Time-to-expiration for evidence bodies of medication harms (TIME): survival analysis from a retrospective cohort

Systematic reviews and meta-analyses are cornerstones of evidence-based medicine, yet their conclusions can be revised or overturned as new evidence emerges. Prior research has mapped how systematic‑review conclusions evolve, but no empirical study has quantified how evidence on drug harms shifts over time. We aimed to examine the stability of safety evidence. We conducted a before‑and‑after comparison study using the SMART Safety dataset, a well-established empirical resource for evidence synthesis of medication-related harms. For each eligible meta‑analysis, we aggregated trial‑level 2 × 2 data annually and applied the IVhet model to generate cumulative effect estimates over time; Kaplan–Meier survival analysis was applied to estimate time to signals of evidence expiration and convergence, with the first-year cumulative estimate as the starting point. We defined signals of evidence expiration as the first change in effect direction, statistical significance, or a ≥ 50% relative change in the pooled odds ratio (OR), and convergence as a ≤ 0.10 absolute difference between consecutive cumulative ORs over five annual iterations. Median times to expiration and convergence were then estimated across eligible meta‑analyses. We included 505 meta-analyses, spanning 310 harm outcomes, with a median of 328 adverse events per meta-analysis. During follow-up, signals of evidence expiration occurred in 356 (70.5%) of these early-stage evidence, with a median time of 3 years (95% CI 3 to 4). Among expiration events, 117 (32.9%) were triggered solely by a ≥ 50% relative change in effect magnitude, and 89 (25.0%) met two or three criteria simultaneously. Subgroup analyses showed similar patterns, with medians ranging from 2 to 7 years. For the 416 with I² < 50%, 278 (66.8%) met the expiration definition, with a median of 4 years (95% CI 3 to 4). Only 27 early-stage evidence (5.3%) met the convergence criterion during follow‑up; the 5th and 25th percentiles for time to convergence were 9 and 31 years, respectively, while the median was not estimable. Current evidence suggests that early meta‑analytic evidence on harms may expire within a relatively short time, while signals of convergence tend to appear much later. These observations point to cautious interpretation of meta‑analytic evidence on harms and suggest that timely updating should be considered.