Conditional Moment Tests for Parametric Duration Models

This paper develops and compares specification tests for parametric duration models estimated with censored data. The tests are based on generalized residuals (the integrated hazard), which is exponentially distributed if the model is correctly specified. I present several conditional moment tests based on the generalized residuals: a raw moments test, a test based on Laguerre polynomials, and a Lagrange multiplier (LM) test. The LM test extends Lancaster’s (1985) test by allowing an arbitrarily precise approximation of the likelihood under the alternative. The raw moments test implemented via an auxiliary regression is examined using both asymptotic and bootstrap critical values. Monte Carlo evidence indicates that no one test dominates the others in all situations in terms of size, power, and ease of use. When the data are not censored, the Laguerre test appears to be the best choice. When there is censoring in the data, the Laguerre test is still at least as powerful as the other tests, but the raw moment test may be more convenient to perform. For the convenience of the practitioner the explicit forms of the tests for exponential and Weibull duration models are presented. ∗This paper is based on chapter 3 of my dissertation (Prieger, 1999). I thank seminar participants at UC Davis, and especially Colin Cameron, for helpful comments. Earlier versions of the paper benefitted from comments from Alan Marco, Enrico Moretti, Roger Studley, and seminar participants at UC Berkeley. The inevitably remaining mistakes are inevitably mine. PRELIMINARY: PLEASE DO NOT CITE WITHOUT PERMISSION.