Truncation Bias
33 Pages Posted: 31 May 2016
Date Written: May 28, 2016
Abstract
In the case of truncation, which is the widespread phenomenon plaguing the majority of all fields of empirical research, the observed data distribution function is truncated and related to participants' covariates only, rendering Heckman's seminal and known correction procedure not implementable. Thus, for the correction of selectivity bias propagated by truncation we introduce a new methodology that recovers the unobserved part of the data distribution function, using only its observed truncated part. The correlation patterns among the non-participants' covariates (which are all functions of the recovered non-participants' density function) are recovered as well. The rationale underlying the ability to recover the unobserved complete density function from the observed truncated density function relies on the fact that the latter is obtained by conditioning the former on the selection rule. Once this unobserved part is recovered one can estimate the selection rule equation for the hazard rate calculation as if the full sample consisting of both participants and non-participants is observable. Monte-Carlo simulations attest to the high accuracy of the estimates and above conventional √(n) consistency.
Keywords: Selectivity bias correction, Truncated Probit
Suggested Citation: Suggested Citation