Change-Point Detection Does Not Detect Emerging Adverse Events Earlier than Disproportionality Analysis at an Equal Evidence Threshold: A Corrected Re-analysis of Australia's DAEN

Version 4 correction. Versions 1 to 3.3 of this preprint reported that ensemble CPD detected emerging adverse events in established drugs about three quarters earlier than cumulative disproportionality analysis (the pre-registered H4 subgroup result), and presented this as the paper's deployable finding. That conclusion does not hold. Analyses added in this version (Sections 2.11, 3.2 and 3.4) show that, in this subgroup, CPD detections were dated at or before the quarter of the pair's first report in 12 of 14 pairs, when a median of one report had been received; the disproportionality rules required at least three. The confirmation criterion of at least 10 reports was assessed on each pair's whole-series total, so it drew on reports received after the detection quarter. Held to the same evidence floor as the comparators, CPD detected these signals no earlier (median difference 0 quarters). The prospective simulation recorded back-dated change-point quarters rather than the quarters in which an alarm would have been raised; corrected alarm times are reported in Section 3.4. The title, Key Points, Abstract, Sections 1, 2.4, 2.6, 2.7, 2.9 to 2.11, 3.2 to 3.4, 3.7, 3.8, 3.11, the Discussion, the Conclusions and the captions of Figures 1 and 2 have been revised. The validation-tier performance figures, cross-database results and the retraction of the confidence-PPV inversion are unchanged. Background. Disproportionality analysis (DPA) underpins spontaneous-reporting pharmacovigilance. Applied cumulatively, it flags when a drug-event association crosses a threshold, not when reporting changed. Change-point detection (CPD) scans time series for distributional shifts and has been proposed as a faster complement, but head-to-head timing comparisons must hold the evidence available to each method constant. Methods. We applied an ensemble of three CPD algorithms (PELT, CUSUM, and Bayesian Online Change Point Detection, BOCPD) to quarterly drug-specific adverse event reporting proportions in Australia's Database of Adverse Event Notifications (DAEN; 664,747 reports, 2004 to 2025), and compared detection quarters with cumulative PRR and, post hoc, cumulative BCPNN-IC. The pre-registered primary hypothesis (H1) was the overall paired CPD-vs-PRR difference on Tier A confirmed pairs; the directional secondary (H4) predicted a larger CPD advantage for established drugs developing a new adverse event (genuine emergence) than for newly marketed drugs. Validation used 32 Therapeutic Goods Administration (TGA) regulatory actions (Tier A; the 10-report confirmation threshold was derived on this set), 40 author-curated associations (Tier B), and 230 OMOP/EU-ADR reference pairs (Tier C, the only independent benchmark). Post hoc, we ran CPD on the disproportionality trajectories themselves, counted the reports each method had received when it fired, compared methods at an equal floor of 3 received reports, and recomputed prospective BOCPD at the quarter the alarm would have been raised. Parallel FAERS series were built for the 32 TGA pairs. Results. Sensitivity was 74% on Tier A (in-sample threshold), 95% on Tier B and 21% on Tier C; specificity was 89% (10 false positives among 87 negative controls). H1 was null against both comparators (median 0 quarters; p = 0.47 vs PRR, p = 0.53 vs BCPNN-IC). As originally computed, H4 appeared supported: in the Tier A genuine-emergence subgroup (n = 14), CPD was earlier than PRR by a median of 3.5 quarters (p = 0.015) and than BCPNN-IC by 3.0 quarters (p = 0.018). CPD applied to the cumulative IC trajectory retained this difference (+3.0 quarters, p = 0.002), excluding the input series as its cause. The cause was the evidence at detection: CPD fired with a median of 1 report received (comparators 4), at or before the first report in 12/14 pairs, and none had reached the 10-report confirmation threshold at detection. At an equal 3-report floor the difference was 0 quarters (p = 0.67 vs PRR; p = 0.97 vs BCPNN-IC). Prospective BOCPD quarters had been back-dated by a median of 8 quarters; at true alarm quarters with a 3-report floor, BOCPD was no earlier than PRR (median 0 quarters; 2 earlier, 8 later, 12 tied; p = 0.071). Pooling DAEN and FAERS accelerated detection over DAEN alone by 8 quarters (p < 0.0001) but not over FAERS alone (p = 0.28): the pooled quarter equalled the FAERS quarter in 27/31 pairs. A previously reported confidence-PPV inversion could not be reproduced and remains retracted. Conclusions. Ensemble CPD did not detect emerging adverse events earlier than cumulative disproportionality analysis in DAEN once both were held to the same evidence floor, overall or in established drugs. An apparent subgroup advantage arose because CPD was allowed to fire on less evidence and was confirmed with hindsight. Earlier-detection claims for signal-detection methods should report the evidence each method had received at the moment of detection, confirm only on data available at that moment, and compare alarm times rather than retrospectively located change-points. Cross-database benefit reduces to access to the larger database.