FIML estimation of treatment effect models with endogenous selection and multiple censored responses via a Monte Carlo EM Algorithm

We formulate a Monte Carlo EM algorithm to estimate treatment effect models involving multiple censored responses. The algorithm has at least three advantages with respect to traditional methods. First, it does not require integrating the unobserved information out from the likelihood function, which reduces the estimation time dramatically and permits to solve problems involving a high number of latent variables. Second, it reduces the estimation of the vector of slopes to the calculation of a GLS estimator, and numerical techniques are required only to estimate the elements in the disturbance covariance matrix. Third, it has low sensitivity to the selection of starting values and fragile identification. We illustrate the method by estimating a 3-equation treatment model; then we compare the performance of our algorithm against a quasi-Newton optimization that uses numerical integration.