Classifying Implicit Memory Tests

24 Pages·2003·2.14 MB·English

JOURNAL OF MEMORY AND LANGUAGE 29,389-412 (1990)
Classifying Implicit Memory Tests: Category Association and
Anagram Solution
KAVITHA SRINIVAS AND HENRY L. ROEDIGER III
Rice University
In three experiments we investigated the nature of two implicit memory tests, free asso-
ciating to category names and solving anagrams, by manipulating several variables during
study. Additional implicit and explicit tests were included for comparison (word fragment
completion and free recall). Primed category association responded like free recall by show-
ing an advantage from generating words relative to reading them out of context, a levels of
processing effect, and no effect of study modality (visual or auditory). The results of these
variables on anagram solution were less clear cut, but tended to resemble the effects of
primed word fragment completion (a modality effect, only a slight levels of processing
effect, and little priming from pictures). Dissociations were obtained between explicit and
implicit tests (e.g., free recall and primed fragment completion), but more importantly
between implicit tests (e.g., primed fragment completion and category association). Because
implicit tests can be dissociated, no single system underlies their performance; we empha-
size instead their processing requirements to explain dissociations among tests.
01990
Academic Press, Inc.
Implicit memory tests are defined as tests comes to mind. Implicit memory is re-
that do not require conscious recollection vealed to the extent subjects are more
of a prior study episode for their successful likely to complete the fragment with the
completion (Schacter, 1987); nevertheless, word “donkey” after having studying the
they show a benefit in performance from item than if they had not studied the item.
the previously studied episode. Although Some examples of tasks considered implicit
the conscious status of a subject during a memory tests are completing word frag-
test is arguable, the implicit nature of these ments, such as d _ n _ e _ (e.g., Tulving,
tests is operationalized through instructions Schacter, & Stark, 1982), completing word
at testing (Gardiner, Dawson, & Sutton, stems, such as don (e.g., Graf,
1989; Schacter, Bowers, & Booker, 1989). Squire, & Mandler, 1984), identifying
For example, on an implicit word fragment briefly presented words (e.g., Jacoby &
completion test, subjects are instructed Dallas, 1981), or making word/nonword de-
to complete a word fragment such as cisions to targets (e.g., Kirsner, Milech, &
d-n-e- with the first response that Standen, 1983). Explicit memory tests are
presumed to require conscious recollection
Part of this research was supported by NIH Grant of a prior episode (Schacter, 1987). In ex-
ROl HD15054 and was presented at the 1988 meetings plicit measures of memory, test instruc-
of the Psychonomic Society in Chicago, and the 1989 tions refer to a particular spatial or tempo-
meetings of the Midwestern Psychological Association
ral context in a subject’s personal history.
in Chicago. Experiments 1 and 2 (in part) constituted a
Examples of these tests are the standard
master’s thesis submitted to Purdue University by the
first author, under the direction of the second author. measures such as free recall, recognition,
We thank committee members Mark McDaniel and and cued recall.
Eliot Smith for their advice, and Brad Challis and The distinction between explicit and im-
Larry Jacoby for comments on the manuscript. Cor-
plicit memory tests is interesting because
respondence concerning this article can be sent to the
the two classes of tests exhibit different
authors at the Department of Psychology, Rice Uni-
versity, Houston, TX 77251-1892. patterns of results as a function of certain
389
0749-S%Xl90 $3.00
Copyright 0 1990 by Academic Press, Inc.
All rights of reproduction in any form reserved.
390 SIUNIVAS AND ROEDIGER
independenta nd subject variables. For in- posed a transfer appropriatep rocessinga c-
stance, compared to normals, amnesic pa- count of these dissociations based on four
tients show impaired performance on ex- assumptions: (a) Memory tests benefit to
plicit memory tests, but their performance the extent that operations required at test
on implicit memory tests is equivalent to recapitulate operations required at study;
that of normals (e.g., Graf & Mandler, (b) implicit and explicit memory tests typi-
1984).A lso, study variables such as levels cally require different retrieval operations,
of processingt hat influence explicit mem- and consequently benefit from different
ory tests in one way frequently do not affect types of processing during learning; (c)
implicit memory tests in the same way most implicit tests rely heavily on the
(e.g., Jacoby & Dallas, 1981). match of perceptual features between
These functional dissociations between learning and test episodes, or data-driven
explicit and implicit memory tests have led processing; and (d) most explicit tests re-
some researcherst o conclude that they are quire the encodedm eaning of concepts for
tapping two different memory systems: one successful recollection, or conceptually-
that is impaired in amnesiaa nd anothert hat driven processing.
is preserved in amnesia. Squire’s (1986, An important corollary to these assump-
1987)t heory is an exemplar of this view. He tions is that explicit memory tests can de-
argues that dissociations between explicit pend on data-drivenp rocessinga nd implicit
and implicit memory tests are evidencef or tests on conceptually-driven processing;
two different memory systems: the declar- there is no necessaryc orrelation between
ative and the procedural systems. The de- the explicit-implicit distinction on the one
clarative system is responsible for con- hand, and the conceptually-driven data-
scious accesst o facts and past experiences driven distinction on the other. Also, the
and is necessary for performance on ex- proposed distinction between data-driven
plicit memory tests. The procedurals ystem and conceptually-driven tests is not in-
records the processing operations of the tendeda s a dichotomy, but rather as repre-
system as they are modified by events, but senting end points on a continuum. Tests
not their explicit description. Procedural may involve both types of processes. In-
memory is revealedb y performanceo n im- deed, a more useful assumption is to de-
plicit tests. Similarly, Tulving (1985, 1987) scribe two continua, one for each type of
discusses dissociations between explicit processing( Weldon, 1988),t o acknowledge
and implicit tests as evidence for multiple that these two modes of processingc an be
memory systems. According to him, ex- varied orthogonally (i.e., they need not
plicit tests tap the episodic memory system, trade off against one another, as implied if
and implicit tests tap either the semantico r only a single continuum is postulated).
procedural memory systems. Data-driven and conceptually-driven
An alternate view of these dissociations processinga re operationallyd efinedb y Roe-
is basedo n the assumptiont hat memory is diger et al. (1989b)t hrough study manipu-
revealedt o the extent that processingo per- lations used by Jacoby (1983). Jacoby
ations at study and test overlap (the princi- (1983)h ad subjects study words in one of
ple of transfer appropriatep rocessing,M or- three conditions. In the No Context condi-
ris, Bransford, & Franks, 1977). By this tion subjectsr ead words aloud without con-
view, dissociations between explicit and text (e.g., XXX-COLD), and the condition
implicit memory tests occur becauset hey was assumed to involve maximal data-
typically require different modes or types driven processing. That is, subjects must
of processinga t test (Jacoby, 1988;K olers processt he visual data, the given letters of
& Roediger, 1984; Roediger, Weldon, & the word, to read it aloud. In the Generate
Challis, 1989b).R oedigere t al. (1989b)p ro- condition, subjects produced words from
CLASSIFYING IMPLICIT TESTS 391
semantic clues (e.g., HOT-???), and this The memory systems view and the trans-
condition was assumedt o involve maximal fer appropriate processing approach make
conceptually-driven processing, because differential predictions about the presence
no “data” were given to guide the bottom- of functional dissociations among memory
up production of the target word. Instead, tests. The systems view implies that func-
production of cold involved top-down, as- tional dissociationss hould not generally oc-
sociative processes,b ecause subjects had cur among explicit memory tests or among
been told to produce oppositesi n the Gen- implicit memory tests, because they are
erate condition. In the Context condition, presumably tapping the same memory sys-
subjects read words with the semantic clue tem. The transfer appropriate processing
(e.g., HOT-COLD); this condition assumed view predicts dissociations among explicit
to involve both data-driven and conceptu- tests or among implicit tests whenever they
ally-driven processing. Given these as- require different types of processing. Un-
sumptions, a memory test can be classified fortunately, most of the researchi n the area
as data-drivenw hen better performanceo c- has confoundedt he implicit/explicit nature
curs in the No Context condition than in the of the test with the type of processing re-
Generatec ondition; it can be classified as quired by the tests (Roediger & Blaxton,
conceptually-driven when performance is 1987b).T ypically, one explicit test (which
better in the Generatec ondition than in the usually requires conceptually-driven pro-
No Context condition. cessing)i s compared with one implicit test
Roediger et al. (1989b) also suggested (which usually requires data-driven pro-
several convergingo perationsi n the classi- cessing), and when a functional dissocia-
fication of memory tests as conceptually- tion is found between the tests, different
driven or data-driven. Briefly, data-driven memory systemsa re postulated. In order to
tests should be more affectedb y changesi n unconfoundt he two, one has to compare at
surface information between study and least two explicit tests with different pro-
test, such as changesi n modality (auditory cessingr equirements,a nd two implicit tests
or visual), symbolic form (pictures or with differing processingr equirements.
words), or language for bilinguals (say, The first such investigation was made by
Spanish and English). On the other hand, Blaxton (1985, 1989)w ho compared five
they should be relatively immune to manip- memory tests in all-three explicit memory
ulations involving conceptual elaboration tests and two implicit memory tests. Two of
such as the levels of processingm anipula- the explicit tests were conceptuallydriven
tion (Craik & Lockhart, 1972),f orming im- tests (free recall and cued recall with se-
ageso f words’ referents (Paivio, 1986),o r- mantic cues),a nd one of them was designed
ganizing words into higher order units to be data-driven (cued recall with graphe-
(Tulving, I%@, and other forms of elabo- mic cues). Of the two implicit tests, one of
rative processing that have large positive them was data-driven( word fragment com-
effects on conceptually-driven tasks such pletion), and the other was designedt o be
as free recall. Conversely, conceptually- conceptually-driven (answering general
driven tests should be little affectedb y ma- knowledge questions). Thus, if the word
nipulations of surface information (unless HEMLOCK was presented in the study
these manipulations also engender differ- phase, subjects were asked to recall all the
ences in conceptual processing; e.g., pic- studied words (free recall), to use the cue
tures vs. words), but should be strongly af- POISON to remember the studied word
fected by manipulationso f conceptuale lab- (semantic cued recall), to use the word
oration. Roediger et al. (1989b)d iscussed HAMHOCK to remember the studied word
various explicit and implicit tests in terms (graphemic cued recall), to solve the frag-
of this classification. ment HE _ _ 0 _ K (word fragment com-
392 SRINIVAS AND ROEDIGER
pletion), or to answer the question “What low associate to the category name (e.g.,
did Socrates drink at his execution?” (gen- stool) prior to the test will facilitate re-
eral knowledge). Performance on these sponding with the associate on the subse-
tasks was compared as a function of quent test relative to a nonstudied baseline.
whether words were studied in a Generate, Gardner, Boller, Moreines, and Butters
Context, or No Context condition, as in Ja- (1973) and Graf, Shimamura, and Squire
coby (1983). (1985) have shown preserved priming in
Blaxton (1989, Experiment 1) found bet- amnesics with this test. We thought that
ter performance in the Generate condition category association would likely require
compared to the No Context condition on conceptually-driven processing, because
the conceptually-driven tasks (regardless of there is no perceptual match between the
whether they were explicit or implicit), and studied item and the test cue, and perfor-
better performance in the No Context com- mance on this test is guided by associative
pared to the Generate condition on the processes.
data-driven tasks (regardless of whether The anagram solution test requires the
they were explicit or implicit). These data subject to unscramble letters strings in the
provide evidence favoring the transfer ap- test phase to form words (e.g., otosl for
propriate processing account over the “stool”). The advantage or priming accru-
memory systems account of functional dis- ing from having studied the word on the
sociations; they suggest that functional dis- solution of the anagram relative to a non-
sociations occur when tests require differ- studied baseline is the index of retention on
ent modes of processing, regardless of this test. A priori, it was thought that ana-
whether the tests are explicit or implicit. gram solution would require data-driven
They also illustrate the critical need to com- processing, since solution of the anagrams
pare several different memory tests to draw seems to depend on the match between the
secure conclusions regarding dissociations. perceptual features at study and at test.
The present experiments were designed Also, Jacoby and Dallas (1981) noted
as further tests of the transfer appropriate briefly that parallel effects were obtained
processing account and the memory sys- between anagram solution and perceptual
tems account of functional dissociations identification, although only the latter re-
among memory tests. They were also sults were presented in their article. Perfor-
meant to test the predictions of the transfer mance on these tests was compared to per-
appropriate processing approach with two formance on the word fragment completion
seldom used implicit tests. Specifically, the test, which is an implicit test known to be
aim in these experiments was to classify largely data-driven (Blaxton, 1989; Roedi-
these tests according to their processing re- ger & Blaxton, 1987a).
quirements using the converging operations Three experiments were conducted to
described above. compare performance on the three implicit
Two implicit tests were selected for com- measures. Experiment 1 was an attempt at
parison, one appearing to depend on con- classifying the tests using the operational
ceptually-driven processing (a category as- definitions described previously for con-
sociation test), and the other seeming to ceptually-driven and data-driven tests.
depend on data-driven processing (an ana- Thus, in Experiment 1, we observed the ef-
gram solution test). Briefly, the category fects of generating versus reading a word
association test requires the subject to free on test performance for the three implicit
associate to a given category name (e.g., tests. Free recall was also included as a
Articles of Furniture) for a specified time conceptually-driven explicit measure. Ex-
during the test phase. The measure of inter- periments 2 and 3 were attempts at obtain-
est in this test is whether studying a ing converging operations in the classifica-
CLASSIFYING IMPLICIT TESTS 393
tion of these tests. In Experiment 2, the frames were used to guide the generation of
modality in which a studied item was pre- 40 target words. In each case, a sentence
sented (auditory or visual), and the level of context was prepared so that the target
processing of the studied item (orientation word could be completed by the subject at
to its meaning or its appearance) were var- the end of the sentence (for e.g., “Heroin is
ied. In Experiment 3, the effect of changing related to c “). The sentence con-
the symbolic form of the studied items (pic- texts were selected so that they specified
tures or words) was observed for the word only the critical word. No mention was
fragment completion and anagram solution made of the category to which the target
tests. word belonged. These materials were
normed with 105 subjects in order to ensure
EXPERIMENT 1 that subjects generated the correct words
Method 95% of the time. Four sets of 10 items each
Subjects were then created by randomly assigning
the items to each set. These sets were ro-
One hundred and fifty-two Purdue Uni-
tated through the four study conditions to
versity undergraduates participated in the
create four study lists that completely
experiment in partial fulfillment of a course
counterbalanced conditions and materials
requirement.
across subjects. Thus, the same item (e.g.,
Design thunder) was presented in the three studied
A 4 (Study Conditions: Generate, Con- conditions as follows: (a) Generate: Light-
text, No Context, or Nonstudied) x 4 (Test ning is associated with t . (b) Con-
Conditions: Free Recall, Category Associ- text: Lightning is associated with thunder.
ation, Word Fragment Completion, and (c) No context: thunder. The fourth set of
Anagram Solution) mixed factorial design nonstudied words was used as a baseline
was used in the experiment. Study Condi- measure for the implicit tests.
tions were varied within-subjects, while Test materials were prepared as follows
Test Conditions were varied between-sub- for implicit memory tests. Word fragments
jects. For the free recall and word fragment were prepared by haphazardly deleting cer-
completion tests, 32 subjects were assigned tain letters of the word; for instance,
to each test, whereas 44 subjects received “thunder” was presented as t h _ _ _ e r.
the category association and anagram solu- Most word fragments had one solution, but
tion tests. a few had more than one. All word frag-
ments were presented in lowercase to en-
Materials courage a perceptual match between the
The study and test stimuli were 40 cate- study and test conditions. These fragments
gory exemplars drawn from two different were normed with 35 subjects so that the
sets of category norms (Battig & Montague, baseline rate of performance on the frag-
1969; Hunt & Hodge, 1971), such that each ments was about 30%. Similarly, anagrams
stimulus word was a category exemplar of a were prepared by haphazardly scrambling
particular category. These stimuli were se- the letters of the word stimuli presented in
lected to be medium to low frequency as- the study phase. Thus, for instance.
sociates of the category names. In order to “thunder” was presented as “tderhun.”
minimize differences in solving the stimuli Again, the anagrams usually had one solu-
on the anagram solution task, words were tion, but some had more than one. Ana-
selected from a narrow range of word grams were presented in lowercase letters
length (&7 letters). The complete set of ma- and they were also normed with 30 subjects
terials appear in Appendix A. so that baseline rates averaged about 30%.
In the Generate condition, sentence For the category association test, category
394 SRINIVAS AND ROEDIGER
names used at test were identical to the respectively, as a function of study condi-
names used in the norms from which the tion. Each test procedurew ill be described
critical words were drawn (Battig & Mon- briefly.
tague, 1969;H unt & Hodge, 1971). In the free recall test, subjects were
A set of 20 filler items and nine practice askedt o recall the studiedw ords on a blank
items were also constructed. These were sheeto f paperf or 7 min. For all the implicit
transformed appropriately for each test memory tests, the cover story used was
(i.e., fragments, anagrams, or category that the experimenterw as interested in de-
names were presented). veloping some materials for a future study.
In the category association test, subjects
were told that they would be presentedw ith
Procedure
category names, and they had to name all
Subjects were tested individually. In the the things belongingt o that category within
study phase,t hey were given a booklet that the given time limit. (This was similar to the
contained a mixed list of items from differ- presidentsa nd capitals tasks). The category
ent conditions. They were instructed to namesw ere presentedo n an IBM computer
read the critical target item, which was un- screen for a period of 30 s. Subjects’ re-
derlined, out loud in each of the three con- sponsesw ere recorded on a tape recorder.
ditions. Dependingo n the study condition, The experimenter also discreetly recorded
subjectse ither completedt he sentencew ith the critical words that the subject pro-
the target word (Generate condition), or duced.
read the underlined word in a sentence In the word fragment completion test,
(Context condition) or read the word in iso- subjectsw ere asked to complete fragments
lation (No Context condition). In rare cases of words that were presented on an IBM
when subjects failed to generate a target computer. Each word fragment was dis-
item correctly, the experimenter said the played on the screen for 30 s or until the
correct response. subject responded.S ubjects completed the
Subjects were instructed to pay attention fragments verbally. Voice keys were not
to the words that they read aloud. They used to record the latencies to complete
were told that they might be given a mem- word fragments (or anagrams), because
ory test later, but the nature of the memory subjects tended to verbalize while solving
test was left unspecified. A signal recorded anagramso r word fragments. Experiment-
on tape was used to pace the subjects ers were trained to hit a key as soon as the
throught he study task at 12 s/item. A cover subject responded, and latencies for re-
sheet was also used to ensuret hat subjects sponsesw ere recorded in ms on the com-
spenta n equal amount of time on eachi tem. puter. The solution words given by the sub-
Following the study phase, subjectsw ere jects were also recorded on the computer
given two filler tasks. For the first 5 min, by the experimenter. Nine practice trials
subjects wrote down the names of all U.S. were given before starting the actual test.
presidents they could remember, and for Following the practice session, 60 word
the next 5 min, subjects wrote down the fragments were presented to the subjects
nameso f U.S. state capitals. and the ratio of studied words to nonstud-
In the test phase,s ubjectsr eceivede ither ied words was 1:1 . The procedureu sed for
a free recall test, a category association the anagrams olution test was the same as
test, a word fragment completion test, or an the procedure used for the word fragment
anagram solution test, depending on the completion test, except that subjects saw
group to which they had been assigned.I n anagramso n the IBM computer screena nd
all four test conditions, the variable of in- tried to solve them verbally. The latencies
terest was the proportion of items correctly to respondf or each item, as well as the so-
recalled, produced, completed, or solved, lution words given by the subjects, were
395
CLASSIFYING IMPLICIT TESTS
recorded on the computer by the experi- The finding of better performance in the
menter. Generate condition compared to the No
Context condition is consistentw ith results
Results and Discussion from previous experiments (e.g., Blaxton,
Becaused ifferent memory tests typically 1989; Smith & Branscombe, 1988). How-
have different baselinest hat complicate the ever, unlike other findings (e.g., Jacoby,
interpretation of cross-test comparisons, 1983),t he results suggestt hat subjectsw ere
the results for the four different memory significantly better at recalling words that
tests (free recall, word fragment comple- were studied without context (.32) than
tion, anagrams olution, and category asso- they were at recalling words studied in con-
ciation) will be discussedi n different sec- text (.21). A suggestiono f this samep attern
tions. Performance on different tests is occurred in Blaxton’s (1989)f ree recall re-
comparedi n a final section. The overall re- sults, but her 3% difference between the
sults of Experiment 1 are summarized in Context and No Context conditions was
Table 1, which displays the proportion of not significant. This anomalous finding is
target words correctly produceda s a func- probably due to the natureo f the generation
tion of the different study conditions on materials used in this study. The target
each of the four tests. Due to missing ob- words in the Context condition were em-
servationsi n the responset ime data for the bedded in sentences,a nd thus may have
word fragment completion and the anagram beenl ess salient than words in the No Con-
solution tests, the responset ime results for text condition. Alternatively, items studied
theset ests are not presentedh ere. The data, in the Context condition may suffer input
however, paralleled the results with pro- interferencef rom the study of more mate-
portion correct as the dependentm easure. rial. Becauseo ther studies using sentences
The level of significance for all the results to evoke generationo f items did not include
reported in this paper was set at .05. a No Context condition (e.g., Kane &
Free recall. Words studied in the Gener- Anderson, 1978), comparison of findings
ate condition were free recalled better than across studies is hazardous.W hatever the
words studied in the No Context (12% dif- reasonf or the inferiority of the Context to
ference)o r Context conditions (23% differ- No Context condition, the advantageo f the
ence), and words studied in the No Context Generate over No Context condition ap-
condition were recalled better than words pears even more impressive, because the
studied in Context condition (11% differ- processo f generationo vercame the inhibi-
ence). A repeatedm easuresA NOVA con- tion (from input interference or whatever)
fumed these observations, with a signifi- engenderedb y items appearing in a sen-
cant main effect of study condition, F(2,62) tence context. .
= 22.89,M Se = .02, and the least signifi- This advantage in free recall confirms,
cant difference (LSD) for comparisonsb e- with the presentm aterials, that free recall is
tween means of .07. a conceptually-driven test. Now we may
TABLE 1
PROP~RTIONOFRESPONSESCORRECTLYPRODUCEDORSOLVEDASAFUNCTIONOFSTUDYANDTESTTYPE
IN EXPERIMENIT
Study condition
Test type Generate Context No context Nonstudied
Free recall 0.44 0.21 0.32 -
Category association 0.33 0.25 0.23 0.16
Word fragment completion 0.35 0.41 0.45 0.21
Anagram solution 0.57 0.58 0.62 0.49
396 SRINIVAS AND ROEDIGER
ask if category association is also a concep- dition implicates a lexical or semantic com-
tually-driven test, in which case we should ponent, too (Weldon, 1988).
find the same pattern of results with this Anagram solution. The results showed
test (i.e., superior performance in the Gen- priming effects on the Generate, Context,
erate condition to that in the No Context and No Context conditions (8, 9, and 13%,
condition). respectively) with a 5% advantage of No
Category association. Results showed Context to the Generate condition. Re-
that words studied in the Generate, Con- peated measures ANOVA indicated a main
text, and No Context conditions were pro- effect of study condition, F(3,129) = 6.14,
duced more often on the category associa- MSe = .02. The LSD for comparisons be-
tion test than were the Nonstudied words tween means was .07.
(17, 9, and 7%, respectively). As in free re- The critical comparison between the No
call, words studied in the Generate condi- Context and Generate conditions (although
tion were produced more often than those in the predicted direction) was not signifi-
in the Context (an 8% advantage) and No cant, despite having more observations on
Context (10%) conditions. However, on this test than on the word fragment comple-
this test, the difference between the Con- tion test (440 observations as opposed to
text and No Context tests was not signifi- 320). Although these data are suggestive,
cant. they do not permit us to conclude that the
Repeated measures ANOVA revealed a anagram solution test is data-driven.
significant main effect of Study condition, Comparison of measures. To determine
F(3,129) = 8.16, MSe = .02. The LSD for whether performance on study conditions
comparisons between means was .07. The was a function of the type of test, an
advantage of Generate to No Context con- ANOVA was performed with study condi-
ditions in the category association test in- tion as a within-subjects factor, and test
dicates that it is also a conceptually-driven as a between-subjects factor. The critical
task. Test x Study interaction was significant,
Word fragment completion. Data for the F(3,450) = 12.98, MSe = .03, suggesting
word fragment completion test indicated that four different tests showed different
significant priming effects on the Generate, patterns of results across the study condi-
Context, and No Context conditions (14, tions.
20, and 24%, respectively). Compared to To examine these different patterns, sep-
the free recall and category association arate Study x Test interactions were per-
tests, the pattern on this test was reversed formed for each combination of the three
so that subjects were better at solving word implicit tests (free recall was not included
fragments when they had read the words because there is no nonstudied baseline
without context compared to when they measure in free recall). The Study x Test
had generated them (a 10% difference). An interaction with word fragment completion
ANOVA revealed a significant main effect and anagram solution tests was not signiti-
of study condition, F(3,93) = 9.00, MSe = cant, F(3,222) = 1.48, MSe = -03. This
.04, and LSD for comparisons between suggests that both anagram solution and
means was .10. word fragment completion show similar
The advantage of No Context to Gener- patterns of results across the study condi-
ate conditions suggests that word fragment tions, which would be expected if both
completion is largely a data-driven test, and tasks require data-driven processing. The
this is consistent with earlier findings (e.g., Study X Test interaction with the category
Blaxton, 1989; Smith & Branscombe, association and the anagram solution tasks
1988). However, the fact that significant was significant, F(3,258) = 2.98, MSe =
priming also occurred in the Generate con- .02. Despite the fact that both are implicit
CLASSIFYING IMPLICIT TESTS 397
tests, the data reveal a functional dissocia- study will affect test performance on con-
tion between category association and ana- ceptually-driven tests, because the typical
gram solution. Finally, the Study x Test levels of processing manipulation involves
interaction between the word fragment conceptual elaboration, while holding con-
completion and category association tests stant the display of data. On the other hand,
was also significant, F(3,222) = 5.00, MSe the modality of presentation of the study
= .03, indicating another functional disso- material should affect performance on data-
ciation between two implicit tests. driven tests, because variation in modality
The finding of dissociations between im- affects the perception of data, but not their
plicit measures of memory tends to under- conceptual elaboration. From the findings
mine the argument that dissociations be- in Experiment 1, one might predict that the
tween explicit and implicit measures neces- levels of processing manipulation will affect
sarily imply the existence of different performance on the category association
memory systems, if one assumes that a sin- test, but that the modality manipulation will
gle system underlies performance on all im- not. If these predictions are fulfilled, the
plicit tests (Roediger, Srinivas & Weldon, results will provide converging evidence
1989a). On the other hand, the results of suggesting that the category association
Experiment 1, in general, support the pro- test is conceptually-driven. On the other
cessing view of dissociations between ex- hand, if anagram solution is a data-driven
plicit and implicit memory tests. The cate- task, modality of presentation, but not lev-
gory association test showed a different els of processing, should affect perfor-
pattern of results from the anagram solution mance. If anagram solution responds to
or word fragment completion tasks because both types of processing, as the results of
it required conceptually-driven processing Experiment 1 seem to indicate, then one
whereas the other two (especially word might see slight effects of both modality
fragment completion) were data-driven. and levels of processing, because anagram
Although findings in Experiment 1 are solution may reflect a mixture of the two
broadly consistent with the transfer appro- processes. Finally, for purposes of valida-
priate processing approach, one problem tion, the word fragment completion test
emerged. Performance on the anagram so- was also included. In accordance with pre-
lution test showed a data-driven pattern vious findings, performance on this test
(No Context > Generate), but the differ- should be little affected by the levels of pro-
ence was not significant. Experiment 2 was cessing manipulation (e.g., Roediger, Wel-
designed to explore further the nature of don, & Stadler, 1987; but see Squire, Shi-
the two new implicit memory tests. mamura, & Graf, 1987), but should be af-
fected by the modality of presentation
EXPERIMENT 2
(Blaxton, 1989; Roediger & Blaxton,
In Experiment 2 we employed the levels 1987a).
of processing and modality manipulations
to provide converging evidence for the con- Method
clusions drawn from Experiment 1. Exper-
Subjects
iment 2 was also aimed at resolving the na-
ture of the anagram solution test, because One hundred Purdue University under-
unambiguous classification of the test as ei- graduates and 50 Rice University under-
ther data-driven or conceptually-driven graduates participated in the experiment in
was not possible based on the results of Ex- partial fulfillment of a course requirement.
periment 1. The Rice subjects were all assigned to the
The Roediger et al. (1989a) approach pre- word fragment completion condition, as
dicts that the level of processing during this group was tested last. Thus, compari-
398
SRINIVAS AND ROEDIGER
son between this group and the other two For each block of 24 items, subjects were
must be viewed with caution. given different instructions. For the seman-
tic orienting task, subjects were told to
Design think about the meaning of the word that
Four of the five within-subjects study they saw or heard, and rate it along a scale
conditions used in the experimentw ere ob- of pleasantnessth at rangedf rom 1 (very un-
tained by crossingl evels of processing( se- pleasant)t o 7 (very pleasant).T he task was
mantic or physical orienting task) with the demonstrated with an example. For the
modality of presentation (auditory or vi- physical orienting task, subjects were told
sual). A fifth condition-a group of non- to count the number of consonantst hat the
studied items-was included to provide a word contained, and were asked to circle
baseline measure. Items processedi n dif- the appropriate number on the sheet that
ferent ways were presentedi n two blocks. contained numbers from 1 to 7. Thus, in
Thus, half of the subjects first processed either task the nature of the overt response
the words for meaning in the semantic was the same.
orienting task, whereas the other half Within each block, items were presented
performed the physical orienting task first. visually via slide projector for a period of 6
Finally, test type was varied between- s each, or read aloud by an experimenter
subjects, so that subjectsw ere given either twice during a 6-s period. The second pre-
the word fragment completion test, the cat- sentationi n the auditory condition was in-
egory association test, or the anagram so- tended to help till the 6-s interval, to mini-
lution test. mize the difference from the visual condi-
tion in which the word was presented
Materials throughout the period.
The materials for this experiment were Following the study phase, the experi-
drawn from the same norms as thoseo f Ex- menter gave instructions for the test phase
periment 1. However, 60 new category ex- for about 2 min. Subjects were told that in
emplars were selected so that they were the upcomingp haseo f the experiment, they
each 7 letters in length to help equate the would help researchersd evelop materials
difficulty of items on the anagrams olution for a future study. Subjects were given test
test. The materialsw ere againc hosent o be booklets that either contained a set of 70
low frequency associateso f the category word fragments, 70 category names, or 70
names. A complete set of the materials ap- anagrams,i n each case 60 target items (48
pears in Appendix B. studied, 12 not studied) and 10 filler items.
Twelve items were randomly assignedt o Thus, the ratio of studied items to nonstud-
five different sets, and these were rotated ied items was 48:22f or all three tests. Sub-
through all the study conditions. This re- jects were given 30 s for each item in the
sulted in five different study lists for com- test booklet. Depending on the task, they
plete counterbalancing. In addition, be- were required to solve either the word frag-
cause the order of the orienting task was ment or anagrama s quickly as possible, or
varied, 10 different study lists were cre- to write down as many exemplars to the
ated. category name as possible. The data were
then scored for the proportion of correct
Procedure
responsesp roduced on the test as a func-
Subjects were tested in small groups of tion of the study condition.
3-7 subjects, with a total of 50 subjects as-
signed to each of the three test conditions. Results and Discussion
In the study phase, subjects were required An overview of the results of Experiment
to read and hear a set of 48 target words. 2 is presentedi n Table 2, which shows the