Outliers in Binary Choice Models

Résumé

This paper focuses on the problem of outliers in binary choice models. It is show that identifying outliers as observation with a residual close to one in absolute value might be misleading and outlier detection procedures should rely on influence measures on the fit. Two scalar measures are derived to evaluate the influence of each observation as well as the influence of a group of observations on i) the vector of estimated parameters and ii) the vector of estimated probabilities. Also, the method proposed by Peña and Yohai (1995) to treat the problem of masking in linear models has been generalized to the case of binary choice models. A small Monte Carlo study analyzes the performance of a11 measures and an empirical application presents a diagnostic strategy for detection of outliers.