Estimating Linear Probability Functions: A Comparison of Approaches

Abstract

A linear probability function permits the estimation of the probability of the occurrence or non-occurrence of a discrete event. Nerlove and Press (p. 3–9) outline several statistical problems that arise if such a function is estimated via OLS. In particular, heteroskedasticity inherent in such a regression model leads to inefficient estimates of parameters (Amemiya 1973, Horn and Horn). Moreover, without restrictions on the conventional OLS model, probability estimates lying outside the unit (0–1) interval are possible (Nerlove and Press). Goldberger and Kmenta suggest two approaches for alleviating the heteroskedasticity problems inherent in the OLS regression model. Logit analysis will also alleviate heteroskedasticity problems and ensure that estimated probabilities will lie within the unit interval (Amemiya 1974, Hauck and Donner, Hill and Kau, Horn and Horn, Horn, Horn, and Duncan, Theil 1970).