Two-phase response-dependent sampling designs for time-to-event analysis

Abstract

Measuring expensive covariates for all subjects within a cohort may not be a feasi- ble option due to a study0s budgetary or logistical constraints. As a result of such limitations, we need to consider sampling designs that account for subjects that have missing data. To design a study allowing incomplete covariate data for some subjects, it is better to employ a cost-efficient sampling design, which balances the efficiency of parameter estimates and power of association tests with the sample size. Response- dependent sampling is a cost-efficient sampling design. In this design a subset of subjects is selected from a cohort, based on the response variable (and inexpensive covariates), which has already been gathered for all subjects in the cohort. In our study, we focus on response-dependent two-phase sampling designs. During phase I of the sampling design, all members in a cohort are measured for the response variable and the inexpensive covariates. In phase II, a subset of the cohort is selected, based on the response variable obtained in phase I, and the expensive covariate(s) are measured only for those selected. In our study, the response variable that determines which in- dividuals are selected for phase II is a continuous time-to-event variable; wherein this type of the response variable maybe subject to censoring. The most common response-dependent sampling design for time-to-event data is the case-cohort sam- pling design. We explore variations of the case-cohort design which give more efficient association estimates for a given sample size. We stratify cases and non-cases based on the observed time-to-event values and apply basic stratified sampling. Modify- ing the proportion of observations selected from each strata changes the efficiency of association estimates.