Drug Safety Evidence Has a Shelf Life: Most Harm Findings Expire Within Three Years

Every prescription a doctor writes rests on a quiet assumption: that the evidence describing a drug’s harms is still true. A new study published in BMC Medicine suggests that assumption deserves far more scrutiny than it usually receives. An international team of researchers led by Xinyi Wang of the Chinese Evidence-based Medicine Center at West China Hospital, Sichuan University, and Chang Xu of Shanghai Jiao Tong University School of Medicine, set out to answer a deceptively simple question with major consequences for patient safety: how long does the evidence base behind a medication harm signal actually remain stable before it changes in a meaningful way? Their answer, drawn from one of the largest empirical investigations of its kind, is sobering. For most meta-analyses of drug harms, the early evidence effectively expires within roughly three years, while the point at which estimates truly settle down may take decades to arrive, if it arrives at all within a researcher’s career.

The study, formally titled Time-to-expiration for evidence bodies of medication harms, or TIME, is the first empirical attempt to quantify how safety evidence on medications shifts over time. Systematic reviews and meta-analyses are the cornerstones of evidence-based medicine, pooling results from many randomized controlled trials to produce a single, more precise estimate of an effect. But these pooled conclusions are not carved in stone. As new trials are published, the cumulative estimate can drift, sometimes changing direction, sometimes losing or gaining statistical significance, and sometimes swinging dramatically in magnitude. Prior research has mapped how conclusions of systematic reviews evolve in general, but no one had systematically measured this phenomenon specifically for drug harms, an area where instability can directly affect prescribing decisions, regulatory warnings, and patient counseling.

To build their cohort, the researchers turned to the SMART Safety dataset, a well-established empirical resource designed for evidence synthesis of medication-related harms. From this resource they identified 505 eligible meta-analyses covering 310 distinct harm outcomes, with a median of 328 adverse events captured per meta-analysis. Rather than treating each meta-analysis as a static snapshot, the team adopted a dynamic, cumulative approach. For each meta-analysis, they aggregated trial-level two-by-two data year by year, and applied the IVhet model, an inverse-variance heterogeneity model designed to produce robust pooled effect estimates in the presence of between-study heterogeneity. This generated a running sequence of cumulative odds ratios over time, starting from the first-year cumulative estimate and extending forward as new evidence accumulated annually.

The analytical framework borrowed a tool from a very different corner of medicine: survival analysis. Just as oncologists track time to death or disease progression, the researchers tracked time to two distinct statistical events. The first was evidence expiration, defined as the first occurrence of any of three signals: a change in the direction of the effect estimate, a change in statistical significance, or a relative change of at least 50 percent in the pooled odds ratio. The second was evidence convergence, a more demanding criterion requiring that consecutive cumulative odds ratios differ by no more than 0.10 in absolute terms across five consecutive annual iterations. Using Kaplan-Meier survival curves, the team then estimated the median time to each event across the entire cohort of meta-analyses, treating each evidence body as a subject whose stability could be followed over time.

The headline finding is striking. During follow-up, signals of evidence expiration occurred in 356 of the 505 meta-analyses, or 70.5 percent of the early-stage evidence bodies examined. The median time to expiration was just three years, with a 95 percent confidence interval of three to four years. In other words, within roughly three years of the first pooled estimate being formed, the majority of harm evidence had already shifted in at least one fundamental way. Subgroup analyses across different categories of evidence showed similar patterns, with median times to expiration ranging from two to seven years depending on the subgroup. Even among the 416 meta-analyses with low heterogeneity, defined as an I-squared statistic below 50 percent, 278, or 66.8 percent, still met the expiration definition, with a median time of four years. Statistical homogeneity, it turns out, offers only modest protection against the erosion of early conclusions.

Drilling into the mechanics of expiration reveals which signals did the damage. Among the 356 expiration events, 117, or 32.9 percent, were triggered solely by a relative change of at least 50 percent in the effect magnitude, meaning the pooled odds ratio more than doubled or more than halved without necessarily flipping direction or significance. Meanwhile, 89 events, or 25.0 percent, met two or three of the expiration criteria simultaneously, indicating that a substantial minority of evidence bodies underwent wholesale transformations rather than subtle adjustments. These are not trivial statistical tremors. A 50 percent swing in an odds ratio describing a drug’s adverse effect could mean the difference between a harm considered clinically negligible and one demanding regulatory attention, or vice versa. The fact that such swings commonly occur within the first few years of an evidence body’s existence challenges the comfortable assumption that early meta-analytic findings on safety can be treated as settled.

If expiration is fast, convergence is glacial. Only 27 of the 505 early-stage evidence bodies, a mere 5.3 percent, met the convergence criterion at any point during follow-up. For those that did converge, the 5th percentile of time to convergence was nine years and the 25th percentile was 31 years, while the median could not even be estimated because so few evidence bodies reached the criterion. The practical implication is stark: the point at which a harm estimate can be considered genuinely stable, with consecutive annual estimates differing by no more than 0.10 across five consecutive years, is likely to arrive decades after the first evidence appears, if it arrives at all. Clinicians, guideline developers, and regulators are therefore operating in a prolonged window in which the evidence they rely on is both unstable and far from converged, a combination that demands explicit acknowledgment in how safety conclusions are framed and communicated.

The authors draw a direct and cautious conclusion from these patterns: early meta-analytic evidence on medication harms may expire within a relatively short time, while signals of convergence tend to appear much later, and timely updating of safety reviews should therefore be actively considered. This has cascading implications across the evidence ecosystem. Living systematic reviews, which are updated continuously as new trials emerge, suddenly look less like an administrative luxury and more like a methodological necessity for drug safety. Guideline panels that cite harm meta-analyses may need to attach explicit currency dates and reassessment triggers to their recommendations. Pharmacovigilance systems, which monitor drugs after market approval, could use expiration-style metrics to flag when accumulated evidence about a known adverse effect has shifted enough to warrant re-evaluation. The study’s framing of evidence as something with a measurable survival curve, analogous to a patient’s disease-free survival, offers a portable conceptual tool for all of these applications.

There are, of course, important nuances to keep in mind when interpreting these results. The expiration signals are statistical events, not proof that any particular clinical conclusion was wrong; a change in significance or a 50 percent shift in an odds ratio may reflect the play of chance in small early evidence bases, genuine new information from better-designed trials, or both. The convergence criterion, requiring five consecutive years of near-identical estimates, is deliberately stringent and would classify some genuinely stable evidence as not yet converged. The study also examines evidence bodies at the meta-analysis level rather than tracking individual clinical decisions, so the downstream impact on prescribing behavior remains to be studied. Still, the sheer scale of the cohort, 505 meta-analyses and 310 harm outcomes, and the consistency of the findings across subgroups lend considerable weight to the central message. Evidence on drug harms behaves less like a finished building and more like a structure under continuous renovation, and anyone who steps inside should check the inspection date.

For a field that has invested enormous effort in standardizing how evidence is produced, through PRISMA reporting guidelines, prospectively registered protocols, and GRADE certainty ratings, the TIME study adds a dimension that has largely been missing: time itself as a quality attribute of evidence. A meta-analysis published this year may carry a confidence interval, a heterogeneity statistic, and a risk-of-bias assessment, but it does not yet carry a shelf life. This research suggests it should. As the authors and their international collaborators, spanning institutions from the University of Queensland to the University of York to the University of North Carolina, make clear, the goal is not to paralyze decision-making by declaring all early evidence unreliable. It is to calibrate confidence appropriately, to update systematically and promptly, and to recognize that in the specific and consequential domain of medication harms, yesterday’s pooled odds ratio may already be yesterday’s news.

Subject of Research: Temporal stability and expiration of meta-analytic evidence on medication harms

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

Article References: Wang, X., Furuya-Kanamori, L., Doi, S. A., Gu, Z., Loke, Y., Li, S., Mayo-Wilson, E., Lin, L., Golder, S., Chu, H., Vohra, S., Guo, X., & Xu, C. (2026). Time-to-expiration for evidence bodies of medication harms (TIME): survival analysis from a retrospective cohort. BMC Medicine. https://doi.org/10.1186/s12916-026-05296-8

Image Credits: AI Generated

DOI: 10.1186/s12916-026-05296-8

Keywords: medication harms, drug safety, meta-analysis, systematic reviews, evidence expiration, evidence convergence, survival analysis, pharmacovigilance, IVhet model, odds ratio, evidence-based medicine, BMC Medicine