Time-to-Event, Competing Risks, and Longitudinal Outcomes Without Wishful Assumptions

Cox models and beyond. The Cox model remains a workhorse, but proportional hazards (PH) should be assessed rather than assumed. Plot Schoenfeld residuals or time-varying effects; when PH is dubious, report RMST differences or fit flexible parametric survival models. For time-varying exposures (dose titration, line switches), use extended Cox models or g-methods tailored to dynamic strategies.

Competing risks. Death or treatment cessation can preclude the outcome. Decide whether interest lies in cause-specific effects (hazards when competing events are censored) or in cumulative incidence (probability of event type by time). Implement cause-specific hazards for etiologic questions and Fine–Gray subdistribution models when predicting absolute risks that mix events matters for decisions. Report both where feasible and reconcile their interpretations in plain language.

Recurrent events. Many outcomes recur (exacerbations, hospitalizations). Choose models that match the clinical mechanism and estimand. Andersen–Gill treats recurrences as a counting process (marginal, order-agnostic); Prentice–Williams–Peterson conditions on prior events (gap-time or total-time); Wei–Lin–Weissfeld fits strata by event order. When burden over time is the target, compare mean cumulative functions; when rate ratios are policy-relevant, estimate marginal rates with robust variance that respects within-person correlation. Pre-specify grace windows that define distinct events to avoid artifact counts in rapid sequences.

Longitudinal responses. For repeated measures (e.g., lab trajectories, symptom scores), choose generalized estimating equations (GEE) for population-averaged effects with robust sandwich errors, or mixed models for subject-specific inference and handling of irregular visit times. State the working correlation (exchangeable, AR-1) and verify sensitivity to that choice. For PROs, follow instrument-specific missingness rules and avoid ad-hoc imputation that violates scale properties.

Intercurrent events and censoring. When events like treatment switching or discontinuation are common and informative, standard censoring produces bias. Use inverse probability of censoring weights (IPCW) or joint models for longitudinal and survival data when trajectories and hazard are entwined. Explain the causal contrast (treatment policy vs. hypothetical no-switching) and show weight diagnostics and effective sample sizes so fragility is visible.

Calibration and discrimination. When predictions guide coverage or safety monitoring, evaluate both discrimination (C-index, time-dependent AUC) and calibration (calibration-in-the-large, slope, and flexible calibration plots). Transport models cautiously: re-calibrate across systems or countries when coding and care patterns differ. File code, parameter hashes, and performance tables with the cut manifest.

Multiplicity and fragile effects. Exploratory subgroup forests invite false positives. Limit to prespecified modifiers with clinical rationale; control familywise error or false discovery rate where confirmatory claims are implied; and present shrinkage or hierarchical partial pooling for many-cell comparisons. Always pair subgroup ratios with absolute risk differences and counts to prevent over-interpretation of small denominators.

Confounding Control, Variance Estimation, Missing Data, and Inference Under Complex Designs

Propensity score (PS) toolset. Use active-comparator, new-user cohorts whenever feasible; then deploy PS methods to balance observed confounders: matching (with calipers and ratio choices), stratification, inverse probability of treatment weighting (IPTW), and overlap or matching weights when positivity is weak. Report pre/post standardized mean differences (target <0.1), PS overlap plots, and the effective sample size (ESS = (∑w)²/∑w²) to reveal variance inflation under weights.

Doubly robust and targeted estimators. Combine PS with outcome models (augmented IPTW, targeted maximum likelihood) to retain consistency if either model is correct. Use cross-validation and simple, interpretable transformations (splines, bins) and keep variable importance summaries. Where machine learning aids fit, log algorithm versions and seeds in the manifest to preserve reproducibility.

Variance that matches the design. Under weighting or matching, default model-based standard errors are misleading. Use robust sandwich variance (with stabilization for small samples), replicate-weight methods (e.g., bootstrap, jackknife) that respect the design (pairs bootstrapping for matched sets; cluster bootstrap for site-level correlation), or M-estimation frameworks implemented with empirical influence functions. In cluster-correlated data (sites, practices), use cluster-robust variance or hierarchical models; declare the level of inference (cluster vs. individual) in the SAP.

Missing data. Separate missing covariates from missing outcomes. For covariates, prespecify multiple imputation using chained equations with passive imputation for derived variables and include outcome and exposure where appropriate to meet congeniality. Combine imputation with weighting carefully (impute first, then compute PS/weights; average treatment effects across imputations with Rubin’s rules, re-computing weights within each). For outcome misclassification (common in claims/EHR), use validation subsamples and probabilistic bias analysis to propagate plausible sensitivity/specificity through effect estimates.

Time-varying confounding. When disease status, adherence, or care intensity both affect outcome and future treatment, standard regression may control away the causal path or open colliders. Use marginal structural models (stabilized IPTW), the parametric g-formula for explicit dynamic regimes, or structural nested models. Present weight distributions and truncation thresholds; test identification assumptions with negative controls where possible.

External controls and borrowing. When integrating registries or literature comparators, diagnose exchangeability (balance metrics, overlap) before combining. If borrowing information, cap influence via robust mixture or commensurate priors (Bayesian) or calibration weighting (frequentist). Simulate operating characteristics (bias, variance, coverage, type I error) under prior-data conflict and weak overlap and include both best-case and adversarial scenarios in the technical appendix.

Distributed networks. In federated analyses, harmonize code lists and model specifications centrally; run locally; and meta-analyze site-level effects with random-effects models when practice patterns differ. File per-site manifests (terminology versions, software, algorithm hashes). Stratify negative controls by site to expose data idiosyncrasies masked by pooling.

Small samples and rare events. For sparse outcomes, consider exact or penalized likelihood (Firth) to reduce small-sample bias; use profile likelihood CIs. In survival with few events, prefer RMST; in logistic settings with near separation, penalization stabilizes inference. Always report event counts per parameter and avoid overfitting through parsimony and shrinkage.

Diagnostics, KRIs/QTLs, Packaging, and a 30–60–90 Plan for Inspection-Ready Biostatistics

Diagnostics that drive action. Dashboards should show: covariate balance by subgroup; PS overlap and extreme weights; IPCW/PS weight distributions and ESS; cluster correlation diagnostics; missingness patterns; negative-control results; and sealed-cut reproducibility status. Each tile must click to proof—tables, code-hashes, manifests, and, when needed, chart-validation artifacts. Numbers without provenance are not inspection-ready.

Key Risk Indicators (KRIs) and Quality Tolerance Limits (QTLs). Examples of KRIs: poor overlap (≥10% of weighted mass at PS <0.05 or >0.95); unstable weights (≥2% beyond truncation); unresolved negative-control signals; persistent PH violations without alternative summaries; or sealed-cut mismatches. Candidate QTLs: “post-adjustment SMD >0.1 for any prespecified confounder,” “ESS <50% of treated cohort after weighting,” “unresolved missingness >10% in critical covariates,” “retrieval pass rate <95%,” or “RMST and HR disagree materially without explanation.” Crossing a limit triggers containment, a dated corrective plan, and owner assignment.

Packaging for regulators, HTA, and journals. Provide a compact dossier that includes: the estimand, a target-trial table, cohort criteria, code lists with versions, exposure and outcome algorithms, confounding strategy, model specifications, variance approach, diagnostics, and sensitivity and quantitative bias analyses. Tables should pair relative and absolute effects; survival outputs should include RMST differences; subgroup tables should show counts and shrinkage-aware estimates. File code and environment hashes; keep sealed-cut identifiers in table footers.

Reproducibility by design. Freeze data, code, and parameters as sealed cuts; store manifests with hashes for inputs, transformations, and outputs; capture random seeds for all resampling and ML fits; and rehearse five-minute retrieval drills that regenerate a key table live. In distributed networks, capture per-site environment summaries and align version bumps with change-control notes that explain impact.

30–60–90-day implementation plan. Days 1–30: define estimands and target-trial tables; inventory outcomes and exposure data; draft SAP with model classes, variance methods, and diagnostics; set up sealed cuts and manifests; prepare code shells for balance checks, overlap, PH tests, and RMST. Days 31–60: build active-comparator, new-user cohorts; execute PS models; finalize weighting/matching choices; run negative controls; implement time-to-event and recurrent-event frameworks; establish multiple imputation pipelines; and pilot federated runs if applicable. Days 61–90: finalize primary and sensitivity analyses; compile diagnostics; simulate operating characteristics for borrowing or complex weights; lock the dossier; and conduct retrieval drills with leadership and statisticians who will face reviewers.

Common pitfalls—and durable fixes.

• Vague time zero or estimand drift. Fix with a target-trial table and lock windows before code runs.
• Assuming PH by habit. Fix with tests, plots, and RMST or flexible models when PH fails.
• Variance that ignores the design. Fix with robust/replicate-weight variance and design-aware bootstraps.
• Positivity violations hidden by averages. Fix with overlap weights, trimming, or redesigned comparators.
• Missingness hand-waved. Fix with principled imputation and outcome misclassification analyses.
• Machine learning without provenance. Fix with logged versions, seeds, and interpretable summaries.
• Unreproducible results. Fix with sealed cuts, manifests, code hashes, and scheduled regeneration tests.

Ready-to-use biostatistics checklist (paste into your SAP template).

• Estimand defined; target-trial table completed; time zero anchored.
• Exposure, outcomes, and follow-up windows prespecified with versioned code lists.
• Confounding plan (matching/weighting/overlap/doubly robust) and diagnostics locked.
• Variance methods design-aware (robust/replicate weights); cluster correlation addressed.
• Time-to-event framework selected; PH assessed; RMST reported when informative.
• Recurrent-event model chosen with grace windows; mean cumulative functions reported as needed.
• Missing-data strategy defined; misclassification assessed via validation subsamples and bias analysis.
• Negative controls specified; quantitative bias/E-value or tipping-point analyses planned.
• Sealed data cuts, manifests, program hashes, and seeds archived; retrieval drills passed.
• KRIs/QTLs monitored; containment playbooks rehearsed with owners and due dates.

Bottom line. RWE biostatistics is a disciplined system: design-anchored estimands, models that respect time and competing risks, confounding control with transparent diagnostics, variance and missing-data methods that match the design, and an evidence chain that explains itself. Build that once—tables, manifests, diagnostics, and retrieval drills—and your estimates will travel across regulators, HTA bodies, journals, and time with confidence.