The number of primary events per variable affects estimation of the subdistribution hazard competing risks model

Objective — To examine the effect of the number of events per variable (EPV) on the accuracy of estimated regression coefficients, standard errors, empirical coverage rates of estimated confidence intervals and empirical estimates of statistical power when using the Fine-Gray subdistribution hazard regression model to assess the effect of covariates on the incidence of events that occur over time in the presence of competing risks.

Study Design and Setting — Monte Carlo simulations were used. We considered two different definitions of the number of EPV. One included events of any type that occurred (both primary events and competing events), while the other included only the number of primary events that occurred.

Results — The definition of EPV that included only the number of primary events was preferable to the alternative definition, as the number of competing events had minimal impact on estimation. In general, 40 to 50 EPV were necessary in order to ensure accurate estimation of regression coefficients and associated quantities. However, if all of the covariates are continuous or are binary with moderate prevalence, then 10 EPV are sufficient to ensure accurate estimation.

Conclusion — Analysts must base the number of EPV on the number of primary events that occurred.