Radin’s 2011 review of 75 years of forced-choice precognition experiments noted that the four major classes of studies (forced-choice, free-response, psychophysiological, and implicit decision) each produced cumulative odds against chance far exceeding conventional significance thresholds, approximately 10²⁴, 10²⁰, 10¹⁷, and 10⁷ respectively, yet individual studies frequently failed to reach significance.⁵ This pattern is precisely what a small true effect with high variance would produce: meta-analytic aggregation reveals the signal that individual underpowered studies cannot. Sequential analysis is one methodological response to this problem, designed to extract structural information from data that standard hit-rate analysis discards.

Radin’s 1990 paper in the European Journal of Parapsychology reported a replication and extension of sequential analysis applied to psi experiment data, demonstrating that the method could statistically enhance the detectability of effects beyond what standard analysis revealed.³ The specific artifact addressed was the assumption of trial independence: if trials are not independent, because a participant’s mental state persists across trials, then treating them as independent discards information. The sequential method explicitly models transition probabilities between trial outcomes, testing whether the observed transition matrix departs from the independence baseline. The strength of ruling-out for the independence assumption was partially addressed: the method demonstrates non-independence but cannot fully distinguish psi-mediated from ordinary psychological sources of sequential dependency (e.g., response strategies, fatigue).

Early work by Radin, May, and colleagues explored informational models of psi applied to RNG experiments, asking whether the structure of psi effects was better described by information-theoretic measures than by simple effect sizes.⁶⁷ A 1988 experiment with a single subject tested whether precognition predicted actual or probable futures, finding evidence for precognition (p = .04) and support for the probable-futures hypothesis (p = .001), with a negative relationship between Shannon information in the target set and psi information transmitted (p = .03, two-tailed).⁸ These results informed the sequential analysis program by suggesting that psi effects carry information-theoretic signatures that standard hit-rate analysis may not capture.

The specific sequential metric tested was p₁: the probability that a trial is correct given that the immediately preceding trial was also correct. With a five-target forced-choice protocol, the chance-expected rate for the sequential metric p₁ — the probability of a correct guess immediately following a correct guess — is p₀ = 0.32, and the observed p₁ across the combined dataset was 0.320502 ± 0.000044, yielding z = 11.28, p = 1.7 × 10⁻²⁹.¹ Control tests were conducted to address the artifact of optional stopping and multiple comparisons: the sequential metric was specified in advance as the primary secondary analysis, and control analyses of non-psi sequential patterns in the data were used to assess whether the effect was specific to the psi-relevant transition or a general property of the dataset. The ruling-out of optional stopping was partially addressed by the fixed-N design (data collection ended by calendar date, not by result). The ruling-out of multiple comparisons was partially addressed by the pre-specification of the sequential metric, though the exploratory nature of the secondary analysis means this remains an unresolved concern. The researcher’s preferred interpretation is that the sequential structure reflects a genuine psi effect that manifests as trial-to-trial dependency rather than mean elevation. Alternative interpretations include systematic response strategies by participants that create sequential dependencies through ordinary psychological mechanisms.

The primary non-psi explanation for sequential dependencies in forced-choice psi data is response strategy: participants may develop habitual patterns of target selection that create trial-to-trial correlations through ordinary psychological mechanisms such as the gambler’s fallacy (avoiding repetition) or its inverse (perseveration). Radin addressed this artifact in the 19-year dataset by analyzing the direction of the sequential effect: if participants were avoiding repetition, p₁ should be below chance; if perseverating, above chance. The observed p₁ was above chance, consistent with perseveration or with a psi effect, but the two cannot be fully distinguished without independent measurement of participants’ response strategies.¹ A second non-psi explanation is that the online protocol allowed participants to develop feedback-based learning strategies over many trials, creating sequential dependencies through reinforcement learning rather than psi. The fixed-N design and automated protocol mitigated but did not eliminate this possibility, as participants could accumulate feedback across sessions.

Radin’s 1993 paper in the Journal of Scientific Exploration reported neural network analyses of consciousness-related patterns in random sequences, using new data and new network configurations to replicate a previously reported study.²⁹ Eight analyses confirmed the presence of person-specific signatures. The artifact most directly addressed was the possibility that the neural network was detecting differences in participants’ response strategies or timing rather than differences in the RNG output itself: the network was trained and tested on the RNG data sequences, not on behavioral data. The ruling-out was partial: the network could detect person-specific patterns, but whether those patterns reflected psi-mediated influence on the RNG or ordinary individual differences in how participants interacted with the experimental interface (e.g., timing of button presses affecting which random bits were sampled) was not fully resolved. The researcher’s preferred interpretation is that the signatures reflect genuine mind-matter interaction; the alternative interpretation is that they reflect sampling artifacts from individual differences in response timing.

Mossbridge and Radin (2021) analyzed data from thousands of participants across three iOS tasks.⁴ The primary finding was that psi performance was often in the direction opposite to participants’ conscious intentions, a pattern the authors called “expectation-opposing” (previously termed “psi-missing” in the literature). Gender and psi belief were found to be related to performance outcomes. The artifact of sampling bias was partially addressed by the large sample size and the diversity of the opportunity sample drawn from app store users, but the self-selected nature of participants who download a psi-testing app introduces selection bias that was not fully controlled. The artifact of multiple comparisons across 149 demographic and personality variables was addressed by reporting effect sizes and confidence intervals alongside p-values, but the exploratory nature of the analysis means that some significant findings may reflect Type-I error. The researcher’s preferred interpretation is that the psi-missing pattern reflects a genuine phenomenon in which conscious intention interferes with an underlying psi process; alternative interpretations include response bias and regression to the mean in participants who initially performed above chance.