An ERP-based, control-question lie detector analog: Algorithms for discriminating effects within individuals' average waveforms

ABSTRACTTwo experimental, P3‐based analog control question tests were run. In both, guilty subjects were presented with a set of seven phrases describing antisocial acts of which they were innocent, plus one phrase describing a guilty act (the analog relevant question), and one act to which a “yes” response (yes‐target stimulus) was required to assure attention. Innocent subjects (run only in Experiment 1) saw all innocent acts plus the yes‐target act. Thus nine acts were seen by guilty and innocent subjects. In both experiments, all subjects had to selectively review their guilty acts privately. Also in both experiments, all subjects were especially questioned about four acts of which guilty subjects were known to be innocent of all but one, and of which innocent subjects were known to be innocent of all. (These falsely accused acts were regarded as control question analogs.) In Experiment 1, the private review and rehearsal took place on the same day as the main test. In Experiment 2, one subgroup (delay‐only) of guilty subjects was run as in Experiment 1, except that the private review‐rehearsal was separated from the main run by 7–14 days. Another subgroup (delay‐rehearsal) of guilty subjects was run just as was the subgroup delay‐only, except that the delay‐rehearsal subgroup additionally received a non‐selective additional interrogation/rehearsal on the delayed main run day. Parietally maximal P3 responses were obtained to yes‐target items in all groups. In Experiment 1, only in the guilty group was the relevant‐minus‐control P3 amplitude difference significant. In Experiment 2, the difference was significant only in the delay‐rehearsal subgroup. A four‐step algorithm (involving relevant‐control amplitude differences and relevant target vs. control‐target cross‐correlations) was used to assess effects within individuals. In Experiment 1, 12 of 13 guilty subjects and 13 of 15 innocent subjects were correctly diagnosed. In Experiment 2, 3 of 8 delay‐only subjects and 7 of 8 delay‐rehearsal subjects were correctly diagnosed. In Experiment 2, the relevant‐minus‐control group P3 amplitude difference was significant in the delay‐rehearsal but not in the delay‐only subgroup. The results suggest that temporally proximal, non‐selective rehearsal procedures are sufficient to activate personal knowledge of a salient (oddball). P3‐generating stimulus phrase, and that even selective rehearsal of guilty acts is not sufficient without temporal proximity.

Two experiments using mock antisocial acts as stimuli, with guilty subjects (n=13 in Exp 1, n=16 in Exp 2) shown one relevant item and innocent subjects (n=15 in Exp 1) shown none. ERPs were recorded while subjects viewed phrases, with a bootstrapping algorithm comparing P3 amplitude differences and cross-correlations between relevant, control, and target items.

Background & Context

This 1991 study by Rosenfeld and colleagues represents a pioneering attempt to adapt event-related potential (ERP) methodology to mimic the Control Question Test (CQT) polygraph format used in pre-employment screening. Prior to this research, P300-based deception detection had focused primarily on the Concealed Information Test (CIT, also known as the Guilty Knowledge Test), which detects recognition of crime-specific details.

This was the first attempt to use P300 methods to diagnose deception in a scenario that modeled the employee screening Control Question Test—a test format that was the major application in U.S. field polygraph practice until Congress banned most employee screening tests in the Employee Polygraph Protection Act of 1988. The research filled a critical gap by exploring whether brain-based measures could provide a more objective alternative to traditional autonomic measures in the CQT format.

The study introduced a novel four-step bootstrapping algorithm to analyze individual subjects' ERP waveforms—addressing the statistical challenge that traditional t-tests on single-sweep ERP values are too insensitive for within-subject diagnostics.

Research Design & Methodology

In both experiments, guilty subjects were presented with seven phrases describing antisocial acts of which they were innocent, plus one phrase describing a guilty act (the analog relevant question), and one act requiring a "yes" response (yes-target stimulus) to assure attention. Innocent subjects (run only in Experiment 1) saw all innocent acts plus the yes-target act, for a total of nine acts.

The design cleverly simulated a CQT by having subjects: selectively review their guilty acts privately, and be especially questioned about four acts of which guilty subjects were known to be innocent of all but one. These falsely accused acts functioned as control question analogs.

Experiment 1 included 13 guilty subjects and 15 innocent subjects, with private review/rehearsal occurring on the same day as the main test. Experiment 2 investigated the critical role of temporal proximity and rehearsal in memory activation. One guilty subgroup (delay-only, n=8) had 7-14 days separating private review from the main test. Another subgroup (delay-rehearsal, n=8) received the same delay but with an additional non-selective interrogation/rehearsal on the delayed main run day.

ERPs were recorded while subjects viewed phrases presented every two seconds. The diagnostic algorithm compared relevant-minus-control P3 amplitude differences and relevant-target versus control-target cross-correlations using a bootstrapping resampling procedure with 100 iterations.

Results & Key Findings

The P300 component demonstrated diagnostic utility, but with important constraints related to memory activation. Key accuracy rates showed substantial variability across conditions:

• Experiment 1: 12 of 13 guilty subjects (92.3%) and 13 of 15 innocent subjects (86.7%) were correctly diagnosed
• Experiment 2 delay-only group: only 3 of 8 subjects (37.5%) correctly diagnosed
• Experiment 2 delay-rehearsal group: 7 of 8 subjects (87.5%) correctly diagnosed

At the group level, parietally maximal P3 responses appeared to yes-target items in all groups. Critically, significant relevant-minus-control P3 amplitude differences emerged only under specific conditions: in the guilty group of Experiment 1, and in the delay-rehearsal subgroup of Experiment 2.

The delay-only condition—where subjects reviewed guilty acts 7-14 days before testing without additional rehearsal—showed poor detection rates, demonstrating that selective rehearsal alone is insufficient without temporal proximity.

Discussion & Significance

This research demonstrated that temporally proximal, non-selective rehearsal procedures are sufficient to activate personal knowledge of salient stimuli that generate P300 responses. The findings suggest that even selective rehearsal of guilty acts is not adequate for reliable P300-based detection without temporal proximity to the test session.

The study advanced P300-based deception detection methodology by showing that the CQT format—unlike the CIT—presents unique challenges when adapted to brain-based measures. The CQT relies on differential salience of personally relevant questions versus control questions, whereas the CIT relies on recognition of specific crime details. The P300's sensitivity to both stimulus probability and meaningfulness makes it theoretically suited to both formats, but practical implementation requires careful attention to memory activation factors.

The development of the four-step bootstrapping algorithm was methodologically significant, as it provided a statistically valid approach for making diagnostic decisions at the individual level rather than relying only on group averages. This algorithm became influential in subsequent P300-based deception detection research.

Limitations & Considerations

Several important limitations constrain interpretation and generalizability:

• Mock crime paradigm: The "guilty acts" were imagined antisocial behaviors checked on a private list, not actual committed acts with genuine consequences or emotional weight
• Small sample sizes: Particularly in Experiment 2, subgroups contained only 8 subjects, limiting statistical power
• Laboratory setting: Subjects faced no real-world stakes, likely reducing motivation to conceal compared to actual criminal or employment screening contexts
• Limited ecological validity: The rehearsal manipulation may not reflect how memories consolidate and remain accessible in real forensic or screening situations

The study did not include countermeasure testing, which later research (including by Rosenfeld's own lab) showed could compromise P300-based detection methods.

Practical Applications

This research has direct implications for any attempt to use ERP-based methods in pre-employment or security screening contexts. The findings suggest that effective P300-based testing in a CQT-like format requires recent activation of the relevant memories—information that may be logistically challenging in field applications where scheduling constraints exist.

For practitioners, the research highlights that brain-based measures are not simply "drop-in" replacements for autonomic polygraph channels. The temporal dynamics of memory activation create unique operational requirements. The bootstrapping algorithm approach demonstrated here laid groundwork for computer-automated scoring methods that have become standard in modern P300-based concealed information testing, though most contemporary applications have returned to the CIT format rather than pursuing the CQT analog approach explored in this study.