Research in Parapsychology, 1975, PP 22-25

Psychophysiological Studies

EEG CORRELATES OF TRIAL-BY-TRIAL PERFORMANCE IN A TWO-CHOICE CLAIRVOYANCE TASK: A PRELIMINARY STUDY

Edward F. Kelly and James Lenz, Duke University


This experiment represents our first attempt to study trial-by-trial EEG correlates of performance in an ESP task. The test device was a binary electronic random number generator (RNG) built by Fritz Klein of Duke University Medical Center. The subject’s task on each trial was to guess which side of the device had been selected as the target by the RNG, and to register his decision by pushing a response lever in the corresponding direction. Trials and hits were accumulated separately by counters mounted on the test device; thus, immediate trial-by-trial feedback was always available.

The single subject was Lalsingh (Sean) Harribance. We have so far had just one opportunity to work with Sean. The procedure was as follows: Sean was fitted symmetrically with monopolar parietal leads (Grass gold cup electrodes at C3 and C4 referenced to the corresponding mastoids) and led to our shielded experimental room. There he was introduced to the test device and allowed to practice freely while in the adjacent control room we carried out standard calibration procedures. Then the experiment began. Sean was requested to do approximately 100 trials, keeping his eyes closed. (He still received auditory feedback from the hit counter on the test device.) He was urged to relax as much as possible between trials and a minimum interval of five seconds between trials was enforced. The two channels of EEG data, code signals from the test device indicating hit or miss and side chosen, and a voice track of experimenter commentary were recorded on analog tape. The EEG signals were acquired by a Grass 79B polygraph, with low-pass cutoff at 35 Hz and a time constant of 0.3 seconds.

The session lasted 45 minutes and 106 trials were recorded. The raw data were subsequently digitalized and a standard file created for the FILMAN data-processing system [see pp. 15-8]. Each output data record (i.e., the data record for each trial) contained the hit/miss and side-chosen information in its grouping-variable fields, and two seconds of EEG sampled at 256 per second for each channel in its data fields.

Overall, Sean obtained only 48 hits in the 106 trials, corresponding to a negative CR of less than one—not a promising beginning. Simple EEG averaging, as expected, did not permit discrimination of hitting from missing trials, nor did it reveal any consistent waveform associated with preparation to respond. In this experimental situation we anticipated that any EEG factors linked to success would not be time-locked to the response, but distributed more diffusely through the pre-response interval; therefore power spectrum analysis was the primary data reduction procedure. Raw frequency spectra was .5 Hz resolution were computed for the two seconds preceding each response for each of the two EEG channels. Thus, for each trial and each EEG channel, we had one estimate of a spectrum which consisted of a vector of numbers, each representing the estimated power at a particular point in the frequency spectrum. Our general plan was to use vectors of spectral estimates as criteria (dependent variables) for a two-way multivariate analysis of variance (MANOVA), with hit or miss and side chosen as the factors.

Averaging and plotting the raw spectra, we found that missing trials were characterized visually by higher power throughout the region from about 6 to 14 Hz. Because our analysis programs are so limited in the size of problems they can presently handle, we could only use MANOVA to looks statistically at a small segment of the complete, fully resolved spectrum at one time. Since the effects obtained in the first analysis using 1 Hz segments were consistent in form, however, we could begin to block over frequency components within trials and thus include wider segments of the spectrum, simultaneously improving the distributional characteristics of the individual estimates. Two other MANOVAs were therefore done blocking the spectrum first into 2 Hz segments, and next into 4 Hz segments roughly equivalent to the delta, theta, alpha, and beta EEG bands.

The results of the MANOVA analyses indicated that the power spectrum of the pre-response EEG appeared to discriminate, to a statistically significant degree, between hitting and missing responses, confirming our visual observations mentioned above. (The side-chosen factor was always entirely nonsignificant, as was the interaction of the two factors.) Although the first MANOVA (1 Hz segments) showed a nonsignificant overall effect of hitting versus missing, there were consistent trends and several strong individual components. Blocking the data in the second and third MANOVAs increased these trends and yielded marginal significance even in the overall tests of the entire spectrum.

The main source of significant discrimination was concentrated in the excess of power on missing trials, especially in the frequency region around 12-13 Hz. It was found in both hemispheres and was independently significant in both. It appeared somewhat larger, perhaps, on the right side. The effect also appeared to be simultaneous in the two hemispheres, rather than alternating in some fashion, since the cross-spectra do not discriminate between hits and misses.

We stress that these results are highly preliminary; we will have to see them again before taking them altogether seriously. Nevertheless, we suggest that they could make gross physiological sense. On the assumption that in this particular session ESP information was present on some of the missing trials (if at all), the excess activity in the 12-13 Hz range on missing trials could be grossly construed as reflecting the processing of that information. Interestingly, when we plot the distribution of power in the 12-13 Hz region across trials, separately for hitting and missing trials, we find two distributions that are rather similar except for a small secondary peak at the right-hand end of the distribution for the misses. It may be that we will be able, through this kind of analysis, to find clusters of physiologically definable trials corresponding to "real" (versus chance) ESP hits and misses; in short, it looks as though there are a lot of interesting possibilities for developing prediction and classification schemes in which we use the results from one segment of data to predict into another, using multivariate and probabilistic criteria of group membership. We eagerly await Sean’s next visit, to see if we can replicate and extend these results in an identical test environment.