A Pre-Registered Test of a Correlational Micro-PK Effect: Efforts to Learn from a Failure to “Replicate”

Maier, M. A., & Dechamps, M. C. (2022). A pre-registered test of a correlational micro-PK effect: Efforts to learn from a failure to “replicate”. Journal of Scientific Exploration, 36(2), 251–263. https://doi.org/10.31275/20222235

AI Assessment

A pre-registered micro-PK replication that returned a clean null, made unusually instructive by a data error the authors caught in the original finding. Using a quantum random number generator and Bayesian analysis, the study tested whether observers could bias quantum outcomes and whether a previously reported correlation between two conditions would re-appear. Both standard micro-PK tests favored the null, and the correlation the study was built to replicate turned out to rest on a coding error in the original dataset, which the authors discovered and reported in full. Every figure on this page was verified verbatim against the open-access primary article.

Provenance

DOI. 10.31275/20222235. The article is Platinum Open Access under a Creative Commons CC-BY-NC 4.0 license (submitted 24 June 2021, accepted 22 February 2022, published 22 August 2022).

Study type. Pre-registered Bayesian micro-psychokinesis (micro-PK) study using a quantum random number generator (qRNG), run online with a within-subjects subliminal-priming design. The two authors share first authorship and are both at Ludwig-Maximilians-Universität München.

Funding. The article states no external funding source. Data collection was carried out by the survey company Kantar, and the procedure was approved by the ethical board of the Department of Psychology at LMU Munich and at Kantar.

Data availability. The study was preregistered at OSF (osf.io/a47g2), and the preregistration, the experimental data, and the analysis scripts are openly accessible at osf.io/tbha6 (the stimulus material is excluded for rights reasons).

Source basis. Every figure on this page was confirmed against the primary article as published in the Journal of Scientific Exploration, Volume 36, Number 2 (Summer 2022), pp. 251–263.

What the paper reports

Maier and Dechamps set out to replicate two findings from an earlier series of three micro-PK studies (Dechamps et al., 2021, Studies 1 to 3): a standard micro-PK effect in which observers appear to bias a qRNG toward positive images, and a post-hoc correlation found between the two within-subjects conditions across the combined original dataset (total n = 12,254).¹ In the present study (final N = 2,052), the two standard micro-PK tests both favored the null hypothesis: the experimental (positive-priming) condition yielded a Bayes factor of BF₀₁ = 11.00 (mean positive-image score M = 9.96, SD = 2.26, against a chance expectation of 10 out of 20), and the control (neutral-priming) condition yielded BF₀₁ = 13.95 (M = 9.94, SD = 2.26), each indicating strong evidence for the absence of an effect. The pre-registered correlation between the two conditions was r(2,052) = .01 with a final BF₀₁ = 8.03, moderate evidence for the null.²

The most consequential development was not a psi result but a data error. After the present data had been analyzed, the authors discovered that the strong evidence for the original correlation, r(12,254) = .032, BF₁₀ = 36.46, had been produced by an erroneous coding of eight participants in the original Study 3, whose repeat-participation scores were retained when they should have been deleted. After correcting the error, the original correlation became r(12,254) = −.01 with a final BF₀₁ = 49.48, very strong evidence that the correlation the present study had been pre-registered to replicate never existed. The pre-registered predictions were left in place and the data analyzed exactly as specified, but the authors reframe the study as an unsystematic case report on experimenter expectations: they had held a strong, documented expectation of an effect that was objectively absent from the original data.

This study’s results indicate no evidence for the existence of a correlational (and standard) micro-PK effect. In other words, the actual correlational data did not meet the experimenters’ conscious expectations.

How it was run

• Design. A within-subjects online experiment with two conditions, an experimental (positive-priming) and a control (neutral-priming) condition, each participant viewing 40 trials built from 20 matched positive/negative target-image pairs, each pair used twice.
• Priming procedure. On each trial a prime was masked and presented three times (55 ms each, with 110 ms forward and backward masks) to keep it subliminal. In the control condition the prime was a 50/50 mixture of the two target images; in the experimental condition the positive image was made progressively dominant across the three presentations (50/50, then 60/40, then 70/30).
• Quantum randomness. After the priming sequence a Quantis qRNG by ID Quantique, connected directly to the server without a buffer, generated a single random bit per trial that selected the positive or negative target image. The device passed the DIEHARD and NIST randomness test batteries. Assignment of condition and trial order used a separate pseudo-RNG.
• Outcome measure. The two dependent variables were the mean number of positive images (the number of 0 bits) selected by the qRNG in the experimental and control conditions, with a chance expectation of 10 out of 20 per condition.
• Sample. German participants recruited and tested by Kantar between January and February 2021, aiming for about 100 completions per day. The final sample was N = 2,052; demographic data were available for 1,990 (49.20% male, 50.55% female, 0.25% diverse; mean age 44.22, SD 13.84).
• Stopping rule. A pre-specified Bayesian stopping rule of BF = 10 (strong evidence for either hypothesis), with a supplementary floor of about 1,000 participants and a provision to extend further if a clear trend toward BF₁₀ = 10 was visible at n = 1,000. Analyses used the R ‘BayesFactor’ package, with a correlation prior of ρ ~ Beta(0.1) and one-tailed t-tests using an informed prior δ ~ Cauchy(0.05, 0.05).

Results, as reported

The Bayes factors are reported without confidence intervals, consistent with the Bayesian framework the authors adopt. None of the three Change-of-Evidence analyses reached significance at the alpha = 5% level, so the variation in the sequential Bayes factor of the correlation was not statistically distinguishable from random fluctuation.

Eleven-dimension audit

Pre-registration

The study was pre-registered at OSF (osf.io/a47g2), and the authors state that data collection, reporting, and analyses followed the exact protocol in the preregistration, including the predicted re-appearance of the correlation when standard micro-PK effects were absent. The unusual feature is that the pre-registered hypothesis itself was grounded in a finding later shown to be an artifact of a coding error; the preregistration was honored, but its motivating premise was retrospectively invalidated, which the authors report openly rather than revising the analysis.

Randomization

Target selection used a hardware quantum RNG (Quantis by ID Quantique) generating one bit per trial directly from quantum state reduction, described as passing the DIEHARD and NIST test batteries, with care taken that each participant received an individual bit and that the device worked without a buffer. A separate pseudo-RNG handled the non-critical assignment of condition and trial order. The randomization source for the dependent measure is therefore well specified.

Sensory leakage

Micro-PK has no sender, so the classical leakage concern does not apply; the relevant analogue is whether the subliminal primes were genuinely outside awareness, since a consciously perceived prime could bias choices through ordinary cognition rather than any anomalous influence. The primes were masked and briefly presented to keep them subliminal, but the authors acknowledge as a limitation that prime efficacy was not verified per participant with a signal-detection task, and that occasional conscious identification of a prime cannot be ruled out.

Blinding

The task was a passive online watching task with no experimenter in contact with participants during testing, and outcomes were determined by the qRNG rather than by any rater, so condition-blinding in the conventional sense is not the operative issue. The study’s own framing makes experimenter expectation the variable of interest: both authors held, and pre-registered, a strong expectation of a correlation, which is the opposite of blind, and they treat that fact as the central interpretive caveat rather than concealing it.

Optional stopping

Stopping followed a pre-specified Bayesian rule (BF = 10) with a supplementary floor of about 1,000 participants and an extension provision. The authors report transparently that the BF = 10 criterion for the correlation was met very early (n = 24, BF₁₀ = 13.90) but that they judged the subsample underpowered and continued to N = 2,052, stopping when funds were exhausted. The Change-of-Evidence analyses were post-hoc and not part of the preregistration, which the authors state explicitly.

Outcome measure

The confirmatory outcomes were defined in advance: the mean number of positive images per condition (tested against a chance value of 10 of 20 with one-tailed Bayesian t-tests) and the Bayesian correlation between the two conditions’ mean scores. The priors were pre-specified (ρ ~ Beta(0.1) for the correlation; δ ~ Cauchy(0.05, 0.05) for the t-tests). The primary endpoints are unambiguous and match the preregistration.

Effect size

Effect sizes are reported in the natural Bayesian and correlational terms: correlation coefficients (final r = .01; original corrected r = −.01) and mean scores essentially at chance (9.94 and 9.96 against 10), with Bayes factors quantifying the strength of evidence for the null. The magnitudes are uniformly small and centered on no effect, consistent with the null conclusion the authors draw.

Multiple comparisons

The pre-registered analyses comprise two one-sample t-tests and one correlation. The three Change-of-Evidence analyses (MaxBF, BF energy, FFT) are additional, explicitly labelled post-hoc, and were each evaluated against 1,000 simulations from the same qRNG design rather than against an uncorrected threshold. The authors note that none reached the alpha = 5% level and caution that a statistical trend appeared in two of the three, with one close to the 5% boundary, which they decline to interpret as confirmation.

Internal replication

The two conditions provide a within-study cross-check, and both returned the null. The study is itself a replication attempt of Dechamps et al. (2021), whose three-study series (Studies 1 to 3) had produced a positive micro-PK result in Study 1 (n = 4,092) that did not replicate in Studies 2 and 3; the present null is consistent with that decline. There is no fresh internal replication of the correlation beyond the single pre-registered test reported here.

External replication

The correlational effect failed to replicate, and the original effect it targeted was shown to be a coding artifact, so on this specific claim the external record is negative by the authors’ own correction. The broader micro-PK literature the paper situates itself within includes two meta-analyses reporting an overall significant effect (Bösch et al., 2006; Radin & Nelson, 1989),³⁴ as well as a recent high-power study reporting an active-intention micro-PK effect that replicated across two datasets (Mossbridge & Radin, 2021).⁵ The present study contributes a well-powered null to that mixed record.

Transparency

Transparency is a particular strength of this paper. It is open access; the preregistration, data, and analysis scripts are openly posted; the null results are reported without spin; and, most notably, the authors voluntarily disclose a coding error in their own prior dataset that invalidated the premise of the present study, then explain how it reshapes the interpretation. The authors also enumerate three limitations of their own: that the efficacy of the subliminal primes was not verified per participant with a signal-detection task; that participant compliance and attentional focus during the passive watching task were not controlled, which they note could have deteriorated any effect; and that the target and prime images were selected by expert valence ratings rather than from sets with normative ratings of valence and arousal, a possible confound. No data point is concealed to preserve the original narrative.

The adversarial record

• Precursor work. The design is an exact replication of the group’s own Dechamps et al. (2021) Studies 1 to 3, which used the same subliminal-priming and qRNG paradigm, and it sits within the long qRNG micro-PK tradition initiated by Schmidt in the early 1970s and developed at the PEAR laboratory (Jahn et al., 1997). The change-of-evidence machinery (MaxBF, BF energy, FFT) was developed in the authors’ earlier methodological work.
• Contested-literature context. Micro-PK is among the most disputed areas of parapsychology. The two meta-analyses the paper leans on report an overall effect, but the Bösch, Steinkamp, and Boller (2006) meta-analysis is itself well known for concluding that the effect, while statistically present, is not distinguishable from a small-study or publication bias once heterogeneity is modeled. The paper’s reframing toward “experimenter psi” invokes a hypothesis (Rabeyron, 2020) that many regard as unfalsifiable, since it can absorb both positive and null outcomes.⁶
• Reading against the study. The entire correlational target was an artifact of a coding error, so the study cannot bear on whether that correlation is real, and the authors concede as much. The experimenter-expectation interpretation they offer is explicitly anecdotal and post-hoc: none of the Change-of-Evidence analyses reached significance, the study was not designed as a test of experimenter psi, and the “trend” in two of three analyses is the kind of marginal pattern the same authors built the Change-of-Evidence tests to guard against. As a test of standard micro-PK, the result is a clean, well-powered null.
• Source-fidelity note. The page reproduces the paper’s own correction trail faithfully, including both the erroneous original statistic (r = .032, BF₁₀ = 36.46) and the corrected one (r = −.01, BF₀₁ = 49.48). The early-stopping value (BF₁₀ = 13.90 at n = 24) is reported once in the text as BF₁₀ = 13.90 and once in the MaxBF analysis as BF_max = 13.90; these refer to the same quantity.