Linear Probability Models: A Treatment Worse than the Disease?

Average Marginal Effects, Linear Probability Models and the Incomparability of Coefficients in Logistic Regression

Abstract:

While being a standard instrument in the toolbox of statisticians in the life sciences for a long time, the adoption of logistic regression for social science research was long delayed by the difficulties in the interpretation of its parameters, which led some scholars to not feel "comfortable" with using it (e.g. Franklin, Mackie, and Valen 1992: 438). More recently, the use of logistic regression was questioned as a means of describing effects of independent variables on binary predictors (Mood 2010) on grounds that coefficients of nested models are incomparable even when estimated with the same sample. Such comparisons are necessary to assess omitted variable biases and to distinguish between direct and indirect effects of independent variables. This has led many scholars to turn to linear probability models instead of logistic regression. Using analytical techniques and Monte Carlo simulations, the paper examines whether the comparability problem can be solved or avoided by the use of linear probability models. It further discusses the KHB-technique (Karlson, Holm, and Breen 1992), which directly addresses the comparability problem, but has not yet gained the same attention in political science that it has in sociology.