Model specification tests and the estimation of treatment effects: An application with random and non-random administrative records

In addition to existing model specification tests based on over identifying information from pre-program periods, this paper conducts a new nonparametric test based on the concept of unconditional bounding. This test has the advantage that it does not require pre-program information on participants. The nonparametric bounding test yields similar information about model misspecification to the standard test based on over identifying information. As an illustration, we apply the test to administrative records from the Jobs First experiment in Connecticut. Records on other welfare recipients are used to provide quasi-experimental panel regression estimates of the impact of time limiting benefits on labor market earnings.