DOCUMENT RESUME ED 036 858 Judd, Wilson A.; Glaser, Robert ... · WILSON A. JUDD AND ROBERT GLASER2...

ED 036 858

AUTHORTITLE

INSTITUTION

SPONS AGENCY

BUREAU NOPUB DATECONTRACTNOTE

EDRS PRICEDESCRIPTORS

ABSTRACT

DOCUMENT RESUME

24 CG 005 157

Judd, Wilson A.; Glaser, Robert.Response Latency as a Function of Training Method,Information Level, Acquisition, and Overlearning.Learning Research and Development Center ReprintNumber 52.Pittsburgh Univ., Pa. Learning Research andDevelopment Center.Office of Education (DHEW) , Washington, D.C. Bureauof Research.BR-5-0253Aug 69OEC-4-10-15832p.

EDRS Price MF-$0.25 HC-$1070*Behavior Patterns, *Learning, Learning Theories,*Paired Associate Learning, *Reaction Time,*Reactive Behavior, Recall (Psychological) ,*Response Mode, Time Factors (Learning)

Response latency was studied as a measure of

associative strength or degree of learning and possible basis forinstructional decision making in computer-assisted instruction.Latency was investigated in a paired-associate task as a function oftraining procedure and information transmission requirements duringacquisition and overlearning. The magnitude and variability oflatency measurements were independent of training during acquisition,but both were reduced by the recall paradigm during overlearning.Latency was a function of number of response alternatives duringacquisition and overlearning. During acquisition, prior to the trialof last error (TLE) , latency remained constant and did not differbetween correct and incorrect responses. There was a substantial dropin latency following TLE. Latency, as a rote verbal task, may be asensitive measure of strength of learning during the overlearningphase, but not during initial learning. (Author)

N)

UN

IVE

RS

ITY

OF

PIT

TS

BU

RG

H LE

AR

NIN

GR

& D

CE

NT

ER

RE

PRIN

T 52

RE

SPON

SE L

AT

EN

CY

AS A

FUN

CT

ION

OF T

RA

ININ

G M

ET

HO

D,

INFO

RM

AT

ION

LE

VE

L, A

CQ

UISIT

ION

, AN

D O

VE

RL

EA

RN

ING

WIL

SON

A. JU

DD

AN

DR

OB

ER

T G

LA

SER

9589£003

U.S. DEPARTMENT Of HEALTH, EDUCATION & WELFARE

OffICE Of EDUCATION

THIS DOCUMENT HAS BEEN REPRODUCED EXACTLY AS RECEIVED FROM THE

;Jr\ PERSON OR ORGANIZATION ORIGINATING IT. POINTS Of VIEW OR OPINIONS

STATED DO NOT NECESSARILY REPRESENT OFFICIAL OffICE Of EDUCATION

4) POSITION OR POLICY

pr\CD

RESPONSE LATENCY AS A FUNCTION OF TRAINING METHOD,

INFORMATION LEVEL, ACQUISITION, AND OVERLEARNING1

Wilson A. Judd2

and Robert Glaser

Learning Research and Development Center

University of Pittsburgh

Reprinted from the Jc.._2L.n-Eti.2fktcationa]tysLgyjmioloMonograph,h,

1969, 60 (4), Part 2.

1 Published by the Learning Research and Development Center at theUniversity of Pittsburgh, supported in part as a research anddevelopment center by funds from the United States Office of Edu-cation, Department of Health, Education, and Welfare. The research

reported herein was supported under Contract Nonr-624(18) with thePersonnel and Training Branch, Psychological Sciendes Division,Office of Naval Research.

2Now in the Department of Educational Psychology, University of

Texas at Austin.

Journal of Educational Paychologli1969, Vol. 60, No. 4, 1-30

JOURNAL OF

EDUCATIONAL PSYCHOLOGYMONOGRAPH

.11gyimallm.../Vol. 60, No. 4, Part 2

August, 1969MOMMONONIMI4

=11=0111=11110111.11.1111111M A. A111110111111,

RESPONSE LATENCY AS A FUNCTION OF TRAININGMETHOD, INFORMATION LEVEL, ACQUISITION,

AND OVERLEARNING'WILSON A. JUDD AND ROBERT GLASER2

University/ of Pittsburgh

Response latency was studied as a measure of associative strength ordegree of learning and as a possible basis for instructional decisionmaking in computer-assisted instruction. Latency was investigated ina paired-associate task as a function of training procedure (a compari-son of the anticipation and recall paradigms) and information trans-mission requirements (a comparison of two, four, and eight responsealternatives to an eight-item stimulus list) during both acquisitionand overlearning. The magnitude and variability of latency measure-ments were independent of training method during acquisition, butboth were reduced by the recall paradigm during overlearning. La-tency was an increasing function of the number of response alterna-tives during both acquisition and overlearning. During acquisition,prior to the trial of last error (TLE) for each item, latency remainedrelatively constant and did not differ between correct and incorrectresponses. There was a substantial drop in latency following theTLE. Pre-TLE latencies were independent of S learning rate,while post-TLE latencies were an increasing function of learningrate. The latency of the first correct response to an item was foundto be shorter if there were no subsequent errors on that item. Ingeneral, the study suggests that latency, at least in a rote verbaltask, may be a sensitive measure of strength of learning during theoverlearning phase, but not during initial learning.

The development of instructional strat-egies is currently an area of considerableinterest to educators and psychologists.

The research reported herein was performedpursuant to Contract Nonr-624(18) between thePersonnel and Training Branch, Psychological Sci-ences Division, Office of Naval Research and theLearning Research and Development Center, Uni-versity of Pittsburgh. Additional support wasprovided under Contract OE-4-10-158 with theOffice of Education, United States Department ofHealth, Education and Welfare. The authors ex-press their gratitude for the guidance and con-structive criticism offered by William W. Cooley,Paul M. Kieldergaard, J. Trevor Peirce, and JamesF. Voss.

Request for reprints should be sent to: RobertGlaser, Learning Research and Development Cen-ter, University of Pittsburgh, Pittsburgh, Pennsyl-vania 15213.

1

VII01111

With the availability of computer-basedinstructional systems, the educator nowhas the potential ability to constructinstructional sequences which will adaptto the particular requirements of individualstudents. Such a sequence would vary thecontent and order of the instructional ma-terials presented to the student as a func-tion of the student's responses. The majorproblem facing the psychologist interestedin this area is that the nature of the func-tions relating the student's responses tooptimal presentation schemes is not wellknown. The computer engineer has pro-vided the educator with the ability tomake extremely fast, relatively sophisti-cated decisions concerning individual stu-dents, and the psychologist needs to be able

Copyright Q) 1989 by the American Psychological Association, Inc,

2 WILSON A. JUDD AND ROBERT GLASER

to provide him with the basis for makingsuch decisions.

A frequent requirement is simply todetermine when a response has gained suf-ficient strength to allow the lesson toproceed to other material. Defining "suffi-cient strength" is of course a problem, buta more immediate problem is that ofsimply measuring the existing strength ofthe behavior. One approachperhaps themost common in practical group instruc-tional situationsis simply to give allstudents that amount of practice whichpast experience has shown to be sufficientfor the average student. This, of course,ignores differences between individuals.One practical solution is to continue prac-tice until the student reaches some behav-ioral criterion which is judged to beadequate. The common behavioral measureemployed is response frequency. This isalso the most reasonable measure in mosteases since the goal of the instruction canusually be defined in terms of an increasein the frequency of the correct or appro-priate response.

Frequency measures, however, may losetheir sensitivity as response probabilityapproaches asymptote. It is desirable thatthe student retain what he has learned,and it is known that retention is a functionof the degree of learning. In this respect,experimental work suggests that responselatency may be useful as a supplement tofrequency measures, and latency is quiteeasily measured during computer-assistedinstruction. A disadvantage of latencymeasures is their wide variability betweenSs and from trial to trial forfore the same S.The conditions contributing to this vari-ability may need to be controlled in orderto render latency measurements useful as abasis for instructional decisions. The pur-pose of the present study was to examineresponse-latency behavior throughout thecourse of learning in a paired-associatelearning task and to investigate certaintask variables which may influence themagnitude and variability of responselatency.

Latency as a Measure of ResponseStrength

Hull, in his Principles of Behavior(1943), stated that habit strength mani-fests itself in the length of the timeelapsing from the onset of the stimulus tothe onset of the associated response. Thisstatement was based on a study of paired-associate response latency done by Sim ley(1933). Sim ley's results indicated thatlatency decreased as a function of practiceafter the associative strength of the itemshad reached the "threshold of recall." Itwas further shown that during the sequenceof correct responses, latency was a positivefunction of the number of trials before thefirst correct response. A similar experimentreported by Peterson (1965) obtainedcomparable results. Numerals were associ-ated with a set of 10 consonant-consonant(CC) bigrams. Each S received 20 trialsbut only those Ss who had 10 successiveerrorless trials were evaluated. A signifi-cant decline in latency was found over thesequence of 10 successive errorless trials.

If response frequency and latency areboth indexes of associative strength, oneshould expect to find reasonable correla-tions between the two measures. Beck,Phillips, and Bloodsworth (1962) andJohnson (1964) presented Ss with non-sense syllables previously scaled for asso-ciation value by Archer (1960) and meas-ured the latency of the Ss' first freeassociations. The obtained correlationsranged from .19 to .70. Johnson thenconstructed paired-associate lists in whichthe response items were the consonant-vowel-consonant (CVC) trigrams forwhich first-response latencies had been ob-tained. The obtained correlation betweenlearning rate and Archer's a (association)value was .41 while the correlation betweenlearning rate and response latency was .36.The multiple correlation using both a andlatency as predictors of learning rate was.57. It would appear that response fre-quency and latency may measure tworelatively distinct factors, both of whichare related to associative strength. Such apossibility is supported by a study byWilliams (1962) in which paired-associate

RESPONSE LATENCY

lists were learned by the anticipationmethod. Knowledge of results was pre-sented for either 1 or 4 seconds. The longerknowledge-of-results exposure caused asignificant reduction in the number oftrials required to reach criterion but didnot produce a comparable reduction in re-sponse latencies; that, is, the response-fre-quency measure was altered but the latencymeasure was unchanged.

Shiffrin and Logan (1965) hold that la-tency is not a measure of associativestrength but a defining property of differentresponses; that is, a fast response and aslow response to the same stimulus are ac-tually two different responses, capable ofbeing differentially reinforced. To demon-strate their point, time authors conducted apaired-associate learning experiment inwhich minimum response latency was lim-ited to either .75 or 2.85 seconds. After Ssreached criterion, they were instructed toidentify as many pairs as possible in 2minutes, the next pair appearing imme-diately following S's response. Under theseconditions, the fast-practice group madesignificantly more responses than did theslow-practice group.

Latency Trends during the LearningProcess

In the past few years a number of stud-ies have investigated response latenciesthroughout the course of paired ,associatelearning. The impetus for conducting mostof these studies has come from questionsderived from attempts to construct mathe-matical models of the associative learningprocess. The data obtained have been exam-ined from new viewpoints which have re-vealed trends of considerable interest. Auseful technique employed for investigatinglatency over the course of paired-associatelearning is to align all item protocols on thebasis of the trial of last error (TLE). TheTLE for a particular item is defined as thelast trial on which an incorrect responsewas made prior to the point at which thatitem reached a criterion of n successiveerrorless trials in which n is some predeter-mined value. Item records are aligned sothat the TLE serves as a point of origin

3

from which all trials, both prior to andafter the TLE, are counted. When suchTLE-based pi otocols are averaged, the re-sult is analogous to a backward learningcurve. All responses falling on a particularTLE relative trial are representative of asimilar stage of learning in that each isequidistant, from the point at which thecriterion is attained.

Decrease in latency following the trialof last error. The manner in which Sim-ley (1933) evaluated his data made hisanalysis very similar to the recent analy-ses which took the TLE as a point of origin.Simley's finding that latency declinedrapidly after the TLE has been supportedin all of the recent studies mentionedbelow.

Latency trends prior to the TLE. Mill-ward (1964) used 12 two-digit numerals asstimuli in a 20-trial paired-associate learn-ing task in which the required responsewas to press one of two buttons, six stim-uli being associated with each button.The latencies on the first trial were rela-tively short, suggesting that Ss simplyguessed on that trial. Latencies increasedrapidly over the next few trials and thenremained relatively constant until theTLE. After the TLE there was a rapiddecrease in latency which did not appearto have reached asymptote after nine suc-cessive errorless trials. Suppes, Groen, andSchlag-Rey (1966) measured latencies ina paired-associate task in which the stim-ulus items were a set of 12 CVC nonsensetrigrams and the required response was topress one of three buttons. They found avery sharp rise in latency over the firstfew trials and concluded, as had Millward,that this was due to Ss making randomguesses on the initial trials. After this ini-tial rise, both correct and incorrect responselatencies were relatively constant up to thetrial before the TLE. Kintsch (1965) ranSs in a paired-associate learning task inwhich stimulus items were 12 nonsensesyllables with Glaze association values of.60, and the required response was tovocalize the numbers one or two. Kintsch'sresults differed from those of Millwardand Suppes et al. in that after the sharp


initial rise, latencies continued to increasesomewhat as a function of trial number upto, and including, the TLE.

Latencies of correct and incorrect re-sponses. Related to the question of response-latency stationarity prior to the TLE is thequestion of the relationship between cor-rect and incorrect response latencies. Suppeset al. (1966) specifically investigatedthis point and found no significant dif-ference between the response speeds of cor-rect and incorrect responses. Millward(1964) presented a plot of the latencies ofcorrect and incorrect responses prior to theTLE. Of the 18 trials at which comparisonsmay be made, incorrect response latencywas longer than correct response latency 14times but the variability of both curves wasso great that one cannot draw any firmconclusions.

Dimas and Zeaman (1963) ran a "min-iature experiment" in which Ss were givenone practice trial during which they couldexamine the stimulus-response (S-R) pairas long as they wished. They were thengiven two test trials during which only thestimulus item was presented and no knowl-edge of results was given. This was followedby a second practice trial and two moretest trials. The stimuli consisted of 12 CCCtrigrams of high association value and theresponse elements were the numbers 1-12.If the latencies of the response sequencesconsisting of four incorrect responses areexamined, it is noted that these responsespeeds are much slower than the latencies ofcorrect responses on the corresponding trials.While this experiment was not directly com-parable with the Suppes et al. experiment,it did suggest that correct and incorrect re-sponse latencies may not always be equiv-alent.

Latency as a function of learning rate.The purpose of Sim ley's (1933) study wasto demonstrate that the rate of learningbelow and above the "response threshold"is a function of the same factors. The rateof learning below threshold was measuredby the number of promptings requiredbefore S could provide the correct responsewhile learning rate above threshold washeld to be indicated by the decrease in re-sponse latency as a function of practice

following the point at which the correctresponse reached threshola, . Sim ley con-cluded that he had demonstrated his pointsince the data for each of the three Ss dis-cussed clearly showed that eesponse la-tency during overlearning was a positivefirmtion of the number of trials requiredto reach threshold, Items which werelearned slowly had higher latencies duringoverlearning. This result was even morepronounced when the recent innovation ofcomparing trials equally distant from theTLE was applied to the data. The moredifficult items, as defined by the number oftrials required to reach the TLE, had longerresponse latencies even when the itemswere equated as to the number of trials ofoverlearning.

This same effect was found in the datapresented by Suppes et al. (1966). Whenmean response times were correlated withitem difficulty (measured in terms of thenumber of trials required to reach theTLE) the Spearman rank-difference corre-lation coefficients for the two experimentalsessions discussed were .68 and .77. In con-trast, mean response time after the TLEwas unrelated to item difficulty in the datapresented by Kintsch (1965). Millward(1964) noted that in his data, latency wasin general a positive function of the num-ber of trials required to reach the TLEthroughout the learning task. He attributedthis to the fact that slow learners wouldrequire more trials to reach the TLE andpostulated that slow learners may havelonger response latencies than fast learn-ers. This hypothesis was not evaluated inthe article. Millward's results suggest thatlatency may be a positive function of itemdifficulty prior to the TLE when level oflearning is held constant by equating thenumber of subsequent trials required toreach the TLE. This suggestion was sup-ported by the Suppes et al. study. Whensubgroups were ranked according to itemdifficulty and the subgroups' mean laten-cies for correct responses prior to the TLEwere also ranked, the Spearman rank-dif-ference correlation coefficients for two ses-sions were .71 and .69.

Latency on the trial of last error. TheTLE itself may have interesting properties.

RESPON6E LATENCY 5

Suppes et al. (1966) found that, in general,response latency on the TLE was consider-ably longer than either the preceding errorlatencies or the preceding success latencies.The frequency with which the TLE la-tency was greater than either the imme-diately preceding or subsequent responselatencies was significantly greater thanchance. The authors pointed out that thesame phenomenon was to be found inKintsch's (1965) data and in an unpub-lished study by W. K. Estes and D. Horst.

First correct response latency and sub-sequent errors. The paired-associate learn-ing experiment reported by Williams(1962) used a list of 25 pairs of four-letter words. Her analysis of the latencydata was based on the trial of the firstcorrect response for a given item as op-posed to the TLE. Latency data were pre-sented for the trial of the first correct re-sponse and for the ninth trial thereafter.Item sequences were classified into twogroups on the basis of whether or not in-correct responses were made following thefirst correct response. Latency measuresfor item sequences containing incorrectresponses declined from 2.02 to 1.50 secondswhile the sequences of all correct items de-clined from 1.87 to 1.40 seconds. The de-cline in latency over the nine trials is inagreement with the data previously dis-cussed. The more interesting finding wasthat first-correct-response latencies werelonger if S made subsequent errors. This isin agreement with the conception of re-sponse latency as an index of associativestrength, If the associative strength wererelatively low at the time of the first cor-rect response, the response latencies wouldbe expected to be longer and it would alsobe expected that there would be a higherprobability of one or more subsequentincorrect responses. This explanation is con-tradicted, however, by the Eimas and Zea-man (1963) experiment. To test the hy-pothesis that correct-response latency isindicative of response strength, Eimas andZeaman classified all items answered cor-rectly on Test Trial 1 as either fast or slow.The slow items were not recalled particu-larly slowly on Test Trial 2 nor did theytend to be recalled incorrectly. Instead,

they showed a significantly greater incre-ment in speed on Test Trial 2 than did theitems classifed as fast on Test Trial 1.

Summary of latency trends. These re-cent studies of latency trends duringpaired-associate learning have raised ques-tions of interest in five different areas. First,what is the nature of the latency curve priorto the TLE? Is there always a sharp initialrise following the first, relatively fast re-sponses? After the occurrence of such arise, if any, do response latencies remainconstant or do they continue to increase upto the TLE? Secondly, there is the questionof whether or not response latency, on agiven pre-TLE trial, is a function of thecorrectness of the response. Third, thereare several questions concerning the rela-tionship of response latency to item diffi-culty and/or individual differences in learn-ing rate. Is latency during overlearning afunction of the number of trials requiredto reach a criterion of errorless responding?Does this relationship hold during the earlystages of learning, prior to the TLE? Ifthese effects do exist, are they attribut-able solely to item difficulty or are theyalso a function of individual differences inlearning rate? Fourth, is the occurrence ofmaximum response latency on the TLE areliable phenomenon? Finally, does the re-sponse latency of the first correct responseto a given item predict the probabilityof occurrence of subsequent errors on thatitem?

DEFINITION OF THE PIOBLENI

While the recent studies concerning theattributes of response latency during thecourse of learning have interesting implica-tions for latency as a basis for instructionaldecisions, it appears that the effects ofassociative strength on response latencyare relatively complex and dependent on anumber of task-related variables which areunrelated to learning per se. These task-related variables are related to performancerather than learning; during learning thesevariables act as parameters which influencethe magnitude of the effect of learningvariables. Hence, while the present experi-ment was concerned with systematic changesin latency during the course of learning,


equal consideration was given to parametricinvestigation of task variables which wereconsidered likely to influence the relation-ship of latency and learning. If the taskvariables which influence latency can beidentified and brought under experimentalcontrol, the investigator will be in a muchstronger position to evaluate the funct:ronsrelating latency to associative strength.Two such task variables were investigated:training method and information trans-mission. The anticipation and recall para-digms were compared and the effects of in-formation transmission were explored byvarying the amount of information re-quired for errorless responding.

Training MethodIn his investigation of response latencies

after the TLE, Peterson (1965), using ananticipation procedure, was unable to deter-mine the presence of any systematic trendsin the latency data prior to the TLE andsuggested that the variability appeared tobe so great as to render any attempt atanalysis futile. He postulated that at leasta partial cause of the high degree of varia-bility might be interference effects due toincorrect responses and suggested that theinterference might be alleviated by usinga recall paradigm. Faster learning undera recall paradigm, as contrasted with ananticipation paradigm, has been demon-strated in a series of experiments conductedby Battig (Battig & Brackett, 1961; Battig& Wu, 1965). In general it was found thatrecall procedures resulted in a consistentlyhigher percentage of correct responses pertrial and required fewer trials to reach cri-terion. Battig and Brackett suggested thatrecall procedures may be superior due totheir separation of the two behavioral proc-esses of producing a previously learnedcorrect response and learning new S-Rassociations. If the temporal contiguity ofthe associative- and response-productionprocesses does produce interference effectswhich retard the rate of learning, it wouldappear quite likely that these effects wouldalso tend to increase the magnitude andvariability of response latencies. If the useof a recall paradigm does reduce these inter-ference effects, the variability of response

latency during the early stages of learningshould be reduced.

In this study, the anticipation and recallparadigms were, therefore, contrasted andit was hypothesized that response latenciesprior to the TLE wou:d be shorter and lessvariable under the recall paradigm thanunder the anticipation paradigm. In addi-tion, the magnitude and variability of re-sponse latencies after the TLE were con-trasted under the two training paradigms.

Information Transmission

If one is interested in determining thevariables which influence latency in a ver-bal learning task, one of the most fertilerelated areas of investigation would ap-pear to be the study of choice or disjunc-tive reaction time (DRT). If viewed in thecontext of information theory, it bvomesevident that the behaviors required in apaired-associate learning task and in aDRT task are similar forms of informationprocessing. The major difference is thatthe S-R associations are well known to Sin the DRT task while they must belearned in the paired-associate task. Oncethe strength of the S-R associations hasrisen above "threshold," the two tasks be-come quite similar. Since the latency ofresponses which occur after the TLE maybe of special interest, it would appear tobe worthwhile to examine the relevance ofthe findings of the DRT studies to paired-associate learning.

For almost 70 years following the workof Merkel (1953), the generalization waswidely held that DRT was some positivefunction of the number of response alterna-tives. With the application of informationtheory to the study of reaction time, theproposition was refined to a quantitativestatement. Hyman (1953) varied theamount of information in the stimulus dis-play in a DRT task which had a one-to-one correspondence between stimuli andresponses and concluded that DRT is apositive linear function of the number ofbits of information in the stimulus display.A previous study by Hick (1952), however,had shown that this relationship did nothold if Ss were allowed to make errors.Subsequent studies by Bricker (1955) and

RESPONSE LATENCY 7

Rabbitt (1959) supported Merkel's originalcontention that it is the number of theresponse alternatives and the relative prob-ability of any particular response whichinfluences DRT. The currently acceptedfunction is DRT = a + b Ht, where Htis the amount of information, expressed inbits, transmitted by S per S-R event.

In this study, it seemed quite likely,therefore, that response latency after theTLE would be a positive function of theamount of information transmitted. SinceS's performance was essentially errorlessduring this period, the amount of informa..tion transmitted was equal to the log2 ofthe number of response alternatives. Therewas a possibility that response latencyprior to the TLE would also be related tothe amount of information transmitted. DuP,to the occurrence of errors, informationtransmission was not directly related to thenumber of response alternatives but it wasconsidered that the relationship might besufficient to have some influence. It was,therefore, hypothesized that response la-tencies prior to the TLE would increaseas some function of the number of responsealternatives.

Learning

The current experiment attempted toanswer a number of questions concerningsystematic changes in latency during thecourse of learning. The five points in-vestigated were the following:

1. What is the nature of the function re-lating latency to practice prior to the TLE?Millward (1964) and Suppes et al. (1966)found a sharp initial rise on the first fewtrials but latency then remained relativelyconstant until the TLE. Kintsch (1965),on the other hand, found a steady increasein latency up to, and including, the TLE.

2. Do correct-response latencies differfrom incorrect-response latencies priorto the TLE? Suppes et al. (1966) foundno difference between these latencies butdata from Eimas and Zeaman (1963) indi-cate that under some conditions, incorrectresponses are slower than correct re-sponses.

3. Is latency before and/or after theTLE a function of learning rate as defined

by the number of trials required to reachthe TLE? This effect was found after theTLE in the data presented by Sim ley(1933), Millward (1964), Kintsch (1965),and Suppes et al. (1966). It was foundprior to the TLE in the data obtained byMillward, Kintsch, and Suppes et al. Is thiseffect a between-S variable (slow learnersare slow responders) as suggested by Mill-ward or a within-S vas:able (difficult itemshave long response latencies) as suggestedby Sim ley's data?

4. Is response latency on the TLE reliablygreater than the latency of immediatelypreceding incorrect responses or imme-diately subsequent correct responses? Ab-normally long latencies on the TLE werefound in the data presented by Suppes et al.and Kintsch. Millward did not find thephenomenon.

5. Is the latency of the first correct re-sponse to an item longer if S makes subse-quent errors on that item? Williams (1962)reported that this was the case but her find-ings were contradicted to some extent bythe data reported by Eimas and Zeaman(1963).

METHOD

The experiment consisted of 16 replications of a2 X 3 factorial design. The two training methods,anticipation and recall, were contrasted and threelevels of information transmission were investi-gated by pairing two, four, or eight response al-ternatives with the members of an eight-itemstimulus list. Different Ss were used in each of thesix treatment grctips.

Subjects

The Ss were drawn from introductory psychol-ogy classes. A total of 103 Ss were run,. Of these,six failed to complete the experiment as a resultof equipment failures and one was rejected dueto an error on the part of E. The remaining 96 Sswere evenly divided between males and females.Eight males and eight females were assigned toeach of the six treatment groups. The Ss were notgiven a formal vision test but were required toread the stimuli aloud during the first trial of thewarm-up task. No S experienced difficulty in cor-rectly identifying the stimuli.

Materials

The stimuli used in the tasks were consonant-vowel-consonant (CVC) trigrams of 20-30% as-sociation value as determined by Archer (1960).The four trigrams VAH, VAQ, VEH, and VOZwere used in the warm-up task. The main task


HOME /POSITION/

Fic. 1. Full-scale diagram of response positionsfor Information Level 3.

PILOT LAMPSEFORF EDRACK

0/\

ES ?Oka U.Y.0

stimulus list was ZAB, ZAF, ZEF, ZEG, ZIK,ZIX, ZOK, and ZOX. These stimuli were selectedso as to be highly similar in terms of the composi-tion and placement of the letters. The responsesto be associated with the above stimuli were keypositions on a response panel. The Ss were re-quired to associate two, four, or eight key positionswith the eight stimuli. This corresponded to thetransmission of one, two, or three bits of informa-tion when all of an S's responses were correct.

Apparatus

The experiment was controlled by an on-linePDP-7 digital computer. The system presentedthe stimuli and knowledge of results, processed S'sresponses, maintained records of each S's progress,and timed response latencies and interitem inter-vals. Response latencies were measured, with anaccuracy of ±.001 seconds. All other timing wascontrolled to within --t.02 seconds. A completerecord of each S's stimuli, responses, and responselatencies was punched out on paper tape duringthe course of the experiment. These paper tapeswere fed back into the computer, and a limitedamount of data reduction was done while the nextS was being run. The output of this data-reductionprogram was a printed summary of each S's record.

The stimulus trigrams were presented on thescreen of a cathode-ray tube 3 inches high X 4inches wide. The trigram letters were all uppercase. Each letter was generated as a set of pointsselected in a 7 X 5 point matrix. The letters were1/2 inch high X 3/13 inch wide. The Ss were requiredto indicate their responses by pressing buttonsmounted on one of three response panels. Thepanels rested on a table in front of S and could bemoved about to maximize ease of responding. Two,four, or eight push-button microswitches weremounted at the top of each panel to correspondto the three levels of information transmission.The pushbuttons had 1/2 -inch diameter caps whichextended 1/2 inch above the panel. A force of 5ounces over a distauce of 1/2 inch was required toactuate the switches.

On all the panels, the switches were mounted

on a semicircular arc with a 2-inch radius. Thisallowed a center-to-center distance of 3/4 inchbetween the switch caps on the eight-key pane).The switch caps were not numbered or otherwiseidentified for Ss. The keys were assigned arbitrarynumerical identities for the purpose of programcontrol and response recording but Ss had noaccess to this information. The identities of thekeys are illustrated by the diagram of the eight-key panel in Figure 1. On the four-key responsepanel, Keys 1-4 were located in the same positionsas the correspondingly numbered keys on theeight-key panel. Only Keys 1 and 2 were presenton the two-key response panel. This arrangementassured that Keys 1 and 2 were in the same rela-tive position on all three response panels and thatKeys 3 and 4 were in the same positions on thetwo more complex panels. The correct responsekeys were indicated by illuminating a red pilotlamp next to the correct key. A white ring markedthe center of the arc on which the switches weremounted. The Ss were instructed to respond withthe index finger of their preferred hand and tokeep their finger on the ring between responses. Abuzzer was used to warn Ss at the start of a testtrial and to indicate the end of a task.

Randomization Procedures

The experiment consisted of 16 replications ofthe 3 X 2 factorial design. The treatment condi-tions were administered in random order withineach replication. When an S arrived at the labora-tory, he was assigned to the next available treat-ment condition in the replication currently as-signed to his sex. The assignment of responses tostimuli was varied randomly over the 16 replica-tions, and the order of item presentation duringeach trial was randomized. The randomness of theorders was constrained to the extent that the sameitem was never presented twice in succession bybeing the last item on one trial and the first itemon the next trial. All Ss responded to items in thesame order under both the anticipation and recallprocedures. The sequence of orders was repeatedonce every 30 trials.

Experimental Procedure

The Ss were run one at a time. After the in-structions were given, S was alone in the roombut could be observed through a one-way visionwindow. The computer control system was locatedin a separate room. After S was seated at the con-sole, he was given typical paired-associate learn-ing-task instructions which were varied as little aspossible between the anticipation and recall proce-dures. All the experimental conditions included awarm-up period. During this period, Ss weretrained by the same method (anticipation or re-call) and with the same response panel (two, four,or eight buttons) that they would use in the maintask. The training list consisted of only four itemswhich were always associated with Buttons 1 and2 regardless of the number of buttons on S's re-

RESPONSE LATENCY 9

sponse panel. The warm-up task was paced byallowing Ss only 3 seconds in which to respondfollowing the presentation of the stimulus. Failureto respond was counted as an error. Responsetimes were unlimited during the main task. Thisprocedure was intended to deter Ss from adoptinga strategy of rehearsing each item a number oftimes before proceeding to the next item. Re-sponse times were unlimited in the main task toprevent the occurrence of a truncated latency dis-tribution. As far as it was possible to determine,this strategy was successful.

In both the training task and the main task, Sswere drilled until they reached a criterion. Thiswas not the usual criterion for the entire list but acriterion for each of the items in the list. Re,.sponse records were maintained for the individualitems. When a series of six successive errorlesstrials was completed for a given item, the controlprogram noted that that item had reached cri-terion. The program continued to present the itemon subsequent trials but the occurrence of errorswas irrelevant to how long the drill was continued.Drill on the list was terminated a set number oftrials after the trial on which the last item reachedcriterion. This final drill period was 2 trials longin the training task and 10 trials in the main task.The trial prior to the first trial in the series of sixsuccessive errorless trials was designated the TLEfor that item. This schedule fo° determining thepoint at which drill was term:Ilated assured thatall items would have at least 16 trials after theTLE. The procedure did not assure the absenceof errors after the TLE, but Suppes et al. (1966)noted that the few incorrect responses which didoccur after items had reached a similar criterionappeared to be careless mistakes, the latencies ofwhich were consistent with the short latencies ofwell-learned responses.

The experimental conditions of the anticipationand recall procedures were equated as far as possi-ble. Under the anticipation procedure, the onset ofa .5 second auditory warning signal occurred 1.5seconds hefOre the beginning of a trial, where atrial was one presentation of the complete list.The stimulus item was presented and remained onthe screen until the occurrence of S's response.The pilot lamp next to the correct key was thenilluminated. The word on the screen and the lightstayed on together for 2 seconds. The screen wasthen erased and the light was turned off at thesame time. After a 1.5-second interitem interval,the next word was presented on the screen. Suc-cessive list presentations were separated by a4-second intertrial interval and the warningbuzzer always preceded the first item of each trialby 1.5 seconds.

The recall procedure incorporated successivetraining and testing phases within each trial. Dur-ing the training phase, each trigram-light pair waspresented together for 2 seconds. Following eachpresentation, the screen was erased and the lightwas turned off for a 1.5 second interitem interval.At the time of the end of the presentation of the

last trigram-light pair in the training phase of thetrial, the warning buzzer was turned on for .5second. The first trigram in the test phase was pre-sented on the screen 1.5 seconds after the onset ofthe buzzer. During the test phase, the stimulusitem remained on the screen until S responded.The screen was then erased and remained blankfor a 1.5-second interitem interval. No knowledgeof results was given during the test phase of thetrial. Following the completion of the test phase,the screen remained blank for a 4-second inter-trial interval.

The relative position of the keys on the responsepanel raised a problem of experimental control. Itwas considered a definite possibility that sincethey had relatively distinctive positions, the keyson the two-key response panel and the end keyson the four- and eight-key response panels mightbe subject to a serial position effect; that is,items for which these keys were the correct re-sponses might be learned more quickly than theitems for which the responses fell into the middleof the key array. In addition, it was possible thatthe perceptual-motor task of locating and pressinga key which was distinctive in that it was isolatedor at the end of the line of keys might be fasterthan an equally well-learned response to one ofthe keys in the middle of the key array. It was infact found that items for which the correct re-sponses were the end keys in the four- and eight-key tasks had shorter latencies and required fewertrials to reach the TLE as compared with keysin the middle of the array. This finding suggestedthat the two-key task might be qualitatively dif-ferent from the four- and eight-key tasks and thatwithin the latter tasks, the different responsesmight not be analogous. If this were the case, thespecific relationships under investigation mightbe a function of key nosition, and informationderived from an analysis which treated the datafrom all keys as homogeneous might be mislead-ing.

Hence, for each experimental hypothesis orquestion to which this problem was consideredrelevant, a preliminary analysis of variance wasconducted which included key position as an addi-tional within-S variable. The question of interestwas not whether latency varied as a function ofkey position but whether there was any significantinteraction between the key-position variable andthe variable of interest in that particular analysis.When such an interaction was found, two separateanalyses of the data were conducted: the firstcovering Keys 1 and 2 for all three informationlevels and the second covering Keys 3 and 4 foronly the two higher order information levels. Ifno interaction was found, only one analysis, cover-ing all key positions and all information levels,was conducted.

Since the failure to find a significant interactionin this preliminary analysis would result in ig-noring possible differences between key positions,an alpha error was considered to be of smallerconsequence than a beta error. For this reason, a


100

80

cc 60cc0U

cc 40wa.

20

09 17

..."'"'

.- . .- -- -- _...--,e.

---..... ---- -PERCENTCORRECT

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

/

LATENCY

//

//

25

2500

2000

1500 icLoi

2

1000 z

1

500

033

TRIAL NUMBER

FIG. 2. Correct response probability and response latency as a function of trial number.

high probability level of .10 was selected as a sig-nificance criterion. Latency scores were positivelyskewed, and transformations which yielded thedistribution most normal in appearance were usedin the analysis for a particular problem. Thesquare root, log, and reciprocal transformationswere evaluated. Almost all of the statistical testsemployed were analyses of variance of either astraight factorial design or a factorial with splitplot design. For the statistical tests which weredirectly concerned with the experimental hy-potheses and questions, as opposed to the key-position-variable tests, the significance criterionselected was .05. Latency data presented in tablesand graphs in the results section are raw scoremeans in units of milliseconds.

RESULTS

Changes in Latency over the Course ofLearning

Figure 2 illustrates the relationship be-tween correct response probability and re-sponse latency over the first 33 trials ofthe main task. All experimental treatmentconditions have been grouped together.Correct response probability increased asthe usual negatively accelerated functionand approached an asymptote at about98% correct. Over the same period, re-sponse latency rose slightly over the firstfew trials and then began a decline whichcontinued through Trial 20. The relia-bility of the data points represented by the

curves decreases after Trial 16 since thetraining period terminated for differentSs at different points after this trial.

Figure 3 illustrates these same relation-ships with the protocols aligned on thebasis of the TLE. Again, all experimentaltreatment conditions have been groupedtogether. Twenty-five percent of the data,or 192 responses, are represented on TrialTLE-8.3 This percentage increasesuntil all of the 768 responses are representedin the data points on Trial TLE+1 and alltrials thereafter. Correct response proba-bility increased from what would be ex-pected by chance on Trial TLE-11 to 54%correct on Trial TLE-1. Correct responseprobability was, of course, zero on the TLEand one on Trials TLE+1TLE+6 sincethe criterion for defining the TLE was sixsuccessive errorless trials. Error rate onTrials TLE+7TLE+16 was never greaterthan 4%. In general, response latencies re-mained fairly constant prior to the TLEand then decreased in a negatively acceler-ated curve which did not appear to havereached an asymptote at the point atwhich training was terminated, at TrialTLE+16.

Latency prior to the TLE. Millward

°Eight trials before the TLE.

100--

RESPONSE LATENCY 11

60-

, I /---1/ /--/

v N.,

/ // t( 1

PERCENT 1

CORRECT11

-16 8 TLETLE RELATIVE TRIAL NUMBER

FIG. 3. Correct response probability and response latency as a function of TLE relativetrial number.

LATENCY

2500

2000

--1500

,-1000

500

16

0

(1964) and Suppes et al. (1966) notedthat the latency of the first response to anitem was short, relative to the next fewtrials, These observations were made underan anticipation paradigm in which Ss weresimply guessing on the first trial, and it wassuggested that the short initial latenciesreflected this guessing behavior. In the cur-rent experiment, the first response latenciesof the anticipation Ss were 146 millisecondsshorter than the corresponding responses of

3000-

2500

ANTICIPATION

2000

/~N,

1500-1,/

1000

INFORMATION LEVEL 1" 2. 3 .

the recall-treatment Ss who had alreadybeen shown one correct pairing of the items.This would appear to be indicative of theguessing behavior of the anticipation Ss.

Figure 4 illustrates changes in latencyover the first eight trials of the experimentaltask. Millward and Suppes et al. also foundthat latencies increased sharply over thefirst few trials. This was not a consistentfinding in the current experiment. Onlythe two more difficult eight-response tasks

2500

2000

RECALL

\ //iNN"'`--Ns.

15001 N,7L-- ..... , ...

... ,,..,

1000

, I f 1 I 01 2 3 4 5 6 7 8 1 2 t 16 17

TRIAL NUMBER TRIAL NUMBER

FIG. 4. Response latency of the six treatment groups over the first eight trials of theexperimental task.

8

12 WILSON A, JUDD AND ROBERT GLASEit

TABLE 1TREATMENT GROUP MEANS FOR PRE-TLERESPONSE LATENCIES (IN MILLISECONDS)

FOR ALL KEY POSITIONS AS A FUNCTIONOl' PRACTICECOMPARISON OF THEFIRST AND SECOND HALVES op"rHE

Pitm-TLE TRIAL SERIES

Treatment ts First half Second half Increase

hiforznationLevel

1 12 1,488 1,673 1852 12 1,834 1,999 1653 12 2,050 2,522 472

MethodAnticipation 18 1,914 2,277 363Recall 18 1,668 1,853 185

Total 36 1,791 2,065 274

Note.Abbreviated: TLE = trial of last error.

(Information Level 3) could be said tohave demonstrated substantial increases inlatency over the early trials. Reference toFigure 3 illustrates that, in general, re-sponse latency prior to the TLE remainedrelatively constant. The curve may be some-what distorted, however, since as one movesfrom the left toward the TLE, an increas-ingly higher proportion of less difficultitems is encountered and, as will be dis-cussed later, the less difficult items tendedto have shorter latencies. The curve, there-fore, may not reflect the true relationshipbetween latency and practice prior to theTLE.

To investigate this possibility, the fol-lowing question was posed: Does responselatency change as an item receives practiceprior to the TLE? It appeared reasonable toexpect that a change in response latencywould be more likely to occur and would bemore meaningful for those items which hada substantial number of trials prior to theTLE. For each S, the item which had thegreatest number of trials prior to the TLEwas selected. In the case of ties, the tieditems were averaged together. Only itemswhich had six or more trials prior to theTLE were considered and Ss who had noitems with at least six pre-TLE trials wereexcluded from the analysis. Since, for theanticipation-treatment groups, the initialresponse to each item was made before Sshad observed the presentation of the correct

S-R pair, the first trials of these groups werenot included in the criterion-trial count norin the analysis. The series of trials for eachitem was split into a first and second half.Shown in Table 1 are the means determinedfor each half of the trial series. These meanscomprised the scores which were examinedas a within-S variable in the analysis. Theresults of the analysis indicated that re-sponse latencies during the two halves ofpractice did not differ significantly, Thetable of means indict 'Les that there was sometendency for the response latencies to in-crease, but this tendency was apparentlynot consistent. It may be concluded that, ingeneral, response latencies remained con-stant or increased slightly. There wasclearly no reduction in response latencyprior to the TLE.

Latency after the TLE. Figure 3 indi-cated that response latency declined fol-lowing the TLE. To evaluate this finding,post-TLE trial number was incorporatedas a within-S variable in an analysis ofvariance in which information-transmis-sire level and training method were be-tween-S variables. Group means are pre-sented in Table 2. Averaged across alltreatment conditions, response latency de-creased by 572 milliseconds from TrialTLE+1 to Trial TLE-1-16. This was asignificant reduction (F = 33.62, df =15/1,350, p < .001).

Summarizing the changes in latency overthe course of learning, it can be statedthat following a possible increase on thefirst few trials (the probability and mag-nitude of which was dependent on the taskcharacteristics) response latency remainedrelatively constant prior to the TLE and

TABLE 2RESPONSE LATENCY (IN MILLISECONDS) FOR ALL

KEY POSITIONS AFTER THE TRIAL OF LASTERROR (TLE) AS A FUNCTION OF TRIAL

NUMBER AND TRAINING METHOD

Method n OverallM

M forTrial

TLE+1

M forTrial

TLE-1-16De-

cline

Anticipation 48 1,629 1,968 1,433 535Recall 48 1,323 1,744 1,135 609

Total 96 1,476 1,856 1,284 572

RESPONSE LATENCY

then decreased substantially over the first16 trials following the TLE.

Intratrial Variability in Response LatencyVariability in response latency was in-

vestigated under each of the experimentalconditions and in the two major stages oflearning, pre- and post-TLE. The scoresobtained represent variability in responselatency within individual trials. For eachS, a variance was calculated for each trialprior to, and including, the TLE for thelast item to reach criterion. These intra-trial scores were then averaged together toobtain a mean pry -TLE variance for eachS. The same procedure was followed forthe first 16 trials after the TLE. A sum-mary of the group mean standard deviationsis presented in Table 3. The most strikingfeature of the scores in Table 3 is thatthe post-TLE standard deviations areconsistently smaller than the pre-TLEscores. It will also be noted that thevariability of the three information-levelgroups increased as a positive function ofthe amount of information which S was re-quired to transmit both prior to and afterthe TLE.

TABLE 3TREATMENT GROUP MEAN STANDARD DEVIATION

SCORES (IN MILLISECONDS) REPRESENTINGINTRATRI AL VARIABILITY IN RESPONSE

LATENCIES FOR ALL KEY P0141..tyc Ns

Treatment

Pre-TLEtrials

Post-TLEtrials

n SD n SD

Information Level 1 (A)Anticipation (B) 16 792 16 431Recall (C) 15 716 16 351

Information Level 2 (D)Anticipation 16 1,322 16 642Recall 16 1,078 16 440

Information Level 3 (E)Anticipation 16 1,339 16 927Recall 16 1,769 16 782

A X BC 31 755 32 391D X BC 32 1,200 32 541E X BC 32 1,554 32 854ADE X B 47 1,151 48 667ADE X C 48 1,198 48 524

Total 95 1,174 96 595

Note.Abbreviated: TLE = trial of last error.

13

TABLE 4EFFECTS OF INFORMATION TRANSMISSION LEVELS

AND PRESENTATION METHODS ON RESPONSELATENCIES FOR ALL HEY POSITIONS PRIOR

TO THE TRIAL OF LAST ERROR

Treatment n Latency M(in msec)

Information Level1 28 1,5002 28 1,9723 28 2,384

MethodAnticipation 42 1,926Recall 42 1,983

Total 84 1,954

It had been hypothesized that recallvariance would be less than the varianceof latencies obtained under the anticipa-tion paradigm. To test this hypothesis, aMann-Whitney U test (Siegel, 1956) wasconducted which compared the standarddeviation scores of the anticipation andrecall treatment Ss both prior to and afterthe TLE. Prior to the TLE, no differencewas detected between the two treatments.After the TLE, a significant proportionof the recall-treatment scores were smallerthan the anticipation scores (z = 2.90,p < .004). It may be concluded that priorto the TLE, intratrial response latencyvariability was not influenced by trainingmethod but after the TLE, variability wassignificantly less under the recall paradigm.

Response Latency as a Function ofTraining Method

Pre-TLE latency. All pre-TLE responselatencies (with the exception of initial re-sponses to anticipation items) were av-eraged together to obtain a mean pre-TIElatency for each S for whom pre-TLEdata were available. Mean pre-TLE la-tency was treated as a between-S variablein a factorial analysis of variance whichexamined the effects of both trainingmethod and information level. The resultsof this analysis are given in Table 4. Pre-TLE latencies did not differ significantly asa function of training method. The twomeans representing the anticipation andrecall treatments differed by less than 60milliseconds. It may be concluded that re-


2500

2000

Ut(;1 15002z>-U

ANTICIPATION

RECALL

uj 1000I.

500

0 Jill I 1 I I I I I

1 4 8 12 16

POST-TLE TRIAL NUMBER

Da. 5. Response latency of anticipation- andrecall-method Ss as a function of post-TLE trialnumber,

sponse latency prior to the TLE was in-dependent of the training method.

Post-TLE latency. Rather than group-ing all post-TLE data together, the effectsof post -TLE practice were evaluated bytreating the first 16 post-TLE trials as awithin-S variable in a factorial analysis ofvariance in which information level andmethod were between-S variables. Sincepost-TLE practice was experimentally con-trolled, data were available from all of the96 Ss.

As shown in Table 2, mean post-TLEresponse latencies, averaged across the 16trials, were 300 milliseconds faster underthe recall paradigm than under the anticipa-tion paradigm. This difference was statisti-cally significant (F = 22.23, df = 1/90,p < .001). In addition, there was a signifi-cant interaction between the method andtrials variables (F = 4.03, df = 15/1,350,p < .001). The effect of this interaction maybe seen in Figure 5. While post-TLE la-tency declined under both paradigms, themagnitude of the reduction was slightlygreater (74 milliseconds) for the recall paradigm than for the anticipation paradigm.It will be recalled that prior to the TLE,response latencies did not differ as a func-tion of training method. On the TLE itself,the anticipation and recall treatment groupmeans differed by only 10 milliseconds. It

was only after the TLE that recall laten-cies became substantially faster than thecomparable anticipation latencies, and themagnitude of this difference continued toincrease during overlearning.

Response Latency as a Function of Infor-mation Transmission Requirements

Pre-TLE latency. Disregarding the initialresponse to ant',3ipation items, the meanresponse latency prior to the TLE wascalculated for each S. A factorial analysisof variance was conducted in which infor-mation level and training method were be-tween-Ss variables. Group means are pre-sented in Table 4. Pre-TLE latency in-creased as a positive linear function of thenumber of bits of potential informationinvolved in the task, that is, the amount ofinformation which S was required to trans-mit for errorless responding. This relation-ship was statistically significant (F =13.48, df = 2/78, p < .001). Since pre-TLEresponding had a high error rate, Ss wereactually transmitting less than the potentialinformation in the task, but latency didappear to be a linear function of potentialinformation rather than transmitted infor-mation.

Since all key positions were included inthis analysis, the question arises as towhether this trend was not simply the re-sult of the higher order information levelshaving keys which resulted in longer la-tencies because of their position in the mid-dle of the keyboard array. That this was notthe case was demonstrated by the failureof the preliminary key-position analysis tofind a significant interaction between thekey-position and information-level var-iables. In addition, if the data from the endkey positions (Keys 1 and 2) are examined,the three latency means for the threeinformation levels are 1,528, 2,079, and2,323 milliseconds, respectively.

Post-TLE latency. The preliminaryanalysis which incorporated key positionas an additional variable found that afterthe TLE there was a significant interac-tion between the information-level andkey-position variables (F = 36.28, df =1/1,860, p < .001). For the two-bit, four-response task, response latency was in-

RESPONSE LATENCY

dependent of key position but for thethree-bit, eight-response task, the inner keypositions (Keys 3 and 4) had longer laten-cies than the outer key positions (Keys1 and 2). One analysis was, therefore,conducted for only Keys 1 and 2 acrossall three information levels and a secondanalysis treated only Keys 3 and 4 in thetwo higher order information levels. Infor-mation level and method were between-Svariables in the factorial analyses. Latencyscores for each S on each of the first 16post-TLE trials formed a within-S trialsvariable. Means of these data are presentedin Table 5, and curves representing changesin latency over trials are illustrated inFigures 6 and 7.

Response latency increased as a func-tion of information level for both sets ofkey positions, and in both cases the analy-ses indicate that these increases were sig-nificant (Key Positions 1 and 2: F =7.60, df = 2/90, p < .001; Key Positions3 and 4: F = 21.58, df = 1/60, p < .001).If the mean latencies for Keys 1 and 2are examined, the function relating latencyto the number of bits of information trans-mitted does not appear to be linear. Theincrease in latency as a result of movingfrom one to two bits of information is lessthan half the increase resulting from mov-ing from two to three bits.

TABLE 5TREATMENT GROUP MEANS FOR RESPONSELATENCY (IN MILLISECONDS) AFTER THE

TRIAL OF LAST ERROR (TLE) AS AFUNCTION OF INFORMATION

TRANSMISSION LEVELS

Informationlevel Overall M for

TrialTLE+l

M forTrial

TLE+16Decline

Key Positions 1 and 2

1

23

Total

32 1,244 1,496 1,057 43932 1,369 1,555 1,261 29432 1,653 2,426 1,302 1,12496 1,422 1,826 1,207 619


23

Total

323264

1,3771,9271,652

1,6862,8602,273

1,1811,6981,440

5051,162

833

15

INFORMATION LEVEL 12

4 8POST-TLE TRIA L. NUMBER

FIG. 6. Post-TLE response latencies of Informa-tion Levels 1,2, and 3. (Key Positions 1 and 2 only.)

The manner in which the function relat-ing latency to information level differed forthe two different sets of key positions maybe seen by contrasting the magnitude of theincrease in response latency from Infor-mation Level 2 to Level 3 in the two analy-ses. For the end key positions, Keys 1and 2, the effect of changing from four toeight response alternatives was to increasemean latency by 284 milliseconds. For Keys3 and 4, which were internal components ofthe key array, the comparable effect wasto increase the mean latency by 550 milli-seconds. The effect is also evident in thedistance separating the curves represent-ing Information Levels 2 and 3 in Figures6 and 7.

These figures also indicate that the de-cline in response latency over the firstfew post-TLE trials was much greater forInformation Level 3 than for the twolower . order levels. This difference did notpersist over the full 16 trials, however. TheInformation X Trials interaction was notsignificant in either of the two analyses.

Pre-TLE Latency of Correct and IncorrectResponses

Prior to the TLE, 45% of the responseswere correct. For each 5, the mean pre-TLE latencies of correct and incorrectresponses were determined for each item.The preliminary analysis incorporating

16

3000

2500

2000

1500

1000

WILSON A. JUDD AND ROBERT GLASER

INFORMATION LEVEL 2 PM111M WW1.

.4 8 12

POST-TLE TRIAL NUMBERib

Flo. 7. Post-TLE response latencies of Informa-tion Levels 2 and 3. (Key Positions 3 and 4 only.)

key position as a variable found a signifi-cant interaction between the key positionand correctness variables (F = 3.15,df = 1/96, p < .10). Therefore, one analy-sis treated only Key Positions 1 and 2, anda second analysis treated only Key Posi-tions 3 and 4. Means are presented inTable 6. Correct and incorrect responsesprior to the TLE were not significantly dif-ferent. This finding was independent of keyposition and the effects of the experimentalvariables of training method and informa-tion level.

Since the scores on which these analyseswere based were the means of all the pre-TLE responses, it was possible that whilethe pre-TLE averages might not differ, therelationship between correct and incorrectresponse latencies may have changed as afunction of practice. For example, correctresponses might have become faster as anitem approached the TLE while incorrectresponses became slower at approximatelythe same rate. If the correct and incorrectlatency curves crossed, this could causethe mean value of the curves to be approxi-mately the same. To evaluate this possi-bility, the latencies of correct and incorrectresponses were plotted over the pre-TLEpractice period. The average curve forall treatment groups and key positions isshown in Figure 8. It is evident from thisplot that the relationship between correct

and incorrect response latencies did notchange systematically over trials. Bothcurves tended to approximate a constantlatency of about 2 seconds. It may be con-cluded that prior to the TLE, correct re-sponse latencies did not differ from the la-t encies of incorrect re:-,ponses.

Response Latency as a Function of ItemDifficulty

On the basis of Simley's (1933) study, itwas expected that for a given 8, the moredifficult items would have longer latenciesthan that S's less difficult items. Responselatencies prior to and after the TLE wereinvestigated separately with respect tothis question. In general, it was found thatthose items which were associated with themiddle key positions had longer latenciesand also required a greater number of trialsto reach criterion than did the items asso-ciated with the end keys. While thisrelationship was in the direction anticipatedby the contention that the more difficult

TABLE 6TREATMENT GROUP MEANS FOR CORRECT AND

INCORRECT RESPONSE LATENCIES (INMILLISECONDS) PRIOR TO THE TRIAL

OF LAST ERROR

Treatment n

Response latencies

Correct Incor-rect

Incor-rect-

correct


Information Level1 22 1,587 1,625 382 22 2,051 2,296 2453 22 2,156 2,638 482MethodAnticipation 33 1,852 1,862 10Recall 33 2,012 2,511 499Total 66 1,932 2,187 255


Information Level2 24 1,812 2,134 3223 24 2,601 2,167 434MethodAnticipation 24 2,127 1,947 180Recall 24 2,287 2,354 67Total 48 2,207 2,151 56

RESPONSE LATENCY

items would have longer latencies, it islikely that the positive correlation wassimply an artifact. That is, items whichhad responses assigned to the end keyswere probably learned more quickly be-cause of the distinctive position of the keys,and this distinctive position may have alsomade the perceptual-motor task of locat-ing and pressing a key easier awl hencefaster for an end key. Any analy of therelationship between item diflicilty andresponse latency which treated all keypositions as analogous would be bi..sed byhaving a greater number of inner it( y posi-tion items in the difficult-item group and amajority of items associated with Keys 1and 2 in the less-difficult-item group. Twoseparate analyses were therefore conductedwhich treated Key Positions 1 and 2 andPositions 3 and 4 separately.

Item difficulty was measured by count-ing the number of trials required for anitem to reach Trial TLE+1. Thg mostdifficult item and the least difficult itemwere selected for each pair of key positionsfor each S on this basis. Scores were cal-culated for these items by computing themean latency for all responses prior to theTLE for the pre-TLE analysis and forthe first 16 trials after the TLE for thepost-TLE analysis. In the case of ties forthe most or least difficult item, the tieditems were averaged into a single score.Thus, for each set of analyses, each Sfor whom data were available was repre-sented by a pair of latency scores for hismost difficult item and his least difficultitem for the key positions relevant fo thatparticular analysis.

Pre-TLE latency. Latency data priorto the TLE and the number of trials re-quired to reach Trial TLE+1 are shownin Table 7. For the pre-TLE analysis, onlythose items were considered which had atleast one trial after the first presentationof the S-R pair and prior to the TLE. Rela-tively few Ss were available for this analy-sis since, in some cases, all of an S's itemshad the same number of pre-TLE trialsand more often, an S had an inadequateamount of pre-TLE data on the appropriatekey positions. The amount of data avail-able was not sufficient for a complete fac-

30001

2500

17

TLE RELATIVE TRIAL NUMBER

Ma. 8. Correctprior to the TLE.

and incorrect response latencies

torial test including all treatment condi-tions. Only 21 of the 48 recall-treatmentSs had pre-TLE data on Keys 1 and 2.Since the method factor did not appearto alter the relationship between the la-tencies of the most and least difficult items,anticipation and recall Ss were groupedtogether. Analysis of the data shown inTable 7 indicated that pre-TLE responselatency did not differ as a systematic func-tion of item difficulty for Key Positions 1and 2.

There were at least 8 Ss available in eachexperimental condition for the inside KeyPositions 3 and 4 and this was consideredto be sufficient for a complete factorialanalysis. Analysis of the data in Table 7indicated that the response latencies of thedifficult items were significantly longerthan the latencies of the responses to thoseitems which Ss found to be least difficult(F = 13.41, df = 1/28, p < .001). In addi-tion, this difference was more pronouncedfor the four-response, two-bit task than forthe eight-response, three-bit task. This re-sulted in a significant two-way interaction(F = 7.97, df = 1/28, p < .01).

Post-TLE latency. After the TLE, datawere available for all items over at least16 trials. The most and least difficult itemsfor each key position were selected foreach S by the procedure described above.


TABLE 7TREATMENT GROUP MEANS FOR PRE-TLE RESPONSE LATENCIES OF MOST AND LEAST DIFFICULT

ITEMS FOR EACH SUBJECT

TreatmentMost difficult item Least difficult item

Latencya Trials Latency4 Trials

Most-least difficult item

Latency°. I Trials


Information Level1 13 1,817 11.1 1,586 2.3 231 8.82 13 1,808 13.3 2,060 3.6 252 9.73 13 2,369 8.8 2,138 4.4 231 4.4Total 39 1,998 11.1 1,928 3.4 70 7.7

Kz.y Positions 3 and 4

Information Level2 16 2,046 10.4 1,516 3.4 530 7.03 16 2,116 10.6 2,034 6.4 82 4.2MethodAnticipation 16 2,150 12.9 1,716 5.6 434 7.3Recall 16 2,013 8.1 1,834 4.2 179 3.9Total 32 2,081 10.5 1,775 4.9 306 5.6

Latency Ms given in milliseconds.

A small number of Ss were lost becauseall their items had the same number ofpre-TLE trials. Data for Key Positions1 and 2 and Positions 3 and 4 are presentedin Table 8. In the case of both analyses, themost difIcult items had longer latenciesthan did the least difficult items. The dif-ferences were small, on the order of 100 to200 milliseconds, but significant in bothcases (Key Positions 1 and 2: F = 5.03,df = 1/72, p < .05; Key Positions 3 and 4:F = 6.59, df = 1/52, p < .05). In neithercase was there any significant interactionwith the information-level or method varia-bles.

Response Latency as a Function of SubjectLearning Rate

In the previous section, variation in re-sponse latency was investigated as awithin-S function of item difficulty. Theanalyses discussed in this section dealt withvariation in response latency as a functionof individual differences in S learning rate.Responses prior to and after the TLE weretreated separately. The measure of S learn-ing rate employed was the total numberof item presentations across all items andall key positions prior to and including the

TLE. From each treatment group of 16Ss, the 4 Ss who had the highest suchscores were classified as slow learners andthe 4 Ss who had the lowest scores wereclassified as fast learners. In the pre-TLEanalysis, 3 of the fast learners had no dataprior to the TLE and had to be replacedwith Ss who were slightly slower learners.

Pre-TLE latency. A single pre-TLEresponse-latency score was calculated foreach key position for each S. This scorewas the mean of all response latenciesafter the first presentation of the S-R pairand prior to the TLE. It was found thatkey position was not a significant variablein this context, and a single analysis in-cluding all key positions was conducted.The mean response latencies and the meannumber of pre-TLE trials per item for eachtreatment group are presented in Table 9.Analysis of variance indicated that the slowand fast learners differed only slightly intheir pre-TLE response latencies. Neitherthe learning-rate variable nor any inter-action between learning rate and the infor-mation-level or method variables was sig-nificant. It may be concluded that slowand fast learners did not differ in responselatency prior to the TLE.

RESPONSE LATENCY 19

TABLE 8TREATMENT GROUP MEANS FOR PAST -TLE RESPONSE LATENCIES OF MOST AND LEAST DIFFICULT

ITEMS FOR EACH SUBJECT11,.....

Treatment nMost difficult item

Latency* Trials

Least difficult item

Latency* I Trials

Most-least difficult itn

Latency* I Trials


Information Level1 26 1,251 10.0 1,162 0.4 892 26 1,420 10.0 1,356 2.3 643 26 1,732 7.7 1,550 3.3 182

MethodAnticipation 39 1,611 13.1 1,455 3.5 156Recall 39 1,325 5.4 1,258 0.5 67

Total 78 1,464 9.2 1,356 2.0 108

9.67.74.4

9.64.97.2

Key Positions 3 and 4.11,i...1Information Level

2 28 1,429 10.8 1,266 2.4 1633 28 1,963 10.9 1,817 6.1 146

MethodAnticipation 28 1,890 13.8 1,656 5.7 234Recall 28 1,502 7.9 1,428 2.8 74

Total 56 1,696 10.8 1,542 4.3 154

8.44.8

8.15.16.5

Latency Ms given in milliseconds.

Post-TLE latency. A single post-TLEresponse-latency score was calculated foreach key position for each S. This scorewas the mean of the first 16 post-TLE re-sponses. The preliminary key-position anal-ysis of variance indicated a significantinteraction between key position and thelearning-rate variable (F = 3.37, df =1/24, p < .10). Therefore, two separateanalyses for Key Positions 1 and 2 andPositions 3 and 4 were conducted. The

TABLE 9TREATMENT GROUP MEANS FOR PRE-TLERESPONSE LATENCIES OF SLOWEST AND

FASTEST LEARNERS FOR ALLKEY POSITIONS

Treatment

Slowlearners

Fastlearners

Slow-fastlearners

La-tency Trials La-

tency Trials La-tency Trials

Information Level1,500 7.1 1,525 1.1 -25 6.0

2 1,878 11.4 2.3 51 9.12,318 11.7 2,004 4.1 -3286 7.6

Me3thodAnticipation 1,866 12.9 2,000 3.5 -134 9.4Recall 1,932 7.3 1,771 1.5 161 5.8

Total 1,899 10.1 1,885 2.5 14 7.6

mean response latencies and the mean num-ber of pre-TLE trials per item for eachtreatment group are presented in Table 10.The treatment-group latency means inthese tables indicate that for both sets ofkey positions, slow learners had slightlylonger response latencies during the post-TLE trials but this difference was signifi-cant only for the outside key positions,Keys 1 and 2 (F = 6.79, df = 1/36, p <.05). There were no significant interactionsbetween the variables of S learning rateand information level or method in eitheranalysis.

Response Latency in the Area of the TLEResponse latency in the area of the TLE

itself is of special interest. Suppes et al.(1966) found response latencies on theTLE to be consistently longer than the la-tencies of the immediately preceding andfollowing trials. In addition, preceding sec-tions have shown that while latencies priorto the TLE remained constant and inde-pendent of practice, post-TLE latenciesdecreased rapidly as a function of practice.


TABLE 10TREATMENT GROUP MEANS FOR POST -TLE

RESPONSE LATENCIES OF SLOWESTAND FASTEST LEARNERS

Slowlearners

Fastlearners

Slow-fastlearners

TreatmentLa-tency Trials La-

tency Trials La-tency Trials


Information Level1 1,245 7.1 1,049 1.0 196 6.12 1,547 11.4 1,227 2.0 320 9.43 1,732 11.7 1,887 3.8 -135 7.9MethodAnticipation 1,673 12.9 1,486 3.5 187 9.4Recall 1,344 7.3 1,276 1.0 68 6.3

Total 1,508 10.1 1,381 2.3 127 7.8


Information Level2 1,521 11.4 1,264 2.0 257 9.43 1,971 11.7 1,883 3.8 88 7.9

MethodAnticipation 1,934 14.3 1,805 4.4 129 9.9Recall 1,558 8.8 1,342 1.4 216 7.4

Total 1,746 11.6 1,574 2.9 172 8.7

The TLE thus appears to be a point ofsome importance in the systematic changesin latency over learning.

Changes in latency from Trial T LE -1to the ME . Only those items were con-sidered which had an incorrect responseon Trial TLE-1 which also were not aninitial anticipation response. All such itemswhich were available were averaged to-gether to yield two mean scores for each S,one representing the latency of an error re-sponse on trial TLE-1 and the other repre-senting the TLE latency. These scores weretreated as a within-3 variable in a factorialanalysis of variance in which informationlevel and method were between-S variables.The group-mean data of this analysis arepresented in Table P. The increase in in-correct response latency from Trial TLE-1to the TLE, averaged over all experimentalconditions was only 102 milliseconds. Thisdifference was not significant nor werethere any significant interactions with theexperimental variables. It may be con-cluded that the latency of the TLE re-sponse was not substantially greater thanthe latency of immediately preceding in-correct responses.

Changes in latency from the TLE toTrial T LE +1 . This analysis investigated

the question of whether or not there was asignificant drop in latency from the TLEto the first trial past the TLE. A pair ofscores representing Trials TLE and TLh'+1were computed for each S who had atleast one available protocol by taking themean of all available response pairs, Thesescores were treated as a within-S variablein a factorial analysis of variance in whichinfer nation level and method were be-tween-S variables. The group-mean dataof this analysis are presented in Table 12.Responses on Trial TLE+1 were, on theaverage, 251 milliseconds faster than re-sponses on the TLE. This was a significantreduction (F = 13.38, df = 1/84, p < .001).The magnitude of the reduction in latencywas a significant function of the informa-tion levels with the middle, two-bit taskdemonstrating the greatest change (F =4.38, df = 2/84, p < .05). There was, inaddition, a significant three-way interac-tion between information level, method,and trial relative to the TLE (F = 5.25,df = 2/84, p < .01). This may be attributedto the finding that while the decline wasgreater for the anticipation than for therecall treatments on the one-. and three-bit tasks, the two-bit, recall-treatmentgroup demonstrated a reduction in latencyof 925 milliseconds that was much greaterthan that of the comparable anticipationgroup. Examination of the scores of theindividual Ss in two-bit recall conditionindicated that this was a consistent trend.Of the 15 Ss, 13 demonstrated a reduction in

TABLE liTREATMENT GROUP MEANS FOR COMPARISONOF INCORRECT RESPONSE LATENCIES ON

TRILLS TLE-1 AND TLE FOR ALLKEY POSITIONS

Treatment TrialTLE-1

TrialTLE

TLE-TLE-1

Information Level1 16 1,605 1,770 1652 16 2,323 2,525 2023 16 2,516 2,455 -61

MethodAnticipation 24 2,163 2,214 51Recall 24 2,134 2,287 153

Total 48 2,148 2,250 102

Note.-Abbreviated: TLE= trial of last error.

RESPONSE LATENCY

latency from the TLE to the next trialand of these 13, 5 had a reduction in la-tency that was greater than 1 second. Itmay be concluded that while, in general,responses were faster on Trial TLE+1 thanon the TLE, this effect was negligible forthe eight-response, three-bit task. Further-more, the particular combination of con-ditions present in the two-bit recall taskresulted in an unusually large decrementin response latency.

Latency trends in the area of the TLE.Following the completion of the previoustwo analyses it was felt that a more ade-quate description of changes in latency inthe area of the TLE was needed. All itemprotocols which had at least three pre-TLEtrials were selected. Approximately 50% ofthe items met this criterion. Mean correctand incorrect response latencies for TrialsTLE-3-TLE+3 were calculated for eachS who had at least one item protocolwhich met the criterion. The scores of all

21

TABLE 12TREATMENT GROUP MEANS FOR COMPARISON OF

RESPONSE LATENCIES ON TRIALS TLE ANDTLE+1 FOR ALL KEY POSITIONS

Treatment TrialTLE

TrialTLE+1

TLE-TLE+1

Information Level 1 (A)Anticipation (B) 15 1,914 1,852 262Recall (C) 15 1,547 1,470 77

Information Level 2 (D)Anticipation 15 1,890 1,808 282Recall 15 2,508 1,580 925

Information Level 3 (E)Anticipation 15 2,877 2,820 51Recall 15 2,224 2,321 -97

A X BC 30 1,730 1,581 189D X BC 30 2,199 1,594 605E X BC 30 2,550 2,573 -23ADE X B 45 2,227 2,029 198ADE X C 45 2,093 1,790 303Total 90 2,180 1,909 251

Note. -- Abbreviated; TLE = trial of last error.

the available Ss were then averaged to-gether to obtain treatment-group means.The resulting data for each treatment groupare presented in Table 13. Although nostatistical tests were made on these data,

TABLE 13CORRECT AND INCORRECT RESPONSE LATENCIES IN THE AREA OF THE TRIAL OF LAST ERROR FOR

ALL KEY POSITIONS

Treatment nCorrect -Incor-

rectTLE-3 TLE-2 TLE-1 TLE TLE+1 TLE+2 TLE+3

Information Level 1 (A)Anticipation (B) 14 C 1,813 2,269 1,655 1,439 1,569 1,414

I 1,806 1,334 2,092 1,990Recall (C) 14 C 1,542 1,309 1,616 1,565 1,405 1,260Information Level 2 (D) 1,723 1,955 1,625 1,621Anticipation 16 C 1,683 1,941 1,511 1,660 1,883 1,521

I 1,521 2,127 2,013 1,907Recall 13 C 2,685 2,047 2,297 1,755 1,590 1,359

I 1,843 2,736 3,172 2,425Information Level 3 (E)

Anticipation 16 C 2,672 2,923 2,641 2,692 2,337 2,117I 2,893 2,379 2,642 2,897

Recall 15 C 2,435 2,249 1,882 1,916 1,839 1,721I 2,193 2,031 2,485 2,120

A X BC 28 C 1,689 1,852 1,636 1,502 1,487 1,337I 1,769 1,595 1,920 1,805

D X BC 29 C 2,107 1,987 1,848 1,703 1,752 1,448I 1,642 2,371 2,510 2,144

E X BC 31 C 2,561 2,608 2,287 2,317 2,096 1,925I 2,579 2,224 2,567 2,521

ADE X B 46 C 2,081 2,417 1,954 1,952 1,945 1,696I 2,126 2,006 2,260 2,277

ADE X C 42 C 2,239 1,913 1,929 1,749 1,617 1,455I 1,963 2,246 2,491 2,051

Total 88 C 2,152 2,191 1,943 1,855 1,789 1,581I 2,057 2,108 2,359 2,169


there do appear to be several trends whichmay be promising for future investigation.First, it will be noted that for this particularselection of data, the latency of the TLEresponse was, in general, shorter than thelatency of an immediately preceding in-correct response. This trend held for five ofthe six treatment groups but was morepronounced for the recall training methodconditions. The most accurate descriptivestatement that can be wade based onthis data is that incorrect response laten-cies tended to reach a peak during thelast few trials prior to and including theTLE. There did not appear to be any re-liable sharp division points.

Secondly, it will be recalled that it waspreviously demonstrated that, in general,the latency of correct and incorrect re-sponses prior to the TLE did not differ.For the sample of data shown in Table13, there were no systematic differencesbetween correct and incorrect response la-tencies on Trials TLE-3 and TLE-2, buton Trial TLE-1, correct response latencieswere consistently shorter than the latenciesof the corresponding incorrect responses.While the two-bit recall and three-bitanticipation tasks had only negligibledifferences between correct and incorrectrespoase latencies, the other four task con-ditions demonstrated substantial differences.It may be the case that response latenciesdo become indicative of the correctness ofthe response as the item approaches thepoint at which it is finally learned.

Finally, the data represented in Table 13suggest that the decline in correct responselatency evident after the TLE may beginprior to the TLE. Under each of the threeanticipation training method conditions,mean correct response latency declinedfrom Trial TLE-2 to Trial TLE-1. Underthe recall paradigm, only the three-bit taskdemonstrated a reduction in latency acrossthese trials but for all three informationlevels, latency on Trial TLE-2 was lessthan the latency of the preceding trial.

Since these trends were postulated aposteriori with reference to a particularsample of data, the use of statistical teststo evaluate the frequency of their occur-rence would not appear to be justified. They

are presented only as an attempt to clarifyfor future study the nature of latency be-havior during an important phase of thelearning process.

First Correct Response Latency as a Func-tion of Subsequent Errors

This final analysis reverses past pro-cedure by examining latency on the trialof the first correct response rather than ontrials relative to the TLE. The latency ofthe first correct response was determinedfor each item for each S. All such firstcorrect responses were then divided intotwo groups on the basis of whether or notS made any errors after the first correctresponse and prior to reaching the criterionof six successive errorless trials. To re-main consistent with the terminology sug-gested by Williams (1962), those items onwhich subsequent errors did occur weretermed break items. Items for which therewas no subsequent error were termed non-break items. For each S, a pair of scoreswas computed which consisted of the meanlatencies of the first correct responses ofall the break and nonbreak items for thatS. These scores were treated as a within-Svariable in a factorial analysis of variancein which information level and trainingmethod were between-S variables. The Sswhose data did not include both break andnonbreak items were excluded from con-sideration. At least 12 Ss were available ineach treatment group. A summary of thegroup-mean data is presented in Table 14.

First correct responses to break itemshad longer latencies than the correspond-ing nonbreak items for five of the sixtreatment groups. This difference was sig-nificant (F = 19.44, df = 1/66, p < .001).The magnitude of the difference increasedas a positive function of information levelfor the anticipation-procedure groups butwas a nonlinear, U-shaped function ofinformation level for the recall-proceduregroups. These varying relationships, shownin Table 14, resulted in a significantthree-way interaction (F = 3.29, df =2/66, p < .u5). To relate the findings of thisanalysis to the previously discussed analysestaking their point of origin from the TLE,it may be pointed out that the first correct

RESPONSE LATENCY

TABLE 14TREATMENT GROUP MEANS FOR LATENCIES OF

FIRST CORRECT RESPONSES FOR ALL KEYPOSITIONS DIVIDED ON THE BASIS OF THEOCCURRENCE OF SUBSEQUENT ERRORS

Treatment Breakitems

Non-breakitems

Break-non-

break

Information Level 1 (A)Anticipation (B) 12 1.746 1,647 99Recall (C) 12 1,544 1,248 296

Information Level 2 (D)Anticipation 12 1,436 1,285 151Recall 12 2,707 1,589 1,118

Information Level 3 (E)Anticipation 12 2,774 2,367 407Recall 12 1,846 2,119 273

A X BC 24 1,645 1,447 198D X BC 24 2,071 1,437 634E X BC 24 2,310 2,243 67ADE X B 36 1,085 1,766 219ADE X C 36 2,032 1,652 380

Total 72 2,009 1,709 300

responses of break items occurred prior tothe TLE while the first correct responses ofnonbreak items occurred on TrialTLE+1. The results of this analysis are,therefore, consistent with the previousfindings that correct responses remain rela-tively slow on the trials prior to the TLEand are then considerably faster on TrialTLE+1. In addition, however, this analysisdoes suggest that when a nonbreak itemwas learned, as indicated by the factthat no more errors were made on that item,the latency of the response was imme-diately shortened. Although all the pre-ceding responses had been in error, S'svery first production of the correct responsewas substantially faster than the preced-

responses.

CONCLUSIONS AND DISCUSSION

Training Method

It was expected that the recall paradigmwould result in shorter and less variablelatencies prior to the TLE than would theanticipation method, following through thesuggestion made in previous work byPeterson (1965). The present investigatorsfound, however, that pre-TLE responselatencies did not differ between the twoparadigms in either duration or variability.The recall paradigm did result in fasterlearning and a lower postcriterion errorratea finding which is in agreement withprevious research (Battig & Brackett,

23

1961; Battig Sr Wu, 1965). The recall para-digm did, therefore, have some instruc-tional advantage and this finding in itselfmight lead one to expect that pre-TLElatencies would be shorter under the recallparadigm. If the recall training methodwas more efficient because interference ef-fects were reduced under this paradigm, theresponse-latency measures must have notbeen sensitive to the interference effects.

After the TLE, the recall paradigm re-sulted in shorter and less variable latenciesthan did the anticipation paradigm. Thereduced variability was probably not mean-ingful in itself, but a function of the shortermean latency for recall. Averaged overall post-TLE latencies, the ratio of thestandard deviation to the mean was .40 forboth the anticipation and recall paradigms.The smaller magnitude of the recall-methodlatencies could have been the result of oneor both of two separate factors. First, re-sponses may have been effectively pacedat a higher rate in the recall paradigm thanin the anticipation paradigm. The minimumpossible elapsed time between recall re-sponses was 1.5 seconds. Due to the knowl-edge-of-results presentation, the minimumtime between anticipation responses was3.5 seconds. The different rates at whichthe items were presented may be analogotrto the situation discussed by Williams(1962). Williams' Ss learned word pairsby the anticipation method, and knowledgeof results was exposed for either 1 or4 seconds. The longer exposure timeresulted in a significant reduction in thenumber of trials required to reach cri-terion but did not produce the expectedreduction in latencies. Williams sug-gested that the slower presentation rateresulting from the longer knowledge-of-results exposure may have specifically in-creased latencies. The anticipation-para-digm latencies may have been increasedfor the same reason in the post-TLEperiod of the current experiment but ifthis were the case, it might be expectedthat the effect would have also been pres-ent in the data prior to the TLE. If, onthe other hand, response latencies afterthe TLE are indicative of the degree ofoverlearning and if the recall paradigm is a


more efficient training method during over-learning as well as during early learning,the shorter post-TLE latencies may reflectthe higher response strengths of the itemslearned under the recall paradigm. It willbe recalled that the difference between thepost-TLE recall and anticipation laten-cies increased as a function of practice.This would be expected if the anticipation-recall difference were due to suprathresholdresponse strength increasing at a fasterrate under the recall paradigm. However,the same effect might be expected if the dif-ference were due to a discrepancy in therate at which the two tasks were paced.A comparison of the two methods in whichinterresponse interval was held constantacross the two paradigms should differen-tiate between hypotheses.

In summary, the recall paradigm wasthe more efficient training procedure interms of response probability. Responselatency prior to the TLE was independentof the training method. As practice wascontinued after the TLE, recall-treatmentlatencies became increasingly shorter thanthe corresponding anticipation responselatencies. If post-TLE latencies are in-deed indicative of suprathreshold responsestrength, recall may also be the more effi-cient training procedure during overlearn-ing.

Information Transmission RequirementsAs was anticipated from the reaction-

time literature, response latency increasedas the number of response alternatives in-creased. These results support the findingsof Bricker (1955) and Rabbitt (1959) thatvariation in the number of response alterna-tives rather than the number of stimuli issufficient to alter the information charac-teristics of the task. In the current experi-ment, all three information-level groupshad eight-item stimulus lists. Only thenumber of response alternatives differed.The latencies of the different information-level groups became ordered on the basisof the number of response alternatives verysoon after the start of the main task. Pre-TLE response latencies, averaged acrosstraining methods, closely approximated alinear function of the amount of potential

information in the task, that is, a log2 func-tion of the number of response alternatives.Since Ss were making errors during thisperiod, however, the amount of informa-tion actually being transmitted was muchsmaller. The average amount of informationtransmitted per response in the two-, four-,and eight-response-alternative tasks wasapproximately .01, .13, and .36 bits, re-spectively. It would be expected from thei'eaction-time literature that response la-tencies would be a linear function of theamount of information actually beingtransmitted but the latency data appearedto be more a linear function of the amountof potential information in the task.

On the TLE itself, latency was approx-imately a linear function of the potentialinformation in each task. There was aslight tendency toward concavity but thisshould probably be discounted in view ofthe subsequent results. Following the TLE,the amount of transmitted information ap-proximated the potential information ineach task. Immediately after the TLE, thefunction became positively accelerated inthat the eight-response task had excessivelylong latencies. As post-TLE practice con-tinued, however, the function became morelinear. This was true for both the outsidekey positions (Keys 1 and 2 alone) and forall key positions taken together. Responselatencies on the eight-key task decreasedmore quickly for the outside keys and forthese keys, the function was essentiallylinear for the post-TLE Trials 8-16. Thereis no reason to expect that the functionwould not have become linear for all keysif practice had been continued.

Several summarizing conclusions may bedrawn. Response latency increased as thenumber of response alternatives was in-creased. The S's latency behavior reflectedthe different numbers of response alter-natives very soon after the beginning ofthe task. During the early stages of learn-ing, prior to and including the TLE, la-tency was a linear function of the potentialinformation in the task. Immediately fol-lowing the TLE, when the transmitted in-formation approximated the potential in-formation, latency became a positivelyaccelerated function of the amount of in-

RESPONSE LATENCY

formation in the task but as practice con-tinued, the function became more linear.

Response Latency Prior to the TLEAn S's first response to an item was

slightly faster under the anticipation para-digm than under the recall paradigm. Thisprobably reflected the fact that the antic-ipation Ss were simply guessing on thefirst trial while the recall Ss, having onceviewed the correct pairs, had some basisfor making a decision. There was a tend-ency for the latency of difficult items toincrease over practice prior to the TLE.This increase was more pronounced underthe anticipation paradigm and for the eight-response tasks but the trend was not sig-nificant for any of the treatment groups.There was no evidence of a reduction inlatency as a function of pre-TLE practiceuntil the last one or two trials prior to theTLE. During the period in which correctresponse probability increased from 30% to54%, response latencies remained essen-tially constant. It would appear that la-tency was not indicative of responsestrength during the pre-TLE period.

Correct response probability averagedacross all treatment groups would be ex-pected to be .25 by chance alone. The ob-tained probability, averaged over the en-tire pre-TLE period, was .45. Therefore,roughly half of the correct responses ob-served during this period could have beendue to factors other than chance. If thesefactors had any effects on the latency ofcorrect responses, they were effectivelymasked by chance responding. In the eight-response tasks, the observed correct re-sponse probability was .40 while the prob-ability expected by chance alone wouldbe only .125. Thus, approximately 70% ofthe correct responses could have been dueto factors other than chance and it wouldbe expected that this higher proportion ofresponses might override any masking ef-fects due to chance responding. On theeight-response task, the relationship of cor-rect to incorrect response latency was afunction of key position. For items assignedto outside keys, correct responses were 482milliseconds faster than incorrect responses.For items assigned to inside keys, correct

25

responses were 434 milliseconds slowerthan incorrect responses. A portion of thisdiscrepancy may be attributed to differencesin average response speed between thekeys. In general, responses to outside keyswere faster than responses to inside keys.For items assigned to outside keys, all thecorrect responses would be to the faster,outside keys but some portion of the incor-rect responses would be to the slower, in-side keys. The situation would be reversedfor the items assigned to the inside keys.When allowance is made for this influence,there would not appear to be any strongsystematic difference between correct andincorrect responses in even the eight-re-sponse tasks. It seems most parsimonious toconclude that the influence of the learningfactors evident in correet response prob-ability prior to the TLE could not be de-tected on the basis of the latencies of cor-rect and incorrect responses.

The latency data were quite variableduring the pre-TLE period and the hypoth-esized stabilizing effects of the recall-paradigm training method did not occur.It may be the case that the variabilityin response latency is inherent in the re-sponse-production process in the earlystages of learning and is not attributable tothe 'postulated interference effects of theanticipation paradigm. As was discussedin the previous 'section, one of the few fac-tors which did have a significant influenceon latency during the pre-TLE period wasthe number of response alternatives.

The effects of variation in item difficultyare rather difficult to assess during the pre-TLE period. On the outside keys, there wasa considerable range of difficulty as meas-ured by the number of trials required toreach the TLE but the corresponding dif-ference in latency was negligible. There wasa comparable range of difficulty for theinside keys and in that case the corre-sponding difference in latency was highlysignificant with the most difficult itemsbeing 306 milliseconds slower than theeasiest items. There was also a significantinteraction with information level forthese keys, and the Item Latency X Diffi-culty relationship was almost solely due tothe data from the four-response tasks. There


is no obvious reason why this particularsituation should have resulted in a signifi-cant relationship. The latency differencesbetween hard and easy items, ambiguousas they were, do srggest that there wassome increase in latency if an item con-tinued over a considerable number of trialswithout being learned. It may be that Sswere able to recognize the stimulus com-ponent as being a member of a difficultitem before they were able to supply thecorrect response member.

Response Latency in the Area of the TLEIt was anticipated that the TLE would

be a point at which sharp, systematicchanges in latency measurements wouldoccur. This anticipation was supported bythe finding in the current experiment,as well as in previous studies, that re-sponse latencies remained fairly constantprior to the TLE and then began to de-crease immediately following the TLE.The results of the current experiment sug-gest that the change in latency associatedwith the TLE may not be as discrete aswas suggested by the findings of Suppeset al. It was found that latency on the TLEwas not significantly greater than the la-tency of the immediately preceding in-correct responses. For items which had anumber of trials prior to the TLE, theredid appear to be some increase in incorrectresponse 'latency as items approached theTLE but the maximum latency tended tooccur one or two trials prior to the TLE asoften as it occurred on the TLE itself. Thiswas true for the two- and four-responseanticipation tasks, the tasks which weremost similar to the conditions employedby Suppes et al., and it is not apparentwhy the results differed between the twoexperiments.

There was a large and significant dropin latency from the TLE to Trial TLE+1although the significant interaction be-tween information level, method, and trialnumber is difficult to explain. Since re-sponse latencies were, in general, constantprior to the TLE, the significant reductionin latency immediately following the TLEsuggests that the post-TLE decline beganon Trial TLE+1. For some items, the de-

cline did not begin until a few trials afterthe TLE but since correct responses couldhave occurred by chance alone after theTLE, correct response probability may nothave reached asymptote for these items untilone or two trials after the TLE. If latenciesdid begin to drop only after correct responseprobability reached asymptote, the impli-cation is that probability and latency areboth measures of the same process butlatency only becomes a sensitive measurewhen the probability measure has reachedasymptote. Closer examination of the data,however, suggests a lower correlation be-tween correct response probability andresponse latency. By definition, correctresponse probability did not reach asymp-tote until after the TLE but the decline inlatency appeared to begin prior to theTLE for some items. Although the conclu-sions are speculative, it appeared to be thecase that in some instances, correct re-sponse latencies became shorter while in-correct response latencies remained con-stant, or increased slightly, on the last oneor two trials prior to the TLE. No suchtrends are evident in the data availablefrom previous studies but if the tendenciesdetected in this experiment are indicative ofthe underlying processes, latency may be-come a sensitive measure before the pointat which correct response probabilityreaches asymptote. It might further beinferred that response probability and la-tency are measuring two different factors.

Rather than viewing the TLE as a pointat which distinct changes occur in bothcorrect response probability and resnonselatency, it may be more fruitful to con-sider the trials immediately prior to andafter the TLE as an area of transition.The area may be analogous to a psycho-physical threshold in that while distinctlydifferent situations hold on either sideof the threshold, the behavior in questionis highly variable and probabilistic in thearea of the threshold itself. The most ac-curate statement that can be made at thistime is that latency begins to decline oneor two trials before or after the point atwhich the probability of a correct responsereaches asymptote.

RESPONSE LATENCY

Response Latency during Over learningResponse latency after the TLE was a

negatively accelerated, inverse function ofpractice. The decrement curves could havebeen the result of an increasingly large pro-portion of items undergoing a sudden,discrete reduction in latency, but examina-tion of the response curves of individualitems indicated that the observed declineresulted from a gradual decrement in la-tency for all items. The major portion of thedecline occurred on the first few trials afterthe TLE but there was no indication thatthe curve had reached asymptote at thepoint at which practice was terminated. Itis difficult to predict at what point the la-tencies would have reached asymptote.None of the previous studies continuedpractice beyond 10 trials past the TLE. Asindicated, both the information-level andmethod variables had significant effectsduring overlearning. Post-TLE interitemvariability was considerably smaller thanpre-TLE variability. The mean standarddeviation, over all treatment conditions,was approximately half as large after theTLE as it was prior to the TLE. Part ofthis reduction may be attributed toshorter post-TLE latencies but the ratioof the standard deviation to the mean was.53 prior to the TLE and .40 after the TLE.

It was found that post-TLE responselatencies were a significant positive func-tion of item difficulty as defined by thenumber of trials required to reach the TLE.While the data obtained by Millward(1964) and Suppes et al. (1966) indicatethat longer pre-TLE response protocols hadslower post-TLE responses, it was not pos-sible to determine from the available datawhether this effect was due to item difficultyor individual differences in learning rkte.Post-TLE latency differences between themost and least difficult items were not largebut it must be remembered that these dif-ferences were derived from mean latencyscores averaged over the first 16 post-TLEtrials; it may be the case that the differ-ences are more pronounced immediatelyafter the TLE. The interesting aspect ofthis relationship between item difficultyand post-TLE latency is that during the

27

post-TLE period, the reTonse strengthsof the most and least difficult items wouldhave been defined as equivalent on thebasis of correct response probability andthe number of trials of overlearning whicheach item had received. An obvious areafor future research would be to attempt todetermine if some indicant of responsestrength such as retention or transferwould confirm the latency measure sug-gestion that the response strengths of thetwo types of items still differed. It is pos-sible, the work of Shiffrin and Logan(1965) suggests, that post-TLE differencesin response latency between the most andleast difficult items were not an indicant ofresponse strength but simply due to theitems being practiced with different re-sponse speeds during the pre-TLE period.The current study did find that difficultitems had longer latencies prior to the TLEand the practice effect alone could accountfor the post-TLE difference.

In addition to intra-S differences inlatency as the result of item difficulty,post-TLE latencies were demonstratedto be related to S learning rate. Slow learn-ers tended to be slow responders duringthe post-TLE period. The difference be-tween fast and slow learners was significanton only the outside positions, Keys 1 and2. The mean scores on the inside keys, al-though not significantly different, demon-strated a tendency toward the same rela-tionship. It does not appear that theslower, post-TLE latencies were a functionof pow response practice prior to the TLEsince the mean latencies of the slow andfast learners were essentially equal priorto the TLE. If the post-TLE differenceswere a function of S learning rate and ifpost-TLE latencies are indeed a measure ofoverlearning, this suggests that the numberof trials required for an S to reach a re-sponse probability criterion may be corre-lated with the rate at which the associa-tions are strengthened during overlearning.To achieve the same degree of retention,a slow learner may require more over-learning practice than a fast learner. Post-TLE response latencies may provide ameans of determining the amount of over-learning practice which would be required


for a particular S to assure a given degreeof retention.

Some Suggestions concerning Measurementof the Learning Process

With respect to the measures of correctresponse probability and response latency,the learning process appears to have twovery distinct periods: early learning, theperiod prior to the point at which responseprobability reaches asymptote, and over-learning, the period during which latencyundergoes its greatest systematic varia-tion. The bulk of experimentation inverbal learning has dealt with only onemeasure, response probability, and onlythe first phase of the learning process.

It would appear that latency is not a sen-sitive measure of associative strength dur-ing the pre-TLE period. During this pe-riod, response probability seems to be themost accurate measure available. Responseprobability was sensitive to differences inthe training method in that recall Ssreached a response probability criterionwith fewer trials than did the anticipation-method Ss. Pre-TLE response latencieswere insensitive to the training method.Latencies did not reflect the increase inassociative strength indicated by the in-crease in correct response probability froma chance level to .54 just prior to the TLE.Finally, latencies did not differentiate be-tween correct and incorrect responses dur-ing this period.

Response latencies were indicative ofthe complexity of the task during the pre-TLE period in that latency was a functionof the number of response alternatives.This might imply that latency would beuseful as a measure of other task param-eters which influence learning. Latencymay be a more sensitive measure of theearly stages of learning in more complextasks. The learning task in the currentexperiment was intentionally kept very sim-ple in that response learning was mini-mized. If the task had required an appre-ciable amount of response integration, theresults obtained might have been quitedifferent in that pre-TLE latencies mightwell vary systematically as a function ofresponse learning. In such a learning situa-

tion, latency may be a useful supplementto response probability measures.

During overlearning, after the TLE, therelative utility of the response probabilityand latency measures is reversed. The re-sponse probability measure becomes in-sensitive because it has reached asymp-tote and at about the same time, responselatency seems to become a sensitive meas-ure of associative strength. One cannot besure that latency is measuring associativestrength during this period until latencymeasures are checked against some othermeasure such as retention but there areseveral indications that this is the case.The rapid post-TLE decline in latency,of course, suggests the continued develop-ment of associative strength. Just as re-sponse probability was sensitive to dif-ferences in training method prior to theTLE, the post-TLE reduction in latencywas more pronounced under the recall para-digm than under the anticipation para-digm. Post-TLE latencies appeared to besensitive to differences in learning ratewhich were determined by response prob-ability measures earlier in learning. Thiswas true for both individual differencesin S learning rate and differences in itemdifficulty. Latencies were a positive func-tion of the number of response alternativesthroughout the post-TLE period, as wellas prior to the TLE, but the rate at whichlatency declined did not differ significantlybetween the different information levels.

While it is evident that response-prob-ability measures are more useful thanlatency measures prior to the TLE andthat latency may well be a useful measureafter the TLE, the utility of latency meas-ures in the transition area around the TLEis not at all clear. Is there an abrupt changein latency at the point at which responseprobability reaches asymptote or are thetwo transition points only roughly equiv-alent? The situation might be clarified byusing more precise measurement techniquessuch as using a greater number of responsealternatives to reduce chance effects. Ifthis were done, the TLE would be a moreaccurate indication of the point at whichcorrect response probability reached asymp-tote. An extensive investigation of individ-

RESPONSE LATENCY 29

ual item protocols should also prove to beuseful. It may well be the case, however,that response measures in this area de-monstrate the instability characteristic ofother thresholds.

In summary, the following statementscan be made about latency measures in thetask used in the current experiment: (a)prior to the TLE, latency was a measureof task complexity but did not measure thedevelopment of associative strength; (b)during overlearning, response latency didappear to measure the continued develop-ment of associative strength. These resultsconcerning latency measures provide a hostof experimental questions for future re-search. The most obvious of these is whetheror not the post-TLE decline in latency isactually indicative of the growth of asso-ciative strength. Can the degree of retentionbe controlled by training to a latency cri-terion? Do response latencies reach a stableasymptote and if so, does this asymptotehave any implications for retention? Howdoes latency change over the course oflearning in more complex tasks such asconcept formation? Do latency measureshave utility for instructional decisionsin the early stages of such tasks?

Latency Measurement in InstructionSince this experiment concerned a rote

drill situation, the generality of the resultsand conclusions are somewhat limited withrespect to other types of instructional situa-tions but one broad statement may bemade. It would appear that response la-tencies can be accurate measures of thelearning process and hence can form anadequate basis for instructional decisionsbut their applicability may be limited tospecific stages of learning. In a rote drillcontext, latencies early in learning, priorto the TLE, appear to contain little in-formation of value for instructional deci-sions. Latencies may be quite useful in asituation in which instructional materialshave been carefully programmed so thatcorrect response probability is always rela-tively high but they would seem to beleast useful in situations in which prob-ability of a correct response is very low.

After the TLE, on the other hand, at the

time when response probability measureshave reached asymptote in a rote drillsituation, response latencies demonstratetheir largest and most systematic varia-tion. Interresponse variability also de-creases during this period and this wouldincrease the reliability of latency meas-ures. These findings may have definite im-plications for instructional decision making.In a spelling drill, for example, the goal ofthe instruction is not only that the studentlearn the association but also that theassociations be retained. While it is knownthat overlearning increases retention, theamount of overlearning to be provided hasalways been a relatively arbitrary de-cision. The capability of measuring responselatency may provide a means of determin-ing the optimal amount of overlearningpractice for a particular student and aparticular group of words. If it were foundthat response latencies were not shortenedby an instructional program designed tobring a student to a high level of profi-ciency in a certain skill, the utility of theinstructional procedures might be ques-tionable.

REFERENCES

ARCHER, E. J. A re-evaluation of the meaningful-ness of all possible CVC trigrams. PsychologicalMonographs, 1960, 74 (10, Whole No. 497).

BATTIG, W. F., & BRACKETT, H. R. Comparison ofanticipation and recall methods in paired-associ-ate learning. Psychological Reports, 1961, 9,59-65.

BATTIG, W. F., & Wu, R. D. Comparison of recalland anticipation paired-associate procedureswithin mixed aurally-presented lists. Psycho-nomic Science, 1965, 3, 233-234.

BECK, R. C., PHILLIPS, W. R., & BLOODSWORTH,W. D. Associative reaction times as a function ofassociation value of nonsense syllable stimuli.Psychological Reports, 1962, 10, 517-518.

BRICKER, P. D. Information measurement and reac-tion time. In H. Quastler (Ed.), Informationtheory and psychology. Glencoe, Ill.: Free Press,1955.

Ems, P. D., & ZEAMAN, D. Response speedchanges in an Estes' paired-associate "miniature"experiment. Journal of Verbal Learning andVerbal Behavior, 1963, 1, 384-388.

HICK, W. E. On the rate of gain of information.Quarterly Journal of Experimental Psychology,1952, 4, 11-26.

HULL, C. L. Principles of behavior. New York :

Appleton-Century-Crofts, 1943.


HYIVIAN, R. Stimulus information as a determinantof reaction time. Journal of Experimental Psy-chology, 1953,45, 188 -196.

JOHNSON, R. C. Latency and association value aspredictors of rate of verbal learning. Journal ofVerbal Learning and Verbal Behavior, 1964,3,77-78.

KINTSCH, W. Habituation of the CSR componentof the orienting reflex during paired-associatelearning before and after learning has takenplace. Journal of Mathematical Psychology,1965,2, 330-341.

MERKEL, J. Die zeitlichen Verhaltnisse der willen-statigkeit. Philosophical Studies, 1365, 2, 73-127.Cited by R. Hyman, Stimulus information as adeterminant of reaction time. Journal of Experi-mental Psychology, 1953,45, 188-196.

MILLWARD, R. Latency in a modified paired-associ-ate learning experiment. Journal of VerbalLearning and Verbal Behavior, 1964,3, 309-316.

PETERSON, L. B. Paired-associate latencies after the

last error. Psychonomic Science, 1965, 2, 167-168.

RABBITT, P. /VI. A. Effects of independent variationsin stimulus and response probability. Nature,1959,183, 1212.

SIiIFFRIN, R., dr LOGAN, F. A. Performance speed asa function of practice speed. Journal of VerbalLearning and Verbal Behavior, 1965,4, 335-338.

SIEGEL, S. Nonparametvic statistics. New York :McGraw-Hill, 1956.

SINILEY, 0. A. The relation of subliminal to supra-liminal learning. Archives of Psychology, 1933,No. 146.

SUPPES, P., GROEN, G., & SCHLAG-REY, M. A modelfor response latency in paired-associate learning.Journal of Mathematical Psychology, 1966, 3,99-128.

WILLIAMS, J. P. A test of the all-or-none hypothesisfor verbal learning. Journal of ExperimentalPsychology, 1962,64, 158-165.

(Received August 16, 1968)

DOCUMENT RESUME ED 036 858 Judd, Wilson A.; Glaser, Robert ... · WILSON A. JUDD AND ROBERT GLASER2...

Documents

Transcript of DOCUMENT RESUME ED 036 858 Judd, Wilson A.; Glaser, Robert ... · WILSON A. JUDD AND ROBERT GLASER2...