An evaluation of variable selection methods in competing risks with one rare event: a simulation study.

backgroundIn the competing risk settings, the association of predictor variables with different event types is a crucial task. Accurate variable selection resolves the issue of confounding in research on causes and enables unbiased estimation of probabilities in studies on prognosis, but it becomes complicated when the disease or event of interest is rare.

objectivesIn this paper, we aimed to evaluate the performance of penalization methods in rare conditions.

methodsA variety of scenarios were used to compare the Fine and Gray, LASSO, Adaptive LASSO, SCAD, MCP, and stepwise Fine & Gray methods in competing risk settings with one rare event. The performance of the variable selection methods was assessed using four measures: the mean number of zero coefficients correctly identified as zero, the mean number of incorrect non-zero coefficients mistakenly identified as true, the mean square error and bias of cumulative incidence function.

resultsAccording to the results, SCAD and MCP were the best methods for variable selection and Adaptive lasso failed to identify the correct covariates or eliminate the unimportant ones. All four penalized methods performed equally in terms of prediction accuracy, while stepwise methods were worse in precision. In Fine and Gray method, the problem of non-convergence was seen, especially when the sample size was small.

conclusionIn conclusion, traditional stepwise selection methods are not powerful enough to handle variable selection in rare conditions. Among the penalization methods, SCAD and MCP followed by LASSO were the most capable of selecting the best covariates with the greatest influence on the CIF, while Adaptive LASSO was the least effective.