Diagnostic Analysis and Computational Strategies for Estimating Single Spell Discrete Time Duration Models – A Monte Carlo Study

This extensive Monte Carlo study re-examines properties of the nonparametric maximum likelihood estimator of discrete duration models with unobserved heterogeneity and unknown duration dependence. Alternative specifications and computation strategies are compared. We find: i) The inherent complexity of mixture models poses the major estimation hurdle. ii) It is important to choose a flexible specification for duration dependence. Polynomial specifications with a fixed number of terms can lead to systematically biased estimates of the hazard. iii) In estimation, simulated annealing is found to dominate all other optimization algorithms. Common applied research problems, such as near-boundary false optima are eliminated. iv) A bootstrap procedure is suggested to help choose the number of support points of unobserved heterogeneity. v) Likelihood ratio tests may still be appropriate especially when there are many common parameters in the components of the model. vi) Gateaux derivatives do not appear to help optimize the likelihood function