Flexible Models for Causal Inference in Medicine and Economics

Resumo

The aim of the present work is the study of empirical aspects of a flexible regression procedure designed to perform causal inference, known as the Nonparametric Triangular Simultaneous Equations Model. This procedure helps to mitigate a problem that arise when the model regressors do not fulfill the exogeneity assumption. The main contributions emerge from two empirical applications, in Medicine and Economics, and a new bayesian estimator which is evaluated by Monte Carlo simulation. The first application involves an implementation of the triangular simultaneous equations model to assess the effects of a treatment, defined as time delay to catheterization, on the outcome, defined in terms of survival and cardiac health, for patients with non ST-segment elevation Myocardial Infraction. The main methodological contribution consists on modeling the treatment as a continuous variable, instead of using a dichotomous variable indicating early versus late intervention, and using a flexible Generalized Additive Model for estimation and inference. The second application pursue an estimation of the class size’s effect on schooling achievement (measured by Literature’s test-scores), for students from sixth grades of the primary school in Uruguay. Main innovations consist on both, implementation of a flexible additive model that enables us to take into account nonlinear effects of control variables, and perform an adequate trimming of outlier observations, which are usually ignored in similar applications. The bias caused by these outliers is illustrated by a Monte Carlo simulation exercise. Finally, the simulation study addressees the problem of weak identification in the nonparametric instrumental variable framework. In particular, it assess the performance of two alternative non-parametric estimators of the Triangular Simultaneous Equations Model when weak instruments are present. Two estimators are compared, the Two Stage Generalized Additive Model (2SGAM) and a new Bayesian Nonparametric Instrumental Variables (BNIV) estimator. Simulation results support the advantages of BNIV over 2SGAM when instruments are weak. Specifically, when the concentration parameter ranges between 10 and 16, BNIV outperform 2SGAM in terms of variance. The mentioned efficiency advantage of BNIV does not imply an increment in bias.