Skip to content

The Silence Experiment - Phase 0 Statistical Power Analysis and Study Design Table

FieldValue
JournalZenodo (CERN European Organization for Nuclear Research)
Published2026-08-03
AuthorsClifton Bacon
DOI10.5281/zenodo.21778190

Contrast 1 reaches 86% power with 8 participants and 30 sessions per condition at τ = 0.10, while 12 participants at 30 sessions reach only 77% at τ = 0.20. Participant count is the binding constraint, and the unknown heterogeneity parameter τ must be estimated from pilot data before the stopping rule can be locked. Contrast 2 remains underpowered in Phase 0, so its null result should be treated as unresolved.

Contrast 1 compares the two deepest conditions, C6 and C7, with six shallower cells: baseline and C1 through C5. The analysis assumes a standardized effect d = 0.3 and a one-sided per-test threshold α = 0.00167. Bonferroni adjustment applies that threshold across 3 tests from a family-wise α = 0.005.

The model sums a condition mean, participant intercept, participant-specific deep-condition deviation, and per-session noise. Total variance is scaled to 1. Per-session noise has σe = √0.5 ≈ 0.71, with an intercept ICC of 0.5. The participant intercept cancels within participant. The participant-specific deviation follows N(0, τ²), where τ is the SD of effect heterogeneity across participants.

For Contrast 1, per-participant contrast variance is τ² + (2/3)σe²/S. For Contrast 2, it is τ² + 2σe²/S. Here, S is sessions per condition per participant. Power uses a noncentral t distribution with df = P - 1. Every table was also checked against Monte Carlo simulation.

The empty-chamber baseline is part of Contrast 1. The thermal-dummy control is excluded from Contrast 1, Contrast 2, the Jonckheere-Terpstra test, and the secondary protective-versus-suppression test. It therefore contributes to none of the reported power results.

Power at S = 30 varies sharply with τ and participant count.

Contrast 1 power at d = 0.3 and α = 0.00167. Values at or above 80% meet the target.

Participantsτ = 0.05τ = 0.10τ = 0.15τ = 0.20
40.230.140.080.05
50.500.310.180.10
60.760.530.310.18
80.980.860.620.39
101.000.970.840.61
121.001.000.940.77

At τ = 0.05, 6 participants reach 0.76 and 8 reach 0.98. At τ = 0.10, 8 reach 0.86 and 10 reach 0.97. At τ = 0.15, 10 reach 0.84. At τ = 0.20, 12 reach only 0.77, so none of the tabulated sample sizes reaches 80%.

The result prevents a single participant target from being set before τ is known. Pilot data must supply that estimate, and the table then determines P. This is the note’s principal design finding.

At τ = 0.10, 4 participants gain power from 0.06 at S = 10 to 0.16 at S = 40. The 5-participant row rises from 0.13 to 0.37. Neither sample reaches the 80% target.

At P = 8, power rises from 0.48 at S = 10 to 0.64 at S = 15 and 0.75 at S = 20. It reaches 0.86 at S = 30 and 0.91 at S = 40. At P = 10, S = 15 already gives 0.86, while S = 40 gives 0.99.

Each added session reduces the noise term, but it cannot remove participant-level heterogeneity. Under moderate heterogeneity, the note concludes that roughly 8 or more experts are needed. Participant recruitment therefore determines feasibility more strongly than total session volume.

Contrast 2 compares C7 with C6, so its per-participant noise variance triples relative to Contrast 1. At τ = 0.10, P = 8 produces 0.14 power at S = 10, 0.48 at S = 30, and 0.59 at S = 40. P = 10 reaches 0.70 at S = 30 and 0.82 at S = 40. P = 12 reaches 0.86 at S = 30 and 0.93 at S = 40.

The author retains Contrast 2 in the primary family at the corrected alpha but labels it underpowered. A null is uninformative rather than disconfirming.

The reported Contrast 2 powers are upper bounds. They use the full 30-minute C7 period, whereas the pre-registered primary form uses a duration-matched 15-minute C7 window. The shorter window has less averaging time. The note does not recompute the values because its conclusion does not change.

The secondary protective-versus-suppression test is not separately powered. It is expected to fall between Contrast 1 and Contrast 2 in sensitivity, and closer to Contrast 2. Its estimate and confidence interval are reported in all cases.

At S = 30, the minimum detectable d is the smallest effect that reaches 80% power. Under the note’s criterion, a null at P = 8 and τ = 0.10 rules out d ≥ 0.28. At P = 5 and τ = 0.15, it rules out only d ≥ 0.64. At P = 6 and τ = 0.05, the threshold is 0.31, just above the target d = 0.30.

The note warns against assuming that Phase 0 constrains consciousness-T2 coupling at the d ≈ 0.05 to 0.10 scale reported in prior consciousness-RNG work. That work measures RNG output statistics, not T2 decoherence. Phase 0 cannot reach those effect sizes in either channel.

The protocol should pre-register P and τ, not only the total session count. Sessions per condition should be chosen so simulated Contrast 1 power is at least 0.80 at d = 0.3 across the pre-registered τ range, given the enrolled P.

A short pilot should estimate τ and σe. The recommendation is 2 to 3 practitioners with 5 to 10 sessions each, using the same chamber and sensor. Pilot data should remain separate from confirmatory data. Pilot participants should be excluded from the confirmatory sample when the pool permits, and pilot data should never be pooled into Phase 0. The simulation should be locked before confirmatory data collection.

The fixed session count should be replaced by a rule based on P and the pilot-confirmed τ. The protocol should also set a minimum S floor, with S ≥ 15 given as an example. It should report the MDE at the final P and τ. The paper states that d = 0.3 is a feasibility-driven target, not a theory-derived prediction.

If P falls below 8 at τ = 0.10, the Bayesian track should become the primary readout, with the frequentist test reported for completeness. Bayes factors are likelihood ratios and do not depend on degrees of freedom or critical-value inflation. At P = 5, the frequentist test faces a critical value of about 6.3, which may be structurally unreachable at d = 0.3.

If the required P for 80% power exceeds the recruitable pool, the Bayesian track becomes primary, the MDE is recalculated, and Phase 0 becomes a pilot for the empirical τ and noise model needed by Phase 1. Definitive C7-versus-C6 resolution is deferred to Phase 1, a larger pool, and the transmon sensor.

Analytic values were checked against Monte Carlo simulation with 20,000 replications per cell. Under the null at d = 0, the empirical false-positive rate was 0.0015, close to the 0.00167 target.

For Contrast 1 at P = 8, S = 15, and τ = 0.10, analytic power was 0.642 and Monte Carlo power was 0.643. The P = 10, S = 20 design point had analytic power 0.924 and was confirmed by simulation. At P = 6, S = 30, and τ = 0.05, analytic power was 0.760 and was also confirmed.

Before confirmatory collection, the protocol should lock P, pilot estimates of τ and σe, the required S, simulated Contrast 1 power, and the MDE. If fewer than 8 participants are available at τ = 0.10, the Bayesian track becomes primary and Phase 0 should be reframed as a pilot for Phase 1.

View Original Abstract

Two supplementary documents supporting The Silence Experiment, a pre-registered protocol testing the von Neumann-Wigner interpretation of quantum mechanics against environmental decoherence theory, using EEG-verified observer cognitive state as the independent variable and T₂ decoherence on an NV-center diamond sensor as the dependent variable. This record contains the Stage 1 study design table and the Phase 0 statistical power analysis. They are deposited together because each depends on the other: the design table’s sampling plan and stopping rule are derived from the power analysis, and the power analysis is computed against the design the table specifies. Study design table. Prepared to the PCI Registered Reports Stage 1 template, landscape format. The table maps each research question to its hypothesis, sampling plan, analysis plan, and pre-registered interpretation under every anticipated outcome, covering four hypotheses and three auxiliary assumptions. H1 (Contrast 1), confirmatory: T₂ at [C6, C7] exceeds T₂ at [chamber baseline, C1-C5]. Because the empty-chamber baseline sits in the comparison group, the predicted direction commits the protocol to a coherence-protective contribution (Γcog < 0) rather than merely to the absence of a decohering one. H1b, secondary: [C6, C7] against the empty-chamber baseline alone, the only pre-registered test that separates the protective prediction from a suppression pattern in which the deep states merely return the sensor to its empty-chamber value; no primary test distinguishes them. H2 (Contrast 2): C7 versus C6, retained in the primary family but underpowered in Phase 0, computed on a duration-matched 15-minute C7 window anchored to EEG-verified state entry. H3: Jonckheere-Terpstra ordered test across C1-C7. A1 covers the frozen EEG classifier and its pilot firewall. A2 covers environmental isolation and the metrology co-product. A3 covers the pre-registered positive control, which establishes that the measurement chain is sensitive rather than merely quiet, and without which a null could not be reported as an upper bound. A further row specifies conventions common to all hypotheses: randomisation with a lodged seed and allocation concealment, permuted blocks balanced across time-of-day strata, the frozen T₂ estimation pipeline, exclusion rules at exactly three levels with no outlier rule on T₂ values, and missing-data handling. The table is design-level and host-independent; host-dependent details are supplied in the Stage 1 manuscript body once a host institution is secured. Phase 0 power analysis. A technical note computing achievable power across the Phase 0 design space and resolving the dependency between the protocol’s power section and its stopping rule. The core finding is that participant count, not session count, is the binding constraint. Each participant’s contrast estimate is modelled as a group-level effect plus a participant-specific deviation with standard deviation τ; because τ is unknown until pilot data exist, power is computed across τ ∈ {0.05, 0.10, 0.15, 0.20}. All figures use the non-central t distribution with df = P − 1, treating the participant as the unit of inference, cross-checked against Monte Carlo simulation. At moderate heterogeneity (τ = 0.10) with S = 30 sessions per condition, 8 participants yield power of 0.86 for Contrast 1 and 10 yield 0.97; at τ = 0.15, 10 yield 0.84 and 12 yield 0.94. No realistic session count rescues a four-to-five person study. Contrast 2 is substantially less powered at every design point, because comparing one condition against one triples the per-participant noise variance; even P = 12, S = 40 reaches only 0.93. A null on Contrast 2 is therefore uninformative rather than disconfirming. Low power qualifies the interpretation of a null; it is not grounds for moving a pre-registered directional test out of the confirmatory family after the design is fixed. The note also reports what a null buys: at P = 8 and τ = 0.10, a null rules out d ≥ 0.28; at P = 5 and τ = 0.15, only d ≥ 0.64. Phase 0 cannot reach the d ≈ 0.05-0.10 effect sizes reported in prior consciousness-RNG literature, which in any case tests a different measurement channel. Seven recommendations follow, covering pre-registration of P and τ, a firewalled pilot, a power-derived stopping rule, statement of the minimum detectable effect, use of the Bayesian track at small P, a contingency for unreachable P, and retention of Contrast 2 in the primary family with an explicit underpowered label. Type I error is validated under the null at an empirical false-positive rate of 0.0015 against the 0.00167 target, with 20,000 replications per cell. Common statistical framework. Family-wise alpha 0.005 one-sided across the three primary tests; Bonferroni per-test alpha ≈ 0.00167. The secondary discriminator is reported at alpha 0.005 outside the correction. Target minimum detectable effect d = 0.3 on Contrast 1, feasibility-driven rather than theory-derived: no current theory predicts an effect size for a consciousness-dependent T₂ shift, nor its sign. A thermal-dummy control cell is recorded alongside the eight analysed cells and excluded from every confirmatory test, and therefore from every power figure.