In this paper we adapt an alternative approach, abandoning the regression framework in favor of binary recursive tree estimation, a technique potentially well suited to identifying both threshold effects and cross-dependencies in a wide range of potential explanatory variables. Based on a sorting of countries into a fast- and a slow-growing group, the tree analysis searches across a set of potential explanatory variables to produce a sequence of criteria (in essence, a decision tree) which help determine the likelihood that a country will fall into each group. Since the sequence of criteria can depend upon previous branchings of the tree, the algorithm can readily accommodate cross-dependencies between the explanatory variables. The technique also establishes a hierarchy among the explanatory variables, based on their ability to discriminate between groups, thus providing a natural criterion for deciding which determinants belong in the ‘”core” and which are of secondary importance. Finally, because the algorithm uses interior thresholds, it is by construction extremely robust to outliers, unlike regression analysis.
We construct trees for an annual data set of per capita GDP growth rates and a wide range of potential explanatory variables, covering all member countries of the International Monetary Fund over the period 1960-1996. We find, in line with most previous studies, that physical investment is the most important variable determining growth performance. Yet low investment does not condemn countries to low growth: high human capital and low inflation can partly compensate. Nor is high investment sufficient to generate high growth: high inflation renders rapid growth unlikely even in the presence of high investment. Investment, the key discriminant, is of course hardly an exogenous variable. Applying the methodology to countries ranked by their investment ratio reveals that openness of the economy is the key discriminant, with relative income, terms of trade variability and fiscal variables also having strong predictive power. Consistent with arguments made by Barro, a striking non-linearity emerges with respect to the public sector share: both a very low, and a very high tax revenue to GDP ratio is associated with low investment.
The rest of the paper is organized as follows. Section 2 discusses binary recursive trees. Section 3 describes our data. Section 4 presents the main empirical results. Section 5 provides some brief concluding remarks.