Berger’s published work spans at least four distinct areas: (1) automated ganzfeld instrumentation and the autoganzfeld experimental series at PRL, conducted with Charles Honorton and colleagues including Mario Varvoglis, Ephraim Schechter, Marjorie Quant, Patricia Derr, George Hansen, and Diane Ferrari;^[5] (2) the PsiLab II platform and associated micro-PK experiments using hardware RNGs;^[8][10] (3) forensic methodological audits of the Blackmore and Spinelli datasets;^[2][3] and (4) geomagnetic correlates of psi effect sizes, conducted with Michael Persinger.^[6] A later collaboration with Dick Bierman and Richard Broughton re-examined the randomization properties of the PRL autoganzfeld target sequences.^[11] His work at Brooks Air Force Base on radiofrequency radiation bioeffects is a separate research program not connected to his parapsychology work.^[1]

Gerontology Research Center, NIA/NIH: Designed transportable microprocessor-controlled automated psychological testing machines for studying memory and aging in home settings, programming microprocessors in assembly language.^[1] Virginia Commonwealth University / Medical College of Virginia: Doctoral research and a hardware interface project for the MCV Department of Gerontology.^[1] Psychophysical Research Laboratories, Princeton: Developed the autoganzfeld system and coordinated PsiLab II; also ran Princeton Microware Consultants in parallel, producing commercial Apple II software and a turnkey laboratory automation system for Johnson & Johnson.^[1] Science Unlimited Research Foundation, San Antonio: Published psi-feedback experiments and forensic dataset audits.^[1] OpTech Corporation / Brooks AFB: Senior Research Scientist for US Air Force radiofrequency bioeffects research; designed five automated experimental testing systems and created the first hardware/software system for visualizing a microwave field in three dimensions.^[1] Innovative Software Design (owner, ongoing): Web development, SEO, internet marketing, and currently AI-assisted protocol development.^[1]

The autoganzfeld system used an Apple II Plus computer with a Cavri Interactive Video Interface to control a JVC VCR. Receiver and sender were in separate, sound-attenuated, electrically shielded rooms. The receiver judged by moving a pointer on a 40-point rating scale displayed on a color TV monitor; the computer checked for tied ratings and converted ratings to ranks automatically. A target selector who had no contact with the experimenter or receiver loaded the correct videocassette; the VCR was shrouded to prevent visual cues about tape position.^[5]

The full series comprised 11 experimental series (3 pilot, 8 formal) with 241 participants contributing 355 sessions. The overall hit rate was statistically significant (z = 3.89, p = .00005). Effect size (Cohen’s h) was .20. The 95% confidence interval for the hit rate ran from 30% to 39%. Ten of 11 series yielded positive outcomes. All eight experimenters had positive effect sizes; a chi-square homogeneity test showed results were not significantly non-homogeneous across experimenters.^[5]

Dynamic targets: 190 sessions, 77 direct hits (40%), z = 4.62, p = 1.9 × 10⁻⁶. Static targets: 165 sessions, 45 hits (27%), z = 0.59, p = .276. The point-biserial correlation between target type and series effect size was .663 (t = 3.65, 17 df, p = .002).^[5]

Randomness tests: 12 control samples from the hardware RNG showed a uniform distribution of values across the full target range (Kolmogorov-Smirnov test, p = .577). The experimental target distribution was also uniform (chi-square = 0.86, 3 df, p = .835).^[5]

A potential auditory leakage path from the VCR soundtrack to the receiver’s headphones was investigated. After a circuit modification confirmed by audio spectrum analysis, the last 62 sessions showed a hit rate of 44% (h = .39, z = 3.06), with dynamic target sessions at 50%, higher than before the modification, ruling out auditory leakage as an explanation for the dynamic target advantage.^[5]

Comparison with the Honorton (1985) meta-analysis of 28 earlier nonautomated ganzfeld studies: the two datasets showed consistent mean z scores and effect sizes (t-tests of difference, p = .748 and p = .892 respectively). The combined z for all 39 studies was 7.53 (p = 9 × 10⁻¹⁴). Rosenthal’s file-drawer statistic indicated that 778 additional null-result studies would be needed to reduce the combined database to non-significance.^[5]

Subject-based analysis: mean Stanford z for 241 participants was .21 (SD = 1.04), t(240) = 3.22, p = .00073. Effect size (Cohen’s d) = .21, nearly identical to the trial-based effect size, indicating the result was not driven by a few exceptional subjects.^[5]

Signal detection theory provides the appropriate framework for evaluating 4-alternative forced-choice hit rates against statistical baselines in free-response paradigms of this kind.^[15]

The PsiLab II RNG was a noise-based board that plugged into an Apple II peripheral expansion slot. Two avalanche diodes provided analog noise voltages; LM311 comparators quantized the signals; two independent bitstreams were combined through an exclusive-or circuit to reduce first-order imbalance. Independent data bits were available at 32 kHz for bits and 4 kHz for independent 8-bit bytes. The system software performed an RNG integrity check before each session; failure prevented data collection. Software was password-protected; premature termination triggered a sonic alarm and a coded record on disk.^[4]

In the Psi Invaders game, each button press yielded two yoked 100-bit RNG samples: a feedback sample (linked to game play) and a silent sample (collected but not displayed). The order of feedback and silent samples was pseudorandomly determined at the game’s beginning. A hardware RNG decision randomized each game into a feedback-with-silent-saved (FSS) condition or a feedback-with-silent-destroyed (FSD) condition, maintaining subject and experimenter blindness to condition assignment.^[4]

Extended simulations, 50 simulated experiments each composed of the same number of runs as a complete experiment, were conducted after all four experiments were finished, providing an empirical baseline against which experimental outcomes could be compared. All four silent experimental outcomes exceeded the .05 level; 2 of 50 (4%) of both Psi Invaders and Volition extended simulations were significant at the .05 level.^[4]

Experiment 1 (Psi Invaders, experimenter as sole subject, 600 games): Silent run-score distribution, chi-square(34) = 51.41, p = .03. FSS feedback data, chi-square(34) = 31.40, p = .60. FSD feedback data, chi-square(34) = 27.49, p = .78. Simulation data showed no departures from chance in any condition.^[4]

Experiment 2 (Psi Invaders, 10 subjects, 475 games): Silent run-score distribution, chi-square(34) = 57.65, p = .007. A significant deficit of run scores at the theoretical mean run-score value of 50 contributed to the overall effect (chi-square(1) = 9.84, p = .002). FSS and FSD feedback data were not significant. Simulation data showed no departures from chance.^[4]

Experiment 3 (Volition, 10 subjects, 500 games): Silent run-score distribution, chi-square(34) = 50.53, p = .03. Feedback data not significant. With the experimenter’s data removed, the silent chi-square became 47.98 (p = .06).^[4]

Experiment 4 (Volition, experimenter as sole subject, 400 games): Silent run-score distribution, chi-square(34) = 63.47, p = .002. Feedback data not significant. Simulation data showed no departures from chance.^[4]

Exact binomial probability of 4/4 silent outcomes exceeding the .05 level: 6 × 10⁻⁶ (where p = .05, q = .95, one-tailed).^[4]

The randomization of games into FSS vs. FSD conditions was itself performed by the hardware RNG. In Experiment 1, 323/600 (53.8%) games were assigned to the FSS condition; in Experiment 2, 253/475 (53.3%). Combined, 576/1,075 games were assigned to the condition in which silent data were saved (z = +2.35, p = .01, one-tailed). Simulation data showed normal randomization (526/1,075 assigned to silent-saved, z = −0.70).^[4]

The Bösch, Steinkamp, and Boller (2006) meta-analysis of RNG studies included two of Berger’s PsiLab II experiments in its database.^[16] The mainstream replication-crisis literature provides important context for evaluating small effect sizes in this paradigm: when true effects are small relative to noise, statistically significant studies can systematically overestimate effect magnitude even without publication bias.^[17]

The dependent measure was the percent of significant runs per game-fifth epoch, averaged across 25-game series. Each subject contributed two series (template and replication). For each subject, 20 computer-simulated experimental series were generated as decoys. A correlation matrix was produced and the rank of the correlation with the subject’s own replication data was the primary outcome (chance = 1/21 = .0476).^[18]

Of 8 replication subjects, 2 showed significant template-replication correlations (exact binomial p = .057 at alpha = .05). A third showed a marginal correlation (p = .076); post-hoc, 3/8 at this threshold had exact probability .018. Matched simulation data showed chance numbers of significant correlations.^[18]

Blind matching: the experimenter’s data yielded rank 1; the second exploratory subject’s data yielded rank 2. Among 8 replication subjects, 3 yielded rank 1 (exact binomial p = .005, where p = .0476).^[18]

The unexpected common patterning, a “U” curve with hitting near the game’s beginning, missing in the middle, and recovery toward chance at the end, meant that a summary measure alone would have concluded “no effect,” as consistent hitting and missing cancel each other out. This illustrates a general methodological point about the importance of temporal analysis in RNG data.^[18]

Two “First-Timers” series (FT1 and FT2) each enrolled 50 novice participants for a single ganzfeld session using the PRL automated system. Participants completed the Myers-Briggs Type Indicator (Form F) and a Participant Information Form covering demographics, attitudes toward psi, and self-reported psi experiences. Alpha was set at 5% one-tailed; both direct hits and sum-of-ranks tests were pre-specified as primary measures.^[19]

FT1 overall results were non-significant on both indices. FT2, with 23 of 50 sessions completed at time of reporting, showed both indices approaching significance (direct hit rate 43.5%). The two series differed significantly in recruitment source: 46% of FT1 participants were recruited through ads or media vs. 11% of FT2 (chi-square = 7.32, 1 df, p = .007).^[19]

Feeling-type participants (N = 28): 50% direct hits, p = .0038; mean Z = .55, t = 2.37, 27 df, p = .012. Thinking types (N = 22): 18.2% direct hits, p = .84; mean Z = −.08. Feeling types with dynamic targets (N = 12): 75% direct hits, p = 4 × 10⁻⁴; mean Z = 1.02, t = 3.18, 11 df, p = .0044.^[19]

Number of types of self-reported ESP experiences correlated positively with psi Z-score (r = .306, t = 2.44, 58 df, p = .018). Belief in psi correlated suggestively with Z-score (r = .189, t = 1.47, 58 df, p = .074). Multiple regression indicated that number of ESP experience types accounted for most of the variance contributed by these two variables.^[19]

An interaction between extraversion/introversion and sender type was found: extraverts tended to opt for friend-senders and performed better with them; introverts tended to prefer lab-assigned senders and performed better with them (F = 5.05, 1, 46 df, p = .027).^[19]

The dataset comprised 185 experiments published between 1882 and 1939, primarily card-guessing studies tabulated by Rhine, Pratt, Smith, Stuart, and Greenwood (1940). Effect size was defined as the z score divided by the square root of the number of trials. The median effect size per publication year was the primary measure; maximum and minimum effect sizes per year were computed for comparison. Annual aa (antipodal) geomagnetic index values were obtained for the publication year and for each of the five preceding years.^[6]

Lag correlations between median effect size and aa values: same year, r = −.28 (all years) and −.35 (years with more than one experiment); one year before, r = −.40 and −.56 (p < .01); two years before, −.30 and −.35; three years before, .02 and .15. For years with more than one experiment, correlations for maximum and mean effect sizes with the previous year’s aa values were −.62, −.55, −.60, and −.53 respectively (all p < .01). Partial correlations controlling for temporal order did not alter the magnitudes.^[6]

The authors noted that the effect magnitude (eta approximately .40) and direction (negative) were consistent with prior analyses of spontaneous psi experiences and experimental dream telepathy studies, suggesting that experimental and spontaneous psi experiences may be modulated by a common geomagnetic factor.^[6]

The Blackmore dissertation reported 29 experiments conducted between October 1976 and December 1978. Of these, 21 were eventually published across five peer-refereed papers. Berger’s audit identified the following categories of discrepancy:^[2]

(1) Reordering of published experiments: In two of the five publications, the chronological order of studies as listed in the dissertation’s “Schedule of Experiments” differed from the order presented in the journal papers, creating false impressions of logical progression from preliminary to main experiments. In one case, the “Main Experiment” had actually been conducted before three of the five “Preliminary Experiments” it was said to follow.^[2]

(2) Omission of significant results: In one published paper, a significant correlation (r = 0.286, z = 2.0, p = 0.045) reported in the dissertation was absent from the published version. In another, a probability value was misreported as .52 when the correct value was .052.^[2]

(3) Asymmetric application of methodological criticism: Significant results were dismissed as uninterpretable due to design flaws; the same or comparable flaws were not invoked when studies produced non-significant results. Berger argued that this asymmetry is logically inconsistent: a flaw that can produce a false positive can equally produce a false negative, and the meaningfulness of a study is determined by how well the dependent measure was operationalized, not by whether the result fits prior expectations.^[2]

(4) Misreporting of the database scope: Blackmore’s public claim of “34 independent significance tests” across approximately 34 experiments traced to a dissertation passage that explicitly noted the experiments were not all independent and were derived from 8 experiments in a single chapter, not 34 independent experiments.^[2]

(5) Subject self-scoring: In most experiments in three of the five publications, subjects scored their own or neighboring students’ data, a procedure Blackmore herself acknowledged introduced the possibility of cheating, but which was not disclosed in the published versions.^[2]

Spinelli’s dissertation reported ESP experiments with children across 10 chronological age groups. The three youngest groups (ages 3, 4, and 5–8) showed effect sizes many times larger than any reported in contemporary adult parapsychology research. The seven oldest groups scored very close to chance.^[3]

Berger calculated effect sizes from the hit rates reported in Spinelli’s tables. For the three youngest groups combined (all children under 8), the effect size was approximately 0.39 (z = 21.59, N = 3,000). For subjects over 8, the effect size was approximately 0.009 (z = 0.747, N = 7,000). The consistency of effect sizes across independent subsamples, including 500 subjects randomly removed from the larger groups, was itself statistically remarkable.^[3]

The response-distribution matrix that Spinelli presented to refute randomization concerns showed a chi-square of 4.22 with 80 degrees of freedom (p < 10⁻³¹), meaning the data were so close to perfectly uniform that the probability of this occurring by chance was less than one in 10³¹. Berger noted that psychological research consistently shows humans cannot generate unbiased random sequences, making this level of uniformity implausible.^[3]

First-and-last trial data (from the preliminary 1976 convention paper only) showed significant effects for every age group, including the oldest groups that showed no overall psi. The combined odds against chance for this pattern exceeded 10¹⁵ to one. The middle trials for the oldest groups showed significant psi-missing, perfectly cancelling the first-and-last effects, a pattern Berger described as self-cancelling in a way that would be missed by summary measures.^[3]

Errors were found in the response-distribution matrix: five of ten rows had unequal numbers of targets sent and received, yet correct row totals, a pattern inconsistent with simple typing or tabulation errors, which would have produced errors in the total column rather than symmetrical opposite errors across two independent conditions.^[3]

A later collaboration with Dick Bierman and Richard Broughton re-examined the randomization properties of the PRL autoganzfeld target sequences. The analysis distinguished between a random target selection procedure (which in principle excludes sequential dependencies) and the resulting target sequence (which may contain peculiarities).^[11]

A conservative correction for subject response biases and unequal target frequencies reduced the overall z-score from the reported z = 2.89 to a conservative estimate of z = 2.47. The correction also reduced the apparent difference between dynamic and static target scoring rates from a 10% differential to a non-significant 6.8%, leading the authors to characterize the dynamic target advantage as “suggestive at best” after correction.^[11]

A post-hoc finding: one target set (Set 20, containing targets 77–80) was used approximately three times more often than expected by chance. The binomial probability of this occurring was smaller than 1.25 × 10⁻⁵. Monte Carlo simulations of 100,000 experiments indicated this frequency would occur once in every 2,000 experiments. The over-representation was distributed reasonably evenly across the eight experimenters. No mechanism in the RNG target selection procedure could explain the excess production of random numbers in the 77–80 range. The authors concluded that this anomaly, while unexplained, could not have contributed to the excess of apparent psi hits.^[11]

The publication-bias literature provides relevant context for evaluating whether unequal target frequencies and response biases can account for apparent psi effects in free-response paradigms.^[20]

The Honorton, Berger et al. (1990) autoganzfeld paper is cited in: Bem and Honorton (1994) Psychological Bulletin;^[26] Bösch, Steinkamp, and Boller (2006) Psychological Bulletin meta-analysis of RNG studies;^[16] Utts (1999) Journal of Scientific Exploration;^[30] Radin’s Entangled Minds;^[31] Cardeña et al.’s Parapsychology: A Handbook for the 21st Century;^[28] Carpenter’s First Sight;^[32] and Irwin’s An Introduction to Parapsychology.^[27]

Berger’s Blackmore audit (1989) is cited in Carpenter’s First Sight (listed as “Berger, R. E. (1989). A critical examination of the Blackmore psi experiments. Journal of Parapsychology 51, 296–309”).^[32] Note: the citation in that source gives a different journal volume and page range than the primary paper in this corpus, which is published in the Journal of the American Society for Psychical Research, Vol. 83.

The Berger and Persinger (1991) geomagnetic paper is cited in Braude’s Immortal Remains and in Irwin’s An Introduction to Parapsychology.^[33][27]

The Research in Parapsychology 1988 volume edited by Henkel and Berger is cited as a source for multiple papers in Bem and Honorton (1994), in Carpenter’s First Sight, and in the Associative Remote Viewing handbook.^[26][32][29]

The mainstream replication-crisis literature provides a useful frame for prioritizing these questions. The Open Science Collaboration (2015) found that only a minority of significant psychological effects replicated in independent labs.^[22] For small effects like those in the ganzfeld literature, Gelman and Carlin (2014) show that statistically significant studies systematically overestimate effect magnitude even without publication bias, a Type M error problem that makes the true effect size uncertain even when the direction is reliable.^[17] Bayesian sequential analysis methods provide a framework for designing studies that accumulate compelling evidence efficiently rather than relying on fixed-N designs with low power.^[34]

For the silent-data paradigm specifically, the key design question is whether subject personality variables can be prospectively specified as predictors of silent-data outcomes before data collection begins. The exploratory correlations in Berger (1988), belief in psi and reported PK experiences, are candidates for preregistration in a new series.^[4]