关键词: 处置效应, 模型设定, 可信性, 子集分割

Abstract: In treatment effect studies for observational data, the robustness of model specification is a keystone for credibility. This paper constructs the Model Misspecification Standard Error (MMSE) basing on sample splitting. In certain situation, it shows that if the model setting is correct, the ratio of MMSE and SE converges to zero in probability. Monte Carlo experiments show that, the choice of splitting point does not influence MMSE’s asymptotic properties, but it will influence its finite sample behavior. Furthermore, the extent of misspecification will influence the convergence of the ratio. Applying MMSE to a case study, it finds that it could efficiently identify model misspecification.

Key words: Treatment Effect, Model Specification, Credibility, Sample Splitting