The log of gravity revisited

• Published in: Applied economics Vol. 45; no. 3; pp. 311 - 327
• Main Author: Martínez-Zarzoso, Inmaculada
• Format: Journal Article
• Language: English
• Published: London Routledge 01.01.2013; Taylor and Francis Journals; Taylor & Francis Ltd; Taylor & Francis (Routledge)
• Series: Applied Economics
• Subjects: Bias; constant-elasticity models; Economic analysis; Economic models; Economic theory; Estimating techniques; Estimation; Estimation bias; Feasible Generalized Least Squares (FGLS); Gravity; heteroscedasticity; Least squares method; maximum likelihood; Maximum likelihood method; Methodology; Monte Carlo simulation; Poisson regression; Quantitative Finance; Regression analysis; Simulation; simulations; Studies; Trade models; Trade theory; Social Sciences & Humanities
• ISSN: 0003-6846; 1466-4283
• DOI: 10.1080/00036846.2011.599786

This article evaluates the performance of alternative estimation methods for gravity models with heteroscedasticity and zero trade values. Both problematic issues, recently addressed by Santos Silva and Tenreyro in an influential paper, are re-examined here. We use Monte Carlo simulations to compare the Pseudo Poisson Maximum Likelihood (PPML) estimator recommended by Santos Silva and Tenreyro, a Gamma Pseudo-Maximum-Likelihood (GPML), a Nonlinear Least Squares (NLS) estimator and a Feasible Generalized Least Squares (FGLS) estimator with more traditional techniques. Additionally, estimates of the gravity equation are obtained for three different data sets with the abovementioned methods. The results of the simulation study indicate that, although the PPML estimator is less affected by heteroscedasticity than others are, its performance is similar, in terms of bias and SEs, to the FGLS estimator performance, in particular for small samples. GPML presents however the lowest bias and SEs in the simulations without zero values. The results of the empirical estimations, using three different samples containing real data, indicate that the choice of estimator has to be made for each specific dataset. It is highly recommended to follow a model selection approach using a number of tests to select the more appropriate estimator for any application.