IPCW Reporting and Diagnostics Checklist

What it is

- This guideline is the checklist layer for inverse probability of censoring weighting (IPCW). The concept explains the estimator; this guideline states what a study must report when censoring weights are used to address informative censoring, treatment switching, artificial censoring, loss to follow-up, or protocol deviations. The goal is to make the identifying assumptions, weight models, diagnostics, truncation, and variance estimation visible enough for a reviewer to judge whether the weighted analysis is credible.

When to use

- Use it for longitudinal cohort, target-trial emulation, per-protocol, as-treated, sustained-treatment, or dynamic-strategy analyses when censoring is plausibly related to future outcomes. Common triggers include treatment discontinuation, switching, add-on therapy, disenrollment, site leakage, missing follow-up assessments, and artificial censoring introduced by clone-censor-weight designs. Apply the checklist before fitting the model, because the numerator, denominator, time scale, covariate history, and truncation rules should be pre-specified.

What it requires / checklist domains

- Define each censoring process separately: administrative study end, loss of observability, treatment switch, protocol deviation, death, care outside the system, and missing assessment. State whether death is censoring, an endpoint, a competing event, or part of a composite estimand. Specify numerator and denominator models for stabilized weights, the time grid, covariate history, and time-varying predictors of both censoring and outcome. Report weight mean, range, percentiles, effective sample size, and practical positivity violations. Pre-specify truncation or winsorization thresholds and report the fraction of records affected. Use robust or bootstrap variance because weights are estimated.

When NOT to use - limitations and common misapplications

- Do not use IPCW as a cosmetic fix for unmeasured reasons for loss to follow-up; the exchangeability assumption still depends on measured predictors. Do not censor death as if it were routine loss to follow-up when the estimand requires a competing-risk, composite, or principal-stratum strategy. Do not hide extreme weights, positivity failure, or truncation behind a single adjusted estimate. Do not fit censoring models using post-censoring information or variables affected by future treatment unless the estimand and time ordering justify it. IPCW can reduce bias from informative censoring, but unstable weights can increase variance and sensitivity to model misspecification.

How it maps to this catalog

- This guideline cross-references `inverse-probability-of-censoring-weighting-rwe` for the estimator, `censoring-mechanisms-rwe` for censoring taxonomy, `clone-censor-weight-per-protocol` for the common target-trial use case, `attrition-and-loss-to-follow-up-rwe` for follow-up diagnostics, `missing-data-pattern-table-rwe` for assessment missingness, `estimands-ate-att-intercurrent-events-rwe` for the estimand decision, and `positivity-overlap` concepts through the weighting and propensity score family where relevant. Use this checklist as the reporting gate around the IPCW concept.