Separable unobserved heterogeneity in duration models: testing and generalisations

Speaker(s): 
Petyo Bonev (MINES ParisTech)
Date: 
Monday, April 20, 2015 - 2:00pm
Location: 
Spandauer Straße 1, Room 23

Separabilty of unobserved heterogeneity is a common way to achieve identification in hazard models. In the first part of this paper, we develop two different frameworks for testing the multiplicative unobserved heterogeneity assumption. In the first framework, we assume that we observe multiple duration variables that are caused by a shared unobserved characteristics component (shared frailty). This assumption comes originally from the peer effects literature. The test statistics is based on a ratio of partial derivatives of the joint survival function. In the second framework, we assume that we observe two different cohorts of agents that are exposed at some common calendar point in time to a comprehensive treatment. The assumption origins in the labor market literature. The test statistics exploits a jump in the hazard caused by the treatment effect. In the second part of the paper, we introduce two different models which relax the multiplicative unobserved heterogeneity assumption. In both models we allow for partial nonseparability in i) observed covariates and time and ii) observed covariates and frailty. We exploit monotonicity in the frailty term as well time variation of covariates as sources of identification. Existing nonparametric estimators for mixture models can be adapted to our framework.