Sufficient Sample Sizes for Discrete-Time Survival Analysis Mixture Models

Abstract

Long-term survivors in trials with survival endpoints are subjects who will not experience the event of interest. Membership in the class of long-term survivors is unobserved and should be inferred from the data by means of a mixture model. An important question is how large the sample size should be to come to accurate conclusions with respect to the effect of treatment. This question is studied for trials with survival endpoints in discrete time by means of a simulation study. Various combinations of sample size, hazard probability, and probability of being a long-term survivor are studied. The results show that hazard probability has a large effect on the accuracy of the treatment effect estimate. A sample size of 200 might be sufficient for large hazard probabilities, whereas large sample sizes of 5,000 cases should be considered for small hazard probabilities.