Replication notes · Zenodo 10.5281/zenodo.22677480

Package: Sadeghi, Pehlivanova & Cozzolino (2026), data and R code for "A large-scale analysis of temporal patterns in children's purported past-life memories". CC BY 4.0. Replicated by Epistemic Systems Lab, October 2026.

Outcome

Computational reproduction succeeded. The authors' R script ran without changes under R 4.3.3 (the script states 4.5.1) and produced the output in r_output.txt. An independent Python re-implementation (replicate.py, pandas/scipy/statsmodels) gave identical values to three decimals for every statistic compared.

The UVA Division of Perceptual Studies announcement reports 1,678 cases and a 40 to 45 per cent drop in median interval between the least and most unexpected deaths. The data give 1,678 verified cases and medians of 21 and 12 months for ratings 0 and 4, a 43 per cent drop. The APA article itself was not available to check line by line.

Main results reproduced (verified cases)

Statistic R (authors) Python (independent)
Unexpectedness × log IT, Pearson r −0.126, N = 1,075 −0.126
Spearman ρ −0.133 −0.133
Naturalness × log IT, r 0.132, N = 1,235 0.132
Age at death × log IT, r 0.042, p = .157 0.042, p = .16
Joint OLS: unexpectedness b, p −0.0306, .027 −0.0306, .027
Joint OLS: naturalness b, p 0.064, .115 0.064, .115
Median IT by rating 0/1/2/3/4 21/20/17/17/12 same

Effect size is small: the joint model explains under 2 per cent of variance in log interval.

Sensitivity checks added (not in the authors' script)

Check Unexpectedness × log IT
Exclude the one negative interval r = −0.124, N = 1,074
Exclude intervals above 20 years (37 cases) r = −0.113, N = 1,038
Exclude same-family cases r = −0.118, N = 695
Same-family cases only r = −0.113, N = 380
No family link only r = −0.170, N = 272
Log constant +13 instead of +4 r = −0.117
Mann-Whitney, rating 4 vs 0–3 medians 12 vs 18 months, p < .001
Bootstrap 95% CI for median difference 3 to 11 months
Median regression on raw months −2.7 months per rating step, p < .001

The direction and significance of the main association held in every check.

Coding observations

  1. Endpoint-only anchors. The codebook defines unexpectedness 0 ("anticipated >1 month") and 4 ("completely unexpected") only. Ratings 1 to 3 have no written definition in the package.
  2. No reliability information. The package does not say who rated unexpectedness, verification or family association, how many raters there were, or what agreement they reached.
  3. Blank missing values. All missingness is coded as blank, so "not applicable", "not recorded" and "unknown" cannot be told apart.
  4. Missing ratings differ on the outcome. Of verified cases with a known interval, 244 lack an unexpectedness rating. Their median interval is 24 months against 14 for rated cases. If ratings were harder to make for some kinds of case, the rated subset is not a random sample of cases.
  5. Overlapping codes. 728 of 838 unnatural deaths are rated 4 on unexpectedness. With both in one model, naturalness loses significance and unexpectedness keeps it at p = .027.
  6. Ordinal scale in parametric tests. Unexpectedness is ordinal; the authors' Spearman check gives the same result as Pearson.
  7. Package housekeeping. The README refers to CORT_IT_analysis.R; the file supplied is Analysis Code.R. The longest interval in the data is 1,800 months (150 years), worth confirming against the source record.

Files