predcumi: A postestimation command for predicting and visualizing cumulative incidence estimates after Cox regression models

In the presence of competing risks, calculation of cumulative incidence should provide a more realistic assessment of the probabilities of the event of interest conditional on covariates than provided by either the Kaplan–Meier failure probabilities or the event probabilities predicted directly from the Cox regression model. Enzo Coviello has previously provided user-written Stata programs that calculate either crude cumulative incidence estimates over time (stcompet) or cumulative incidence estimates over time adjusted to some user-specified values of covariates (stcompadj), which are useful for making between-group comparisons but have limitations for evaluating the individual risk predictions. I will describe the motivation behind a new postestimation command (predcumi) that facilitates the calculation and visualization of cumulative incidence estimates after Cox regression models, calculated based on each individual’s covariate patterns or optionally with flexible adjustment of covariates to user-specified values or means or percentiles of the covariate distribution. The most recently fitted Cox model is assumed to be for the event of interest, and given the user’s specification of the competing event, the cumulative incidence calculations are based on cause-specific hazards estimated from Cox regressions. Examples will be provided and comparisons made with the previous user-written programs and Stata’s official implementation of competing risks models based on the Fine and Gray model formulation (stcrreg).