Forecasting missing endpoint data in clinical trials and choosing remedies from randomised evidence: models and validation on 718 completed trials

Background. Clinical trials lose time and statistical power when required assessments go missing, and the loss is usually discovered at database lock, when nothing can be done. Methods. We describe models that turn a trial's incoming records into decision support: continuous measurement of endpoint completeness against the protocol, the sponsor's expectation and the floor in the statistical analysis plan; a probabilistic forecast of completeness at the next data cut from a hierarchical beta-binomial model with a common shock, change-point refitting and online conformal calibration; attribution of shortfalls to operational causes with a robust over-dispersion factor and family-wise error control; and remediation, in which each candidate solution's effect comes from real randomised evidence or from the gap measured in the trial's own data and is propagated through the forecast by Monte Carlo simulation. No model writes data; every output becomes action only through a signed, versioned rule. The models were tested on 718 completed randomised neurology trials from ClinicalTrials.gov (181,302 participants), with the 100 most recent held out, and on 48 published randomised evaluations of retention strategies and 35 pooled estimates. The demonstration portfolio and a stress test of 1,200 study runs use synthetic records. Results. Completion varied between trials and between arms far beyond binomial noise. The hierarchical model predicted 260 held-out arms with 82% of outcomes inside 80% intervals, where a naive binomial estimate covered 51%, and the cause flag's false-alarm rate on a randomisation-balanced reason was 1.7%. Most light-touch strategies changed completion little (text reminders odds ratio 1.08) while advance notice (1.37) and payment (1.53) helped; the meta-analytic intervals contained 95% of held-out real trials at the 95% level. Grounding the library in this evidence changed the recommended remedy in the worked example from an earlier reminder to a change that removes the cause. Against the standard estimator behind completion dashboards, scored on the same held-out trials, the system's intervals were right about as often as they claimed while the standard's were right about half as often, and its false-alarm rate is at or below published central-monitoring benchmarks. Under five departures from the model's assumptions, forecast skill stayed at 26–34% over persistence. Conclusions. We set out, claim by claim, what is verified, what is validated on real data and what remains an assumption, and the protocol that will validate the rest on real assessment-level records. The system is deployed on a live study, but no result here is a measurement of its effect on one. Companion paper: https://doi.org/10.5281/zenodo.23698560