Steven Durlauf, Paul Johnson and Jonathan Temple 1097augmented Solow model estimated by Mankiwet al.(1992).^16 These findings are
extended in Tan (2008), who employs GUIDE (generalized, unbiased interaction
detection and estimation) and finds strong evidence that measures of institutional
quality and ethnic fractionalization define sub-groups of countries which obey
common growth models.^17
Other research has produced additional evidence consistent with multiple
regimes using alternative statistical methods. Some of these study the behavior
of the entire cross-country distribution of per capita income, as we discuss below.
Here we highlight the work of three other researchers. Desdoigts (1999) uses projec-
tion pursuit methods and finds several interesting clusters, which can be described
as the OECD member countries, Africa, Southeast Asia, and Latin America.^18 He
argues that the first of these is identified by variables that proxy for the effects of
initial conditions on subsequent growth behavior. Kourtellos (2003) uses projec-
tion pursuit to construct models of the growth process and finds evidence of two
steady-states. Canova (2004) utilizes a procedure with Bayesian origins that both
estimates the number of groups and assigns countries to groups. The researcher
orders the countries by various criteria (for example, output per capita in the
pre-sample period) and the estimation procedure chooses cluster boundaries and
memberships to maximize the predictive ability of the overall model. He finds that
ordering the data by initial income divides the regions of Europe into four clus-
ters, with statistically and economically significant differences in long-run income
levels and little across-cluster mobility, consistent with the existence of multiple
steady-states. He also finds two clusters among the OECD countries with initial per
capita income again being the preferred ordering variable. There is little mobility
between the clusters, and the implied long-run difference in the average incomes
is “economically large.”
As discussed in Durlauf and Johnson (1995) and Durlaufet al.(2005), studies of
nonlinearity also suffer from identification problems with respect to questions of
convergence. One problem is that a given dataset cannot fully uncover the nature
of growth nonlinearities without strong additional assumptions. As a result, it
becomes difficult to extrapolate those relationships between predetermined vari-
ables and growth to infer steady-state behavior. Durlauf and Johnson (1995) give
an example of a data pattern that is compatible with both a single steady-state
and multiple steady-states. A second problem concerns the interpretation of the
conditioning variables in these exercises. Suppose one finds, as in Durlauf and
Johnson, that high- and low-literacy economies are associated with different aggre-
gate production functions. One interpretation of this finding is that the literacy
rate proxies for unobserved fixed factors, for example, culture, implying that these
two sets of economies will never obey a common production function, and so will
never exhibit convergence. Alternatively, the aggregate production function could
structurally depend on the literacy rate, so that as literacy increases, the aggre-
gate production functions of economies with low current literacy will converge to
those of the high literacy ones. Data analyses of the type that have appeared to
date cannot easily distinguish between these possibilities.