New nonlinear estimators of the gravity equation

• Published in: Economic modelling Vol. 95; pp. 192 - 202
• Main Authors: Mnasri, Ayman, Nechi, Salem
• Format: Journal Article
• Language: English
• Published: Elsevier B.V 01.02.2021
• Subjects: Generalized Heckman two-step; Generalized nonlinear least squares; Gravity model; Heteroscedasticity; PPML; Structural zeros; C63; C13; C01; C15; Structural zeros; Gravity model; F14; Generalized Heckman two-step; PPML; Heteroscedasticity; Generalized nonlinear least squares
• ISSN: 0264-9993; 1873-6122
• DOI: 10.1016/j.econmod.2020.12.011

The gravity model of international trade is often applied by economists to explain bilateral trade between countries. Nevertheless, some estimation practices have been subject to criticism, namely how zero trade values and the heteroskedasticity are handled. This paper proposes new nonlinear estimation techniques to address these issues. In particular, we propose standard and generalized versions of the nonlinear Heckman two-step approach that do not require the log-linearization of the gravity equation and corrects for non-random selection bias, and a generalized nonlinear least squares estimator that can be viewed as an iterative version of the normal family Quasi-Generalized Pseudo-Maximum-Likelihood estimator. Monte Carlo simulations show that our proposed estimators outperform existent linear and nonlinear estimators and are very efficient in correcting the selection bias and reducing the standard deviation of the estimates. Empirical results show that previous studies have overestimated the contribution of variables such as importer’s income, distance, remoteness, trade agreements, and openness. •Some estimation practices of the gravity model of trade are subject to criticisms.•The chief problems are how the zero trade values and heteroskedasticity are handled.•We propose new nonlinear estimation techniques to address these issues.•The new estimators correct the selection bias and reduce the standard deviation.•We show that heteroscedasticity can lead to overestimation of the model’s coefficients.