Total, Direct, and Indirect Effects in Logit Models

It has long been believed that the decomposition of the total effect of one variable on another into direct and indirect effects, while feasible in linear models, is not possible in non-linear probability models such as the logit
and probit. In this paper we present a new and simple method that resolves this issue for single equation models and extends almost all the decomposition features of linear models to binary non-linear probability models
such as the logit and probit. Drawing on the derivations in Karlson, Holm, and Breen (2011), we demonstrate that the method can also be used to decompose average partial effects, as defined by Wooldridge (2002). We present the method graphically and illustrate it using the National Educational Longitudinal Study of 1988