is in the definition of thecomponents: why three components, why these specificcomponents, are they different in kind, and whieh tasksand processes lead to the storage of whieh components?Essentially, the components are meant to represent arange of information stored in memory. Grouping thesetypes into three levels may be a convenient fiction.The important points for the theory of repetitions arethat (1) not all components show differential storagein the same input conditions, (2) performance is afunction of the overlap between the stored informationand the retrieval cue, and (3) this overlap changes asa function of the retrieval environment (essentially,Flexser & Tulving's, 1978, retrieval independence).Within these constraints, and within the constraints ofour present state of knowledge, it makes little differenceif there are three or many more types of components,or the exact nature of the components. Nonetheless,specifying a grouping of the components into threelevels does have its advantages: It allows for ease ofcomparison to other theories of memory, and it providescontact between the theory and various paradigms forease of testing the theory.

The theory does suffer from its vague approach tocontext. The theory makes use of the concept of

COMPONENT·LEVELS THEORY OF SPACING EFFECTS 111

context in deriving predictions about encoding, storage,and retrieval, but nowhere is there a rigorous specifica­tion of what context is, or how it works. Unfortunately,this is a shortcoming of most theories of episodicmemory. What is, perhaps, most strange in component­levels theory is the representation of context as discretecomponents of the episodie trace. This representationwas chosen to allow the same retrieval processes tooperate at all levels. The burden, however, is shifted tothe storage processes. With the assumption that allaspects of the context (Iocal and global) are auto­matieally included in all traces, the theory is forced tocreate extremely cumbersome traces to represent eventhe most trivial events.

One solution to the problem is to assurne that traeesare embedded within a context rather than the reverse.That is, we remember events within the eontext ofother events. These events may be combined intohigher order episodes or schemata (Kintsch, 1978),or in an associative network (Anderson & Bower, 1973)to provide aglobai context. The translation ean becompleted by assuming that accessing the higher orderepisode provides access to the individual events withinthe episode, that the probability of accessing any givenevent is inversely related to the total number of eventsincluded in the episode (cue overload), and that thepoint at which an episode is aceessed or enteredinfluences which events will be retrieved (eue bias).

Although this embedding seems a natural represen ta­tion of context, within component-levels theory it raisesadditional problems. Specifically, this proposal requiresthat each repetition be represented by functionallyseparate traces. This view is contraindicated. Johnstonand Uhl (1976) conducted a test of a predietion froman encoding variability point of view that rules out themultiple-trace hypothesis. They assumed that the greaterthe encoding variability, the more effective therepetition. They used the subject's positive recognitionresponse at the second presentation as an indication thatsimilar encodings were activated at both presentations.If the subject responded negatively at the seeondpresentation, it was interpreted as an indieation of alarge difference in the encodings. Therefore, those itemsnot recognized at the second presentation should bevariably encoded (essentially independently encodedand represented by two traces) and remembered weIl(if one ignores selection artifacts, and the possibilitythat the negative response rnay be due to not storing arepresentation at the first presentation). The oppositeresult was found: The items not recognized at thesecond presentation were recalled very poorly. Similarly,Madigan (1969) demonstrated that only items reeog·nized as old on their seeond presentations eontributed tothe spacing effect in free recall.

Component-levels theory is not embarrassed bythese results. In the theory, encoding variability isconceptualized as the addition of new information to

the same functional locus or representation, not thecreation of functionally separate traces. Nonetheless,this specifieationdoes not completely solve the problem.The theory does not explain how, at the secondpresentation, the trace representingthe first presentationis found so that new information can be added. Onecould speculate about simple, automatie retrievalprocesses when both presentations are in the same list,but these speculations would not account for the resultsof Experiment 2, where the repetitions are betweenlists.

Ultimately, both single- and multiple-trace repre­sentations of repetitions may be necessary. A singlemulticomponent trace may be an adequate representationfor within-lists repetitions and provide an explanation ofJohnston and Uhl's (1976) results. A multiple-tracerepresentation of repetitions may be necessary toaccount for the data generated by between-listsrepetitions.

REFERENCE NOTES

1. Medin, D. L. Paper presented in R. A. Bjork (Chair),Symposium on spacing phenomena in memory. Symposiumpresented at the Mathematical Psychology meetings, Ann Arbor,1974.

2. Wichawut, C. Encoding variability and the effect 0/ spacing0/ repetition in continuous recognition memory (Tech. Rep. 35).Ann Arbor: Vniversity of Michigan, Human Performance Center,1972.

REFERENCES

ASERNETHY, E. M. The effect of changed environmental con­ditions upon the results of college examinations. Journal 0/Psychology, 1940, 10, 293-3Ol.

ANDERSON, J. R. FRAN: A simulation model of free recall.In G. H. Bower (Ed.), The psychology 0/ leaming andmotivation (Vol, 5). New York: Academic Press, 1972.

ANDERSON, J. R., & BOWER, G. H. Human associative memory.Washington, D.C: Winston, 1973.

ANDERSON, J. R., & BOWER, G. H. A propositional theory ofrecognition memory. Memory & Cognition, 1974, 2,406-412.

ATKINSON, R. c., & SHIFFRIN, R. M. Human memory: Aproposed system and its control processes. In K. W. Spence& J. T. Spence (Eds.), The psychology 0/ leaming andmotivation: Advances in researchand theory (Vol. 2). New York:Academic Press, 1968.

BJORK, R. A. Short-terrn storage: The ordered output of a centralprocessor. In F. Restle, R. M. Shiffrin, N. 1. Castellan,H. R. Lindman , & D. B. Pisoni (Eds.), Cognitive theory (Vol. 2).Hillsdale, N.J: Erlbaurn, 1975.

BJORK, R. A., & WmTTEN, W. B. Recency-sensitive retrievalprocesses in long-term free recall. Cognitive Psychology, 1974,6,173-189.

BOWER, G. H. Stimulus-sampling theory of encoding variability.In A. W. Melton & E. Martin (Eds.), Coding processes inhuman memory. Washington, D.C: Winston, 1972.

CRAIK, F. I. M., & LOCKHART, R. S. Levels of processing: Aframework for memory research. Journal 0/ Verbal Leaming andVerbal Behavior, 1972, 11, 671-684.

D'AGOSTlNO, P. R., & DEREMER, P. Repetition effects as a func­tion of rehearsal and encoding variability. Journal of VerbalLeaming and Verbal Behavior, 1973, 12, 108-113.

112 GLENBERG

DEREMER, P., & D'AGOSTINO, P. R. Locus of distributed lageffeet in free recall. Journal of Yerba I Learning and YerbalBehavior, 1974, 13, 167-171.

EARHARD, B., & LANDRY, D. Dual vs. single eneoding inrecognition memory as a function of lag, delay, and availabilityofsemantie information. Memory & Cognition, 1976,4,609-615.

ESTES, W. K. Statistical theory of spontaneous recovery andregression. Psychological Review, 1955, 62, 145-154.

FLEXSER, A. J., & TULVING, E. Retrieval independence inrecognition and recall. Psychological Review, 1978, 85, 153-171.

GLENBERG, A. M. Monotonie and nonmonotonie lag effects inpaired-associate and recognition memory paradigms. Journal ofYerbal Learning and Yerbal Behavior, 1976, 15, 1-16.

GLENBERG, A. M. Influences of retrieval processes on the spaeingeffect in free recall. Journal of Experimental Psychology:Human Learning and Memory, 1977, 3, 282-294.

GLENB ERG , A. M., & BRADLEY, M. K. Mental eontiguity.Journal of Experimental Psychology: Human Learning andMemory, in press.

GREENO, J. G. How associations are memorized. In D. A.Norman (Ed.), Models of human memory. New York: AeademiePress, 1970.

HINTZMAN, D. L. Reeognition time: Effects of reeeney, frequency,and the spacing of repetitions. Journal of ExperimentalPsychology, 1969, 79, 192-194.

HINTZMAN, D. L. Theoretieal implications of the spacing effeet.In R. L. Solso (Ed.), Theories in cognitive psychology:The Loyola symposium. Hillsd ale , N.J: Erlbaum, 1974.

HINTZMAN, D. L. Repetition and memory. In G. H. Bower (Ed.),The psychology oflearning and motivation (Vol. 10). New York:Academie Press, 1976.

JOHNSTON, W. A., & UHL, C. N. The contribution of encodingeffort and variability to the spacing effect on free recall.Journal of Experimental Psychology: Human Leaming andMemory, 1976, 2, 153-160.

KINTSCH, W. Recognition learning as a funetion of the lengthof the retention interval and changes in the retention interval.Journal of Mathematical Psychology, 1966, 3,412-433.

KINTSCH, W. The representation of meaning in memory.New York: Wiley, 1974.

KINTSCH, W. On comprehending stories. In M. A. Just &P. A. Carpenter (Eds.), Cognitive processes in comprehension.Hillsdale, N.J: Erlbaum, 1978.

LANDAUER, T. K. Memory without organization: Properties of amodel with random storage and undirected retrieval. CognitivePsychology, 1975, 7,495-531.

LoCKHART, R. S., CRAIK, S. I. M., & JACOBY, L. L. Depth ofproeessing, recognition, and recall: Some aspects of a generalmemory system. In J. Brown (Ed.), Recall and recognition.London: Wiley, 1976.

MADIGAN, D. A. Intraserial repetition and eoding processes infree recall. Journal of Yerbal Learning and Yerbal Behavior,1969, 8,828-835.

MANDLER, G. Organization and recognition. In E. Tulving &W. Donaldson (Eds.), Organization of memory. New York:Aeademic Press, 1972.

MASKARINEC, A. S., & THOMPSON, C. P. The within-listdistributed praetice effect: Tests of varied context and variedencoding hypothesis. Memory & Cognition, 1976, 4, 741-747.

MELToN, A. W. The situation with respect to the spacing ofrepetitions and memory. Journal ofYerbal Leaming and Yerba IBehavior, 1970, 9, 596-606.

PETERSON, L. R., HILLNER, K., & SALTZMAN, T. Supplementaryreport: Time between pairing and short-term retention. JournalofExperimental Psychology, 1962, 64, 550-581.

PETERSON, L. R., WAMPLER, R., KIRKPATRICK, M.,& SALTZMAN,D. Effect of spacing of presentations on retention of a paired­associate over short intervals. Journal of Experimental Psychol­ogy, 1963, 66, 206-209.

REDER, L. M., ANDERSON, J. R., & BIORK, R. A. A semanticinterpretation of eneoding specificity. Journal of ExperimentalPsychology, 1974, 102,648-656.

SHAUGHNESSY, 1. 1. Persistence of the spacing effect in freerecall under varying incidental learning conditions. Memory &Cognition, 1976, 4, 369-377.

SHAUGHNESSY, J. J., ZIMMERMAN, J., & UNDERWOOD, B. J. Thespacing effeet in the learning of word pairs and the eomponentsof word pairs. Memory & Cognition, 1974, 2, 742-748.

SMITH, S., & GUTHRIE, E. R. General psychology in terms ofbehavior. New York: Appleton-Century, 1924.

SMITH, S. M., GLENBERG, A., & BIORK, R. A. Environmentaleontext and human memory. Memory & Cognition, 1978, 6,342-353.

STEIN, B. S. Depth of proeessing reexamined: The effects of thepreeision of encoding and test appropriateness. Journal of Yerba ILeaming and Yerbal Behavior, 1978, 17, 165-174.

Taros, S. J., & D'AGOSTINO, P. R. Effects of repetition as afunction of study-phase retrieval. Journal ofYerbal Learning andYerbal Behavior, 1976, 15,529-537.

TULVING, E. Theoretical proeesses in free reeall. In T. R. Dixon& D. L. Horton (Eds.), Yerbal behavior and general behaviortheory. Englewood Cliffs, N.J: Prentice-Hall, 1968.

TULVING, E. Ecphorie processes in reeall and reeognition. InJ. Brown (Ed.), Recall and recognition. New York: Wiley, 1976.

UNDERWOOD, B. J. Some eorrelates of item repetition in freerecalllearning. Journal ofYerbal Learning and Yerbal Behavior,1969, 8, 83-94.

UNDERWOOD, B. J., KAPELAK, S. M., & MALMI, R. A. Thespacing effect: Additions to the theoretieal and empiricalpuzzles. Memory & Cognition, 1976, 4, 391-400.

WALLACE, W. P. Clustering in free reeall based upon inputeontiguity. Psychonomic Science, 1969, 14,290-292.

WATKINS, O. C., & WATKINS, M. 1. Buildup of proaetiveinhibition: A eue overload effeet. Journal of ExperimentalPsychology: Human Learning and Memory, 1975, 1,442-453.

WmTTEN, W. B., 11, & BIORK, R. A. Learning from tests:Effects of spacing. Journal of Yerbal Learning and YerbalBehavior, 1977, 16,465-478.

YOUNG, J. L. Reinforcernent-test intervals in paired-associatelearning. Journal ofMathematical Psychology, 1971, 8, 58-81.

NOTES

1. Sehematie representations of similar notions are presentedin Glenberg (1976) and in Landauer (1975). Landauer's scherne,however, illustrates retrieval of at least one of many functionallyseparate traces, each representing a single repetition. The presentscheme illustrates retrieval of one trace that incorporatesinformation from individual repetitions. Glenberg's (1976)illustration can be thought of as representing the descriptivecomponents of the present theory,

(Received for publieation September 1. 1978;revision accepted February 2, 1979.)