Distinguishing true from false memories via lexical decision as a perceptual implicit test

Abstract

Studying a list of associated words (holiday, beach, sun, etc.) produces true memory for on-list items (beach), but also false memory for a nonpresented lure (vacation). I suggest that, because only the true item has been physically presented, true and false memories should be distinguishable if the retrieval task accesses purely perceptual information. This is supported using lexical decision as a perceptual implicit test: at a 3 – 10 min delay, repetition priming was found for physically-presented targets, but there was no semantic priming for lures. This was despite strong false memories for lures in explicit recognition. Given previous findings of lure priming in stem-completion, I argue that to avoid false memories the task must be perceptual, implicit, and produce fast responses.

Distinguishing true from false memories via lexical decision as a perceptual implicit test

In everyday life it is often difficult to distinguish something that really happened from something that might easily have happened, but did not. This confusion between true memories and gist-consistent false memories has been extensively studied in the laboratory using a simple list-learning paradigm. Deese (1959) and Roediger and McDermott (1995) presented lists of related words at study (e.g., holiday, resort, sun, beach, etc.). At test they found high rates of false memory for a lure (vacation) that was implied by the study list items but was never actually presented. Further studies have shown that false memory for the lures is remarkably similar to true memory for on-list items in a number of ways: For example, subjects are highly confident of their false recognitions (e.g., Lampinen, Neuschatz & Payne, 1999), claim to remember the specific event of studying the false item rather than just knowing it was there (remember – know procedure of Tulving, 1985; Roediger & McDermott, 1995), and are willing to assign a false source to the item (e.g., male or female voice at study; Payne, Elie, Blackwell & Neuschatz, 1996).

The issue addressed in the present article is whether it is possible to distinguish true from false memories by finding a test that reveals true memory but not false memory. I suggest that one reason why false memory rates are usually so high is the use of conceptual tasks (recall and recognition) as the standard retrieval tests. Conceptual tasks (Roediger & McDermott, 1993) are sensitive to meaning rather than form (as a simple example, it is easy to recall the meaning of a conversation but hard to recall the specific wording; cf. Anderson, 1974). Thus, high false memory rates would be expected because these tasks tap semantic attributes of traces left by the study list items, with which the lure is highly confusable. In contrast, the use of perceptual tasks, that is, tasks that require reference to form or structure rather than meaning, offers the possibility of reliably distinguishing true from false memories. In the Deese – Roediger – McDermott (DRM) paradigm, both on-list targets and lures are likely to have received strong semantic activation at study (Underwood, 1965); perceptual processing, however, will have occurred only for items that were physically presented (i.e., only for true items). Thus, if the test task retrieves purely perceptual information, this task might show memory for on-list items but elicit no memory for lures.

In addition to the conceptual/perceptual distinction, a further complication is whether the retrieval task is explicit or implicit (Graf & Schacter, 1985). In explicit retrieval, the subject makes conscious and deliberate efforts to recollect the study-phase items (the classic explicit test is free recall). Implicit retrieval, however, does not depend on deliberate reference to the original list. Instead, an influence of prior exposure is revealed indirectly via priming on some behavioural measure (e.g., faster reaction times). If the intention is to assess false memory with a purely perceptual task, then it is important that this task also be implicit. This is because, in explicit tasks, the subject can actively provide themselves with cues to assist their search of memory, and such cues are commonly semantic in nature (e.g., Pressley, 1982).

Two tasks usually classified as “perceptual implicit” are stem-completion and lexical decision. In the stem-completion task, subjects are instructed to complete a stem (e.g., bea____) with the first word that comes to mind, and priming is reflected in higher completion rates with the target word (beach) if this word was seen in the study phase than if it was not. In the lexical decision task, subjects make speeded word – nonword decisions, and priming is reflected in faster reaction times when the target was studied rather than unstudied. As measures of memory, these tasks are considered implicit because they produce priming that is usually independent of explicit recollection (e.g., intact priming in amnesia; Moscovitch, 1982). The evidence that they are perceptual is that (a) the tasks overtly require reference only to orthographic knowledge (i.e., spelling structure of familiar words) rather than to word meaning; and (b) on both tasks, priming is not enhanced by elaborative semantic processing at study (Roediger & McDermott, 1993), and is strongly reduced by study-test mismatch in perceptual modality (auditory vs. visual; e.g., Kirsner & Smith, 1974; Rajaram & Roediger, 1993).

The repetition priming described here in both these tasks corresponds to “true memory” – that is, it reflects the influence of an item that was actually presented at study. “False memory”, on the other hand, would correspond to long-lived semantic priming, defined as an influence of the list items seen at study on processing for the lure at test (e.g., as revealed by faster reaction times for lures from studied lists than unstudied lists). If the logic developed earlier is correct – namely that perceptual implicit tests can discriminate true from false memories – then the prediction would be of normal repetition priming for on-list targets (strong “true memory”), but no semantic priming for lures (no “false memory”).

On the face of it, these predictions might be expected to hold equally well for stem-completion and for lexical decision. Unexpectedly, however, the literature to date suggests that the two tasks might produce very different patterns. In stem-completion, McDermott (1997) and McKone and Murphy (2000) both reported semantic priming for lures at study-test delays of 5 – 10 min, using the standard DRM paradigm. In McDermott's original study (McDermott, 1997) the effect could perhaps have arisen from explicit contamination on the stem-completion test, but this problem was carefully avoided in McKone and Murphy (2000); the latter study employed posttest questionnaires to remove subjects who admitted using explicit retrieval strategies, and also demonstrated two empirical dissociations between memory on the explicit and implicit tasks (one based on modality and one based on number of list presentations). In lexical decision, in contrast, semantic priming is usually extremely short-lived. Priming from a single associate (e.g., doctor on one trial reduces reaction times to nurse on the next; Meyer, Schvaneveldt & Ruddy, 1975) usually survives at most one intervening item between prime and target (approx. 2 s; Dannenbring & Briand, 1982; Davelaar & Coltheart, 1975; Gough, Alford & Holley-Wilcox, 1981; Masson, 1995; Meyer et al., 1975; Neely, 1977). Similarly, semantic priming from social stereotypes (e.g., Asian primes rice) is found in lexical decision with immediate prime-target presentation (Wittenbrink, Judd & Park, 1997), but has not been reported at long delays. Finally, a recent study using the DRM paradigm (delays of 5 + min) found no semantic priming in lexical decision despite using study lists of 10 (English) or 12 (Dutch) related primes for each lure (Zeelenberg & Pecher, 2002). These results suggest that, despite the usual classification of both stem-completion and lexical decision as perceptual implicit tasks, lexical decision may in fact provide purer access to perceptual records of the encoding phase.

The aim of the present study was to assess semantic priming for lures in the lexical decision task, using the same items and procedures that McKone and Murphy (2000) previously used to demonstrate such priming in stem-completion. In McKone and Murphy (2000), we used lists of 15 related primes. Semantic priming (“false memory”) was calculated by taking the proportion of stems completed with lures from studied lists (note that the lure itself was never studied), and subtracting a baseline provided by the proportion of stems completed with lures from unstudied lists. Repetition priming (“true memory”), was calculated in a similar manner for a set of on-list targets – one from each list – that were matched to the lures on several variables (e.g., word frequency). With items matched in this manner, we found that semantic priming for the lures (completion proportion = .18) was not only significant (p < .005), but in fact as strong as repetition priming for the list targets (completion proportion = .17). A separate explicit memory test (stem-cued recall) also confirmed the expected high false memory rates for lures (recall proportion = .38).

In the present experiment the study-phase stimuli and procedure were exactly as in McKone and Murphy (2000; Experiment 1), and the study – test delay was similar. The only difference was in the change to a lexical decision task at test. “Standard” lexical decision procedures were used: That is, the nonwords were word-like, but not so word-like as to make lexical decision responses unusually difficult. Under these conditions, relatively fast reaction times (e.g., 550 – 650 ms) would be expected, arguing that lexical decisions would rely primarily on perceptual (specifically orthographic) information, rather than needing to wait for semantic information to come on line later in processing (cf. Joordens & Becker, 1997). The hope was that, under these circumstances, lexical decision would be able to reliably distinguish items with a recent perceptual record (i.e., true memories) from items that were merely implied in the study phase (i.e., false memories).

Method

Subjects

First-year psychology students of the Australian National University participated during scheduled laboratory sessions. Data were available from 73 subjects who had English as a first language and, prior to the experiment, indicated that their results could be kept for research purposes.

Design

Type of target (lure vs. list item) and list status (studied vs. unstudied) were varied within subjects. Sixteen lists were divided into two sets of eight (Set A and Set B), for which studied versus nonstudied status was counterbalanced across subjects (52% of subjects studied Set A). In the test phase, all subjects first completed the lexical decision task (perceptual implicit test), followed by a standard old – new recognition task included to confirm the presence of false memory for the lures on an explicit conceptual test.

Materials

Materials were taken from McKone and Murphy (2000; stimuli are available from the author). Each list comprised a lure and 15 associates. Eleven lists were Australianised versions of Roediger and McDermott's lists (Roediger & McDermott, 1995, Experiment 2); a further five had been developed in a similar manner. Items had been selected to satisfy various specific constraints of the stem-completion task (see McKone & Murphy, 2000, for details). The 16 list targets and 16 lures were matched on length (M = 5.4 vs. 5.8 letters, respectively) number of syllables (M = 1.7 vs. 1.7), and word frequency (M = 78, SD = 73 vs. M = 84, SD = 97 occurences per million; Kucera & Francis, 1967); the eight items of each type then assigned to Set A and Set B were also closely matched. Note that word frequency is the most important variable affecting unprimed reaction times in lexical decision (e.g., explaining 26% of inter-item variance, Balota, Cortese & Pilotti, 1999), and also substantially affects repetition priming in lexical decision (e.g., Kirsner & Speelman, 1996).

Word stimuli for the lexical decision test included all 32 items of interest: eight list targets from studied lists (i.e., actually seen items), eight lures from studied lists (i.e., items implied by those actually seen), plus eight list targets and eight lures from unstudied lists (i.e., baseline items unrelated to any words studied by that subject). Nonword stimuli comprised 32 items such as feath and revont. These obeyed spelling and phonetic rules of English, and were matched to the real words on number of letters and syllables; they did not, however, include deliberate misspellings of particular real words (e.g., trian), or homophones of real words (e.g., brane).

Procedure

At study, each list was presented on a separate sheet of paper, and the subject was allowed to see only one word at a time through the window of a covering sheet. Subjects were given 1.5 s to learn each word, and a period of 30 s after each list to mentally review the list. Each subject studied eight lists in turn in this way, presented in one of two orders. Within each list, the order in which items occurred was the same for all subjects, with the strongest associates of the lure generally occurring first. List targets did not appear in either the first two or the last two positions of each study list.

Following study, up to 10 students entered individual cubicles to perform the lexical decision task, while any additional students in the class waited quietly. This first group tested (n = 46) had a delay of approximately 3 min between completing the study phase and beginning the test phase (cf. 5 min in McKone & Murphy, 2000, Experiment 1); the delay for the second group to enter the cubicles was approximately 10 min (n = 27). The lexical decision task was presented on IBM compatible PCs (286 model) with reaction times measured through the keyboard and qwktimer software routine. Each item remained on the screen until the subject responded word or nonword, as quickly as possible consistent with being correct. The intertrial interval was 1 s. There were 10 warm-up trials. The 64 experimental trials (32 words, 32 nonwords) were then presented in a different random order for each subject, equating mean delay after the study phase for lures and list targets. The lure and list target from a given list (e.g., vacation and beach) never appeared successively; this ensured that any semantic priming revealed could be attributed to a long-term influence of the 15 primes on the study list, rather than a short-term effect of a single associated prime.

As each group of subjects emerged from the cubicles, they moved on immediately to the explicit old – new recognition task. The 32 target words were printed on a single page in one of four random orders. Instructions were to circle “old” words seen on the original study lists, and only those words. (Subjects were warned to be careful because all words had been seen in the intervening lexical decision task.) Five minutes were allowed to complete the recognition test.

Results

Results are presented in Tables 1 and 2. Analyses of Variance indicated that study-test delay had no influence on memory in either the recognition or lexical decision tasks (ps > .1 for main effects and all interactions). All further statistical analyses were therefore conducted collapsed across the 3-min and 10-min delay conditions (to give the most power, with N = 73), although Tables 1 and 2 include each delay separately for completeness.

Table 1. Percentage of old responses in the explicit old – new recognition task

	List targets (true memory)			Lures (false memory)
	Studied lists	Unstudied lists	S – U	Studied lists	Unstudied lists	S – U
All subjects	77.9	10.6	67.3**	73.8	9.1	64.7**
(N = 73)	(2.0)	(2.0)	(2.7)	(3.0)	(1.6)	(3.3)
3-min delay	79.9	10.3	69.6	73.6	8.2	65.5
(n = 46)	(2.3)	(2.2)	(2.8)	(4.0)	(1.6)	(4.3)
10-min delay	74.5	11.1	63.4	74.1	10.7	63.4
(n = 27)	(3.7)	(4.1)	(5.7)	(4.6)	(3.3)	(5.2)

Note. Means and SEMs (in parentheses) are shown for studied lists (S) and unstudied lists (U). Memory scores are also shown, in each case calculated as S scores – U scores. For comparison of memory scores to zero, *p < .05, **p < .001 two-tailed. Significance tests were conducted only for the “all subjects” data.

Table 2. Reaction times (in ms) in the lexical decision task

	List targets (repetition priming)			Lures (semantic priming)
	Studied lists	Unstudied lists	U – S	Studied lists	Unstudied lists	U – S
All subjects	565	582	17.3*	568	563	– 5.0ns
(N = 73)	(9)	(10)	(8.4)	(8)	(8)	(6.8)
3-min delay	572	591	18.8	578	570	– 7.6
(n = 46)	(11)	(12)	(10.0)	(10)	(8)	(7.5)
10-min delay	553	568	14.8	551	551	– 0.4
(n = 27)	(14)	(16)	(15.5)	(15)	(16)	(13.4)

Note. Means and SEMs (in parentheses) are shown for studied lists (S) and unstudied lists (U). Memory scores are also shown, in each case calculated as S scores – U scores. In calculating lexical decision means, reaction times were excluded if they fell outside the range 300 – 1250 ms, or if the response was incorrect; mean error rates were 2.8 – 3.5% for word conditions, and 6.2% for nonwords. For comparison of memory scores to zero, *p < .05, **p < .001 two-tailed. Significance tests were conducted only for the “all subjects” data.

Old-new recognition

Table 1 shows percentages of old responses in the recognition memory task. Old responses for items from unstudied lists represent genuine “false alarms”, unrelated to any words seen at study. As expected, these rates were low and did not differ for list targets (10.6%) and lures (9.1%), t(72) = 1.03, p > .2. When memory was then calculated as the difference between studied lists and unstudied lists, true recognition for list targets was high (67.3%), as was false recognition for lures (64.7%). This replicates the standard finding of very high false memory rates on explicit recognition (Roediger & McDermott, 1995). Moreover, the present results replicate our earlier finding (stem-cued recall in McKone & Murphy, 2000) that, for these particular lists, explicit false and true memory rates do not differ, t < 1.

Lexical decision

As shown in Table 2, a very different pattern of results emerged in the lexical decision task. Baseline scores from unstudied lists were reasonably similar for list targets (582 ms) and lures (563 ms). Also as expected, significant repetition priming (“true memory”) was obtained, with reaction times for list targets being 17.3 ms faster for studied lists than for unstudied lists, t(72) = 2.06, p < .05. (Note that the amount of this priming is approximately what would be expected given the mean word frequency of the items, cf. Kirsner & Speelman, 1996).

The new result, however, was that there was no evidence at all of semantic priming (“false memory”) in the lexical decision task. Responses to lures from studied lists were no faster than responses to lures from unstudied lists, t < 1; in fact, the effect was in slightly the wrong direction, M = – 5.0 ms, and calculation of a 95% confidence interval (CI) on the priming value ( – 18.3 ms to + 8.3 ms) indicated that anything other than a trivial amount of positive priming for lures was very unlikely. Even if only the shorter delay condition (3 min) is considered, there was still no evidence of priming for lures (M = – 7.6 ms, 95%CI – 22.3 ms to + 7.1 ms). Finally, comparison of lures with list targets confirmed that the amount of semantic priming was significantly different from the amount of repetition priming, t(72) = 2.00, p < .05. Together, these results are consistent with the claim that lexical decision is able to discriminate between items that possess a perceptual record from the study phase (“true” items) and those that do not (“false” items).

Alternative interpretations of the lack of semantic priming?

Two analyses were conducted to discount alternative explanations of the lack of any priming effect for lures. First, in lexical decision, repetition priming is highly sensitive to word frequency: priming is strongest for low frequency words, while very high frequency words sometimes show no priming at all (Kirsner & Speelman, 1996). If semantic priming were to show the same effect, then the overall finding of no priming for lures might be hiding priming for a subset of low frequency items. However, this was not the case. Table 3 shows the lexical decision data broken down into three frequency bands. For list targets, the usual pattern of frequency effects was obtained: low frequency words showed plenty of priming (40 ms); medium frequency words showed an effect half this size (21 ms); and high frequency words showed no priming at all ( – 12 ms). For lures, in contrast, there was no suggestion of priming for items in any frequency band; specifically, even low frequency lures showed no priming ( – 6.3 ms).

Table 3. Lexical decision results (reaction times in ms) broken down by word frequency

	Frequency (/million)	No. items per band	Studied lists	Unstudied lists	U – S	t
List targets	Low (1 – 20)	3	598 (11)	639 (13)	40.2	2.40*
	Med (33 – 75)	8	553 (8)	573 (19)	20.9	1.77^†
	High (100 – 265)	5	546 (10)	534 (11)	– 12.0	< 1
Lures	Low (1 – 20)	5	580 (12)	574 (12)	– 6.3	< 1
	Med (33 – 75)	7	567 (8)	556 (7)	– 11.3	– 1.11
	High (100 – 265)	4	563 (10)	567 (11)	4.1	< 1

Note. SEMs are given in parentheses. Comparison of studied and baseline means is independent samples (df in range 214 – 571), and no SEM on priming can be calculated. This is because, when means are calculated for individual items, error responses lead to missing data. Thus, rather than average on a per-subject basis prior to collapsing over subjects, all valid RTs for items in a given frequency band were averaged regardless of which subject they came from. For comparison of studied and unstudied conditions: *p < .05 two-tailed; ^†significant at p < .05 one-tailed.

A second issue was whether the lack of semantic priming might be attributable to overly fast reaction times (i.e., a ceiling effect). This seemed unlikely because repetition priming effects emerge even on naming tasks, where reaction times are quite a lot faster than in the present experiment (e.g., 495 ms for low frequency words in McKone, 1995). To be cautious, however, I selected the four lure items that produced the slowest baseline reaction times. Mean unstudied reaction time (RT) for these items (603 ms) was noticeably slower than for even the medium frequency list targets (573 ms), but there was still no semantic priming for these “slow” lures ( – 1 ms).

Discussion

The present experiment has clearly dissociated true from false memories in the Deese associative-list paradigm. The old – new recognition task confirmed high rates of false explicit memory. The lexical decision task, however, showed repetition priming for on-list targets (i.e., true memory), but no associative priming for lures (i.e., no false memory), even for low frequency lure items. This result is consistent with the logic that an implicit memory task that is purely perceptual in nature should show priming only for items that left a perceptual trace at study, and not for items that were not physically presented.

In using the same materials and study procedures as McKone and Murphy (2000), the present study has also confirmed that stem-completion and lexical decision produce different patterns of semantic priming under closely matched conditions. Of these two tasks, only lexical decision dissociated true and false memories: stem-completion, in contrast, produced priming for lures that was fully as strong as priming for list targets. There are two possible explanations of why this is so. First, it could be that our earlier stem-completion results simply reflected explicit contamination on the supposedly implicit task. As described in McKone and Murphy (2000, pp. 103 – 104), however, there were several good reasons for believing that this was not the case.

Alternatively, a more theoretically interesting interpretation is possible. An important fact about stem-completion responses is that they are slow, with subjects taking approximately 1 s to generate a completion even when speed is strongly emphasised (Horton, Wilson & Evans, 2001). When responses take this long, there is plenty of time for the stimulus (the stem) to activate the full range of knowledge about words, including orthographic information (likely spellings), phonological information (possible pronunciations) and, critically, semantic information (word meanings). This is not to say that the final response on a stem-completion task is based on conceptual information; indeed, in Experiment 2 of McKone and Murphy (2000) we showed that the semantic priming for lures in fact disappeared when the study modality no longer matched the test modality (ie. auditory-list – visual-lure, rather than visual-list – visual-lure). This suggests a rather complex picture in which, for example, initial activation of orthographic information from the stem feeds forward to the semantic level, where it interacts with semantic activation of the lure left over from the study phase, and then feeds back to strengthen the orthographic representation of the lure as opposed to alternative completions of the stem.

In lexical decision, in contrast, responses are fast. Whenever a task requires a speeded decision, it would be expected that the first useful information to become available will drive performance (Joordens & Becker, 1997). For lexical decision, words and nonwords can usually be discriminated at the orthographic level (i.e., in terms of whether the item has a familiar spelling or not), without having to wait for semantic information to become available later in processing. This leads to fast reaction times. More importantly in the present context, it also produces a response with a very strong focus on perceptual, rather than conceptual, information.

As concrete evidence for the importance of fast responses, several word recognition studies have demonstrated that the way in which items are processed in lexical decision depends heavily on the difficulty of the decision. The “standard” lexical decision procedures as employed here produce fast reaction times consistent with easy decisions based on first-stage (i.e., orthographic) processing alone. It is also possible, however, to tweak the stimulus set to encourage more semantically based lexical decisions. For example, adding very unfamiliar real words or using extremely wordlike nonwords (e.g., pseudohomophones such as brane) increases reaction times for all items on the list (Balota & Chumbley, 1984). In a clear demonstration that this increases semantic contributions to lexical decision, Joordens and Becker (1997) used these types of manipulations to extend semantic priming (using a single prime) out to eight intervening items (10 s), as opposed to the usual 0 – 1 intervening items in this task.

These arguments imply that a task being classified simply as “perceptual implicit” is not sufficient to ensure that it will discriminate false memories from true memories. Instead, I suggest that three criteria must be met, all of which are important in ensuring that the retrieval test is purely perceptual in nature. First, the task must have an instrinsic focus on perceptual rather than conceptual aspects of the stimulus. Second, the task must involve implicit rather than explicit retrieval, to avoid the subject altering this instrinsic focus by strategic means. And third, the task must produce fast responses, to maximise reliance on “easy” first-stage perceptual processing, rather than later-stage semantic information.

Finally, a comment on the possible relevance of the present findings to more applied settings is in order. The critical question in most applied settings – such as eyewitness testimony – is the truth or falsity of one specific memory of one specific person. In the present experiment, the lexical decision task has been able to discriminate reliably between false and true memories, but only averaging over eight words per subject per condition, and then over 73 subjects. Obviously, the lexical decision task will not be useful in allowing a reliable decision about whether any one response (i.e., one item from one subject) is a true or false memory. On the other hand, the present study has at least demonstrated that true memories and false memories can be distinguished on the basis of the way in which memory of a prior event is retrieved (also see Reysen & Nairne, 2002). A common assumption is that it is encoding that is responsible for “producing” false memories (e.g., in the DRM paradigm, see Smith & Hunt, 1998, or Underwood, 1965): this somehow implies that, once encoding is over, it is all too late for telling a false memory apart from a true one. However, the present results indicate that, like all memory, false memory is in fact based on an interaction between encoding and retrieval. Specifically, I have argued that true and false memories are differently encoded at study – both leave strong semantic traces after the study phase, but only true events leave perceptual traces – and that this difference can be identified later as long as the retrieval task taps perceptual, rather than conceptual, aspects of the original event.

Acknowledgments

Supported by Australian Research Council Small Grant F99006. Thanks to John Brown for programming and initial data extraction, and the 1999 PSYC1001 tutors for conducting the experiment.