Stata Tip 128: Marginal Effects in Log-transformed Models: A Trade Application

Since the introduction of the margins command in Stata 11, the empirical literature has increasingly used marginal effects, predictive margins, and adjusted predictions in postestimation analysis. Marginal effects are particularly useful for the interpretation of parameter estimates after logit, probit, poisson, and other nonlinear regression models. If the covariate of interest is in logs, however, obtaining meaningful results from margins, dydx() is not straightforward. In this article, I first illustrate these difficulties in the context of estimation with poisson. I then suggest that a researcher should always compute the derivative of interest and code it manually with margins‘s expression() option. Lastly, I illustrate these problems using the gravity equation from the trade literature.