The Evidence Layer for Scientific AI

 Quick Analysis Plan Copied

 Long Analysis Plan

1. Evidence base and what is already controlled

The PNAS study labeled 5-second pre-probe windows with self-reported states (ON, MW, MB, don't-remember) from 62 adults performing a modified SART with high-density EEG, and tested MB against ON/MW using spectral power, complexity (KC, SE), connectivity (wSMI, PLV), ERPs, and MVPA decoding PNAS mind-blanking methods and controls. Critically, vigilance covariates were already entered in some models, so the lapse-confound test is partially built in, but only "in some models" — your analysis must make it uniform. The bioRxiv precursor (26 adults, same paradigm) independently replicated slower RTs and more misses in MB bioRxiv precursor study of absent-mindedness, though at smaller N with probe-catching limits.

2. Analysis plan (code on the Zenodo dataset)

Download the deposited EEG/behavioral data (doi:10.5281/zenodo.17649277; prior dataset at osf.io/ey3ca). Pipeline, in order:

• Step 1 — Recompute the core effects. Rebuild per-trial features (RT, miss, spectral power in delta-theta-alpha-beta-gamma, KC/SE, wSMI/PLV) and linear mixed models with state as fixed effect and subject/dataset random intercepts, mirroring the paper. Record the raw MB vs MW contrast for every feature.
• Step 2 — Uniform lapse accounting. Add vigilance (pre-probe RT trajectory, local RT variability, time-on-task) as covariates to EVERY model, not just some. Use RT distribution modeling rather than mean RT: HALS data show RT mean-variance relations are task- and sex-dependent and nonlinear at extremes HALS reaction-time mean-variance and sex effects, so mean-RT-only lapse adjustment can leave the extreme-tail (miss-prone) trials confounded. Test: does the MB x feature interaction survive a model including (a) trial-level vigilance, (b) a matched-lapse subsample where MB and MW trials are propensity-matched on RT and miss rate?
• Step 3 — Probe-timing misalignment. The 5 s pre-probe labeling assumes state stability across the window. Falsification check: re-run the analysis with windows of 1, 2, 5, and 8 s, and with a state-dynamics model (hidden Markov over inter-probe trials) that estimates when the state actually began. If the MB physiological signature is an artifact of averaging across a state transition, it should shrink or shift as window length changes; if it is genuine, it should be stable or sharpen at shorter windows.
• Step 4 — Decoding as an independent arbiter. Replicate MVPA face-vs-digit decoding: the paper reports robust decoding for ON and MW but not MB MVPA decoding absent during mind blanking. Because decoding operates on stimulus-locked epochs rather than pre-probe averages, it is largely immune to probe-timing misalignment; persistence of the decoding gap after lapse matching is the strongest single piece of evidence for state distinctness.
• Step 5 — Sensitivity to non-ignorable missingness. MB trials may be systematically probe-ambiguous ("don't-remember" labels). Treat this as a missing-data problem: run complete-case vs sensitivity analyses under plausible label-missingness assumptions, since no single imputation approach is unbiased under non-ignorable missingness Nonignorable missing-data handling study.

3. Known limits and what would change the verdict

Self-report probes are the weakest link; classifier performance was modest on average and varied across participants; MB prevalence is trait-dependent; the "don't-remember" category could absorb mislabeled MB. If propensity-matching on RT/misses eliminates the wSMI and P3b differences, or if the signature is window-length-dependent in the way Step 3 predicts for an artifact, MB distinctness fails. Confidence: high that MB co-varies with lapses and EEG changes (strong evidence, two datasets); moderate that it survives full lapse+timing adjustment (not yet uniformly tested).

Figure: analysis logic — the dashed edge is the confound path (generic lapse) that Step 2-3 must sever before concluding MB is physiologically distinct.