Estimating the gravity model when zero trade flows are frequent

The gravity model is widely used to model international trade flows and to test hypotheses about the impacts of policy reforms. The combination of zero trade flows and heteroscedastic errors has been shown to create potentially serious bias in parameter estimates from this model. This paper evaluates the performance of alternative estimators of the gravity equation when zero trade flows result from several economically-based data generating processes (DGPs) with heteroscedastic residuals and potentially omitted variables. The analysis uses Monte Carlo simulation to assess the ability of different estimators to recover known parameters and empirical analysis of a cross-sectional dataset to assess the properties of the underlying data-generating process and the implications of different estimators in a real-world setting. Some key findings include: The DGP used for analysis of this question should reflect the fact that the presence of zero trade flows is not independent of the explanatory variables used in the model. The combination of many zero trade flows and heteroscedastic errors has serious consequences for estimates of gravity equations. Truncated Pseudo-Maximum Likelihood estimators appear to provide useful initial estimates irrespective of the data-generating processes. Standard threshold-Tobit estimators perform well in a Tobit-based DGP if the heteroscedasticity problem is explicitly dealt with. When the data are generated by a Heckman sample selection model the Zero-inflated Poisson (ZIP) model appears to be best, while the Heckman ML estimator in logs also performs well. When the data are generated by a Helpman, Melitz and Rubinstein (HMR)-type model with heterogeneous firms, both the ZIP estimator and the HMR estimator appear to provide reasonable results. Non-nested testing on one dataset supports the Heckman sample selection model, and hence ZIP-type estimators.