hesela.devInverse probability of censoring weighting
Inverse probability of censoring weighting reweights observed contributions by reciprocal probabilities of remaining uncensored to account for incomplete follow-up.
The denominator is a probability of remaining uncensored, not of being censored. Valid adjustment needs an appropriate censoring model, conditional independence given the included history and positive observation probabilities over the target population and time range. Measured-covariate balance checks cannot establish independence from unmeasured predictors. Balance of covariate means alone can miss nonlinear or interaction-related imbalance. [1]
For a time-dependent Brier score, observed events at or before the horizon and event-free observations through it use weights evaluated at their respective relevant times. Earlier censoring supplies no observed horizon outcome; it is not a nonfailure label. The scikit-survival 0.25.0 implementation estimates censoring with marginal Kaplan-Meier and assumes random censoring independent of features. A computed score does not prove those assumptions or support extrapolation beyond follow-up. [2]
For drive studies, record withdrawal reasons and the history available before withdrawal. Weighting does not reconstruct individual missing failure times or automatically resolve selective retirement. [1]
Relations
- related
- Informative censoring, Right-censoring, Brier score, Prediction horizon, Kaplan-Meier estimator
Citations
- Jackson (2019): Diagnosing Covariate Balance Across Levels of Right-Censoring Before and After Application of Inverse-Probability-of-Censoring Weights
- scikit-survival 0.25.0: time-dependent Brier score, censoring weights and supported evaluation times
Sources & caveats
- The storage application is a methods interpretation, not a fitted fleet result. Jackson (2019), DOI 10.1093/aje/kwz136, concerns statistical methodology rather than drive measurements.