The Gravity of Log

It is common practice to estimate log-linearized equations using Ordinary Least Squares (OLS). To address zeros and estimation bias due to heteroscedasticity, Santos Silva and Tenreyro (2006) proposed a Pseudo-Poisson Maximum Likelihood (PPML) estimator which has become a popular alternative. In this paper, we investigate under which conditions log OLS is superior to PPML estimation. We also demonstrate the implications of our findings using a number of examples from different fields of economic research.