critical issues in interference theory · 2017. 8. 24. · critical issues in the theoretical and...

.1/('lIIory and CognitionJ973. 1"01. J. /9·40

Critical issues in interference theory *

LEO POSTMANUniversity of California. Berkeley, California 94720

and

BENTON J. UNDERWOODNorthwestern University, Evanston. Illinois 6020/

Critical issues in the theoretical and experimental analysis of interference processes in retention are reviewed. Theevolution of classical two-factor theory is traced, and the strengths and weaknesses of the contemporary version of thisposition are examined. Recent critiques of current interference theories by Martin (1971a) and Greeno, James, andDa Polito (1971) are reviewed and examir....d. New conceptualizations of interference proposed by these authors, whichplace major emphasis on retrieval dependencies and on the role of encoding and retrieval processes, are considered andevaluated.

The classical two-factor theory of interference(Melton & Irwin, 1940), which for a long time had fewserious rivals as a framework for the analysis offorgetting, has recently become a focus of importantcriticism and controversy. The purpose of this paper isto assess the present status of the theory and to examinesome of the basic challenges which have been mountedagainst the classical position and its currentmodifications.

In the interest of clarifying the central issues,we willbegin by restating briefly the guiding concepts andassumptions of two-factor theory as it has beeninterpreted by its proponents. The two factors whichgive the theory its name are, of course, unlearning andcompetition. Of these, the process of unlearning, whichwas made accessible to direct measurement by theintroduction of the MMFR test (Barnes & Underwood,1959), has been receiving primary attention in boththeoretical discussions and experimental investigations.In the next section, we trace in broad outline thedevelopment of contemporary views of the conditionsand characteristics of unlearning.

THE CONCEPT OF UNLEARNING

Original Formulations

There are two basic defining characteristics of theconcept of unlearning: (a) The consequence ofunlearning is reduced availability of the first-list responseon a test for retroactive inhibition (RI); and (b) thetemporal locus of the events responsible for unlearning is

*The research of the first author is supported by a grant fromthe National Institute of Mental Health; the research of thesecond author is supported by the Personnel and TrainingResearch Programs of the Office of Naval Research. Portions ofthis paper have benefited greatly from critical appraisals suppliedby Joel Zimmerman and John J. Shaughnessy of NorthwesternUniversity.

during interpolated learning (IL), e.g., the acquisition ofA-C in the A-B, A-C paradigm, rather than at recall. Inthe analysisof unlearning, the major theoretical questionconcerns the nature of the processes that come into playduring IL and serve to lessen the availability of first-listresponses. In the originalformulation of the hypothesisof unlearning, Melton and Irwin tentatively identifiedthe unreinforced or punished elicitation of first-listresponses as errors during the transfer phase as theessential antecedent of unlearning. The occurrence offirst-list intrusions during second-list learning providesdirect evidence for the elicitation of errors that arefollowed by nonreinforcement. Given this fact, Meltonand Irwin concluded that their data "clearly favor atheory that attributes a portion of the RI to a weakeningor unlearning of the original S-R relationship during theinterpolated learning [1940, p. 200] ," Thus, unlearningwas seen as contingent upon the evocation of oldresponses during the acquisition of a new task. WhileMelton and Irwin emphasized overt intrusion errors asdeterminants of unlearning, it was soon recognized thatcovert intrusions rejected by the S as inappropriate maylead to the same.consequences (Thune & Underwood.1943). Subsequently, it was shown that the ratio ofovert to covert intrusions does not influence the level ofRI (Keppel & Rauch, 1966). We shall. therefore.in te rpret the classical hypothesis as attributingunlearning to the consequences of the elicitation of bothovert and covert intrusions during interpolated learning.

While Melton and Irwin were inclined to concludethat unlearning reflected the weakening of S-Rassociations, they also entertained an alternativeinterpretation, applying an hypothesis suggested earlierby Thorndike and by Wendt: "The inhibition or'unlearning' of one response occurs when it isunreinforced or 'punished' because another incompatibleresponse has been fixated [1940. p. 20 I] ." According tothis conception, unlearning of first-list responses occursafter the acquisition of second-list responses. Finally,

19

20 POSTMAN AND UNDERWOOD

Melton and von Lackum subsequently alluded to thepossibility that unlearning is a consequence of responsecompetition during IL. They suggested that the sourceof such competition may be nonspecific, which impliesthat the interference may operate at the list level: "Ifthe unlearning factor is so interpreted, the competingresponses established when an intrusion from theoriginal list occurs without reinforcement during thelearning of the interpolated list are not only the specificresponses involved in the recitation of the interpolatedlist [1941, p. 172f] ." Thus, while several alternativemechanisms of unlearning were considered in theoriginal statements of two-factor theory, each of theinterpretations proceeded from the assumption that theoccurrence of interlist intrusions was the necessaryantecedent condition.

RelationshipBetweenUnlearning and Negative Transfer

Historically, the two-factor formulation was anextension of McGeoch's (1942) transfer theory of RI.McGeoch attributed negative transfer and RI tocompetition of responses during IL and at recall,respectively. In both cases, reproductive inhibition wasseen as responsible for the decrements in performance.With the introduction of the factor of unlearning, theprocesses producing negative transfer and RI could nolonger be considered identical except for their temporallocus of operation. Consequently, the adoption of thetwo-factor position entailed the systematic question ofthe relationship between negative transfer andunlearning. The analysis of Melton and Irwin carried theclear implication that unlearning was contingent uponthe occurrence of negative transfer. This conclusionfollows because the interlist intrusions that are assumedto activate the process of unlearning are a manifestationof negative transfer. While intrusions of A-B areexpected to accompany the acquisition of A-C, it wasdefinitely not assumed that the unlearning of A-B mustprecede the formation of the new A-C association. Onthe contrary, the possibility was left open that theunreinforced elicitation of A-B may follow as well asprecede the acquisition of A-C. Whatever the exacttemporal locus of the intrusions, the process ofunlearning the first list was seen as likely to delay thelearning of the secondlist (Melton& von Lackum, 194I,p. 173). The same point was made by Barnes andUnderwood (1959) when they stated that negativetransfer may result from the interference accompanyingthe unlearning process.

The fact that there is an inevitable interaction offirst-list unlearning and second-list learning does notentail the conclusion that the latter depends on theformer. Such an assertion could not be, and has notbeen, made by proponents of two-factor theory on bothlogical and empirical grounds. The logical objection isinherent in the two-factor formulation. If the acquisition

of the transfer task invariably consisted of thereplacement of A-B by A-C, competition at recall wouldbe entirely eliminated as a source of interference; eitherone or the other of the responses would be available butnever both. The essence of two-factor theory is, ofcourse, that both sources of interference can becomeeffective at the time of recall, even after a high degree ofinterpolated learning which ensuresthe availability of allC responses. In a broader theoretical context, classicalinterference theory can be seen as more compatiblewiththe concept of a habit-family hierarchy than with theall-or-none view of associative learning (Estes, 1960).The latter position does, of course, hold that tworesponses cannot be attached simultaneously to the samestimulus.

There are a number of compelling empiricalobjections to the assumption that the acquisition of A-Cpresupposes the unlearning of A-B. First of all, regardlessof the degree of IL, unlearning is virtually nevercomplete and in fact rarely exceeds about 50%. Therelevance of this fact to the interpretation of unlearningwas emphasized by Barnes and Underwood (1959}whenthey introduced the MMFR procedure of assessing theavailability of first-list responses: "It does not seem thatall items would be extinguished, even with an extremelylarge number of trials on A-C.... Thus, while thepresent data support an extinction hypothesis they donot indicate why items are not, nor are likely not to be,extinguished [po 102]." The differential resistance offirst-list items to unlearning, which has also beendemonstrated in other studies (e.g., Postman, Stark, &Henschel, 1969), presents an important explanatoryproblem. At the same time, however, such findings ruleout the assumption that the acquisition of A-C dependson the unlearning of A-B.

Experiments on probabilistic learning, in which twodifferent responses are learned concurrently to the samestimulus, provide equally strong evidence that theestablishment of A-C is not contingent on theelimination of A-B. As Popp and Voss (1965) haveshown, negative transfer and unlearning do occur underthese circumstances but depend critically on the patternof alternation of the Band C responses: the morefrequent the changes from one response to the other, thelower are the amounts of negative transfer andunlearning. Frequent shifts serve to maintain theavailability of both responses. Using an arrangement inwhich the successive lists were separated, Postman andParker (1970) found that the original associations can bemaintained during transfer learning with little loss inefficiency of performance. There was only a moderatereduction in the absolute speed of A-C acquisition andno significant change in the amount of negative transferrelative to a CoD baseline condition when Ss wererequired to recall both Band C responses to each of theA terms during the transfer phase. Furthermore, thisprocedure virtually eliminated RI on a terminal test ofrecall. Related findings are those of Dallett and

CRITICAL ISSUES IN INTERFERENCE THEORY 21

D'Andrea (1965). who instructed one group of Ss to usethe B responses as mediational aids in A-C learning andanother group to unlearn the B responses in order toreduce associative interference. While the Ss reportedthat they tried to comply with the instructions, theoverall speed of second-list learning was not affected.The attempts at mediation appear to have been largelyunsuccessful. since they produced only a moderateeffect on the amount of RI. The critical point forpurposes of the present discussion is, however, thatdeliberate attempts either to maintain or to unlearnfirst-list associations did not influence the speed oftransfer learning. The studies which we have justreviewed converge on the conclusion that someassociative interference is inevitable whenever newresponses are attached to old stimuli. regardless ofwhether the old responses are maintained or discarded.

On the basis of all the available evidence, it is safe toconclude that the unlearning of A-B is not a necessarycondition of A-C learning. As we have emphasized, suchan assumption has never been part of the hypothesis ofassociative unlearning or two-factor theory. Inspeculations about the mechanisms of unlearning, theguiding assumption has rather been that the acquisitionof A-C provides an opportunity for the unreinforcedelicitation of A-B. Whether and how frequently suchelicitations occur, determining differential resistance ofindividual items to RI, is a question to whichexperimental analyses have so far failed to yield asatisfactory comprehensive answer,We will return to thisproblem below when we discuss experimental tests ofthe elicitation hypothesis.

The Extinction Analogy

What Melton and Irwin called Factor X andtentatively identified as a process of unlearning sooncame to be seen as functionally analogous to theexperimental extinction of conditioned responses.Failure to reinforce a conditioned response results inextinction, and the same may be conveniently assumedfor verbal associations (Underwood, 1948a, b). It hasbeen recognized, however, that the interferenceparadigm A-B, A-C is more comparable to the operationsof counterconditioning. in which a new conditionedresponse is substituted for the old one. than toexperimental extinction where the reinforcement iswithheld (Barnes & Underwood. 1959, p. 97). In anyevent. extinction and associative unlearning have beenused as interchangeable concepts in most recentdiscussions of the mechanisms of RI. In spite of itsobvious limitations and inadequacies (cf. Keppel. 1968,p. 194ff), the analogy has proved heuristically useful as aguide to the systematic exploration of the functionalproperties of unlearning. The investigation of thespontaneous recovery of unlearned associations is themost important case in point. Since such studies werefirst initiated. the evidence for recovery has been far

from consistent (for summaries see Keppel, 1968;Postman. Stark. & Fraser, 1968). Recent experimentsindicate,however, that rises in first-listrecall are likelyto be observed after intervals of the order of half anhour (Forrester, 1970; Kamman & Melton, 1967; Martin& Mackay, 1970; Postman, Stark, & Fraser, 1968;Postman, Stark, & Henschel, 1969; Shulman & Martin,1970). The conditions under which recovery may beexpected after intervals of a day or more (e.g., Abra,1969; Ceraso & Henderson, 1965, 1966; Silverstein,1967) remain to be fully specified. Taken together, thedemonstrations of the phenomenon have beensufficiently frequent, and in the most recent studiesconsistent and predictable, to support the historical andpragmatic usefulness of the extinction analogy.

The extinction analogy also provided a theoreticalunderpinning for the elicitation hypothesis which hasremained an important point of departure in the analysisof the mechanism of unlearning. If the unreinforcedelicitation of errors is, indeed, the essential antecedentof unlearning, then some process akin to extinction mustbe postulated to account for the weakening of theintruding associations. In the next section, we considerthe implications of recent tests of the elicitationhypothesis.

Tests of the Elicitation Hypothesis

The elicitation hypothesis represents an explicitstatement of the assumption that unlearning is theconsequence of the intrusion of first-list responsesduring the acquisition of the transfer list. From thebeginning, the hypothesis has suffered from the apparentweakness that the absolute frequency of overt interlistintrusions is typically quite low, so that the occurrenceof covert intrusions had to be postulated to account forthe magnitude of the observed retroactive effects. Testsof the elicitation hypothesis had of necessity to beindirect. The approach adopted in many investigationswas to manipulate conditions which could reasonably beassumed to influence the frequency of covert if not overtintrusions. The general prediction is. of course. that anyvariable that lowers the probability of intrusions shouldreduce the level of RI. Studies of the effects of theform-class similarity of the responses in the successivelists are a case in point: the greater the similarity. themore frequent and persistent the interlist intrusionsshould be and hence the greater the RI. This expectationhas been confirmed (Friedman & Reynolds. 1967:Postman, Keppel. & Stark. 1965): the fact that a shift inform class influences RI under unmixed- but not undermixed-list conditions (Birnbaum. 1968a) adds weight to

the original interpretation of these findings. Thehypothesis also received apparent support from Goggin's(1967) demonstration that interpolated learning by themethod of prompting. which reduces the opportunitiesfor intrusion errors. serves to lower the level of Rl ascompared to the conventional anticipation procedure.


On the other hand, the results have been mixed andinconclusive in experiments testing the prediction that,with the number of transfer trials held constant. interlistintrusions should be more frequent and RI greater underconditions of multiple-list than single-list interpolation(Birnbaum, 1968b; Goggin, 1968; Postman. 1965;Weaver& Danielson, 1969).

For purposes of the present discussion, the mostimportant implication of the elicitation hypothesis isthat RI should be related inversely to the speed ofsecond-list learning. The more rapidly an A-C item isacquired, the less likely it becomes that an erroneousA-B association will be elicited in the course of transferlearning. That is, the presence of a strong A-C protectsA-B from unlearning. This relationship should hold forvariations in speed of learning related to type ofmaterials, S ability, and item difficulty. AlthoughPostman and Stark (1965) found some evidence for theexpected correlation with S ability, other tests haveyielded little or no empirical support for this prediction.Two relevant findings may be cited here. Varying thetype of interpolated materials, Birnbaum (1968c) foundthat transfer lists composed of normative associateswhich were acquired extremely fast did not produce lessRI than lists of weakly associated pairs which werelearned relatively slowly. The results are not consistentwith the elicitation hypothesis, although speed ofacquisition and terminal degree of interpolated learningwere confounded since a fixed number of interpolatedtrials was given.

The same confounding, which is difficult to avoid, ispresent in Runquist's (1957) analysis of the relationshipbetween the degree of second-list learning of individualitems and the level of RI for corresponding first-listitems. The lack in reduction of RI as a function of speedof List 2 learning derives added importance from thefact that for paradigms of stimulus identity such as A·B,A-C there is often a positive correlation between thespeed of first-list and second-list learning ofcorresponding items (e.g., Postman & Warren, 1972;Wichawut & Martin, 1971). When items are pooledacross Ss, individual differences in learning abilitybecome an additional source of positive correlation.These factors should have favored the expectedcovariation between the rank of List 2 items and recallof first-list associations, but failed to do so in Runquist'sresults. On the other hand, Wichawut and Martin (1971)have recently reported a positive relationship between Band C recall when items are ranked according to speed offirst-list learning. There are measurement problems herethat await solution, since full account has to be taken ofthe terminal probability of both responses to evaluatethe extent to which the levels of recall areinterdependent.

Another reason for expecting a positive correlationbetween Band C recall should be mentioned here,although it does not bear directly on the elicitationhypothesis. It has been postulated that the process of

unlearning may delay, albeit temporarily, the acquisitionof A-C. It follows that the more unlearning an A-B itemundergoes, the greater should be the delay in theacquisition of A-C. Consequently, at the end of IL, bothA-B and A-C would be weaker if the period of extinctionhad been protracted than if it had been brief, and thiswould contribute to a positive correlation between A-Cstrength and A·B recall.

It is fair to conclude that the evidence for theelicitation hypothesis, insofar as it entails a positivecorrelation between the absolute frequency of interlistintrusions and the amount of RI, is far from impressive.The original assumption that the development ofunlearning is monotonically tied to the repeatedelicitation of first-list associations is, therefore, in doubt.It should be recognized at the same time that criticaltests of the hypothesis have been difficult to devise,partly because of the rare occurrence of overt intrusionsand partly because assessments of the effects of speed ofacquisition have been confounded with variations indegree of learning. For purposes of the discussion thatfollows below, however, it is important to emphasizethat the elicitation hypothesis has been viewed asentailing a positive rather than inverse correlationbetween the speed of acquisition (and strength whencorrelated with speed) of second-list items and the recallof corresponding first-list items on a test of RI. Theweakness of the evidence for a direct relationshipbetween the frequency of intrusions and the amount ofunlearning suggests that the unreinforced elicitation oferrors may be a condition that triggers otherprocesses-not in a one-to-one fashion-which becomeoperative in reducing response availability. We willreturn to this possibility when we discuss the hypothesisof response-set interference.

Component Analysis of Unlearning

A major step in the verification of the unlearninghypothesis was the introduction of the MMFR test,which made it possible to obtain estimates of' theavailability of first-list responses independently ofspecific competition. At about the same time, thecomponent analysis of transfer, which is based on atwo-stage conception of associative learning, was appliedto the measurement of unlearning (McGovern, 1964).The point of departure is the assumption that theunlearning of responses per se and of specificassociations are in large measure independent processeswhose effects summate to determine the level of first-list"recall. Differences in the amounts of RI observed understandard paradigms of transfer can be satisfactorilypredicted on the basis of this assumption, with the lossof responses as such referred to the unlearning ofcontextual associations and the loss of pairwiseassociations to the unlearning of specific S-Rconnections. Component analysis has made it possible tosubsume a large body of data on RI under the principle


of associative unlearning. As a descriptive and predictivedevice. the hypothesis has, therefore, continued to besuccessful. On the other hand, attempts to specify theexact mechanisms of associative unlearning have metwith only indifferent success. The explanatory power oftwo-factor theory does not, of course, rest on theunlearning hypothesis alone. The analytic usefulness ofthe concept of competition, as applied in conjunctionwith the principle of unlearning, must also beconsidered.

THE CONCEPT OF COMPETITION

Specific vsGeneralized Competition

In the transfer theory of RI formulated by McGeoch(1942). competition was the result of incompatibleresponses being attached to the same stimulus. Bothresponses were assumed to remain available.Competition was seen as leading to the blocking ofcorrect responses and interlist intrusions at recall. Asalready noted, owing to the rare occurrence of interlistintrusions, reproductive inhibition remained largely ahypothetical construct rather than an observable fact.The majority of the errors contributing to RI, even onpaced tests of recall, were failures to respond. Areasonable explanation of the high frequency ofomissions was offered when Newton and Wickens (1956)introduced the concept of generalized responsecompetition, i.e., the tendency to continue to givesecond-list responses on the retention test for the firstlist. Thus, a distinction came to be made betweenspecific and nonspecific response competition-theformer referring to the blocking of individual responsesin the sense described by McGeoch. and the latter to thepersistence of a set to give the responses learned last.

Extension of the Hypothesisof Generalized ResponseCompetition:

The Hypothesis of Response-Set Interference

The hypothesis of response-set interference (Postman.Stark. & Fraser, 1968) represents an extension andelaboration of the basic principle of generalized responsecompetition. According to this hypothesis, unlearningresults from the operation of a mechanism of responseselection which exerts its primary effect on the entireclass of first-list responses rather than on specificstimulus-response associations. The essential steps in theargument are as follows: (a) During the acquisition ofthe first list. a selector mechanism (Underwood &Schulz. 1960) serves to activate the appropriateresponses and to inhibit the occurrence of theinappropriate ones. (b) In the transfer phase. theoperation of the selector mechanism results in theactivation of the newly prescribed responses and theinhibition or suppression of the earlier ones. (c) Theselector mechanism is characterized by a certain amount

of inertia. The tendency to give the responses learnedlast persists on a test.of recall after the end of IL. Theconsequent impairment of the S's ability to shift back tothe repertoire of first-list responses is designated asresponse-set interference. Response suppression duringIL and response-set interference at recall arecomplementary effects of the operation of the selectormechanism during transfer learning. The temporal locusof the suppression of the first-list responses is during IL;the continuing dominance of the second-list repertoireduring recall constitutes response-set interference.Logically, response-set interference can occur even if thefirst-list repertoire is not suppressed as long as there is astrong disposition to give the responses learned last. It isassumed, however, that first-list suppression during ILand second-list dominance at recall are correlated.Furthermore, the degree of suppression and subsequentresponse-set interference are taken to be a function ofthe level of negative transfer, e.g., to be greater for A·Cthan for C·D. The latter relationship is postulated on theassumption that the elicitation of first-list errors servesto intensify and sustain the operation of the process ofresponse selection. (d) Response selection is assumed tobe a reversible process, and absolute rises in first-listrecall are expected as the interval between the end of ILand the test for RI is lengthened. (It is only a fairrecognition of historical developments that theext inction analogy generated the prediction ofspontaneous recovery, whereas the hypothesis ofresponse-set interference and other positions discussedlater made assumptionspostdicting the phenomenon.)

There were a number of experimental findings whichgave apparent support to the hypothesis. The mostimportant among these were: (a) RI under the A-Cparadigm is typically substantial when recall of theresponses is required but is relatively small or absentaltogether when a recognition procedure. whicheliminates the requirement of response recall, is usedduring the acquisition of the successive lists and on thetest of first-list retention (postman & Stark. 1969).(b) Absolute recovery is observed under both the A-Cand CoD paradigms (postman, Stark, & Henschel. 1969).Furthermore, regardless of the initial level of RI. there isevidence of recovery only when there is a change inresponses (postman & Warren, 1972). Thus. at least thatportion of the total RI which is reversible appears toreflect the loss of responses per se rather than theweakening of specific associations. Other findings. e.g..those related to the similarity of successive responseclasses. are consistent with the hypothesis but can alsobe accommodated. as was shown earlier. by a theory ofassociative interference with the aid of some reasonableassumptions.

There have been numerous recent experimentsdesigned to test the hypothesis of response-setinterference. Before these can be discussed. it isnecessary to make explicit the scope and limitations ofthe h~ pothcsis as it was originally formulated. (:.I) The


loss of responses per se is attributed to the operation ofa mechanism of selection rather than to the unlearningof contextual associations. (b) The hypothesis isconcerned entirely with the mechanisms governingresponse availability per se. It is taken for granted thatthe placement of responses at the time of test remainsunder stimulus control. just as it is in acquisition. Infact, the possibility of a mechanism of responsesuppression was initially suggested by the apparentstability of the associative component of recall. (c) Thepossibility that associative losses may contribute to RI isnot being denied in toto. particularly in view of theheavy interference found under the A-Br paradigm.Thus, it was suggested that first-list associations whichcontinue to be elicited during interpolated learningbecause of failures of the selector mechanism may beweakened progressively (postman & Stark, 1969,p. 176). What is being questioned are the generality andextent of associative unlearning under various conditionsof negative transfer. e.g., under the A-C paradigm.(d) Contrary to the usage adopted by Martin (197la),this is not a hypothesis of list differentiation in theconventional sense of accuracy of identification of listmembership. It is likely that the effective operation of aselector mechanism entails a high degree of listdifferentiation. However, the critical factor is not the S'sability to identify the list membership of whateverresponses do occur but rather the mechanism governingthe availability of alternative response repertoires forrecall.

The hypothesis of response-set interference representsa modification or elaboration of two-factor theory, notan alternative to it. The main changes lie in theassumption that response loss results from the operationof a selector mechanism and the relative weight assignedto list-related as opposed to item-specific interference.The term item-specific here refers not only to pairwiseassociations, but also to individual response terms. Acomponent analysis such as McGovern's implies thatthere are associations between each individual responseand the experimental context, and it is the unlearning ofthese multiple contextual associations which results inthe loss of responses per se. By contrast, the principle ofresponse-set interference holds that the mechanismresponsible for response loss operates at the list level. Asfor the unlearning of specific S-R associations, this typeof item-specific interference could obviously not beruled out on the basisof the available evidence.

Recent critics have placed the concepts of associativeunlearning and response-set interference in sharpopposition as hypothetical mechanisms of RI. They do,to be sure, represent different mechanisms, but they arenot mutually exclusive. Even with the incorporation of aprinciple of response suppression, a dual locus ofretroactive effects-during interpolated learning and atthe time of recall-continues to be assumed.Furthermore, associative interference retains its status asa significant antecedent of the degree of responsesuppression.

Tests of the Hypothesis of Response-Set Interference

The evidence accumulated thus far in support of thehypothesis of response-set interference is certainly notconclusive, and there are some experimental resultswhich pose genuine difficulties for the hypothesis. Onthe other hand, there are also a number of findingsoffered in refutation of the hypothesis which cannot beregarded as critical at the present juncture. Withoutattempting to present an exhaustive review, we willconsider some of the relevant findings with reference toa series of systematic points bearing on the principle ofresponse-set interference.

(1) The usual difference in retention lossbetween theA-C and CAD paradigms does not in itself bear on thequestion of the relative weight of response-setinterference and item-specific unlearning in determiningRI. The degree of suppression of the first-list repertoireand consequent dominance of second-list responses areexpected to be related to the amount of negativetransfer. The fact that RI is observed under the CADparadigm is, however, evidence for interference withresponse recall per se. The existence of such a processwas brought out clearly in a recent study by Lehr,Frank, and Mattison(1972). Considerably greater RI andspontaneous recovery in A-B recall were observed afteran interpolation of an A-C list than after free-recalllearning of an equivalent number of C responses.However, the latter treatment did produce a significantamount of interference followed by recovery. Thus,some response-set interference occurs in the absence ofassociative transfer, but the retroactive effects aregreatly enhanced when such transfer is present. The factthat the difference between A-C and CAD is greaterwhenrecall is stimulus-cued rather than free (Keppel,Henschel, & Zavortink, 1969; Postman, Stark, & Fraser,1968, Experiment IV; Weaver, Rose,& Campbell, 1971)is understandable on the assumption that the presence ofthe A terms maintains the selective set established duringinterpolated learning. It must be noted, however, thatcomparisons between the two types of test arecomplicated by the fact that stimuli are lessavailable asimplicit mediators of response recall after CAD than afterA-C interpolation (Weaver et al, 1971), presumablybecause there is more frequent exposure to the A termsin the latter case.

Differences in RI between the two paradigms assumesystematic importance when they are observed, as theyhave been in several studies, under conditions ofmixed-list interpolation (Delprato, 1971; Weaver et al,1971; Wichawut & Martin, 1971). It has been arguedthat such differences should not be found if the entirerepertoire of first-list responses is suppressed. Hence,superior retention under C·D has been interpreted aspointing to item-specific interference. The argument isreasonable, but before it is accepted as definitive, it mustbe determined whether response-set interference can beselective with respect to readily identifiable subgroups ofitems within a list. In a previous study of transfer with a


mixed-list test (Postman, 1966), an analysis of misplacedresponses showed that items with old stimuli wereconsistently differentiated from items with new stimuli.The assertion that the mechanism of response selectionoperates on the entire repertoire of first-list responseswas made with reference to the conventionalarrangement in which all the items in the interpolatedlist conform to the same paradigm. The conditions ofinterference under mixed-list interpolated learningpresent new analytic problems. The question now ariseswhether paradigm differences point to differentialsuppression of subgroups of items or specific unlearning.One obvious approach to this problem is themanipulation of similarity relations betweenparadigmatic subgroups within a list.

A related point is that the incidence of responsesuppression, and in particular its operation at the listlevel, cannot be considered independent of the length ofthe interpolated task. The selection criteria establishedduring interpolated learning will reflect the number andcharacteristics of the items in the transfer list. This pointbears on the experiments of Birnbaum (1970, 1972)which investigated the effects on RI of the omission offirst-list stimuli from the interpolated list. While theresults were variable, the net outcome was that theomission of first-list stimuli served to reduce the amountof RI for corresponding first-list pairs. It is possible thatthese findings imply item-specific interference. Again,the conclusion cannot be definitive until the boundaryconditions of response-set interference are defined moreprecisely. For example, would such interference beexpected if all but one or two of the first-list stimuliwere omitted? The intuitive answer to this questionwould appear to be in the negative. Hence, thequantitative relationship between amount of stimulusoverlap and evidence for item-specific RI needs to beexamined.

(2) While there is obviously no reason to press the nullhypothesis that item-specific interference never occursunder the A-C paradigm, there definitely are conditionswhich minimize such interference. With reference to theprinciple of response-set interference, it is noteworthythat these conditions appear to be related to thedistinctiveness of the successive response repertoires.There are clear indications that unlearning is essentiallyeliminated when the degree of distinctiveness is eithervery low or very high: (a) Using a paired-associateprocedure in which A-B, A-C, and control items wereintermingled during the presentation phase, Da Polito(1966) found no RI on an MMFR test of recall, althoughthere was negative transfer and apparent proactiveinhibition (PI). Under these conditions, a mechanism ofresponse selection obviously cannot operate. (b) Whenacquisition trials on A-B are widely distributed andfollowed by massed practice on A-C, there is not only amarked reduction of PI in the recall of A-C. but also avirtual absence of unlearning of A·B even after veryheavy amounts of interpolated learning (Underwood &

Ekstrand, 1966, 1967). The reduction of PI has beenattributed to the increased differentiation resulting fromthe change in the temporal schedule of practice. Thepresence of highly effective temporal cues may alsofacilitate the shift from one set of responses to anotheron an unpaced MMFR test. (It is interesting to note thatrecall of A-B remained virtually perfect evenwhen someA·B pairs were carried over into the A-C list, althoughthis procedure reduced differentiation and led tosubstantial Plan a paced test for A-C. It was suggestedthat repeated and nonrepeated pairswere set apart in thesecond list, and this factor may account for thecontinued full availability of the A·B responses on anunpaced test.) These failures to obtain RI, whichoccurred under widely different experimentalconditions, make it clear that item-specific RI under theA·C paradigm is far from inevitable.

(3) The observation that specific associations appearto be highly resistant to interference under the A-B, A-Cparadigm has given face validity to the principle ofresponse-set interference, even though the two processesare not mutually exclusive. Some important questionshave been raised recently about the validity of theinference that specific associations are relatively immuneto retroactive interference. The empirical findings atissue can be briefly summarized as follows. Whenacquisition is by a recall procedure and retention istested by associative matching, some RI is typicallyfound, although of a smaller order of magnitude than onan MMFR test (e.g., Delprato, 1971; Garskof, 1968;Garskof & Sandak, 1964; Sandak & Garskof, 1967).Such measures suffer from the difficulty that there is ashift in procedure between acquisition and the test ofretention, which may possibly have more adverse effectson an experimental group learning two lists than on acontrol group learning a single list. If taken at their facevalue, these results show that the unlearning of specificassociations develops slowly as compared with the lossof response availability. When both acquisition andretention are by the multiple-choice method, theamount of RI is likely to be slight and fall short ofsignificance (Anderson & Watts, 1971; Postman & Stark,1969).

It has now been suggested that the high level ofperformance on a matching or multiple-choice test ismediated by intact backward associations which S is ableto utilize even when the forward associations have beenunlearned. Evidence in support of this conclusion wasfirst presented by Merryman (1971), who used amixed-list design in which the conditions of unlearningwere unidirectional (A-C) for some pairs andbidirectional (A·C and B-E) for others. Theinterpretative difficulties posed by such a mixed-listdesign have been discussed elsewhere (Postman & Stark.1972). More recently. Greenberg and Wickens (197:.:!)carried out a similar study. Acquisitionof the successivelists was by the anticipation method. and an associativematching test was used to measure RI. The comparison


of primary interest is between a condition ofbidirectional unlearning (alternating trial blocks of A-Dand BoC or A-D and C·B) and "double unilateral"conditions (A-D and A-E, C·B and E·B. or B-C and B-E)in which the total number of interpolated trials was heldconstant. Performance on the matching test was worse inthe former than in the latter case, although thedifference fell short of significance. The levels ofretroaction were relatively low in absolute terms,namelyvabout 20% and 10%, respectively. As was truefor Merryman's experiment, the procedure wasnecessarily complex, and it is uncertain whether thefactor of directionality as such was adequately isolated.For example, under the bidirectional treatment therewere changes in both the stimuli and the responses inalternating trial blocks, whereas in the unidirectionalcondition one or the other set of terms remainedconstant. What the effects of such recurrent shifts in theitem pool are is not known. It is not apparent how suchproblems of design can be avoided. The interpretation ofthe results is complicated further by the fact thatforward and backward unlearning produced exactlyequal decrements in A·B retention, which does not agreewith the findings obtained under conventionalarrangements (e.g., McGovern, 1964). However that maybe, it is fair to conclude that mediation by intactbackward associations, if it is effective, enhancesperformance only moderately at best. The conclusionthat losses in the accuracy of associative matching underthe A-B, A-C paradigm are relatively small, especiallywhen there is no change in the method of testingbetween acquisition and recall, can be allowed to standfor the time being.

(4) To the extent that RI reflects the suppression ofthe repertoire of first-list responses, the reinstatement ofthese responses prior to the test of retention shouldreduce the level of interference. In a series ofexperiments designed to evaluate this prediction, Cofer,Failie, and Horton (1971) introduced presentations andfree recalls of the first-list responses following IL andduring first- and second-list learning. These proceduresserved to reduce, but not eliminate, RI. The results lendsome support to the response-suppression hypothesis butleave room for item-specific associative unlearning. Asthe authors indicate, it is uncertain whether suchprocedures as free-recall learning are optimal for thereinstatement of the ensemble of responses in apaired-associate list. There was evidence that subjectiveorganization of the responses developed in the course ofthe special training procedures which may haveinterfered with subsequent associative recall. A similarconclusion was reached by Postman, Burns, and Hasher(1970) in an investigation of the effects of responserelearning on the long-term recall of a single list ofpaired associates.

If the suppression of the repertoire of originalresponses contributes to RI, successive recall trials maybe expected to reinstate the set to give first-list responses

and to reduce the level of interference. There is onlymoderate support for this expectation. First-list recallhas been shown to increase over repeated tests, but ingeneral it fails to eliminate RI (Delprato, 1970;Greenbloom & Kimble, 1965, 1966; Postman, Stark, &Fraser, 1968; Richardson & Gropper, 1964; Weiss &Lazar, 1968). The fact that there is improvementindicates that the loss of availability is at least partiallyreversible. The recent demonstration by Adams,Marshall, and Bray (1971) that there are substantialgains in recall as the retrieval time for each item isextended gives further support to this conclusion. Onthe other hand, the persistence of RI over successivetrials shows at the very least that there is resistance tothe reinstatement of the original set. The ease ofreinstatement may, of course, depend on the degree ofresponse suppression during 11. Consistent with thisinterpretation is Delprato's (1970) finding of substantialgains over successive MMFR tests for the CoD paradigmbut of only minor ones for the A-C paradigm.

(5) The hypothesis of response-set interferencespecifies only one of the mechanisms which should beconsidered as potentially operative within theframework of two-factor theory. If the mechanism is,indeed, operative, it reduces response availability; undercertain conditions, it may account for most of theobserved retention loss, with specific associationsremaining largely intact. That is not to deny by anymeans that the second factor, namely, specificcompetition, will come into play under appropriatecircumstances. This point seems to have beenmisunderstood by Anderson and Watts (1971), whoshowed that the juxtaposition of first-list and second-listresponses on a multiple-choice test of recognitionproduced a significant amount of interference, whereasno RI was found when all the alternatives on the testwere from within the list. That finding in no wayinvalidates the conclusion that the specific first-listassociations remained intact. In fact, the absence of RIunder the noncompetitive treatment shows exactly that.What is demonstrated is that a process akin tocompetition (or list differentiation) can influence therecognition of specific associations as well as recall.

It would clearly be premature to draw any definitiveconclusions at the present time regarding the validity ofthe principle of response suppression. The existingevidence is mixed, although some of the argumentsagainst it have been based on failures of predictions notnecessarily entailed by the hypothesis. Some of theinterpretative difficulties stem, of course, from the factthat the characteristics and parameters of thesuppression process have not been specified precisely asyet. The suppression hypothesis, or some equivalentformulation, can probably not be discarded so long asthere is evidence for interference effects that operate atthe list rather than at the item level. If responsesuppression does occur, its weight in determining RIunder a given paradigm and on a particular test of

CRITICALISSUES IN INTERFERENCETHEORY 27

retention remains a quantitative question to be answeredempirically.

THE CONCEPT OF LIST DIFFERENTIATION

We turn now to an examination of the concept of listdifferentiation which is closely tied to two-factortheory, although it is not an integral part of that theory.In the context of analyses of RI and PI, differentiationhas conventionally denoted the discrimination of the listmembership of responses. Underwood (1945) assumedthat differentiation in that sense was a function of theabsolute and relative strengths of the competingassociations and the length of the time interval since theend of interpolated learning. For more recentinvestigations of factors influencing differentiation, seeWinograd (1968) and McCrystal (1970).

Loss of differentiation was considered a condition ofovert specific response competition, e.g., interlistintrusions would occur when the discrimination of thelist membership of responses broke down. The questionof whether degree of differentiation would influence theamount of retention loss was originally left open; itmight determine only the ratio of covert to overtintrusions, i.e., the number of failures to respond.However, in the absence of time pressure, the rejectionof an intrusion might make it possible to emit thecorrect response if it was available. Thus, loss ofdifferentiation was seen as conducive to overt specificresponse competition and as a potential contributor tothe total amount of interference, particularly on pacedtests of recall. The most convincing support for thelatter prediction has come from studies of PI wherestrong temporal cues to differentiation have been shownto eliminate or drastically reduce the level ofinterference in the recall of the most recent list(Underwood & Ekstrand, 1966, 1967: Underwood &Freund,1968).

Whereas a high degree of differentiation reducesspecific competition, the opposite relationship may beexpected to hold for generalized competition. The S'sdisposition to limit himself to responses from the listlearned last will be facilitated when the first- andsecond-list repertoires are clearly differentiated.Generalized competition is conducive to RI but lowersPI because it favors recall of the list learned last (cf.Postman, 1961). Hence, a loss of differentiation shouldlead to a reduction of that component of RI attributableto generalized competition and to a correspondingincrease in PI. The fact that the level of differentiationhas opposing consequences for specific and forgeneralized competition appears not to have been alwaysunderstood. It is also important to recognize. forpurposes of subsequent discussion. that for a given listthe degree of differentiation and response availability arenot necessarily correlated. Undercertain conditions.only a small proportion of first-list responses may beavailable. but these may or may not be easily

differentiated from the second-list responses. dependingon such factors as the relative strength of the alternativeassociations, the length of the retention interval, and thepresence of other effective cues to list membership. Thispoint is well documented by the fact that on MMFRtests showing heavy retroactive losses, the accuracy ofidentification of the first-list responses that remainavailable is almost invariably very high (e.g., Barnes &Underwood, 1959). We do not think that it is useful toextend the connotations of the concept ofdifferentiation too far beyond the original definitionwhich referred to the accuracy of the identification ofthe list membership of responses. Thus, we prefer not toclassify interpretations of interference at the list level asdifferentiation theories (Martin, 1971a).

THE PRESENTSTATDSOF TWO-FACTOR THEORY

Two-factor theory, elaborated within the frameworkof a component analysis of transfer and supplementedprogressively by such principles as differentiation,generalized competition, and response-set interference,has been able to account for a large range of empiricalphenomena, and particularly for paradigmaticdifferences in the level of RI. In stimulatingexperimentation on spontaneous recovery, it has focusedattention on the reversibility of interference over time.In a descriptive sense, the factors of unlearning andcompetition contingent on loss of differentiation havebeen repeatedly verified. The exact mechanisms ofunlearning remain far from clear, e.g., the exactconditions responsible for the loss of responses per se.The relative weight of associative unlearning andresponse loss in determining the retention deficits afterIL under various paradigms of transfer also remains asubject for investigation.

It must be admitted that the theory cannot accountsatisfactorily for the total range of phenomena of PI.According to the classical view,PI is determined entirelyby response competition at recall. Increases in Plovertime have been attributed to the recovery of priorassociations and the loss of differentiation. As we havenoted, there is persuasive evidence that on paced tests ofrecall, the level of PI is critically determined by thedegree of list differentiation. Proactive inhibition has,however. been consistently observed on MMFR testswhere list differentiation is presumably not required(e.g., Ceraso & Henderson, 1965: Houston. 1967b:Koppenaal, 1963; Postman. Stark. & Fraser, 1968).Recovery of the set to give earlier responses may beinvoked to account for such results. but only with theadditional assumption that the simultaneous arousal oftwo response repertoires gives rise to output interference(cf. Postman & Hasher. 1972). The possibility must alsobe considered that in learning the second response to agiven stimulus. the S is forced to rely on less effectivemediational devices than in learning the first response.

~8 POSTMAN AND UNDERWOOD

As a consequence. the second association is less welllearned than the first, or it is less stable. and hence moresubject to being forgotten. One interesting implicationof the latter hypothesis is that PI and RI may not be inall respects complementary manifestations of the sameunderlying processes. The resolution of theseuncertainties is one of the many tasks facing interferencetheory.

MARTIN'S CRITIQUE OFCURRENT INTERFERENCE THEORIES

In spite of the experimental history of two-factortheory which has yielded a not unfavorable mixture ofempirical successes and failures, Martin (1971 a)concluded in a recent review that neither the hypothesesof associative interference and unlearning nor theprinciple of response suppression is-any longer tenableand that the entire problem of the mechanisms ofretroaction and proaction is in need of drasticreformulation. Martin arrived at his conclusion on thebasis of what he considered certain crucial experimentalfindings which he views as invalidating some of the basicassumptions of current views of interference. In the nextsection, we examine these findings and the conclusionswhich they entail according to Martin's argument.

The Independent Retrieval Phenomenon

As Martin points ou t correctly, a basic assump tion ofthe principle of associative interference is that during theformation of a specific A-C association there is aconcomitant weakening of the corresponding A·Bassociation. It is this assumption which he believes tohave been decisively invalidated by what he calls the"independent retrieval phenomenon." His reasoning is asfollows: "If learning an A-C association entailsunlearning the corresponding A-B association, and if thelikelihood of recalling C indexes the A-C associationstrength, then we must expect the following inequality:PCB/C) < P(B/C); that is, it must be that the recall of Bis less likely when C is recalled than when C is notrecalled [197Ia, p. 316]."

In short the probability of recall of Band C on anMMFR test should be inversely related when successivelists conform to the A-B, A-C paradigm. A series ofchi-square tests based on the data of a number ofdifferent experiments (summarized by Greeno, James, &Da Polito, 1971; Martin, 197 Ia) have failed to provideevidence for the expected dependency relationship. Theprobabilities of recall of Band C appear to be, ingeneral, independent. According to Martin, " ... theindependent retrieval phenomenon denies associativeunlearning because the idea of associative unlearningimplies a conditional pairwise relation between Band Cavailability, but the phenomenon itself is that there is nosuch relation [197Ia, p. 319]."

We consider first the deduction from the principles of

associative unlearning which leads to the prediction of aconditional pairwise relation between Band Cavailability. The point of departure in the deductiveargument is the statement that "learning an A-Cassociation entails unlearning of the corresponding A-Bassociation." The term entails signifies that theunlearning of A·B is a necessary or inevitableaccompaniment of the acquisition of A-C. Furthermore,it is postulated that the more strongly A:C is learned, themore thoroughly A-B should be weakened. The latterassumption is obviously necessary to predict aconditional pairwise relation in the recall of Band C. Aswe have made clear in our earlier discussion, a reciprocalrelation between increments in the strength of A-C anddecrements in the strength of A-B is not in fact anecessary implication of the principle of associativeunlearning. Furthermore. we have shown that if anypairwise relation between B and C recall is to beexpected, it is positive rather than negative. The supportfor the latter deduction is, to be sure, weak, but theopposite prediction simply does not follow from currentviews of the conditions and characteristics of unlearning.

We consider next the measurement procedures bywhich Martin and Greeno have attempted to establishthe independent retrieval phenomenon. Their analysesconsisted of chi-square tests of independence of Band Crecall. Each S item was entered into a 2 by 2contingency table showing the frequencies of the fourpossible combinations of recall and nonrecall of B (Band B) and recall and nonrecall of C (C and C). As notedpreviously, the tests based on data from numerousexperiments have, in general. failed to yield any evidenceof interdependence. We will not dwell on the fact thatsuch chi-square tests are not appropriate, since theassumption of independence of observations is violated.As Hintzman (1972) pointed out, Martin was not correctin his assertion that this violation of the assumptions ofthe test can only inflate, and not depress, the value ofthe statistic. Rather, we will address ourselves to thelogic underlying the construction of the contingencytables and to the question of the potential sensitivity ofsuch tests even if the necessary statistical assumptionswere met.

The contingency tables comprise four cells: Be, Be,BC, and Be. According to Martin and Greeno'stheoretical deduction, the frequencies of Be and BCshould be higher than expected. The latter is thecombination derived from the authors' interpretation ofthe sequence of events in unlearning. If all items wereunlearned, presumably all cases would be concentratedin the BC cell. Such is, however, never the case, and theevaluation of the hypothesis of independent retrieval isbased on the configuration of entries in all four cells.Given this state of affairs, we must ask whatpsychological interpretation ~ to b~ven to the cases inwhich C is not recalled, viz, BC and Be. The failures of Crecall can be attributed to two primary reasons: (a) Cwas not learned during the acquisition of the


interpolated list; and (b) C was indeed learned but wasforgotten for some reason outside the range of theexperimental manipulations, e.g., due toextraexperimental interference. If C was not learned,then the interpolated learning failed to have theexpected effect, and there was no opportunity forunlearning. If C was learned and forgotten for reasonsoutside the defining operations of the experiment, thenthe same must be true for some cases of B. Thus, thecells in the contingency table represent frequencies ofevents whose distribution is not uniquely determined bythe postulated dependency relationship between theunlearning of A-B and the acquisition of A-C.Consequently, even if one were to accept the deductionthat as a consequence of unlearning there should be aninverse relationship between Band C recall, thecontingency tables are not in principle suitable fortesting it.

The difficulties do not end here. Apart from thelogical objections, many of the available contingencytests are vitiated by a lack of sensitivity. In mostexperiments on RI in which the A-B, A-C paradigm isused, recall of the second list is typicallyvery high if notperfect when interpolated learning is carried to a degreesufficient to produce a substantial amount of unlearning.Under these circumstances, a dependency relationshipbetween Band C recall could hardly be detected even ifit were present.Martin and Greeno place major emphasisin their discussions on a cumulative distribution ofchi-square statistics obtained from a variety ofexperimental findings. It is not possible to say in howmany of these cases the possibility of a significant chisquare wasall but precluded by the high level of recall ofC. We can, however, comment on some of the tests forwhich the cell frequencies are reported.

Consider first the results of an experiment by Abra(1969) which Martin (1971a, p. 317) presents as aprototype of the tests of the independent retrievalphenomenon. In this study, the MMFR tests were giveneither immediately or 24 h after the end of 11. On theimmediate test, the mean number of C responses recalledwas 8.1/10, i.e., not far from the ceiling of perfect recallwhich would enforce complete independence of BandC. After 24 h the mean recall of the secondlist droppedto 5.6. but at the same time the mean score for the firstlist rose from 5.4 to 7.7 (apparent spontaneousrecovery). Now it is the high level of performance on thefirst list which constrainsdeviations from independence.

A second set of contingency data available fordetailed examinationwas recently reported by Wichawutand Martin (1971). The major independent variable wasthe degree of second-list learning in the A-B, A-(paradigm. This variable was manipulated by varying theproportion of A-( and CoD trials given during thetransfer phase: 4/12, 8/12. or 12/12. A-(, recallincreased, of course. as a function of the number ofinterpolated trials. Interestingly enough. the chi-squarevalues decreased concomitantly from 1.46 to .42 to .00.

At the lowest level of second-list learning there was atrend toward an inverse relationship between Band Crecall. The authors suggested that a retrieval dependencymight arise when both A-B and A-C are weak.They didnot. however, consider the problem of sensitivity whichis here confounded with A-C availability. The validity ofthe tests at the two higher levels of A-C learning wasclearly minimized by the high levels of C recall, viz,79%and 85%. The latter value wasaccompanied by 72% of Brecalls; it is not surprising, then, that the chi-square valuefor that condition was literally zero. These results haveto be considered in light of the further fact thatWichawut and Martin observed a significant positivecorrelation between speed of second-list learning andstrength of corresponding first-list pairs; consequently,the probability of A-C recall must have varied directlywith the strength of A-B pairs. This correlation wouldobviously militate against the emergence of an inverserelationship between Band C recall.

One other set of data inviting comment is thatobtained by Da Polito (1966) in which the independentretrieval phenomenon was first brought to the fore.Under some of the conditions of that experiment, thelevels of both A-B and A-C recall were moderate, so thatthese particular contingency tests are not open to thecriticism of lack of sensitivity. However, the tests werebeside the point so far as an assessment of themechanisms of unlearning is concerned: Da Polito failedto find any Rl whatsoever in his experiment! It is notusual to draw inferences about the mechanismsgoverning a phenomenon from data in which thephenomenon has failed to materialize.

In a recent note, Martin and Greeno (1972) tookcognizance of Hintzman's (1972) objection that thepostulated inverse relationship between Band C recallmay be masked by factors making for a positivecorrelation between the test outcomes for correspondingitems, in particular differences in stimulus difficulty andS ability. Using the measures obtained in the experimentof Wichawut and Martin (1971), they stratified both Ssand items according to the level of first-list performance.Deviations from independence were found to berelatively small and inconsistent for the variouscombinations of classes of Ss and items. Martin andGreeno acknowledged, however, that item and Sdifferences might in principle render the tests

. ambiguous. This state of affairs led them to draw thefollowing conclusion: "Associative interference theory,when coupled with reasonable considerations of subjectand item effects. allows for any relation between BandC response that might occur in the data....Accordingly, this form of associative interference theoryis essentially untestable: it has far lessempirical contentthan is acceptable or warranted by presentknowledge. . .. Without consideration of possiblesubject and item effects. we have every reason to rejectit ... with allowance for possible subject and itemeffects. it is too open-ended to test and hence of little. if


any, theoretical interest [1972, p. 267]."These conclusions appear to us to be based on rather

unusual and untenable premises. Let us grant (for thesake of the argument only) that the hypothesis ofassociative interference does imply a negativerelationship between B and C recall on MMFR tests. Theinherent lack of sensitivity of contingency tests underthe arrangements of RI experiments, and the regrettablebut nevertheless real and demonstrable presence of Sanditem differences, make it difficult at this time to devisean appropriate test of this prediction. When the vitiatingfactors are disregarded in toto, the prediction appears tofail, and the grosser the violation of the requirements ofan appropriate test, the more complete the apparentfailure is. Hence, the theory is to be rejected! Aninsistence that major variables known to influence. theoutcome be considered in the interpretation of the testsis viewed as making the hypothesis too open-ended andhence of little theoretical interest. The posture, then, isthat the viability of the hypothesis must be judgedentirely by the outcome of one gross statistical test ofuncertain validity. Such a position might possibly bedefensible, at least pending the development of moreappropriate tests, if the negative correlation between Band C recall was the only, or by far the most critical,implication of the principles of associative interferenceand unlearning. As we have tried to show, however, suchis not the case; in fact, the deduction is in error. Thus,we cannot accept the conclusion based on the results ofthe contingency tests that "neither proaction norretroaction can be attributed to associative interferenceor unlearning [Wichawut & Martin, 1971, p. 320]." Atthis juncture in the history of interference theory, that isa sweeping conclusion indeed, and to be justified itwould have to be founded on truly crucial experimentaltests. On logical,methodological, and statistical grounds,the contingency tests of the independent retrievalphenomenon fail even to approximate this criterion.

Critique of the Hypothesisof Response-Set Interference

According to Martin, the lack of pairwise contingencybetween Band C recall makes it necessary to look formechanisms other than "stimulus-response associativeinteraction" to account for retroaction, proaction, andspontaneous recovery. What he calls "list-differentiationtheory" (the hypothesis of response-set interference)appeared to provide one possible avenue of approach.(Martin views this approach as continuous, or at least ascongruent, with McGeoch's principles of independenceand response dominance. However, McGeoch's conceptsapplied, for basic systematic reasons, to specificassociative pairings and not to response sets or listrepertoires. Hence, the analogy is misleading.) Theessential assumptions of the hypothesis of response-setinterference have been outlined above. Since themechanism of response selection is seen as exerting its

main effect on the entire class of first-list responses, apairwise contingency of Band C recall is not expected.However, as Martin sees it. another kind of retrievaldependency is implied: "The idea of list suppression,orhow response dominance comes about, entailsresponse-set organization in memory-the Sets Band Care memorial categories [1971a, p. 321]." Given thisformulation, he then generates the following prediction:After A-B, A-C learning, on a test of free recall for theresponses in the absence of the stimuli, there should besignificant clustering according to list membershipreflecting the existence of response-set organization.This prediction was tested in an experiment by Martinand Mackay (1970). Contrary to the deduction,list-determined clustering was not found (although it wasobserved under the CoD paradigm). There was, however,some marginal evidence of stimulus-determinedclustering such that responses paired with the samestimulus tended to be reproduced together. The resultswere taken to signal a critical failure of the"list-differentiation hypothesis" (Martin, 1971a, p. 321;Wichawut & Martin, 1971, p. 320).

We will now consider the logic of the prediction ofMartin and Mackay and the results of their experiment.First, does the principle of response suppression reallyentail the prediction of list-determined clustering in freerecall? The hypothesis at issue is concerned with themechanisms governing response availability per se;predictions about the order of reproduction of theresponses that fail to be suppressed are outside the scopeof the hypothesis. However, for those specificassociations that remain intact, stimulus control overoutput would be expected to be maintained just as itwas during the acquisition of the successive lists. Anyother assumption would in fact not be reasonable, sinceany theory of RI must start with the empirical fact thaton a conventional MMFR test, Ss are frequently able togive in succession both of the responses to a particuiarstimulus. To the extent that Ss are able to generate thestimulus terms implicitly, the situation should be thesame during free recall of the responses. Thus, theprediction tested by Martin and Mackay confuses twoanalytically separable questions: what is responsible forthe reduced availability of a set of responses, and whatare the principles governing the order of output of thoseresponses that remain available?

Furthermore, a measure of clustering cannot be usedas a final criterion for determining whether or not twosets of responses have been differentiated or form"memorial categories." The reason is, of course, thatclustering reflects an output strategy which is potentiallyunder the S's control. It is always possible for a 5 to scanavailable responses and to reproduce them in accordancewith the demands of a particular test of retention. In thesituation used by Martin and Mackay, Ss had justfinished learning two successive lists by thepaired-associate method, undoubtedly were able to recallmany of the stimuli, and hence would be expected to


behave in free recall very much as they would on anMMFR test. In this connection, it is important to note,as Martin (l971a, p. 320f) also did, that accuracy of listidentification typically remains very high after IL. Bythat criterion, the responses in the two lists clearly doform "memorial categories," and there is every reason tobelieve that they could be segregated or "clustered" inoutput under appropriate instructions.

As far as the specific analyses of Martin and Mackayare concerned, there are indications that their measuresof clustering lacked sensitivity. In particular, thisconclusion is suggested by a juxtaposition of theirmeasures of output order and clustering. On theimmediate test of retention, the mean recall frequencieswere 5.5 for B responses and 8.0 for C responses. Themeasures of output order showed that, on the average, Citems preceded B items by 3.1 positions. It is apparentthat a difference of this magnitude could be obtainedonly if at some point in recall a string of C items wasreproduced in succession, perhaps those for which thecorresponding B responses were not available. Themeasure of clustering had, of course, to take account ofthe absolute frequencies of recalls from the two sets, sothat the occurrence of strings of C items would beexpected "by chance." However, such an interpretationis not psychologically meaningful in an analysis ofretroaction where the difference between thefrequencies of Band C is the essential result of theexperimental manipulation. The picture which emergesis that there was clustering by both listsand stimuli andthat the separate analyses considering each type oforganization by itself did not yield convincing evidencefor either.

The experiment by Martin and Mackay has beendiscussed in detail because it has been used as a majorargument against the viability of the hypothesis ofresponse suppression. We believe that the results bearonly tangentially if at all on that hypothesis. Thesituation is thus similar to that encountered in Martin'scritique of the principle of unlearning. An observationregarding retrieval dependencies regarded by the authoras a sufficient basis for the rejection of an entiretheoretical position turns out to have uncertainimplications and to be far from crucial. An overridingemphasis on retrieval dependencies may be ill-advised inpursuing the question of whether a complex theory oftransfer and forgetting with many empiricalramifications is in principle tenable. Correlationsbetween responses to corresponding stimuliare multiplydetermined, and output order is a characteristic ofperformance under the S's control. There is no reason toview retrieval dependencies as more decisive than otherempirical implications derived more directly andunequivocally from the theoretical positions at issue.

THE ROLE OF ENCODING PROCESSESINTRANSFER AND INTERFERENCE

In this final section. we consider alternative

conceptualizations of transfer and interference advancedby critics of current positions, notably Martin (1971a)and Greeno et al (1971). These approaches sharea majoremphasis on the role of encoding processes in creatingconditions conducive to negative transfer andinterference. There are, however, some importantdifferences between the positions outlined by Martinand Greeno, and we will comment on each of them inturn.

Martin'sAnalysis of Interference Processes

Martin's reading of the empirical facts led him to theconclusion that an adequate theoretical explanationmust encompass the following pattern of resultsobserved on tests of retention after A·B, A-C learning:(a) There is a dominance relation between Band Cproducing retroaction and proaction; (b) in view of theindependence retrieval phenomenon, such dominancecannot be attributed to pairwise associative interferenceand unlearning; yet (c) according to the outputdependencies observed in free recall, Bs and Cs clusterby stimuli so that retention losses cannot be attributedto suppression of sets of responses. Martin believes thatthese apparently divergent facts can be understood onthe basisof certain assumptions about the characteristicsof stimulus encoding in the successive stages of the A-B,A-C paradigm.

The main steps in the argument are as follows: (a) Thenominal stimulus A possesses a variety of separate anddistinctive features which differ in their saliency for thelearner and hence are not equally likely to be sampled.(b) The probabilities with which stimulus componentsare sampled are independent; this characteristic ofindependence does not change in the course ofacquisition of either A·B or A-C. That is, stimuluscomponents do not become associated with each otherin the course of associative learning. (c) The featuresthat are most salient and hence most likely to besampled change when the response C is substituted forB. This change occurs because. as Greeno et al (1971)have postulated, the characteristics of the responsesignificantly influence the sampling probabilitydistribution over the feature set. (d) Retroactiveinhibition occurs because the feature-sampling biasestablished during A·C learning persists on the test ofrecall for A-B. As Martin notes. this assumption is anapplication of the concept of generalizedcompetition-but to feature sampling rather than toresponse recall (1971a, p. 328). (It is worth noting thatthis last point may be classified as a version of"list-differentiation theory" insofar as the sampling biasaffects the entire array of first-list stimuli. Somethinglike stimulus-feature suppression seems to be implied')As the feature-sampling distribution reverts to itsoriginal shape. spontaneous recovery is observed.Proaction is to be expected to the extent that some ofthe features dominant during the acquisition of .-\-8continue to be sampled during the recall of A-r. (e) On

32 POSTMAN ANDUNDERWOOD

Table 1Probability of Reca1l of One or Two Additional Stimulus

Elements Given Recall or Nonrecall of the Response(from Postman & Greenbloom, 1967)

Single-Letter Cue

First Second Third

Recall P N P N P !'<

R .673 248 .871 178 .827 179R .408 76 .219 146 .221 145

Note-N = number of cases.

an MMFR test, the S takes random samples of thefeatures of the nominal As. In these samples, thepresence of features relevant to the recallof Bsand Cs isnot correlated: consequently, the retrieval of responsesfrom the two lists is independent. While the samples aresaid to be random, a bias in favor of the second-listfeatures is assumed to operate to account for thedominance of Cs over Bs (biased random sampling?).(f) As for the clustering of responses by stimuli in freerecall (Martin & Mackay, 1970), the assumptions thatare made are best stated in Martin's own words: "If AB 1and AB2 are two stimuli in A,B learning and Ac 1 is theA-C learning version of ABl' then certainlyAB 1 will bemore similar to Ac1 than to AB 2' ... Accordingly, if infree recall ... the response B1 is recalled, then viamediation through AB 1 the subject is more likely tofollow up with recall of C1 than with B2 [1971a,p. 329]." In short, the similarity of the functionalversions of the same nominal stimuli is taken to beresponsible for the observed dependency in the freerecall of the responses attached during acquisition to thesame stimuli.

Stimulus-Component Independence?

Before considering the explanatory and predictivevalue of this theoretical schema, we will focusexplicitlyon the assumption of stimulus-component independencewhich is of pivotal importance in the development of thetheoretical argument. This assumption is based largelyon the results of studiesby Wichawut and Martin (1970)and Martin (1971b). These investigators usedpaired-associate lists in which the compound stimulusterms werethree four-letter nounsandeachresponse termwas another four-letter noun. On the terminal test forstimulus selection, each of the four elements waspresented singly and the S was required to reproduce theremaining three elements. The main finding of interestfor present purposes was that a given stimulus elementelicited other stimulus components with somesubstantial probability (which increased as a function ofdegree of learning) only when the response term wasrecalled. When the response term was not recalled, theprobability of additional stimulus elements beingreproduced was essentially zero. In other respects, the

usual results of stimulus selection studies wereduplicated. e.g., the effectiveness of the componentsvaried with ordinal position. The dependence ofcomponent reproduction on response recall ledWichawut and Martin to the conclusion that (a) stimuluscomponents do not become associated with each other,and (b) in the presence of a given stimulus componentthe retrievability of additional ones is mediated by thecommon response term.

We have two comments on these findings andinterpretations. First, the inference that thereproduction of componentsIs mediated by the responseterm is not entailed by the conditional probabilities.There is a simpler explanation. As the facts of selectionshow, the Ss attended to some elements but not toothers. If a S had attended to a particular element, Ai'then he was able to recall the response and also toreproduce other elements. The latter fact means,descriptively at least, that an association had beenformed between elements. If a S had failed to attend toan element, he could not give either the response orother elements. It is not apparent why a process ofmediation via the common response has to be invoked atall.Conditional probability does not imply causation.

Our second comment concerns the generality of thefindings of Wichawut and Martin.The implication is thatthere is no incidental learning of associations involvingnonselected elements. Such complete absence ofincidental learning after many exposures to a set ofmaterials is a rather rare phenomenon. It appeared likelythat this state of affairs was peculiar to the arrangementused by Wichawut and Martin in which the stimuluscomponents were distinctive meaningful units and alsobelonged to exactly the sameclass of verbal items as theresponse. Under these circumstances, the Ss may havelearned selectively shortened serial chains anchored tothe terminal unit (the response). The situation mightwell be different when stimuli are less segmented andwell differentiated from the responses. In light of theseconsiderations, a reanalysis of the results of a study byPostman and Greenbloom (1967) was undertaken. Inthat experiment, the stimuli were trigrams and theresponses were digits. The design included two levels oftrigram pronunciability, with response recall higher forthe easy than for the hard list. To maximize thesensitivity of the tests, the reanalysis was limited to thehard lists which also produced more single-letter cueselection than did the easy ones. The probabilities ofstimulus-element reproduction conditional on recall ornonrecall of the response were determined for the 18single-letter components of the stimuli in a list (threepositions in each of six trigrams). The analysis thusparallels that of Wichawut and Martin, except that theresponses were not presented as cues.

As Table I shows, the probability of elementreproduction in the absence of response recall wasclearly greater than zero. Thus, the contingencyobserved by Wichawut and Martin for strings of words


does not hold for trigram stimuli paired with digitresponses. The probabilities of recall of elementsand ofresponse covaried as expected. Another feature worthnoting is that the degree of covariation was less when thefirst letter rather than the second or third was used as atest cue. There are probably two factors responsible forthis pattern of conditional probabilities: (a) Selectionwas more likely to remain strictly limited to the salientfirst letter than to the other two; hence, failures toreproduce other letters along with the response weremost common when the first letter was selected; and(b) component associations reflect normal habits ofprocessing verbal units in serial order; consequently, theremaining elements were more likely to be elicited evenin the absence of response recall (indicating nonselectionas a functional cue) by the first letter than by the othertwo letters. In any event, the resultsof this analysis castdoubt on the generality of the findings ofWichawut andMartin. It is, therefore, premature to incorporate ageneral principle of stimulus-component independenceinto a theory of transfer and interference.

Critiqueof Martin'sPosition

We return now to an overall consideration of Martin'sproposals. In his account, the explanation of thephenomena of retroaction and proaction takes its pointof departure from the principle of encoding variability.Martin's (1968) original formulation of this principle ismodified inasmuch as changes in stimulus encoding are,in agreement with Greenoet al (1971), explicitly tied tochanges in the characteristics of the successive responses.With the aid of the assumption of a feature-sampling biascarrying over from interpolated learning to the test ofretention, the retroactive losses in the recall of the firstlist can be understood; persistence of earlier encodingsproduces PI. While this hypothesis may be reasonable,there has been little or no support for the assumedoccurrence of encoding variability in several recentexperiments. Perhaps the most persuasive data againstthe hypothesis were reported by Goggin and Martin(1970). Other studiesyielding essentially negative resultsare those of Williams and Underwood (1970), Weaver,McCann, and Wehr (1970), and Postman and Stark(1971). A shift in the feature-sampling distribution maybe expected on a priori grounds, but the evidence is thatthe encoding of the nominal stimuli typically does notchange between first- and second-list learning. Thesefindings, which negate the assumption that identicalnominal stimuli have a high probability of being recodedin the transfer phase, should not be confused withresults showing a shift in stimulus selection when anentirely new element is first introduced duringsecond-list learning, i.e.. the stimuli are A in the first listand AX in the second list (Houston. 1967a;Merryman &Merryman. 1971). The latter findings are essentiallyirrelevant to the problem at issue. which is to account

for retroaction when the nominal stimuli remainunchanged.

Martin expresses some reservations about theexperimental tests of recoding that have been carriedout, calling attention to the risk of identifying"obvious,external, experimenter-defined stimulus attributes withthe learner's subtle, internal, subjectively definedfunctional encodings [1971, p. 330]." When that isdone, one may conclude erroneously that the encodingadopted during first-list learning was retained in itsoriginal form in the transfer phase. Martin may becorrect in his concern that the tests of recoding used sofar have lacked sensitivity. However, the concept ofrecoding has explanatory value only to the extent that itcan be translated into experimental operations if theentire question is not to be begged. So far, the attemptsto produce changes in encoding under controlledconditions have yielded largely negative results. Thesefindings must stand until and unless other methods aredeveloped and compel new inferences.

There is an internal contradiction, or at least aninconsistency, in Martin's attempt to reconcile theapparent independence of retrieval of first- andsecond-list responses on MMFR tests and the evidencefor clustering by stimuli in the free recall of responses.To account for the latter, Martin invokes the similarityof the two functional encodings of a given nominalstimulus. Specifically, he suggests the operation of amediational chain B1-AB1-AC1-C1. It is difficult tosee why a parallel process does not take place in MMFR.In fact, it remains unclear exactly what is meant by theindependence of stimulus components if differentfunctional encodings sharing common features eliciteach other as they are alleged to do during free recall ofthe responses. It does not seem possible to insist onstrict independence of stimulus components as theyfunction in MMFR and at the same time to link them inan implicit mediational chain under conditions of freerecall.

We have indicated some of the difficultiesencountered in the application of Martin's theoreticalschema to empirical findings which the author regards ascrucial for his own position as well as those of others.There is in addition a more general systematic objectionto Martin's analysis of the process of interference. Theentire argument is developed with reference to the A-B,A-C paradigm of negative transfer. This paradigm has. tobe sure, served as the point of departure in classicaltreatments of interference. It is a fact. however. thatsubstantial amounts of retroaction have been observedunder conditions in which. the nominal stimuli in thesuccessive tasks are unrelated. e.g., the A·B. C·D andA.B, C·B paradigms. These findings have made itnecessary to recognize the existence of components ofinterference that are not tied to the characteristics of thestimulus for the forward association. in particular thoseresponsible for the reduced availability of responses andbackward associations. A general theory of interference


cannot be considered adequate unless it is applicable tothe entire range of paradigms that produce retroactionand proaction. The limited generality of Martin'sanalysis may be attributed to the fact that he followsGreeno et al (1971) in rejecting, without an adequatesubstitute, the two-stage conception of associativelearning. Thus, no provision is made for the analyticdistinction between response learning and associativelearning which can be logically extended to theunlearning of the two components. In placing theexplanatory burden exclusively on changes in stimulusencoding, Martin appears to have reverted to aSingle-stage conceptualization of interference, similar tothat of Gibson (1940) and Osgood (1949) whichpreceded the development of a component analysis oftransfer.

The issue of the generality of Martin's thesishas to bepressed further. It is a well-known fact that RI and PIare not limited to paired-associate learning but have alsobeen observed under quite different procedures such asserial and free-recall learning. It is possible to say, ofcourse, that in all such cases there is a change in thesampling of the features of whatever the functionalstimuli might be in the successive lists. A sampling biaswould then carry over to the test of recall and produceinterference. The application of such a hypothesisWOUld, however, soon encounter serious difficulties. Letus mention only two examples. Spatial and ordinalpositions have been identified as effective functionalstimuli in serial learning; it hardly seems reasonable toassume that such stimuli comprisesufficiently distinctiveand independent features to account for substantialamounts of retroaction. In the case of free recall, theidentification of the functional stimuli or retrieval cuesfor individual items constitutes a major theoreticalproblem. It seems doubtful that interference in freerecall can be subsumed under the A-B, A-C schemaapplied to individual words. Yet RI in free recall is notonly quite heavy, but its onset is also very rapid (e.g.,Postman & Keppel, 1967; Tulving & Psotka, 1971;Tulving & Thornton, 1959). The problem of generalityfaced by any interference theory thus clearly reachesbeyond the confines of the conventionalpaired-associateparadigms. At least some of the concepts of existingtheories are in principle applicable to the analysis ofinterference outside the domain of paired-associatelearning, e.g., the unlearning of contextual, positional,and spatial associations and the operation of amechanism of respense suppression.

GREENO'S THEORY OFTRANSFER AND INTERFERENCE

Stagesof Associative Learning

The point of departure in Greeno et al's (1971)analysis is a new definition of the successive stages ofassociative learning. In the first stage, a record of the

stimulus-response pair is stored in memory as a unit. Inthe formation of such a unit, the encoding of thestimulus is influenced significantly by that of theresponse with which it is paired; conversely. theencoding of the response is influenced by that of thecorresponding stimulus. In the second stage. the learneracquires the ability to retrieve the responses when thestimulus terms are presented on test trials. A retrievalplan is developed, the efficiency of which dependsprimarily on the characteristics of the stimulus terms.Greeno et al propose this schema in lieu of theconventional division of the acquisition process into aresponse-learning and associative stage (Underwood &Schulz, 1960). The basic evidence presented in supportof the revised conceptualization is that the duration ofthe first stage, as estimated by the application of amathematical model, is found to be influenced bystimulus as well as response difficulty. Thus, bothmembers of the pair, not just the response, areimplicated. This evidence is less than compelling. Sincethe measure used to index the duration of the first stageis strongly correlated with the number of errors beforethe first correct response to an item, the outcome is aforegone conclusion. The duration of the first stagecannot but be a function of both stimulus and responsedifficulty, because both response learning and associativelearning must have progressed to an adequate levelbefore the first correct response to a particular stimuluscan be given.

The new definition of the stages of associativelearning formalizes the assumption that the nature of theresponse influences the functional encoding of thestimulus, and vice versa. This formulation may have anadvantage for certain theoretical purposes, e.g., inrelating negative transfer and interference to changes instimulus encoding. At the same time, however, there is aloss of the explanatory power which has accrued fromthe systematic analysis of response processes per se(Underwood & Schulz, 1960). Such phenomena as theeffects of response familiarization on subsequentassociative learning are left unaccounted for. The earlydemonstration by Underwood, Runquist, and Schulz(1959) that Ss are able to recall responses before theycan pair them with the correct stimuli is overlooked.Quite apart from the difficulties which emerge in theanalysis of interference processes, on which we havealready had occasion to comment in Our discussionofMartin's views, the new two-stage formulation also leavesmany unfilled gaps in our understanding of thephenomena of acquisition.

Negative Transfer

Within the theoretical framework proposed by Greenoet al, negative transfer is expected to develop in the firstinstance because of interference with the storage of newpairs. Such interference will develop to the extent thatthe stimuli, the responses, or both are encoded in the

CRITICAL ISSUES IN INTERFERENCE THEORY 3S

same wav as in the successive lists. Greeno takes it forgranted that encodings will, indeed, be carried over fromthe first to the second task: this is the principle ofpersistent encoding to be contrasted with Martin's(1968) principle of encoding variability. (There may, ofcourse, be persistence followed by change, but therelationship between these apparently contradictoryprinciples has not been considered explicitly by eitherauthor.)

A second source of negative transfer is the disruptionof the first-list retrieval plan when old stimuli are pairedwith new responses. There are two ways in which theretrieval difficulties can be overcome so that the masteryof the transfer task becomes possible: (a) The first-listretrieval system is suppressed, although it remains intactand available for future use. Such suppression of thetotal retrieval plan is possible when there are distinctivecues to list differentiation and hence to thedifferentiation of the successive retrieval systems. Thepresence of new stimuli in the C-B paradigm, and of newresponses in the A-C paradigm,presumably serves thisfunction. (b) In the absence of effective cues to listdifferentiation, the development of a retrieval plan forthe transfer task entails the "disorganization orbreakdown" of the earlier system.The A-Br paradigm, inwhich both the stimuli and the responses remain thesame, is a case in point. Presumably the suppression ofan intact retrieval system and the substitution of a newone can be accomplished more readily and will produceJess interference than the breakdown and reorganizationof a previous systemwhich has become inappropriate.

Given these sources of interference, Greeno et alderive certain predictions about the relative difficulty ofthe two stages of acquisition postulated by them undervarious conditions of transfer. Consider the C·D, A-C,CoB, and A·Br paradigms. Persistence of encoding isassumed to interfere with the storage of pairs inmemory. The magnitude of such interference effects isexpected to be of the same order for cases of stimulusand of response identity because the functionalencodings of the two terms are interdependent. The firststage should, therefore, be accomplished more rapidlyunder the CoD than under the A-C or CoB paradigm, withlittle or no difference between the latter two. As forA-Br, the first stage "might be harder" than in A·C orC·B, but only if the stimulus and response effects oninterference with storage are cumulative. Negativetransfer in the second stage reflects disruption of theretrieval plan which is tied to the characteristics of thestimuli.The prediction is that the second stage should beeasier when the stimuli change, as in CoD and CoB, thanwhen they remain the same, as in A-C and A-Br.Second-stage interference should be heavier in A·Br thanin A·C because successive retrieval systems can bedifferentiated more readily in the latter than in theformer case.

If the second stage is viewed as one of associativelearning, then there is no disagreement with the rank

Table 2Mean Number of Errors Before First Correct Response

Under Vuious Transfer Puadigms

Paradigm

C·D C·B A-e A-Br

I. Postman (1962)LowOL 3.64 4.15 5.05 5.22Medium OL 3.14 3.48 4.92 4.67High OL 3.12 2.82 .3.36 5.23

11. Postman & Warren (1972)Condition I 2.66 2.85 3.19 4.32Condition D 3.20 3.34 3.97 4.11

order of paradigms derived from Greeno's analysis. Thepredictions regarding the first stage are, however, opento question. Since the duration of the first stage is highlycorrelated with the number of trials to the first correctresponse, the conventional position would be that itshould reflect both response and associative learning, butwith the former carrying greater weight than the latter inthe early phase of acquisition. Thus, a differencefavoring CoB over A·Brwould be expected. Furthermore,with highly meaningful responses the associativecomponent should come to the fore early, and withprior forward associations, a more effective source ofinterference than backward ones, A·C should be inferiortoC-B.

The measures of the two stages which Greeno et alpresent in support of their predictions (1971, Table 1)were obtained in a number of experiments. The samplingof conditions was quite uneven: only one studycomprised all four paradigms of interest, anotherexperiment included three, and the remaining six sets ofdata were limited to the CoB and A-Br treatments. Theprediction with respect to the difference between thelatter two was equivocal to beginwith; nevertheless, themain emphasis was on this particular comparison. Themeasures, once more estimated from the parameters of amathematical model, were interpreted by the authors asshowing that CoB is superior to A-Br in the second butnot in the first stage of learning. With some exceptions,the differences in the duration of the first stage weresmall (although they favored CoB in five out of sevencases). In the one experiment in which all fourparadigms were included. C-B was superior to both A-Cand A·Br, although the apparent differences were small.

Since the measures considered by Greeno et al werefragmentary and based on parameter estimates ofunknown validity, we present in Table 2 more directmeasures of the index said to be highly correlated withthe duration of the first stage. viz, the number of errorsbefore the first correct response to an item. The first setof values comes from an experiment by Postman (1962)in which all four paradigms were used and transfer wasmeasured after three different degrees of first-listlearning-to a criterion of 6/10. 10(10. and 10/10 + 50C;overlearning. The materials were lists of paired

36 POSTMAN ANDUNDERWOOD

adjectives. At each level of first-list learning. the meannumbers of errors to the first correct response wereclearly lower for C·B than for A-Br and A.('. Thedifferences between the latter two were small. except atthe highest level of first-list learning where theinterfering associations were heavily overlearned. Thus,the pattern of differences is clearly at variance with thatpredicted by Greeno et al.

The second set of values comes from a study byPostman and Warren (1972). The materials were againpaired adjectives. The first list waslearned to a criterionof two successive errorless trials and was followed by thetransfer list after either 1.5 or 20 min (Conditions I andD). There was a tendency for acquisition to be slowerwhen the second list was learned after a delay ratherthan immediately. Under both conditions, thedifferencesamong paradigms showed the same pattern asin the earlier study: the mean number of errors beforethe first correct response was smaller for C·B than forA.(' and A·Br. The scores were once more higher forA·Br than for A.(', although that difference wasminimized after a delay.

The scores shown in Table 2 yield the same alignmentof paradigms as do the overall measures of transferperformance. That was to be expected on the basis ofthe conventional two-stage formulation, since responselearning and associative learning jointly determine theoccurrence of the first correct response to a specificstimulus. In accounting for whatever discrepancies thereappear to be between the findings of Greeno et al andours, we must for the present limit ourselves to thecomment that our measures were determined from theactual experimental observations, whereas theirs wereparameter estimates.

The one-to-one correspondence between the indexesof the first stage and total transfer performance invitesan additional comment on the conceptualization ofGreeno et al. Their basic distinction is between a storageand a retrieval stage. This distinction suffers from aninevitable ambiguity when it is applied to experimentalobservations. Whether or. not an item has been storedcan be determined only by means of a test of retrieval.Hence, any measure of the storage stage, whetherindexed directly by the first correct response orestimated indirectly, can be obtained only when thesecond stage, in which the S presumably learns toretrieve the items, is already under way. Thus, weconclude that the analysis of negative transfer offered byGreeno et al remains open to question on logical,pragmatical, and empirical grounds.

Retroactive Inhibition

Greeno et al attribute RI to the reduced retrievabilityof first-list items. While the stored representationsof theoriginal units presumably remain intact, the adoption ofa new retrieval plan in the transfer phase interfereswiththe S's ability to use the stimuli as cues in the recall of

the first list. Recovery from RI, e.g., during successivetests, is more likely when the original retrieval plan issuppressed but intact, as in A-C, than when it isreorganized, as in A-Br (cf. James, 1968). An importantfeature of this analysis is that the interference with recallis assumed to operate "at the level of several items or thewhole list. rather than at the level of individual items[ 1971, p. 339]." Consequently, the independentretrieval phenomenon is seen as consistent with thetheory and incompatible with a mechanism ofassociative unlearning. We have already tried to show inour discussion of Martin's position why this inference isnot acceptable.

It remains for us to comment on the interpretation ofRI as due to the suppression or degradation of a retrievalplan brought about primarily by a change in stimulusterms. It is not clear to us what exactly is meant by aretrieval plan for a list as a whole in the context ofpaired-associate learning. The retrieval plan is assumed toevolve "mainly because of the relationships that thesubject discovers among the stimuli or stimulus-responsepairs in the list [po 334] ." But how does the discoveryof such relationships generate a plan for the retrieval ofspecific responses? When a list is segmented into discretepairs, with responses to be given to stimuli in anunpredictable order, the idea of a retrieval plan for thelist as a whole would appear to be a contradiction interms. It is possible to conceive of a systemof encodingthe stimuli that would minimize interpair interference,but such a mode of attack on the learning task wouldnormally be considered under the heading of stimulusdifferentiation. Furthermore, the development of suchencoding devices would presumably be part of thestorage process which is assigned to the first, rather thanthe second, stage of learning. Certainly the retrieval plancannot refer to the acquisition of a repertoire ofresponses, since the concept of a response-learning stageis explicitly rejected. In short, in the absence of anyfurther specification or elaboration, the reference to aretrieval plan for an entire list of arbitrarily paired unitsis elusive. To say, therefore, that RI reflects thesuppression or disruption of such a plan does not carryany concrete implications. The assertion reduces to thetau tology that interference with recall reflectsinterference with retrieval. All this is not to deny that RImay be produced by the loss of retrieval cues, but thishypothesis has to be translated into experimentaloperations appropriate for a given learning situationbefore it can be evaluated(e.g.,Tulving & Psotka, 1971).

The question of generality raised by Martin's analysisapplies also to the position of Greeno et ai, although thelatter do consider the implications of the proposedmechanisms for paradigms other than A·B, A-C.Interference with storage is expected whenever there isan overlap of either stimulusor response elements.Thus,the occurrence of negative transfer under the A-B, C·Barrangement could be accommodated. However,retroactive effects are predicted only when there is a


disruption of the retrieval plan which is tied primarily tostimulus overlap between the successive tasks. Thus,little if any RI is expected under the CoD and CoBparadigms. As we have already noted, however, theseexpectations are contrary to fact. Again, this theoreticalaccount provides no purchase on the explanation of RIin other than paired-associate situations, such as seriallearning and free recall. Furthermore, no explicitconsideration has been given to the mechanisms of PI.

EVALUATION OF RECENT CRITIQUESOF INTERFERENCE THEORY

We will now summarize our evaluation of the recentarguments of Martin and Greeno et al for a newconceptualization of the interrelated processes ofnegative transfer, retroaction, and proaction.

(1) The rejection of the principle of associativeunlearning is based on a mistaken interpretation of therelationship between negative transfer and unlearning. Ithas not been, and need not be, assumed that theunlearning of old associations has to "clear the way" forthe learning of new ones. Hence, the independentretrieval phenomenon, quite apart from the objectionsto the measurement procedures by which it has beenestablished, has no critical bearing on the hypothesis ofassociative unlearning.

(2) The hypothesis of response suppression has beenrejected in toto on insufficient grounds.The question ofwhether or not a mechanism of response selectioncontributes to RI cannot be answered by such measuresas indexes of response clustering in free recall. Theevidence pointing to item-specific associative losses, e.g.,under conditions of mixed-list interpolation and underthe A-Br paradigm, does not in and of itself invalidatethe hypothesis. The question at issue has been, andremains, a quantitative one: what proportion of RIunder given conditions of learning and testing can beattributed to response selection?

(3) While the emphasis placed by both Greeno et aland Martin on the characteristics of both stimulus andresponse encoding is timely and useful, too much of anexplanatory burden is placed on these processes. Thisconclusion is based on two major considerations. First,because of the overriding importance attached to thetemporal course of stimulus encoding during successivetasks, the analyses focus narrowly on conditions ofstimulus identity which do not by any means exhaustthe range of paradigms under which retroactive andproactive effects' are observed. Second, the centralassumptions about either the variability (Martin) orpersistence (Greeno et al) of stimulus encoding acrosstasks are essentially ad hoc and either have not beentested or have for the most part failed to be supportedwhere relevant empirical evidence is available. Theconcept of subjective encoding inherently carries a riskof question-begging invocation and should therefore beused cautiously in theory development.

(4) The analyses of Greenoet al and Martin are besetby an internal contradiction with respect to the contrastbetween generalized and item-specific interference. Thecentral emphasis on the encoding of individual pairscarries a prima facie implication that interference effectsshould be item-specific. This implication is made explicitby Martin. who predicts the loss of an item on the testof recall if the relevant features of a particular stimulusfail to be sampled. However, in order to postdict thephenomenon of retroaction, he must appeal to amechanism that operates on the list as a whole, viz, thepersistence of a sampling bias in favor of the featuresselected during the transfer phase. Since the basicmechanism responsible for retroaction is taken tooperate on the entire list, item specificity loses itssystematic meaning. Whenever a processcomes into playat the list level, it will produce effects on specific items.There is no way of predicting,however, which items willbe vulnerable and which items will be immune togeneralized interference.

Similar considerations apply to Greeno's analysis.Again, the source of interference is at the list level, butnow it is the retrieval plan rather than the aggregate ofstimulus features that is suppressed, and the specificencodings of individual stimuli and responses have nodeterminate consequences for recall performance. Insum, Martin has substituted stimulus-featuresuppression, and Greeno has substituted retrieval-plansuppression, for response suppression. It is fair to askwhether either of the new formulations is at the presenttime anchored more securely to experimental evidencethan the principle of response-set interference.The datawe have reviewed do not seem to compel an answer inthe affirmative. The latter can, moreover, claim someadvantage because of its continuity with a large body offacts implying the operation of response-relatedprocesses in acquisition and recall. Whichever variant oflist suppression one favors, the available findings do notmake it possible to rule out altogether item-specificunlearning in the sense of extinction of individualassociations.

We see no reasons at this time to abandon abruptlythe concept of associative unlearning or to jettisonresponse-set interference as a possible contributor toretroaction and perhaps proaction. On the other hand.we do not wish to reject out of hand all the interestingspeculations offered by Greeno et al and Martin. Theyhave called attention to the possible importance ofchanging encoding processes in determining the courseof transfer and interference at recall. Other investigatorshave focused on the same issue in different contexts(e.g.. Tulving & Psotka, 1971: Wood. 1971). Theintroduction of such ideas into the analysis of transferand interference is very much in the spirit of the times.but it does not call for the dismantling of the entireconceptual framework which has evolved over the yearsout of classical interference theory. We see thisconceptual framework as in continuing need of critical


reexamination; but if this enterprise is to be successful,the most urgent task is the resolution of the manyinconsistencies and apparent contradictions in therapidlygrowing body of experimental findings.

REFERENCES

Abra, J. C. List-I unlearning and recovery as a function of thepoint of interpolated learning. Journal of Verbal Learning &Verbal Behavior. 1969,8,494-500.

Adams. J. A., Marshall, P. H., & Bray, N. W. Interference andretrieval. Journal of Verbal Learning & Verbal Behavior, 1971,10, 548-555.

Anderson, R. C.• & Watts, G. H. Response competition in theforgetting of paired associates. Journal of Verbal Learning &Verbal Behavior, 1971, 10,29-34.

Barnes. J. Moo & Underwood. B. J. "Fate" of first-listassociations in transfer theory. Journal of ExperimentalPsychology, 1959,58,97-105.

Birnbaum, I. M. Response-class similarity and first-list recall withmixed and unmixed transfer designs. Journal of ExperimentalPsychology, 1968a, 77, 542-546.

Birnbaum, I. M. Unlearning as a function of second-listdominance. Journal of Verbal Learning & Verbal Behavior,1968b. 7.257-259.

Birnbaum, I. M. List-I recall and associative strength of Iist-2pairs. American Journal of Psychology, 1968c, 81, 442-447.

Birnbaum, I. M. Response selection and retroactive inhibition.Journal of Experimental Psychology, 1970,85,406-410.

Birnbaum, I. M. General and specific components of retroactiveinhibition in the A-B, A-C paradigm. Journal of ExperimentalPsychology, 1972,93,188-192.

Ceraso. J., & Henderson, A. Unavailability and associative loss inRI and PI. Journal of Experimental Psychology, 1965, 70,300-303.

Ceraso, J., & Henderson, A. Unavailability and associative loss inRI and PI: Second try. Journal of Experimental Psychology,1966,72,314-316.

Cofer, C. N., Faille, N. F., & Horton, D. L. Retroactiveinhibition following reinstatement or maintenance of rust-listresponses by means of free recall. Journal of ExperimentalPsychology, 1971, 90, 197-205.

Dallett, K. M., & D'Andrea, L. Mediation instructions versusunlearning instructions in the A-B, A-C paradigm. Journal ofExperimental Psychology, 1965,69,460-466.

Da Polito, F. J. Proactive' effects with independent retrieval ofcompeting responses. Unpublished doctoral dissertation,Indiana University, 1966.

Delprato, D. J. Successive recall of list 1 following list 2 learningwith two retroactive inhibition transfer paradigms. Journal ofExperimental Psychology, 1970,84,537-539.

Delprato, D. J. Specific-pair interference on recall andassociative-matching retention tests. American Journal ofPsychology, 1971,84, 185-193.

Estes, W. K. Learning theory and the new "mental chemistry."Psychological Review, 1960,67,207·233.

Forrester, W. E. Retroactive inhibition and spontaneous recoveryin the A·B, D-C paradigm. Journal of Verbal Learning &Verbal Behavior, 1970,9,525-528.

Friedman, M. J., & Reynolds, J. H. Retroactive inhibition as afunction of response-class similarity. Journal of ExperimentalPsychology, 1967,74,351-355.

Garskof, B. E. Unlearning as a function of degree of interpolatedlearning and method of testing in the A-B, A-C and A-B, CoDparadigms. Journal of Experimental Psychology, 1968, 76,579-583.

Garskof, B. E., & Sandak, J. M. Unlearning in recognitionmemory. Psychonomic Science, 1964, 1, 197-198.

Gibson, E. J. A systematic application of the concepts of

generalization and differentiation to verbal learning.Psychological Review. 1940,47, 196-229.

Goggin. J. First-list recall as a function of second-list learningmethod. Journal of Verbal Learning & Verbal Behavior. 1967.6,423-427.

Goggin, J. Retroactive inhibition with different patterns ofinterpolated lists. Journal of Experimental Psychology. 1968.76, 102-108.

Goggin, J., & Martin, E. Forced stimulus encoding andretroactive interference. Journal of Experimental Psychology.1970,84, 131-136.

Greenberg, S., & Wickens, D. D. Is matching performance anadequate test of "extinction" effects on individualassociations? Psychonomic Science, 1972,27,227-229.

Greenbloom, R., & Kimble, G. A. Extinction of R-5 associationsand performance on recall trials without informativefeedback. Journal of Verbal Learning & Verbal Behavior,1965,4,341-347.

Greenbloom, R., & Kimble, G. A. Improvement in recall onunreinforced recall trials. Journal of ExperimentalPsychology, 1966,71,159-160.

Greeno, 1. G., James, C. T., & Da Polito, F. J. A cognitiveinterpretation of negative transfer and forgetting of pairedassociates. Journal of Verbal Learning & Verbal Behavior,1971, 10,331-345.

Hintzman, D. L. On testing the independence of associations.Psychological Review, 1972,79,261-264.

Houston, J. P. Stimulus selection as influenced by degrees oflearning, attention, prior associations, and experience with thestimulus components. Journal of Experimental Psychology,1967a,73,509-516.

Houston, J. P. Proactive inhibition and competition at recall.Journal of Experimental Psychology, 1967b, 75,118-121.

James, C. T. A consideration and rejection of the associativetheory of interference and forgetting. Unpublished doctoraldissertation, Indiana University, 1968.

Kamman, R., & Melton, A. W. Absolute recovery of first-listresponses from unlearning during 26 minutes filled with aneasy or difficult information processing task. Proceedings of75th Annual Convention, American PsychologicalAssociation, 1967,2,63-64.

Keppel, G. Retroactive and proactive inhibition. In T. R. Dixonand D. L. Horton (Eds.), Verbal behavior and general behaviortheory. Englewood Cliffs, N.J: Prentice-Hall, 1968.

Keppel, G., Henschel, D., & Zavortink, B. Influence ofnonspecific interference on response recall. Journal ofExperimental Psychology, 1969,81,246-255.

Keppel, G., & Rauch, D. S. Unlearning as a function ofsecond-list error instructions. Journal of Verbal Learning &Verbal Behavior, 1966,5,50-58.

Koppenaal, R. J. Time changes in the strengths of A-B, A-C lists:Spontaneous recovery? Journal of Verbal Learning & VerbalBehavior, 1963,2,310-319.

Lehr, D. J., Frank, R. C., & Mattison, D. W. Retroactiveinhibition, spontaneous recovery, and type of interpolatedlearning. Journal of Experimental Psychology, 1972, 92,232-236.

Martin, E. Stimulus meaningfulness and paired-associate transfer:An encoding variability hypothesis. Psychological Review,1968,75,421441.

Martin, E. Verbal learning theory' and independent retrievalphenomena. Psychological Review, 1971a, 78,314-332.

Martin, E. Stimulus component independence. Journal of VerbalLearning & Verbal Behavior, 1971b, 10,715-721.

Martin, E., & Greeno, J. G. Independence of associations tested:A reply to D. L. Hintzman. Psychological Review, 1972, 79,265-267.

Martin, E., & Mackay, S. A. A test of the list-differentiationhypothesis. American Journal of Psychology, 1970, 83,311-321.


~k("rntal. 1. J. List differentiation as a function of time andtest order. Journal of Experimental Psychology, 1970, 83.220-223.

~lcGeoch. J. A. 77,e psychology of human learning. New York:Longmans, Green, 1942.

~lcGovern. J. B. Extinction of associations in four transferparadigms. Psychological Monographs, 1964, 78(16, WholeNo. 593).

Melton, A. Woo & Irwin, J. \1. The influence of degree ofinterpolated learning on retroactive inhibition and the overttransfer of specific responses. American Journal ofPsychology, 1940,53,173-203.

Melton, A. W., & von Lackum, W. J. Retroactive and proactiveinhibition in retention: Evidence for a two-factor theory ofretroactive inhibition. American Journal of Psychology, 1941,54,157-173.

'ferryman, C. T. Retroactive inhibition in the A-B. A-Dparadigm as measured by a multiple-choice test. Journal ofExperimental Psychology, 1971,91,212-214.

Merryman, C. T., & Merryman, S. S. Stimulus encoding in theA-B', AX-B and the A-B'r, AX-B paradigms. Journal of VerbalLearning & Verbal Behavior, 1971, 10,681-685.

Newton, J. M., & Wickens, D. D. Retroactive inhibition as afunction of the temporal position of interpolated learning.Journal of Experimental Psychology, 1956,51,149-154.

Osgood, C. E. The similarity paradox in human learning: Aresolution. Psychological Review, 1949,56, 132-143.

Popp, R., & Voss, 1. F. Acquisition-transfer and retention ofSoRt and S-R 2 associations as a function of S-R t' S-R2

pattern. Journal of Experimental Psychology, 1965, 70,304-308.

Postman, 1. The present status of interference theory. In C. N.Cofer (Ed.), Verbal learning and verbal behavior. New York:McGraw-Hill. 1961,

Postman, 1. Transfer of training as a function of experimentalparadigm and degree of first-list learning. Journal of VerbalLearning & Verbal Behavior, 1962, 1, 109-118.

Postman, 1. Unlearning under conditions of successiveinterpolation. Journal of Experimental Psychology, 1965,70,237-245.

Postman, 1. Differences between unmixed and mixed transferdesigns as a function of paradigm. Journal of Verbal Learning& Verbal Behavior, 1966,5.240-248.

Postman, 1., Burns, S.. & Hasher, 1. Response availability andassociative recall. Journal of Experimental Psychology, 1970,84, 404-411,

Postman, 1., & Greenbloorn. R. Conditions of cue selection inthe acquisition of paired-associate lists. Journal ofExperimental Psychology, 1967,73,91-100.

Postman, 1., & Hasher, 1. Conditions of proactive inhibition infree recall. Journal of Experimental Psychology, 1972. 92.276-284.

Postman, 1., & Keppel. G. Retroactive inhibition in free recall.Journal of Experimental Psychology. 1967, 74, 203-211.

Postman, 1., Keppel, G., & Stark, K. Unlearning as a function ofthe relationship between successive response classes. Journalof Experimental Psychology, 1965,69, 111-118.

Postman, 1., & Parker, J. F. Maintenance of first-list associationsduring transfer. American Journal of Psychology, 1970, 83,171-188.

Postman, 1., & Stark, K. The role of response set in tests ofunlearning. Journal of Verbal Learning & Verbal Behavior.1965.4,315-322.

Postman, L., & Stark, K. Role of response availability in transferand interference. Journal of Experimental Psychology. 1969.79.168-177.

Postman. 1., & Stark. K. Encoding variability and transfer.American Journal of Psychology. 1971, 84. 461-472.

Postman. L.. & Stark. K. On the measurement of retroactiveinhibition in the A-B. A-D paradigm by the multiple-choice

method: Reply to Merryman. Journal of Verbal Learning &Verbal Behavior, 1972, 11,465-473.

Postman, 1., Stark, K., & Fraser, J. Temporal changes ininterference. Journal of Verbal Learning & Verbal Behavior,1968,7,672-694.

Postman, L., Stark, K., & Henschel, D. Conditions of recoveryafter unlearning. Journal of Experimental Psychology. 1969.820, Pt. 2).

Postman, L., & Warren, 1. Temporal changes in interferenceunder different paradigms of transfer. Journal of VerbalLearning & Verbal Behavior, 1972, 11, 120-128.

Richardson, J., & Gropper, M. S. Learning during recall trials.Psychological Reports, 1964, 15,551-560.

Runquist, W. N. Retention of verbal associations as a function ofstrength. Journal of Experimental Psychology, 1957, 54,369-375.

Sandak, J. M., & Garskof, B. E. Associative unlearning as afunction of degree of interpolated learning. PsychonomicScience, 1967,7,215-216.

Shulman, H. G., & Martin, E. Effects of response-set similarityon unlearning and spontaneous recovery. Journal ofExperimental Psychology, 1970,86,230-235.

Silverstein, A. Unlearning, spontaneous recovery, and thepartial-reinforcement effect in paired-associate learning.Journal of Experimental Psychology, 1967,73,15-21,

Thune; 1. E., & Underwood, B. J. Retroactive inhibition as afunction of degree of interpolated learning. Journal ofExperimental Psychology, 1943,32.185-200.

Tulving, E., & Psotka, J. Retroactive inhibition in free recall:Inaccessibility of information available in the memory store.Journal of Experimental Psychology, 1971, 87, 1-8.

Tulving, E., & Thornton, G. B. Interaction between proactionand retroaction in short-term retention. Canadian Journal ofPsychology, 1959, 13,255-265.

Underwood. B. 1. The effect of successive interpolations onretroactive and proactive inhibition. PsychologicalMonographs, 1945,59(3, Whole No. 273).

Underwood, B. 1. Retroactive and proactive inhibition after fiveand forty-eight hours. Journal of Experimental Psychology.1948a, 38, 29-38.

Underwood, B. J. 'Spontaneous recovery' of verbal associations.Journal of Experimental Psychology, 1948b, 38,429-439.

Underwood, B. 1.. & Ekstrand, B. R. An analysis of someshortcomings in the interference theory of forgetting.Psychological Review, 1966,73,540-549.

Underwood. B. 1., & Ekstrand, B. R. Studies of distributedpractice: XXIV. Differentiation and proactive inhibition.Journal of Experimental Psychology, 1967,74.574-580.

Underwood, B. J., & Freund, J. S. Effect of temporal separationof two tasks on proactive inhibition. Journal of ExperimentalPsychology, 1968,78. 50-54.

Underwood, B. J., Runquist, W. N., & Schulz, R. W. Responselearning in paired-associate lists as a function of intra-listsimilarity. Journal of Experimental Psychology, 1959. 58,70-78.

Underwood. B. J., & Schulz. R. W. Meaningfulness and verballearning. Philadelphia: Lippincott. 1960. .

Weaver, G. E .. & Danielson. G. I. Successive interpolation andfirst-list recall in the A-B. A-e and A-B. D-e paradigms.Journal of Experimental Psychology. 1969. 81. 394-396.

Weaver. G. E.. ~IcCann, R. L.. & Wehr. R. J. Stimulusmeaningfulness. transfer. and retroactive inhibition in the .\-B.A-e paradigm. Journal of Experimental Psychology. 1970. 85.255·257.

Weaver. G. E.. Rose. R. Goo & Campbell. N. R. Item-specificretroactive inhibition in mixed-list comparisons of the A-B.A-C' and A-B. D-e paradigms. Journal of Verbal Learning &Verbal Behavior. 1971. 10.488-498.

Weiss. 1' .. & lazar. G. Successive recall as a measure for theunlearning effect. Psychonomic Science. 1968. 12.255-256.


Wichawut, c.. & Martin, E. Selective stimulus encoding andoverlearning in paired-associate learning. Journal ofExperimental Psychology, 1970,85,383-388.

Wichawut, C., & Martin, E. Independence of A-B and A.cassociations in retroaction. Journal of Verbal Learning &Verbal Behavior, 1971, 10, 316-321.

Williams, R. F.. & Underwood, B. J. Encoding variability: Testsof the Martin hypothesis. Journal of ExperimentalPsychology, 1970,86,317-324.

Winograd, E. List differentiation as a function of frequency andretention interval. Journal of Experimental Psychology. 1968.76(2, Pt. 2).

Wood, G. Activity, organization, retroactive inhibition, and freerecall. Journal of Verbal Learnine & Verbal Behavior, 1971,10,116-122. -

(Received for publication October 9. 1972;accepted November 17, 1972.)

critical issues in interference theory · 2017. 8. 24. · critical issues in the theoretical and...

Documents