The application of subliminal priming in lie detection: Scenario for identification of members of a terrorist ring

Transcription

1 Psychophysiology, 46 (29), Wiley Periodicals, Inc. Printed in the USA. Copyright r 29 Society for Psychophysiological Research DOI: /j x The application of subliminal priming in lie detection: Scenario for identification of members of a terrorist ring MING LUI a,b and J. PETER ROSENFELD a a Department of Psychology, Northwestern University, Evanston, Illinois, USA b Department of Psychology, National University of Singapore, Singapore Abstract We studied a lie detection protocol immune to countermeasures. The 4 stimulus conditions were (1 and 2) supraliminal acquaintance name primed by subliminal acquaintance name (A-A) versus subliminal nonacquaintance name (N-A) and (3 and 4) supraliminal nonacquaintance name primed by subliminal acquaintance name (A-N) versus subliminal nonacquaintance name (N-N). In Experiment 1 and replication, principal components analysis-derived event-related potential components revealed significant differences between dishonestly answered supraliminal acquaintance conditions with differing primes (A-A vs. N-A). In Experiment 2 subjects were required to lie in A-N and N-N conditions, in contrast to Experiment 1, in which subjects lied in A-A and N-A conditions. No significant effects were found. In Experiment 3, the lying task was removed and no significant differences were found. We conclude that subliminal primes modulate ERPs in conditions with supraliminal acquaintance name when the task involves lying. Descriptors: Cognition, Learning/memory, Unconscious processes, EEG/ERP There is a recent growth of cognitive neuroscience studies in deception using different experimental paradig, including mock crime scenarios (Kozel et al., 2; Mohamed et al., 26; Lui & Rosenfeld, 28), autobiographical information (Ganis, Kosslyn, Stose, Thompson, & Yurgelun-Todd, 23; Nunez, Casey, Egner, Hare, & Hirsch, 2; Spence et al., 21), guilty knowledge tests (Langleben et al., 22, 2), and malingering tests (Lee et al., 2). Past studies approached deception by investigating the related cognitive processes, including attention, memory, and response generation processes. For instance, a piece of information that a person intends to lie about (guilty information) is usually more attention catching. And one may involuntarily and automatically retrieve the related contextual memory when perceiving the guilty information. Lying is also supposed to pose more demand on executive control than truth telling (Johnson, Barnhardt, & Zhu, 2). Before lying, one must hold and manipulate competing pieces of truthful and false information in working memory. To give a lying response, one needs to suppress the prepotent truthful response. The inhibition (Falkenstein, Hoormann, Christ, & Hohnsbein, 2; Gehring, Goss, Coles, Meyer, & Donchin, 1993) and monitoring This research was supported by the Department of Defense Polygraph Institute Grants DODP198-P-1 and DoDPI4-P-2 awarded to J. Peter Rosenfeld. We thank Andreas Keil and an anonymous reviewer for excellent suggestions regarding an earlier draft of this report. Address reprint requests to: Dr. Ming Lui, Department of Psychology, National University of Singapore, SG1177, Singapore. m-lui@u.northwestern.edu 889 of response conflicts involves executive control (Botvinick, Nystrom, Fissell, Carer, & Cohen, 1999; Carter et al., 1998). Neuroimaging data supported the involvement of executive control brain regions in deception. In a recent study by Ganis and colleagues, there were stronger activations in anterior prefrontal cortices (bilaterally), parahippocampal gyrus (bilaterally), the right precuneus, and the left cerebellum in the lying compared to truth-telling conditions. Enhanced activations in the anterior cingulate cortex (ACC) and the superior frontal gyrus were also found in another fmri deception study (Langleben et al., 22). It is believed that that there is no specific brain region or electrophysiological marker responsible for lying alone; rather, there is a combination of emotional and cognitive processes related to lying. The present study aimed at developing a paradigm to detect lies by capturing the lie-related memory processes at the individual level. Past ERP Studies on Deception Many previous event-related potential (ERP) studies of deception focused on the P3 component. The P3 is a positivegoing component that occurs between 3 and 8 after stimulus onset. It is an endogenous ERP component related to meaningfulness and rareness of stimuli (Donchin & Coles, 1988). In previous studies of deception, diagnoses depended upon the comparison of the P3 amplitude in response to meaningful versus to other (irrelevant) stimuli on the assumption that the former have more salience than the latter (Allen, Iacono, & Danielson, 1992; Farwell & Donchin, 1991; Rosenfeld, Angell, Johnson, & Qian, 1991; Rosenfeld et al., 1988). These P3-based tests are

2 89 M. Lui and J.P. Rosenfeld based on the fact that a rarely presented item of concealed, meaningful information, a probe, which is recognized by the subject (even if behaviorally denied), will elicit the familiar P3 response, whereas other, frequently presented and nonmeaningful ite of information (irrelevants) will not elicit an increased P3 (Donchin & Coles, 1988). Only guilty subjects are supposed to show a significant difference between the guilty (probe) and nonguilty (irrelevant) conditions. Countermeasure to Lie Detection and the Use of Subliminal Stimuli Nevertheless, recent studies (Mertens & Allen, 28; Rosenfeld, Soskins, Bosh, & Ryan, 24) found successful countermeasures to this lie detection paradigm. A countermeasure is anything a subject attempts to do during the test that tends to prevent the detection of concealed information (Honts, Devitt, Winbush, & Kircher, 1996). The countermeasure studies have involved training subjects to make concealed responses (e.g., wiggling the toe) to the nonmeaningful ite, which significantly increased the P3 response to these irrelevant stimuli, and, therefore, no difference was found between guilty and irrelevant stimulus conditions. This makes it virtually impossible to distinguish probe and irrelevant P3s, whose differences would otherwise be usually diagnostic for deception. An obviously important requirement of physiologically based methods for identifying deception is that such methods be resistant to subjects attempts to defeat them, that is, with countermeasures. In view of this, subliminal stimuli are here proposed to be used in lie detection because they should be immune to countermeasure use: If a key stimulus is presented subliminally, because it cannot be consciously perceived, subjects would not be able to apply specific countermeasures to it. The current study uses the paradigm of subliminal priming. Priming is a phenomenon of faster and more accurate response to stimuli that have had prior exposure (e.g., Tulving & Schacter, 199). Priming of semantically related and unrelated words was found to modulate the amplitude and duration of ERP components (e.g., Besson, Kutas, & van Petten, 1992). In the present study, a priming test stimulus is presented subliminally prior to a supraliminally presented stimulus that evokes an ERP. It is expected that the subliminal presentation of key test stimuli subliminally processed uniquely by guilty subjects will affect ERP responses to the supraliminally presented stimuli in such a way as to allow discrimination of guilty versus innocent subjects. (However, another novel approach to P3-based deception detection which is resistant to countermeasures was recently reported by Rosenfeld et al. (28).) Outline of the Present Study We attempt here to model a lie detection test for someone suspected of being a terrorist, the aim being to identify members of a terrorist ring. If stimuli were names of other terrorists in the secret ring, only fellow terrorists (guilty subjects) should show a difference of brain response between terrorists names and nonterrorists names. In our study, college students were recruited as subjects, and stimuli were supraliminal acquaintance names preceded by a different subliminal acquaintance name versus a nonacquaintance name. Other stimuli were supraliminal nonacquaintance names preceded by a subliminal acquaintance name versus a different subliminal nonacquaintance name. It was hypothesized that the subliminal priming of acquaintance and nonacquaintance names would modulate ERP amplitude. Given that task demand was found to modulate subliminal priming in a previous study (Nakamura et al., 26), the current study examined the effect of the task demand associated with lying on subliminal priming. It is hypothesized that lying would orient subjects sensitivities toward the familiarity of the names, which would increase the ERP difference between conditions primed with acquaintance versus nonacquaintance names. In this study, there were five groups of participants. Two of these (Experiment 1 and the replication study) are near replications of each other and attempt to demonstrate a specific subliminal priming effect during deception. Two other groups (Experiments 2 and 3) were intended to control for two differing respective effects that do not involve subliminal priming or deception, but which could mediate putative priming effects in Experiment 1 and its replication. The fifth group was an innocent (nondeceptive) control group that allowed us to estimate the false positive rate. We note that the groups were actually run in the following order: Experiment 1, Experiment 2, Replication of 1, Experiment 3, and, finally, the innocent group. EXPERIMENT 1 Method Participants Fourteen (average age: 19.3 years, 9 men) Northwestern University undergraduate students participated in the present study for the fulfillment of an introductory Psychology course requirement. They were all right-handed and had normal or corrected vision. All signed an Institutional Review Board (IRB)-approved consent form. Procedures Subjects were first asked to provide the last names of five people they knew very well. They were also asked to select 4 names from a list of 2 last names that did not have any personal meaning to them. These 14 9 names were the stimuli they would view subliminally and supraliminally on the ERP test (see Table 1a). They sat about 1 m from a computer screen. On the screen, brief (subliminal) and long (supraliminal) presentations of last names and symbols were shown. As shown in Figure 1 (trial structure), in each trial, a name appeared on screen for 17 (subliminal), preceded by a 1- forward mask and followed by a 17- Table 1a. Experiments 1 and 3 Stimulus Arrangement and Response Requirement Condition Symbol Subliminal prime Supraliminal target Expt. 1 task Expt. 3 task (half of the subjects) 1 A-A Acquaintance Acquaintance No Right 2 N-A Nonacquaintance Acquaintance No Right 3 A-N Acquaintance Nonacquaintance No Right 4 N-N Nonacquaintance Nonacquaintance No Right Asterisks Acquaintance Yes Left

3 Subliminal priming in lie detection 891 Table 1b. Experiment 2 Stimulus Arrangement and Response Requirement Condition Symbol Subliminal prime Supraliminal target Expt. 2 task 1 A-A Acquaintance Acquaintance Yes 2 N-A Nonacquaintance Acquaintance Yes 3 A-N Acquaintance Nonacquaintance Yes 4 N-N Nonacquaintance Nonacquaintance Yes Asterisks Nonacquaintance No backward mask composed of symbols $$##$$##$$##$$. These were followed by another name, which appeared for 1 (supraliminal). Subjects were asked to respond to the supraliminally presented names with a yes or no button press. They were told that yes means I DO know a person by this name, and no means I DO NOT know a person by this name. In the ERP collection session, subjects were instructed to choose only one of the supraliminally presented acquaintance names to respond truthfully to by pressing the yes button. For the other four acquaintance names, they denied knowing them by pressing the no button. For the 4 nonacquaintance names they selected from the list, they responded truthfully by pressing the no button. They were asked to respond as soon as they could following the supraliminal stimulus. There were five stimulus-response conditions with different response requirements and stimulus arrangements (see Table 1a) in this first experiment and the replication: (1) supraliminal acquaintance name preceded/primed by subliminal acquaintance name (A-A), (2) supraliminal acquaintance name preceded/ primed by subliminal nonacquaintance name (N-A), (3) supraliminal nonacquaintance name preceded/primed by subliminal acquaintance name (A-N), (4) supraliminal nonacquaintance name preceded/primed by subliminal nonacquaintance name (N-N), and () supraliminal acquaintance name preceded by asterisks, with no priming (this was simply an attention forcing condition). Again, subjects responded no to the first four conditions, but yes to the fifth. The stimuli were ordered in a way such that no exact repetition occurred in a single trial. That is, in conditions (A-A) and (N-N), the subliminal prime and supraliminal target were always different even though they were both acquaintance or nonacquaintance names (i.e., the priming was conceptual ). Condition served as a condition to maintain subjects attention by varying the response requirement ( yes instead of no, which was the response in the other four conditions); the data from this condition were not included in any analysis. There were 2 practice trials and 36 actual trials, with 8 trials in each of the Conditions 1 4 and only 4 trials for Condition. After all the experimental trials, subjects were given an awareness test (see below) to check the visibility of the subliminal stimuli in individual subjects. Stimuli were similar to those in the experimental trials except the subliminal stimuli were either nonsense character strings (e.g., dfgiaesfr) or acquaintance names. Subjects were required to do a lexical decision task by deciding whether the subliminal stimuli were words or nonwords. Before the awareness test each subject was also asked whether he/she saw any of the subliminal stimuli, and their subjective reports were recorded. In the awareness test, subjects were given 1 practice trials and 12 actual trials. In subsequent material, because the data from the five groups are meant to be compared, some methods used for all five experiments are described together. Statistical Analysis Procedures: PCA and Bootstrapping Methods (Used in All Five Experiments) Spatial principal component analysis (PCA) on various sites on the scalp is a method utilized to identify clusters of electrodes that are highly intercorrelated. It linearly combines the highly correlated electrodes to form a virtual component or factor. The component or factor yields the virtual site data to be used in later analyses. This has the advantage of capturing most of the relevant variance and therefore preserving the information from the many actual electrodes, and at the same time reducing the redundancy among them (Spencer, Dien, & Donchin, 21). Additionally, the linear combination method of PCA can produce orthogonal factors. Follow up, temporal PCA uses the same principles except that the virtual temporal components are formed by grouping of highly intercorrelated time point data within each spatial component. Baseline Recording Mask 14 $$##$$##$ 1 Subliminal Name 17 Mask $$##$$##$ 17 Supraliminal Name Yes or No Response EEG Epoch Figure 1. Timing of mask and stimulus presentations in a single ERP epoch.

4 892 M. Lui and J.P. Rosenfeld The present PCA data sets were obtained from an average of ERPs at each time point within a given window of all trials of the same stimulus type. First, spatial factors were extracted by Varimax rotation, which produces factors with high loadings on a small number of variables and low loadings on other variables (Kaiser, 196). Also, factors remain uncorrelated after Varimax rotation, which prevents the problem of multicollinearity among factors. Standardized factor loadings were the correlations between a variable (original site) and its corresponding factor (virtual electrode). A site variable was selected only if its factor loading exceeded.6 and if such a high loading for this site variable was restricted to one factor. The factor scores were formed by summing the site values multiplied by respective factor score coefficients (Spencer et al., 21). The spatial ERP components were then subjected to a subsequent temporal PCA. For each individual subject, spatial temporal components were formed from the factor loadings obtained from the group spatial temporal PCA. The spatial temporal components were first analyzed by multivariate analysis of variance (ANOVA) on group data with appropriate follow-up analysis to localize effects. The spatial-temporal PCA and ANOVAs were done with SPSS software. Individual diagnoses using bootstrapping procedures and t tests on spatial-temporal components followed. A bootstrapping procedure was applied to the spatial temporal components in each individual. For example, assume there were x and y trials in Conditions 1 and 2, respectively, in one subject. X and y trials were drawn randomly (with replacement) from the actual pool of spatial temporal component data in each of the Conditions, 1 and 2 separately and respectively, and an ERP average was calculated for each randomly selected trial set within each condition. The random selection process is repeated 1 times to get 1 sets of bootstrapped averages in each trial set. The bootstrapping and averaging procedure was done with a MATLAB script written by the first author. T tests were then applied to test possible amplitude differences between mean amplitudes of bootstrapped distributions for various pairs of conditions. We note that in these experiments, PCAs were performed on databases developed in two ways: (1) Within-study PCA: Each group in each experiment yielded one data set for all conditions within that specific experiment. (2) Combined PCA: Data from all five groups were combined into one database, and a single set of spatial-temporal components was determined for all studies. The first method allows analyses optimized for each group, and the variance source is principally the condition differences within the group. The second method allows differences between groups as well to be a source of variance. It was found that the combined-pca method did not perform better than the within-study PCA method in group analyses and in individual diagnoses. We therefore report the statistical results of the within-study PCA only hereafter. Electroencephalogram Recordings (All Five Groups) The electroencephalogram (EEG) data were referentially recorded from 3 tin electrodes in an Electrocap (Electrocap International, Inc). The reference electrode was put on the nose tip, with the forehead connected to the chassis of the isolated side of the amplifier system (ground). A.3-Hz high-pass and a 3-Hz low-pass filter were used (Contact Precision Instruments EEG 8 system). The sampling rate was at 12 Hz, and the EEG signal was amplified by a factor of,. Electrooculogram (EOG) was recorded differentially from two electrodes diagonally placed above and below the left eye, so as to monitor both vertical and horizontal eye movement. Trials containing 8 mvor more deflections in the EOG electrodes were automatically rejected (without subject s knowledge). Also, off-line visual inspection was done on individual trials to remove trials with especially subtle eye movement artifacts. All electrode resistances were maintained at or below ko. As shown in Figure 1, all trials began with a 14- baseline recording window, followed by a forward mask (1 ), a subliminal name (17 ), a backward mask (17 ), a supraliminal name (1 ), and a response window (166 ). The length of an epoch was 248. The latencies of ERPs in analyses were timed from the onset of the subliminal stimulus, as we presumed that the ERP elicited by supraliminal stimulus would be affected by the preceding subliminal stimulus. For all group analyses and displays, single sweeps and averages were digitally filtered off-line to remove higher frequencies; 3 db point 4.23 Hz. Separate sets of group analyses were done on 1-Hz low-pass data. No significant effects were found. 1 We therefore report only the analysis results of the 4.23-Hz low-pass data in the result session. Results Behavioral All subjects responded correctly (i.e., pressing yes to Condition stimuli and pressing no to all others) in more than 97% of trials. There was no significant difference in mean reaction times between all conditions (p4.). Awareness Test The d indices ranged from.32 to Chi-squared test showed that only 1 out of the 1 subjects had a significant difference in response to word and nonword conditions (w 2.17, p.23). This awareness test result was also consistent with the subject s subjective report. The subject s data were excluded from further analysis. ERP Data and PCA Figure 2 shows a graphical illustration of the ERP (pre-pca) grand averages of 4 (Fz, Cz, Pz, and Oz) out of 3 electrode sites for Experiment 1. For the within-study PCA in Experiment 1, a covariance-based PCA was applied on the 3-sites data from to 1396 after stimulus onset for extraction of spatial components. The data matrix consisted of 3 (number of electrodes) 4 (number of conditions) 176 (number of time points per trial) 14 (number of subjects) cases. The spatial PCA extracted four spatial components (accounting for 88.3% of the variance): frontal-central-parietal (34.6%), occipitalparietal (31.4%), frontal (14.7%), and prefrontal (7.8%). The actual sites included in each component were (1) frontal-centralparietal component (FCP: Pz, C3, P3, Fz, Fc1, Fc, Cp1, Cp, C4, Cz, Fc2, Cp2), (2) occipital-parietal component (OP: O1, P7, Oz, P4, O2, P8, Cp6), (3) frontal component (F3, F7, Af3, F4, F8, Fc6) and (4) prefrontal component (Fp2, Fp1, Af4). It is noted 1 In previous P3 studies, low-pass filter was usually set lower than 1 Hz. For instance, Fabiani, Gratton, Karis, and Donchin (1987) filtered (low-passed) P3 averages at 6.29 Hz and single sweeps at 3.13 Hz. We believed that the main ERP component that revealed the difference between conditions in the current study was the P3, which has a mean frequency under 2 Hz (Duncan-Johnson & Donchin, 1979). The inclusion of high frequency noise in 1 Hz low-pass data may have diminished the P3 effect.

5 Subliminal priming in lie detection 893 FZ CZ PZ OZ Figure 2. Experiment 1 raw (pre-pca) grand averages from 4 of the 3 actual sites. that components have names in common (e.g., frontal, central) but these components have different actual electrodes contributing to the component. A temporal PCA was then performed on the four spatial component data. The temporal PCA resulted in five temporal components (accounting for 91.39% of the variance): (12.4%), (2.2%), (13.4%), (16.1%), and (29.4%). Group ERP Analysis and Individual Diagnosis Table 2a d shows the analytic results for the components extracted with the within-study PCA. In each of the five temporal components, a separate multivariate analysis of variance (MANOVA) was done. In each MANOVA, the four spatial components were four dependent variables. Condition (i.e., the stimulus response condition) was the independent variable. The omnibus MANOVAs showed that the effect of condition was significant in the following temporal components: (1) : Wilks l.48, F(12,96) 2.9, p.; (2) : Table 2a. Experiment 1 Omnibus MANOVAs: Main Effect of Condition Temporal component () Wilk s l Fstatistic p.13. nn. nn.27. nn Note: Results based on within-study PCA-extracted components. nn po.1. Wilks l.38, F(12,96) 3., po.1; (3) : Wilks l.47, F(12,96) 2.61, p. (see Table 2a). For the component 33216, the effect of condition was significant in all spatial components (po.). However, Bonferonni pairwise comparisons showed no other significant differences between any meaningful pairwise comparisons excepting a marginal p value for the OP component (see Table 2c). For the component , the effect of condition was significant in the frontal-central-parietal, F(3,39) 1.12, po.1, and occipital-parietal, F(3,39) 14.43, po.1, spatial components. Bonferroni pairwise comparison showed that Condition 2 (N-A) was significantly more positive than Condition 1 (A-A) for the frontal-central-parietal component (p.3) and the occipital-parietal component (po.1). Because the conservative Bonferonni tests revealed some significant pairwise group differences at , we did individual Table 2b. Experiment 1 Univariate ANOVAs p Values Temporal component () FCP n. nn n OP.11 n.28 n. nn n F n.49 n Pre-F.34.9 nn.43 n Note: Results based on within-study PCA-extracted components. FCP: frontal-central-parietal component; OP: occipital-parietal component; F: frontal component; Pre-F: prefrontal component. n po., nn po.1.

6 894 M. Lui and J.P. Rosenfeld Table 2c. Experiment 1 Bonferroni Tests: Condition 1 (A-A) versus Condition 2 (N-A) Temporal component () FCP n 1..6 OP.6.6. nn.1.26 n F Pre-F Note: Results based on within-study PCA-extracted components. FCP: frontal-central-parietal component; OP: occipital-parietal component; F: frontal component; Pre-F: prefrontal component. n po., nn po.1. diagnoses in these cases (and also in some other near significant cases in Table 2c): Bootstrapping of amplitude difference was done between Condition 1 (A-A) and Condition 2 (N-A). Table 2d shows that 11 of 14 (78.6%) subjects have significantly more positive frontal-central-parietal amplitude for Condition 2 (N-A) than for Condition 1 (A-A), whereas 12 out of 14 (8.7%) subjects had significantly more positive occipital-parietal amplitude for Condition 2 (N-A) than for Condition 1 (A-A). It is noted that for A-N versus N-N, no Bonferonni tests were significant, so no follow-up individual diagnostics were performed. For the component , the effect of condition was significant only in the frontal-central-parietal, F(3,39) 3.83, p.17, and the occipital-parietal, F(3,39) 3.99, p.14) components. Bonferroni pairwise comparison showed that Condition 2 (N-A) was significantly more positive than Condition 1 (A-A) for the occipital-parietal (p.26) but not for the frontalcentral-parietal component. Bootstrapping of amplitude differences was done on the occipital-parietal component. Table 2d shows that 11 out of 14 subjects (78.6%) had significantly more positive amplitude in Condition 2 (N-A) than in Condition 1 (A-A). Discussion and Introduction to the Replication Study The results of the first experiment revealed dramatic subliminal priming effects reaching diagnostic utility levels. Experiments 2 and 3 below were directed at controlling and investigating possible confounding factors, but it seemed important to us to be able to replicate the results in the original study. Therefore, Experiment 1 was replicated as fully as possible in a new group of subjects. In the replication, the procedures were similar to those in Experiment 1, except that the new subjects were asked to pick 4 names from a pool of 1 names rather than only 2 names as in Experiment 1. Another difference was that subjects were asked to pick nonacquaintance names that had equal numbers of syllables as the acquaintance names they had written down. The Table 2d. Experiment 1 Individual Diagnosis Based on A-A versus N-A Temporal component () Spatial component Detection rate frontal-central-parietal 11/14 (78.6%) occipital-parietal 12/14 (8.7%) occipital-parietal 11/14 (78.6%) occipital-parietal 11/14 (78.6%) occipital-parietal 1/14 (71.4%) stimulus arrangements, procedures, and task requirements were otherwise identical to those in Experiment 1. REPLICATION STUDY Results PCA Results The spatial, within-study PCA in this replication extracted three spatial components (accounting for 9.7% of the variance): frontal-central-parietal (38.3%), occipital-parietal (31.%), and frontal (2.9%). The actual sites included in each components were frontal-central-parietal (Cp1, C3, Cz, Fc1, Cp2, Pz, Fc2, C4, Cp, P3, Fc), occipital-parietal (O2, Oz, O1, P7, P4, Cp6), and frontal (Fp2, Af4, F8, Fp1, Af3, F4, Fz, F3, Fc6). This result was very similar to that of Experiment 1 (see above). A temporal PCA was then performed on the three spatial component data. The temporal PCA resulted in four temporal components (accounting for 87.9% of the variance): (2.6%), 3248 (27.%), (1.4%), and (24.4%). This was also similar, though not identical, to the results of the original study, Experiment 1 (see Figure 3). Behavioral RT Data The mean reaction time in the replication was 438.4, almost identical to the value from Experiment 1 at There were no differences among conditions. Awareness Test The d indices ranged from.4 to 1.3, and no significant results were found in the chi-squared tests (all p4.), showing that subjects could not discriminate subliminal words from subliminal nonwords. ERP Data The grand average ERPs for the two spatial components and two condition contrasts are shown in Figures 4 and along with data from other experimental groups. Table 3a d shows the detailed analysis results for the replication, just as for Table 2a d for Experiment 1. Only the temporal component at FCP shows a significant Bonferonni result: Again, Condition 2 (N-A) was found to be significantly more positive than Condition 1 (A-A) whereas no difference was found between Conditions 3 (A-N) and 4 (N-N). This time the significant difference was restricted to the temporal component of the frontal-central-parietal spatial component (p.32). This compares reasonably, given slight procedural differences (see above) and a new subject sample, to the result found in the same spatial component in the original study, which had a slightly earlier temporal component at In individual diagnosis, 12 out of 14 (8.7%) of the subjects had significantly more positive N-A than A-A conditions. (It is also the case that, in the Experiment 1, the occipital-parietal component additionally showed significant pairwise effects that were not observed in the replication.) EXPERIMENT 2 As we reported above, in Experiment 1, no priming effect was found in conditions with supraliminal nonacquaintance names (Condition 3 and Condition 4 in Table 1a). Because in Experiment 1, subjects were lying in Conditions 1 (A-A) and 2 (N-A),

7 Subliminal priming in lie detection 89 Figure 3. Diagram indicating statistical significance in different temporal regions. Time is the subliminal stimulus onset. The areas shaded in black are temporal regions with significant effects among experimental conditions. Significant effects were only found in Experiment 1 and the replication study. conditions with supraliminal acquaintance names, but telling the truth in Conditions 3 (A-N) and 4 (N-N), conditions with supraliminal nonacquaintance names, two possible, non-mutuallyexclusive explanations for the findings are viable: It could be that lying is necessary to get the effect, which occurs when the subjects say no to supraliminal acquaintance names but not to nonacquaintance names. However, it may also be the case that less familiar supraliminal nonacquaintance names are not primed. To investigate these possibilities, in Experiment 2, the response requirement was reversed: subjects lied in conditions with supraliminal nonacquaintance names rather than in conditions with supraliminal acquaintance names. The attention demanding condition (Condition ) was also a supraliminal nonacquaintance name (see Table 1b). It was predicted that no priming effect would be found in conditions with supraliminal nonacquaintance names (A-N and N-N) due to the lack of familiarity and longterm memory representation for nonacquaintance names. Methods Participants Thirteen (average age: 19.2 years, 7 men) Northwestern University undergraduate students who were not in Experiment 1 participated in Experiment 2 for the fulfillment of an introductory Psychology course requirement. They were all right-handed and had normal or corrected vision. All signed an IRB-approved consent form. Stimuli and Procedures The timing and construction of stimuli were similar to that in Experiment 1 except that the attention catching condition was a nonacquaintance name instead of an acquaintance name (see Table 1b). Subjects were required to press yes to all conditions and no to one of the nonacquaintance names they chose. In that way subjects were lying to the conditions with supraliminal nonacquaintance names (Conditions 3 and 4), and they were telling the truth to conditions with supraliminal acquaintance names (Conditions 1 and 2). Results Behavioral Data All subjects responded correctly to over 9% of trials. There was no significant difference in reaction times among all conditions. Awareness Test The d indices ranged from.34 to 1.31, and no significant results were found in the chi-squared tests (all p4.), showing that subjects could not discriminate subliminal words from subliminal nonwords. PCA A covariance-based, within-study PCA was applied on the 3- sites data from to 1396 after stimulus onset for extraction of spatial components. The data matrix consisted of 3 (number of electrodes) 4 (number of conditions) 176 (number of time points per trial) 13 (number of subjects) cases. The spatial PCA extracted three spatial components (accounting for 88.7% of the total variance): frontal-central-parietal (36.8%), occipital-parietal (31.2%), and frontal (2.7%). The actual sites included in each components are frontal-central-parietal (Pz, C3, P3, Fc1, Fc, Cp1, Cp, C4, Cz, Fc2, Cp2), occipital-parietal (O1, P7, Oz, P4, O2, T8, P8, Cp6), and frontal (F3, F7, Fz, Af3, Fp2, F4, F8, Fp1, Af4, Fc6). A temporal PCA was then performed on the three spatial component data. The temporal PCA resulted in six temporal components (accounting for 93.7% of the variance): 236 (16.%), (.4%), 3486 (2.7%), (17.2%), (16.2%), and (12.8%). The spatial components found in Experiment 2 are similar to those found in Experiment 1 except that the prefrontal sites (which were grouped to a separate prefrontal component in Experiment 1) were included in the frontal component in Experiment 2. For the temporal PCA, Experiments 1 and 2 obtained similar first two components (see Figure 3), but the later temporal component structure appears different. Group Analysis and Individual Diagnosis Figures 4 and show that, for both the FCP and OP spatial components, differences between wavefor are restricted to later portions on the epoch, after 7. In each of the six

8 896 M. Lui and J.P. Rosenfeld Expt Replication Expt 2 Expt AA vs NA AN vs NN Figure 4. ERPs of the frontal-central-parietal component in Experiments 1 3 and the replication study. temporal components, separate MANOVAs were done as described previously for Experiment 1 and its near replication. In each MANOVA, the three spatial components were three dependent variables, condition was the independent variable. The omnibus MANOVAs showed that the effect of condition was not significant in any analyses. No significant effects were found in univariate ANOVAs. Therefore no further analysis was done. Discussion Results indicated that the effect of subliminal priming was absent even when subjects lied in conditions with supraliminal nonacquaintance names (i.e., A-N and N-N). The lack of longterm memory representation for nonacquaintance names possibly prevented priming from taking effect (discussed further below). Moreover, when subjects were instructed to tell the truth in conditions with supraliminal acquaintance names (i.e., A-A and N-A), the effect of subliminal priming found in Experiment 1 disappeared. Lying is therefore not sufficient to produce the subliminal priming effects seen in Experiment 1 and its replication. But lying may be necessary to produce the effect. Experiment 3 was intended to test this hypothesis, by replicating the conditions of Experiment 1 except with the lying behavior removed.

9 Subliminal priming in lie detection 897 Expt AA vs NA AN vs NN 1 1 Replication Expt 2 Expt Figure. ERPs of the occipital-parietal component in Experiments 1 3 and the replication study. EXPERIMENT 3 Lying was removed in this experiment; hence the possible modulation of task requirement (i.e., deception) on stimulus-driven subliminal priming effects could be observed by comparing the results of Experiments 1 and 3. With the elimination of the lying task and the associated meaning of lies, in Experiment 3 we expected that the processing of acquaintance and nonacquaintance names would be different and the subliminal priming effect would be attenuated. No difference was expected between Conditions 1 (A-A) and 2 (N-A), and no difference was expected for the comparison between Conditions 3 (A-N) and 4 (N-N). Methods Participants Twelve (average age: 19.3 years, 6 men) Northwestern University undergraduate students who were not in Experiment 1 or 2 participated in Experiment 3 for the fulfillment of an introductory Psychology course requirement. They were all right-handed and had normal or corrected vision. All signed an IRB-approved consent form. Stimuli and Procedures The timing and construction of stimuli were identical to that in Experiment 1 except that subjects were required to press a right

10 898 M. Lui and J.P. Rosenfeld Table 3a. Replication Study Omnibus MANOVAs: Main Effect of Condition Temporal component () Wilk s l Fstatistic p nn Note: Results based on within-study PCA-extracted components. nn po.1. button to all conditions and to press a left button to one of the acquaintance names which they chose (see Table 1a). (The left/ right key assignment was counterbalanced across subjects.) That is, there were no yes and no buttons, just right and left buttons. This procedure excluded the admitting and denying of acquaintance name recognition as in Experiment 1. Also, subjects in Experiment 3 were asked to select four names from a list of 1 common names rather than only 2 names as in Experiment 1, and they were asked to pick the nonacquaintance names that have equal numbers of syllables to the acquaintance names they wrote down. Results Behavioral Data All subjects responded correctly (i.e., pressing one button to Condition stimuli and pressing another button to all others) to over 9% of trials. There was no significant difference in RT among all conditions (p4.). Awareness Test The d indices ranged from.34 to 1.42, and chi-squared tests showed none of the subjects responded significantly differently in response to word and nonword conditions. PCA Figure 6 shows a graphical illustration of the ERP (pre-pca) grand averages of 4 (Fz, Cz, Pz, and Oz) out of 3 eletrode sites for Experiment 3. A covariance-based within-study PCA was applied on 3 sites data from to 1396 after stimulus onset for extraction of spatial components. The data matrix consisted of 3 (number of electrodes) 4 (number of conditions) 176 (number of time points per trial) 12 (number of subjects) cases. The spatial PCA extracted three components which accounted for 8.6% of total variance: frontal-central-parietal Table 3b. Replication Study Univariate ANOVAs p Values Temporal component () FCP n.1 nn F OP Note: Results based on within-study PCA-extracted components. FCP: frontal-central-parietal component; OP: occipital-parietal component; F: frontal component. n po., nn po.1. Table 3c. Replication Study Bonferroni Tests: Condition 1 (A-A) versus Condition 2 (N-A) Temporal component () FCP n.187 F OP Note: Results based on within-study PCA-extracted components. FCP: frontal-central-parietal component; F: frontal component; OP: occipitalparietal component. n po.. (36.1%), frontal (28.%), and occipital-parietal (21.4%). The actual sites included in each component are frontal-centralparietal (C3, Cp1, Cz, Fc, Fc1, Fc2, Cp2, Cp, C4, Pz, P3, F3, Fz, T7, Fc6, P4), frontal (Fp2, Af3, F7, F4, T8, Fp1, F8, Af4), and occipital-parietal (O2, P8, O1, Oz, P7, Cp6). A temporal PCA was then performed on the three spatial components. The temporal PCA resulted in four temporal components (92.1% of the variance explained): 348 (2.4%), 3696 (17.7%), (34.%), and (2.%). Spatial components obtained in Experiment 3 are very similar to those in Experiment 2. However, the temporal components obtained in Experiment 3 are different from those in Experiment 1 and its replication and Experiment 2 starting from 6. In Experiments 1 and 2, three components were obtained after around 6 (e.g., , , and ) whereas in Experiment 3 only two components were obtained ( and ; see Figure 3). Group Analysis and Individual Diagnosis Figures 4 and show grand average ERPs from virtual component sites. It appears from visual inspection that the ERPs are not different among all conditions at both virtual sites. In each of the four temporal components, a separate MANOVA was done as before. In each MANOVA, the three spatial components were three dependent variables. Condition was the independent variable. The omnibus MANOVAs resulted in no significant effects in all analyses (p4.). Therefore no individual diagnosis was carried out. Discussion Removal of the deception requirement in Experiment 3 led to an absence of subliminal priming effects, and therefore, an inability to find any significant pairwise comparisons of interest, which would suggest that deception is a necessary element in the protocol of Experiment 1 ( Table 1a) that produces ERP differences between differentially primed, dishonest denials of acquaintance recognition. Table 3d. Replication Study Individual Diagnosis Based on A-A versus N-A Temporal component () Spatial component Detection rate frontal-central-parietal 12/14 (8.7%) Note: Results based on within-study PCA-extracted components.

11 Subliminal priming in lie detection 899 FZ CZ 1 A-A N-A A-N N-N A-A N-A A-N N-N PZ OZ 1 A-A N-A A-N N-N A-A N-A A-N N-N Figure 6. Experiment 3 raw (pre-pca) grand averages from 4 of the 3 actual sites. INNOCENT GROUP Methods Participants Eleven (average age: 18. years, 4 men) Northwestern University undergraduate students who were not in Experiments 1 or 2 or 3 participated in the innocent group experiment for the fulfillment of an introductory Psychology course requirement. They were all right-handed and had normal or corrected vision. All signed an IRB-approved consent form. Stimuli and Procedures The procedures were similar to that in the replication study except that subjects were asked to provide only one acquaintance name (to be used as attention catching or target stimuli) and choose 8 nonacquaintance names from the list of 1 common names. They were asked to choose two 1-syllable names, four 2-syllable names, and two 3-syllable names. Half of the names selected were denoted as pseudo acquaintance names and half of them as nonacquaintance names in the data analyses. The pseudo acquaintance names and nonacquaintance names have the same number of syllables (one 1-syllable name, two 2-syllable names, and one 3-syllable name). The timing and construction of stimuli were similar to that in the replication study. Subjects were required to press a no to all stimuli except the target, which was an acquaintance name subjects provided. The button positions were counterbalanced (half of the subjects pressed the left button for yes response and the right button for no response and half of the subjects responded in the other way). Results Behavioral Data All subjects responded correctly (i.e., pressing yes to Condition stimuli (target stimuli; see Table 1) and pressing no to all others) to over 92% of trials. There was no significant difference among all conditions (p4.) in RT. Awareness Test The d indices ranged from.31 to 1.2, and chi-squared tests showed none of the subjects responded significantly differently in response to word and nonword conditions (p4.). ERP Data For the innocent control study, the spatial temporal components identified in the replication group were utilized to analyze the data in the control group. We could have used the components from Experiment 1, but the treatment of the replication group was more similar in all other ways but guilt and innocence to that of the control group than was the treatment of the Experiment 1 group. (The components extracted from Experiment 1 and the replication were relatively similar anyway, as shown in Figure 3 and its discussion below.) The analytic results showed no significant main effects in MANOVA or in univariate ANOVA tests. Consistent with these results, Figure 7 shows virtual site grand average ERPs for the innocent control group, FCP component (top), and OP component (bottom). At FCP, there is not

12 9 M. Lui and J.P. Rosenfeld Innocent AA vs NA AN vs NN 1 1 Innocent Figure 7. ERPs of the frontal-central-parietal component and occipital-parietal component in the innocent control group data. much difference between A-A and N-A wavefor and, where there is a difference, A-A appears mostly more positive, unlike the priming effects seen in Experiment 1 and the replication. The OP component shows the same null results. For the A-N versus N-N comparison, both FCP and OP show some apparently more negative segments for A-N, but as noted above, none of these effects in the control group reached or approached significance. We did individual diagnostics in this innocent control group on components comparable to those utilized in Experiment 1 and the replication, however, because we wanted to see what the false positive rates would be. These rates varied from 2/11 to /11. In the replication study most comparable to the innocent group, we correctly detected 12/14 (86.7%) subjects using the within-replication study PCA-extracted FCP component, The same spatial and temporal component in the innocent group led to 4/11 (36.3%) false positives. This yields a Grier (1971) A index of test efficiency of.84. (Table 4a,b shows A values for various pairings of hit rates from Experiment 1 and its replication with false positive rates from the innocent control groups.) Summary of Differences among the Five Groups In Experiments 1 and 2, subjects were asked to choose nonacquaintance names from a list of 2 names, and the number(s) of Table 4a. Experiment 1 and Innocent Group Individual Diagnosis, A-A versus N-A, within-study PCA Temporal component () Spatial component Detection rate False positive (innocent group) A frontal-centralparietal 11/14 (78.6%) 4/11 (36.4%) occipital-parietal 12/14 (8.7%) /11 (4.%) occipital-parietal 11/14 (78.6%) 2/11 (18.2%).88 syllables of the acquaintance and nonacquaintance names were not matched. In the replication, Experiment 3, and the innocent groups, subjects were asked to choose nonacquaintance names from 1 common surnames in the United States. The names were selected from a web site (Most Common Names and Surnames in the U.S., n.d.) listing all surnames with over.1% frequency in the U.S. population during the 199 census. They were also instructed to choose names that matched the number of syllables of the acquaintance names they provided. RT Data for All Studies, Averaged across Conditions These results (in milliseonds) were as follows: Experiment 1: 43.77, Experiment 2: , Experiment 3: 44.8, Replication: 438.4, Innocent: It is also noted that there were no significant differences in overall reaction time among the three experiments plus the replication of Experiment 1. However, a 1 ANOVA including the innocent group yielded F(4,63).39, p.1. This suggested that the mean RTof the innocent group was different than the mean of the experimental groups, and, indeed, F(1,62) 14.4, po.1, for this test. Clearly, the lack of lying and related manipulations in the innocent group affected RT uniquely in the innocent group in comparison to the other groups. Table 4b. Replication and Innocent Group Individual Diagnosis, A-A versus N-A, within-study PCA Temporal component () Spatial component frontal-centralparietal Detection rate False positive (innocent group) A 12/14 (8.7%) 4/11 (36.4%).84

13 Subliminal priming in lie detection 91 Qualitative ERP Data in All Studies Figures 4 and show the grand averages for each experimental group in the two spatial components, frontal-central-parietal (FCP, Figure 4) and occipital-parietal (OP, Figure ), extracted from the within-study PCAs, respectively, each component accounting for more than 2% of the variance in the data. The two columns in each figure show superimposed wave for for the A-A versus N-A comparison at the left, and A-N versus N-N at the right. Recall that in Experiment 1 and the replication, priming was expected for the A-A versus N-A comparison, but not the A-N versus N-N comparison. For the FCP comparison in Figure 4, in Experiment 1 and in the replication, the priming effect appears to be a negative shift throughout, but seen mostly between and 8, which is present only in A-A versus N-A, as expected. In the OP component (Figure ), the same differences are seen; the contrasts between A-A versus N-A and A-N versus N-N seem very clear. For Experiments 2 and 3, there are no apparent priming effects for FCP or OP in any contrast, as expected. General Discussion Experiment 1 and its replication demonstrated a protocol in which ERPs to dishonestly answered primed by subliminal nonacquaintance names, supraliminally presented acquaintance probes could be discriminated from primed by subliminal nonacquaintance names, supraliminally presented acquaintance probes. (We are noting only results obtaining in both the original and replication studies.) The discrimination was good enough so that about 8% 86% of the subjects (depending on selection of specific spatial and temporal components analyzed) could be identified in their dishonest denials of acquaintance recognition (e.g., saying no to supraliminally presented acquaintance names). In addition to the difference in dishonest behavior between conditions A-A and N-A, on the one hand, versus A-N and N-N, on the other, is the difference in the nature of supraliminal stimuli in the pairs of conditions (A in A-A and N-A vs. N in A-N and N-N). Experiment 2 was directed at identifying possible key necessary factors for producing the effects seen in Experiment 1 and its replication. Conditions 1 and 2 in Experiment 2 were exactly like those in Experiment 1 except that subjects told the truth and did not lie in Conditions 1 and 2 of Experiment 2 (Table 1, cf. a and b). Subliminal A did not prime supraliminal A in Experiment 2 as it did in Experiment 1, in the absence of deception. (We assume the priming of one stimulus by another requires identical types of stimuli. It also appears that familiar stimulifacquaintancesfbut not unfamiliar stimuli are more effective in priming; see below.) We conclude that the specific nature of the subliminal prime is not sufficient to generate the priming effect. Neither is deception per se sufficient, as Conditions 3 and 4 produced no priming in Experiment 2. Experiment 3 reproduced the stimulus conditions of Experiment 1, but with the deception requirement removed and, again, no priming occurred. Taken together, all these results suggest deception is necessary but not sufficient to produce priming effects. The other necessary factors likely include whether or not the stimuli are familiar (see below) as well as the relationship or interaction of subliminal and supraliminal stimuli. To answer this question, one must perform further parametric studies. Nevertheless, the present results suggest that a countermeasure-resistant protocol for detecting concealed information may be developed around the protocol of Experiment 1. P3 Effects In Experiments 1 3, we have tried to systematically look at the effect of the lying task requirement and stimulus type on subliminal priming. In all three experiments, stimuli consisted of subliminal primes of acquaintance or nonacquaintance names preceding a supraliminal target of different acquaintance or nonacquaintance names. Results of Experiment 1 indicated that the ERP response differed when different subliminal primes preceded supraliminal acquaintance names (A-A and N-A). The condition with a congruent prime target pair (acquaintance preceding acquaintance names) produced a smaller positive amplitude than the condition with an incongruent pair (nonacquaintance preceding acquaintance names) in the temporal segment, which is in the typical P3 region. It was hypothesized that P3 is related to decision making (Kutas, McCathy, & Donchin, 1977) and the end of a decision process (Donchin & Coles, 1988). The context-update hypothesis suggested that P3 signifies the updating of working memory when processing unexpected events (Donchin, 1981). In Experiment 1 and the replication, the supraliminal stimulus in Condition 2 (N-A) was supposed to be less expected than in Condition 1 (A-A) due to the incongruency of the prime target pair and may therefore elicit a larger P3 during the updating of working memory. In fact, a modulation of P3 by subliminal primes was reported by Dehaene, Kerszberg, and Changeux (1998), though the modulation was in ter of latency but not amplitude. The stimuli were numerals 1 9 in Arabic or word form. A supraliminal target numeral was preceded by a subliminal prime numeral. Subjects performed a simple semantic categorization task on the target numeral. Primes and targets were either congruent or incongruent in the response requirement. The ERP component that showed a prime target congruity effect was the central positivity at around 6, which was delayed by around 24 in incongruent trials compared with congruent trials. Top-Down Effects of Long Term Memory The results in Experiments 1 and 2 revealed that, under the same task requirement and stimulus conditions, subliminal neural priming occurs only in conditions with supraliminal acquaintances but not with supraliminal nonacquaintances. The present results of the presence and absence of priming effect in acquaintance and nonacquaintance conditions, respectively, may provide hints about the interaction between perceptual processing of stimuli and preexisting long-term memory representation (Henson, 23). Some studies have found priming for both familiar and unfamiliar words and faces (Bowers, 1994; Goshen-Gottstein & Ganel, 2; Stark & McClelland, 2). However, behavioral priming effects are generally larger for familiar than for unfamiliar stimuli. Also, there were previous studies that found priming only for familiar but not for unfamiliar stimuli (e.g., Ellis, Young, & Flude, 199). The findings in conscious priming generally indicated that priming of stimuli with preexisting representation in long-term memory undergoes different mechanis compared to priming of novel stimuli lacking representation in long-term memory. The present study differed from the aforementioned studies in that subliminal stimuli were used. This may explain why different results of priming were found in supraliminal acquaintance name conditions (A-A and N-A) and supraliminal nonacquaintance name conditions (A-N and N-N). It maybe the case that the signals of subliminal primes were too weak to activate and facilitate the formation of a new