Diagnostic analysis and computational strategies for estimating discrete time duration models—A Monte Carlo study

Abstract

This paper uses Monte Carlo analysis to study important and contentious issues in estimating single-spell discrete time duration models. We find simulated annealing dominates gradient methods for recovering true models. We recommend a partially flexible step function for duration dependence combined with likelihood ratio tests for determining support points of unobserved heterogeneity. We find that ignoring time-changing features of explanatory variables introduces substantial biases in model coefficient and average partial effect estimates. These biases do not diminish as sample size increases.