Robustness of Econometric Variable Selection Methods

Variable selection in cross-country growth regression models is currently a major open research topic and has inspired theoretical and empirical literature, see [6]. There are two categories of research problems that are intimately connected. The first problem is model uncertainty and the second is data heterogeneity. Recent literature aims to overcome the first problem by applying Bayesian Model Averaging (BMA) approaches in finding important, robust and significant variables to explain economic growth. While BMA offers an appealing approach to handle model uncertainty very little research has been undertaken to consider the problem of data heterogeneity. In this paper we analyze the issue of data heterogeneity on the basis of the exclusion of countries, i.e. we will take a closer look at the robustness of approaches when countries are eliminated from the data set. We will show that results of BMA are very sensitive to small variations in data. As an alternative to BMA in the cross-country growth regression debate we suggest the use of “classical” Bayesian Model Selection (BMS). We will argue that there is much in favor of BMS and will show that BMS is less sensitive in the identification of important, robust and significant variables when small variations in data are made. Our empirical results are undertaken on the most frequently used data set in the cross-country growth debate provided by [4].