The Influence of Variability in Measurement Periods on the Results of Discrete-Time Survival Analysis

Abstract

Within-wave variability is common in longitudinal studies. Previous studies show this does not influence slope parameters and standard error in latent growth models. The current study aimed to extend these findings for discrete time-survival analysis in an experimental setting. In an experimental setting, people randomly assigned to either a treatment or control condition are measured to establish a treatment effect. By doing a simulation study, parameters for discrete-time survival analysis are estimated both for a fixed time points distribution as for varying time points distributions: a truncated normal, uniform and truncated chi-square distribution. Using an underlying Weibull function, data was simulated for these distributions, while changing width of the measurement interval (0, 0.25, 0.5, 1). These simulations created 270 scenarios. Results for treatment effect bias showed how only one scenario exceeded the criterion of Hoogland and Boomsma (1991) presenting more than 5% parameter bias when width of a truncated normal distribution was set to 1. In addition, the percentage standard error bias also did not exceed the criterion of 10%. Overall, it can be concluded that increasing measurement intervals and varying shape of measurement distribution do not substantially influence discrete-time survival analysis. This may lead to new standards in the practice of survival analysis.