Estimation of Ordered Response Models with Sample Selection

Abstract

We introduce two new Stata commands for the estimation of an ordered response model with sample selection. The opsel command uses a standard maximum-likelihood approach to fit a parametric specification of the model where errors are assumed to follow a bivariate Gaussian distribution. The snpopsel command uses the semi-nonparametric approach of Gallant and Nychka (1987, Econometrica 55: 363-390) to fit a semiparametric specification of the model where the bivariate density function of the errors is approximated by a Hermite polynomial expansion. The snpopsel command extends the set of Stata routines for semi-nonparametric estimation of discrete response models. Compared to the other semi-nonparametric estimators, our routine is relatively faster because it is programmed in Mata. In addition, we provide new postestimation routines to compute linear predictions, predicted probabilities, and marginal effects. These improvements are also extended to the set of semi-nonparametric Stata commands originally written by Stewart (2004,

Socioeconomic inequalities in self-perceived oral health among adults in Chile

“…We use maximum likelihood estimation methods to jointly estimate the outcome and participation equations. A likelihood ratio test comparing the log likelihood of the joint model with the sum of the log likelihoods for the ordered response and participation models yields a test of the null hypothesis of no sample selection [33]. …”

The Impact of Social Capital on Wages of Rural Migrants and its Gender Difference in China

“…By relaxing assumptions on the joint distribution of e and v, such that the form of FðÁ; ÁÞ is unknown, the approximation methodology from Stewart (2004) and De Luca (2008) can be generalized to deal with the case here. This is the semiparametric estimation of De Luca and Perotti (2011). The sketch of this method is: first to approximate the unknown joint distribution function FðÁ; ÁÞ by integrating a Hermite polynomial expansion approximation of the unknown joint density of e and v, and then to maximize the pseudo log-likelihood function (6) to obtain the semiparametric estimator of the vector of parameters (b; g 0 ; g; l 1 ; l 2 ; l 3 ).…”

Uptake of agri-environmental schemes in the Less-Favoured Areas of Greece: The role of corruption and farmers’ responses to the financial crisis

“…The primary aim of the analysis was to obtain consistent estimates of a vector of parameters ˇ1 and ˇ2, by using observations from the selected sample. Given that both parts of the structured model are non-linear, the maximum likelihood estimation (MLE) is considered the most adequate choice (Luca and Perotti, 2011).…”

“…In MLE, the covariance of the error terms of the two equations is not directly estimated (Luca and Perotti, 2011), but the transformed correlation coefficient ˆ , based on Fisher's z transformation is estimated instead:…”