Inverse Probability Censoring Weighting (IPCW)

IPCW assigns a weight to each uncensored observation equal to the inverse of the estimated probability of remaining uncensored up to that time point.

• Weights are estimated using a model for the censoring process, often a Cox proportional hazards model or Kaplan-Meier estimator with censoring as the event.
• Stabilized weights are commonly used to reduce variability: the numerator is the marginal probability of being uncensored, and the denominator is the conditional probability given covariates.
• The weighted pseudo-population represents what would have been observed in the absence of dependent censoring.

Standard survival methods assume non-informative censoring, where the censoring mechanism is independent of the event time. IPCW relaxes this assumption.

• Dependent censoring occurs when factors that predict censoring also predict the outcome, such as sicker patients dropping out of a clinical trial.
• By modeling the censoring process conditional on observed covariates, IPCW creates a pseudo-population where censoring becomes independent of the outcome.
• This is essential in observational studies and long-term follow-up where loss to follow-up is correlated with prognosis.

IPCW is implemented by maximizing a weighted pseudo-likelihood where each subject's contribution is multiplied by their inverse probability weight.

• For the Cox model, the partial likelihood is modified so that risk set contributions are weighted.
• The resulting IPCW estimator is consistent and asymptotically normal when the censoring model is correctly specified.
• Variance estimation requires robust sandwich estimators to account for the uncertainty in the estimated weights.

Accurate IPCW depends on correctly specifying the censoring model, which predicts the probability of being observed at each time point.

• Common approaches include logistic regression for discrete time points or Cox regression treating censoring as the event.
• All time-fixed and time-varying confounders of the censoring-outcome relationship must be included.
• Misspecification of the censoring model can lead to biased estimates, making sensitivity analyses and diagnostics critical.

IPCW is widely used in comparative effectiveness research and clinical trials where informative dropout threatens validity.

• Treatment switching: When patients in the control arm cross over to the experimental treatment after progression, IPCW can adjust for the resulting dependent censoring.
• Quality-adjusted survival analysis: IPCW corrects for censoring in the estimation of quality-adjusted life years (QALYs).
• Recurrent events: IPCW handles dependent censoring in models for repeated hospitalizations or disease exacerbations.

IPCW relies on the assumption of no unmeasured confounders of the censoring process, which is untestable from observed data.

• Extreme weights: Subjects with very low probability of being uncensored receive very large weights, inflating variance. Weight truncation is often applied as a bias-variance trade-off.
• Positivity assumption: Every subject must have a non-zero probability of being uncensored at each time point, given their covariates.
• Diagnostic tools include weight distribution plots and standardized differences in covariates before and after weighting.