Introducing stipw: inverse probability weighted parametric survival models

Inverse probability weighting (IPW) can be used to estimate marginal treatment effects from survival data. Currently, IPW analyses can be performed in a few steps in Stata (with robust or bootstrap standard errors) or by using stteffects ipw under some assumptions for a small number of marginal treatment effects. stipw has been developed to perform an IPW analysis on survival data and to provide a closed-form variance estimator of the model parameters using M-estimation. This method appropriately accounts for the estimation of the weights and provides a less computationally intensive alternative to bootstrapping. stipw implements the following steps: (1) A binary treatment/exposure variable is modelled against confounders using logistic regression. (2) Stabilised or unstabilised weights are estimated. (3) A weighted streg or stpm2 (Royston-Parmar) survival model is fitted with treatment/exposure as the only covariate. (4) Variance is estimated using M-estimation. As the stored variance matrix is updated, post-estimation can easily be performed with the appropriately estimated variance. Useful marginal measures, such as difference in marginal restricted survival time, can thus be calculated with uncertainties. stipw will be demonstrated on a commonly used dataset in primary biliary cirrhosis. Robust, bootstrap and M-estimation standard errors will be presented and compared.