Abstract
Empirical duration analysis calls for numerical methods to search for a maximum of a non-linear likelihood function. The Newton method is probably a common choice. Yet, it might also be an unwise choice, as we demonstrate fitting a large empirical data set, for which Newton and a quasi-Newton method (Bfgs) perform poorly compared with a Gauss method (Bhhh). We proceed by an extensive Monte Carlo simulation to compare the methods for a variety of proportional hazards duration models with respect to computing time and number of iterations upon convergence. We also study the effects of starting values and step-length methods. We conclude that the Bhhh method is outstanding in terms of speed, particularly for semi-parametric hazards models. We also report on sophisticated starting methods that significantly reduce the number of iterations.
Published Version
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Schedule a call