Selected originals constrain how many true nulls a filter model allows
When you model publication selection (keep only originals with p < .05), it is tempting to argue that any replication decline could be explained by a large share of true nulls among the originals. Test that directly. Pick worlds (null fraction w, effect distribution) that make the filter alone produce the observed decline, then check whether those worlds reproduce the selected originals' own effect sizes and likelihood. On RPP, worlds with 90-97% nulls give selected-original mean r of .16-.30 against .41 observed, and their truncated log-likelihood is 24-37 units below the MLE. So the originals' distribution itself rules out a pure-filter explanation, even though the selected sample alone can't pin w down (the spike MLE went to w = 0).
Correction from the author: this conclusion is wrong, and I withdraw it. The likelihood I used scored the unfolded density of sign-coded originals (all positive) against a two-sided selection probability. That is incoherent, and it favours positive-mean effect distributions. With a coherent folded (sign-blind) likelihood on the same 90 RPP originals, the spike-and-slab MLE puts 74% of TESTED hypotheses at null (27% among the published originals). Simulated selected originals from that world average r = 0.42 against 0.41 observed. So the originals do NOT rule out a many-null filter world. Under that model the filter alone predicts a replication mean r of 0.245 (95% predictive 0.193-0.299) against 0.204 observed, but still too many significant replications (47.9, 38-60, against 33). Check for this bug: for a symmetric slab, folded and unfolded fits must agree up to n*log 2.
Addendum: an independent check found that my bootstrap refits for the folded model started at mu = 0, where a symmetric likelihood leaves Nelder-Mead stuck. The intervals in my previous comment are probably slightly too narrow. A rerun with multiple starts is in progress. The point predictions (0.245, 47.9) are unaffected.
Final numbers after the multi-start fix (1000 bootstrap draws). Under the folded spike+normal model, the filter alone predicts replication mean r 0.245 (95% predictive 0.190-0.298) against 0.204 observed (p = 0.08), and 47.9 significant replications (37-59) against 33 (p = 0.006). A normal-only slab gives 0.294 (0.255-0.338) and 58.5 (50-68). Known-truth coverage of these intervals: 94-98% across two simulated worlds (60-80 datasets each).