The awareness of novelty for strangely familiar words: a laboratory analogue of the déjà vu experience

Participants

Thirty English-speaking participants (20 female, 10 male; mean age = 24.1 years, SD = 6.5 years) were tested and reimbursed at a rate of £5/h for their time. Written consent was obtained from all participants. The protocol was approved by the University Teaching and Research Ethics Committee at the University of St Andrews (approval number PS10697).

Stimuli and materials

Over the course of the experiment, each participant was presented with 24 DRM word lists based on Stadler, Rodiger & McDermott’s norms (1999). The 24 lists comprised the 12 which yielded the highest false alarm rates for critical lures and the 12 useable lists which yielded the lowest false alarm rates. For each study list, 12 words were randomly selected from the 15 words published per list in Stadler et al. For each test list, 3 old words (targets; selected from the previously studied 12 words) were presented alongside 2 semantically unrelated new words (unrelated lures), 2 semantically related new words (related lures; these were randomly selected from the 3 words excluded from study presentation) and the single critical lure (see Appendix S1 for four example study-test lists). All word lists were presented on PCs running MATLAB (The MathWorks Inc., Natick, MA, 2000) and Psychophysics Toolbox (Brainard, 1997).

Two paper questionnaires were administered: a pre-experimental questionnaire; and a post-experimental questionnaire. The pre-experimental questionnaire was completed before the first study-test block and assessed participants’ previous déjà vu experience (yes/no response) and frequency (<1, 1–4 or >4, times a year). There followed an explanation of déjà vu based on Brown’s (2004) definition, “a feeling of familiarity coupled with the knowledge that this familiarity is incorrect”, and an open-ended question requested that participants write a short passage summing up previous déjà vu experiences according to the following instructions: “Please provide a short account of a ‘typical’ déjà vu experience you have had. Try to include some detail concerning the following points: Emotional intensity of a typical déjà vu experience; Duration of a typical déjà vu experience; How a déjà vu experience might typically make you feel about the reliability of your memory”. The post-experimental questionnaire, completed at the end of the experiment, confirmed participants’ experiences of déjà vu during the experiment (yes/no response) and was again followed by an open-ended question requesting a summary of the experimentally-generated experience of déjà vu according to the following instructions: “Please provide a short account of your déjà vu experience(s) during the experiment. Try to include some detail concerning the following points: Emotional intensity of déjà vu experience(s); Duration of déjà vu experience(s); How the déjà vu experience(s) made you feel about the reliability of your memory”.

Design and procedure

We manipulated subjective familiarity by presenting participants with DRM lists yielding the highest or the lowest critical lure false alarm rates according to Stadler, Rodiger & McDermott (1999). Lists yielding high false alarm rates were used in the high familiarity conditions. Lists yielding low false alarm rates were used in the low familiarity conditions. We manipulated objective novelty by varying the string that participants monitored study words for. For each study list, participants were presented with a new 1–3 character string, and indicated the number of words which began with this string once they had seen all 12 words. For the high novelty lists, no words in the study list and only the critical lure in the test list began with the character string. For the low novelty lists, at least one word in the study list (and not the critical lure) began with the character string. During each test list, the number of words indicated by participants as beginning with the string was re-presented to participants, highlighting the novelty of critical lures in only the high novelty condition. Thus, there were two within-subjects list-level factors, novelty (high, low) and familiarity (high, low). These combined to produce four types of list: (1) high novelty-high familiarity; (2) high novelty-low familiarity; (3) low novelty-high familiarity; and (4) low novelty-low familiarity. There was also one within-subjects word-level factor with four levels: targets (previously studied words), unrelated lures (previously unstudied words which were semantically unrelated to the studied words), related lures (previously unstudied words but which were drawn from the same DRM list), and critical lures (previously unstudied words to which all the studied words were semantically related). All four word conditions were presented within each test list.

Over the course of the experiment, six study-test list pairs of each list type (high novelty-high familiarity etc.) were presented (24 study-test list pairs in total). To allow participants to rest periodically, the experiment was split into four study-test blocks. The list composition of each block was randomly assigned such that participants would not inevitably encounter each list type in each study-test block. Each study-test block comprised six consecutive study lists, followed by six consecutive test lists. For each participant, corresponding study and test lists were presented in the same order (i.e., study1, study2, study3, study4, study5, study6, test1, test2, test3, test4, test5, test6).

Figure 1 shows a schematic of a study list and its corresponding test list. The to-be-monitored character string was presented in size 48pt font alongside word stimuli for the duration of each study list. Twelve words were serially presented in size 48pt font, in the centre of the screen, for 3 s each. At the end of the study list, participants were prompted to register the number of words beginning with the character string in that list (0–5 indicated using the keyboard).

Throughout each test list, the previously-monitored character string and the number indicated by the participant at the end of the study list were presented in size 48pt font, at the top of the screen. Eight words (three targets, five lures) were then serially presented in size 48pt font, in the centre of the screen. Test words were presented in a pseudorandom order modelled on the procedure used by Roediger & McDermott (1995). A target always occupied test position 1 and the critical lure always occupied test position 6, 7 or 8. (The three targets selected comprised the word from study position 1, one word selected at random from study position 2–6 and one from study position 7–12.) The remaining targets and lures were allocated at random to the unoccupied test positions. Below each word, the prompt, “old(b) or new(n)?”, presented in size 48pt font, cued participants to indicate whether the word was previously presented at study or not. Once a response had been made, a new prompt, “Confidence: L(1), M(2), H(3)?”, cued participants to indicate their confidence in the previous decision using. All test responses were self-paced and responses were made using the keyboard keys listed in parentheses.

In addition to the old/new and confidence judgments collected for each test stimulus, we also assessed the occurrence of a déjà vu experience and its intensity using an on-screen toggle system. This response system avoided unnecessarily asking participants about their déjà vu experiences repeatedly, leaving them free to report the experience only when it arose. A déjà vu status bar, located at the bottom of the screen had the default status “deja vu(d): none”. Participants reported the occurrence of déjà vu by pressing the ‘d’ key, which would cycle through the intensity options. On pressing it once, the status would change to “deja vu(d): low”. Pressing it again would result in a status change to “deja vu(d): medium” and a third time, to “deja vu(d): high”. If the ‘d’ key was pressed a fourth time, this would restart the cycle at “deja vu(d): none”. Participants could indicate the occurrence of déjà vu at any point during each trial (i.e., during the self-paced windows for old/new and confidence responding). The status reverted to the default “deja vu(d): none” at the start of each new test word presentation. When participants had completed four study-test blocks they completed the post-experimental questionnaire. The entire procedure lasted no longer than one hour for each participant.

n-gram analysis of déjà vu descriptions

n-grams are continuous sequences of n words found to occur within a passage of prose. n-grams with n = 1 are referred to as unigrams, and those with n = 2 as bigrams. Examples of unigrams from within this sentence are “sentence” and “from”, whereas examples of bigrams from within this sentence are “of unigrams” and “are sentence”. We identified differences in the strings of words used to describe naturalistic and experimentally-generated déjà vu by conducting a rudimentary n-gram analysis. The procedures reported here largely mirror those reported in Selmeczy & Dobbins (2014). Prose passages from the pre-experimental and post-experimental questionnaires underwent identical preparation for n-gram analysis: spelling errors were corrected; contractions were completed (e.g., “don’t” became “do not”); symbols with known meanings were written out in full (e.g., “/” became “or”, “=” became “equals”); and “deja vu” replaced with “dejavu”.

Separately for unigrams and bigrams, we counted the number of times each n-gram appeared in the pre- and post-experimental passages. For n-grams with total occurrences (N) across both passages of at least the median (unigrams: 5, bigrams: 4), we used N, the number of occurrences in the pre-experimental passage, and an assumed binomial-p parameter of .5 to calculate a binomial distribution z value and corresponding p value of each n-gram. As we were interested in n-grams which differentiated the two accounts, we set an uncorrected p threshold of .05 and tabulated these n-grams for examination.

Accuracy

To establish that the DRM procedure was indeed generating erroneous familiarity for critical lures, we conducted a one-way repeated measures ANOVA on accuracy according to word condition (critical lure, related lure, unrelated lure, target), collapsed across list conditions. Assumptions of sphericity were violated, χ²(5) = 11.63, p = .040, therefore degrees of freedom were corrected with Greenhouse-Geisser estimates of sphericity, using ε = .819. The effect of word condition on accuracy was significant, F(2.46, 71.28) = 47.43, p < .001, ${\eta }_p^2=.621$, with accuracy lowest for critical lures, as anticipated (see Table 1 for descriptives).

To check whether our list manipulations influenced levels of erroneous familiarity, we next conducted a 2 x (novelty: high, low) x 2 (familiarity: high, low) repeated measures ANOVA on critical lure accuracy. There was a significant main effect of novelty, F(1, 29) = 24.23, p < .001, ${\eta }_p^2=.455$, such that high novelty critical lures, M = .714 [.617, .811], were more accurately responded to than low novelty critical lures, M = .414 [.322, .506]. There was also a significant main effect of familiarity, F(1, 29) = 51.11, p < .001, ${\eta }_p^2=.638$, such that high familiarity critical lures, M = .475 [.394, .556], were less accurately responded to than low familiarity critical lures, M = .653 [.583, .722]. Finally, there was a significant interaction between novelty and familiarity, F(1, 29) = 11.45, p = .002, ${\eta }_p^2=.283$. Focusing on the main effects, it is evident that both list manipulations had the intended effects on levels of erroneous familiarity generated for critical lures—increased novelty salience decreased erroneous responding by drawing participants’ attention to the objective novelty of the critical lure, whilst the lists selected from the Stadler, Rodiger & McDermott (1999) norms for their elevated false alarm rates also demonstrated elevated false alarms in the current procedure.

Déjà vu frequency

Déjà vu was reported at least once by 18 of the 30 participants (60%). Across all participants, the mean number of déjà vu reports was 12.83 [5.85, 19.82]. This value rose to 21.39 [11.41, 31.37] in the subsample who reported at least one déjà vu. In the following analyses, we analyse déjà vu frequency across the whole sample.

Déjà vu occurrence was assessed on a trial by trial basis. We were therefore able to calculate the likelihood with which a word from each condition would yield a déjà vu report (see Table 1 and Fig. 2A). Déjà vu frequency, as a proportion of all words presented within the given combinations of conditions, was assessed in a 2 (novelty) x 2 (familiarity) x 4 (word) repeated measures ANOVA. Assumptions of sphericity were violated for the main effect of word, χ²(5) = 31.17, p < .001, and the interactions between word x novelty, χ²(5) = 92.60, p < .001, word x familiarity, χ²(5) = 54.54, p < .001, and word x novelty x familiarity, χ²(5) = 33.49, p < .001. Degrees of freedom for these effects were corrected using Greenhouse-Geisser estimates of sphericity, using ε = .564, ε = .413, ε = .482 and ε = .555 respectively. Across all effects involving familiarity however, there were no significant differences, all ps > .350, suggesting that our manipulation of DRM strength did not influence déjà vu responses independently of the other factors. We therefore present the remaining effects involving the novelty and word conditions below.

There was a significant main effect of novelty, F(1, 29) = 8.05, p = .008, ${\eta }_p^2=.217$, with a greater frequency of déjà vu reports under high novelty, M = .097 [.051, .142], than low novelty, M = .064 [.027, .101]. There was also a significant main effect of word, F(1.69, 49.10) = 10.28, p < .001, ${\eta }_p^2=.262$, driven by the high frequency of déjà vu reports for critical lures (see Table 1 for descriptives).¹ Both of these findings are consistent with our hypotheses. Although there was no graded response according to familiarity condition, the lists contriving salient novelty generated the most déjà vu reports, likely driven by the greatest conflict between DRM-induced familiarity and novelty. Consistent with this interpretation, we were also able to show that déjà vu was reported more for critical lures than any other word condition.

The novelty x word interaction was also significant, and likely responsible for both main effects presented above, F(1.24, 35.95) = 6.45, p = .011, ${\eta }_p^2=.182$. Figure 2 illustrates the homogeneity of responding within word conditions, which is broken only for critical lures. Crucially, critical lures in the low novelty condition remained comparable to other word conditions in their likelihood of yielding déjà vu reports, in the region of 10%, whereas those in the high novelty condition elicited déjà vu responding around 25% of the time. The presence of salient novelty, whereby participants were made aware that stimuli which they otherwise found to be familiar could not be so, appears important in elevating categorical reports of déjà vu within this procedure.

Our hypothesised elevation of déjà vu responses in maximal clash conditions was predicated upon participants correctly identifying critical lures as objectively new. To establish that the described pattern of déjà vu responses persisted in items to which participants had given correct recognition responses but not incorrect recognition responses, we recalculated déjà vu response likelihoods according to recognition response correctness (Fig. 2B). Given the ineffectiveness of the familiarity manipulation in influencing déjà vu frequency above, we collapsed across familiarity conditions so as to compare only high and low novelty lists by word condition. In a 2 (novelty) x 4 (word) repeated measures ANOVA on déjà vu responses to correctly identified words, assumptions of sphericity were violated for the main effect of word, χ²(5) = 41.30, p < .001, and the interaction, χ²(5) = 105.37, p < .001. Degrees of freedom for these effects were corrected using Greenhouse-Geisser estimates of sphericity, using ε = .565 and ε = .385 respectively. The main effect of novelty was no longer significant, F(1, 26) = 3.63, p = .068, ${\eta }_p^2=.123$, though the main effect of word survived, F(1.70, 44.11) = 8.85, p < .001, ${\eta }_p^2=.241$, again carried by the elevated déjà vu responding to critical lures.² The novelty x word interaction was no longer significant, F(1.16, 30.06) = 2.16, p = .100, ${\eta }_p^2=.077$. In an equivalent ANOVA for incorrect responses there were no main effects of novelty, F(1, 12) = 3.11, p = .103, ${\eta }_p^2=.206$, or word, F < 1, and no significant interaction, F < 1.³ Although the previously described effects are attenuated when split according to recognition response correctness, there is nothing to suggest that the hypothesised elevation in déjà vu responses was driven by responses to critical lures which participants have incorrectly identified as old. There is therefore little indication that the overall findings relating to déjà vu frequency are driven by stimuli to which participants should be reporting no more than baseline levels of déjà vu.

Déjà vu intensity

Déjà vu intensity was measured after the presentation of each word and coded as being rated from 1 (low) to 3 (high). We were interested in whether déjà vu, once reported, differed in intensity according to list condition. We again collapsed across familiarity conditions so as to compare déjà vu intensities across high and low novelty lists. We also restricted our analysis to critical lures, the stimuli in which déjà vu reports were most frequent, to avoid the problem of empty cells decimating the analysis. Fourteen participants reported déjà vu under both novelty conditions. A repeated measures t-test found no significant difference between déjà vu intensity ratings for high novelty critical lures, M = 1.67, [1.31, 2.03], and low novelty critical lures, M = 1.79, [1.38, 2.21], t(13) = −0.69, p = .503 d = −.196. (There were too few participants reporting déjà vu in both novelty conditions to warrant further analysis of the data subdivided according to correct [9 participants] and incorrect [4 participants] recognition responses.) Overall, déjà vu intensity did not vary according to novelty condition.

Déjà vu descriptions

In order to quantitatively explore discursive accounts of naturalistic and experimentally-generated déjà vu experiences, we excluded participants who had never experienced déjà vu or who did not experience déjà vu in the experiment. This left us with a subsample of 15 participants. We conducted n-gram analyses on these participants’ accounts of naturalistic and experimentally-generated déjà vu. Unigrams and bigrams which significantly differentiate naturalistic and experimental experiences of déjà vu are listed in Table 2, wherein positive z values indicate n-grams which were used in the descriptions of previous déjà vu experiences significantly more than the experimental déjà vu experiences. Negative z values indicate n-grams which showed the opposite pattern of correspondence.

A common thread across unigrams and bigrams was the presence of words describing the specificity of the déjà vu experience in question. Experimental reports of déjà vu were characterised by descriptions relating to the stimuli (“word[s]”, “the word[s]”), modality of presentation (“seen”, “had seen”) and the setting (“experiment”, “the experiment”) e.g., “Sometimes, I had a very slight feeling that I had seen a particular word before”. Naturalistic déjà vu experiences were more generalised (“situation”, “experience”) e.g., “Typically a scenario or situation I am in just seems familiar”. In this respect, experimental reports of déjà vu appear to be in response to an experience which is more restricted to certain stimuli within the environment than naturalistic déjà vu experiences.

A related dimension along which there was the suggestion of differentiation was the duration of the experience. “Minutes” was used disproportionately to describe naturalistic experiences e.g., “A typical deja vu experience for me lasts a couple of minutes”. “Seconds” appeared in the unthresholded table, but was not diagnostic of one or other category of experience (naturalistic: 4, experimental: 3, z = .352, p = 1.00) and was used in similar contexts across accounts e.g., naturalistic—“It lasted only a few seconds…,” and experimental—“The deja vu last only a few seconds each time”. In general, it would appear that the experimental experience was restricted to a shorter duration than that to which naturalistic déjà vu experiences can extend, though this generalisation did not fit with all participants’ experiences e.g., the following from a description of the experimental déjà vu: “The duration was longer than previously experienced deja vu but also fainter”.

More generally, these accounts offer insights into the nature of the experimentally-generated experience not afforded by the dichotomous or categorical response options available to them in the experiment itself. Specifically, some participants who used the toggle system to indicate that they had experienced déjà vu were much more cautious about describing the experimental experience as déjà vu when given the opportunity to explain their experience more precisely e.g., “I am unsure whether it was exactly deja vu but in some cases I saw words usually words I was expecting to see in the groups but did not, and it felt as if I had seen them”. and “I am not really sure I had the two deja vu I reported, or if I think I had them only because it was the task of the experiment”. Others reiterated what would be inferred from their categorical responses e.g., “The feeling of deja vu experience is quite strong but I only felt it when certain words came up, when it moved on to the next word… the feeling disappeared…”. These accounts raise the question of whether demand characteristics and response acquiescence are important in influencing categorical responses in studies such as this one. Whilst this may not apply to all participants, there are clearly some participants for whom discursive response options give the experimenter a clearer idea of the inferences they should be making based on participant responses.