A Monte Carlo study on non-parametric estimation of duration models with unobserved heterogeneity

We conduct extensive Monte Carlo experiments on non-parametric estimations of duration models with unknown duration dependence and unknown mixing distribution for unobserved heterogeneity. We propose a full non-parametric maximum likelihood approach, based on time-varying lagged explanatory covariates from observational data. By utilising this data-based identification source, we find that both duration dependence and unobserved heterogeneity can be reliably estimated. Our Monte Carlo evidences show that variation in time-varying lagged explanatory variables contributes to the identification of both duration dependence and unobserved heterogeneity, especially when sample sizes are limited. For limited sample sizes, maximum penalised likelihood with information criteria seems to produce more accurate estimators than pure maximum likelihood. Our approach can be easily extended to multivariate competing risks model with dependent unobserved heterogeneities.