More recent work covering thresholds, path dependence and cross-dependence includes Clark, Murphy, Shleifer and Vishny, David, Arthur, Azariadis and Drazen, Jones and Manuelli, Kre-mer, Aghion and Howitt, Matsuyama, Easterly and Ciccone and Matsuyama, among others. Ex ante, the existence of such threshold effects is not implausible. For instance, it is hard to believe that raising the primary school enrollment rate from 90 to 95 percent has much effect on GDP growth. Yet arguably there is some threshold level of human capital below which GDP growth begins to suffer. Similarly, the effects of inflation on growth may be highly nonlinear, with a positive effect at very low inflation rates (as inflation “greases” the economy), and a negative effect at higher inflation rates (as inflation confuses relative price signals in the economy) [Bruno and Easterly ].
More complex non-linearities enter to the extent that the effect of one growth determinant depends on the level, or presence, of another determinant. Thus, trivially, accumulation of human capital is unlikely to do much for a country ravaged by civil war. Similar, if less extreme, complementarities are likely to arise for many growth determinants. There is no reason to suppose, for instance, that the effect of inflation (or R&D expenditure) on GDP growth is independent of the level of physical or human capital in the country, the development of its financial system and the quality of property rights.
In the presence of such non-linearities, no universal growth recipe exists, rather, the elasticity of growth with respect to a particular factor will differ across countries dependent on their other characteristics. These differences are difficult to capture in the standard regression framework. By definition, the coefficients in a multi-variate regression analysis capture the marginal effect of a change in the explanatory variable, holding constant the other variables, impeding identification of threshold effects or complementarities. In principle, the problem can be overcome by including sufficiently many dummy variables and interactive terms in the regression. But in the absence of clear-cut theoretical predictions about the nature of the interaction, let alone the level of any thresholds, adding such terms soon becomes impractical given that the typical growth regression contains anywhere form eight to fifteen right hand side variables.