gms | German Medical Science

Introduction and issue: When working with time-to-event-data there often are competing risk, i.e. several states that are considered as absorbing for statistical analysis, such as death or transplantation in hepatocellular carcinoma patients. In a setting like this the cumulative incidence, i.e. the proportion of an event type observed up to a certain time, is an easily interpretable quantity and patients can be compared with respect to their cumulative incidence for different competing risks. If there is right censoring, one way to analyze effects on the cumulative incidence is by using a pseudo-value approach, which allows, among others, the application of regression models to estimate the effect of covariates on the cumulative incidence in the presence of competing risk at different time points. We will briefly explain the underlying theory of pseudo-values and adapt this approach for a stepwise boosting algorithm.

Material and methods: Specifically, pseudo-observations for the cumulative incidence at a grid of time points are estimated for a time-to-event setting with competing risk and right censoring. These boosting algorithm provides a stagewise regression modeling approach for estimating the effect of covariates in the course of time, by coupling variable selection across time points but allowing for separate estimates. To illustrate these methods, we apply the algorithm to clinical cancer registry data from hepatocellular carcinoma patients.

Results: The use of pseudo-values in regression models for the cumulative incidence in datasets with right censoring and competing risks is seen to enable the estimation of model parameters that have a straightforward interpretation. Additionally we can fit regression parameters at different time points, and as a result of this time-dependent effects on the cumulative incidence can be judged, as illustrated for the application.

Discussion: The proposed approach for a stagewise regression technique based on pseudo-values is seen to more generally provide variable selection and useful estimates for investigating time-dependent effects on the cumulative incidence.