Terence C. Mills and Kerry Patterson xxviiNaturally, underlying the concept of convergence is an economic model, typi-
cally a neoclassical growth model (with diminishing returns to capital and labor),
which indicates the sources of economic growth and a steady-state which the
economy will (eventually) attain. At its simplest, growth econometrics leads to
cross-country regressions of output growth rates on variables motivated from the
underlying growth model and, usually, some “control” variables that, additionally,
are thought to influence the growth rate. It is the wide range of choice for these
control variables, and the resultant multiplicity of studies, that has led to the, per-
haps pejorative, description of this activity as the “growth regression industry.”
One response has been the technique of model averaging, so that no single model
will necessarily provide the empirical wisdom. A second central convergence con-
cept isσ-convergence. As the notation suggests, this form of convergence relates
to the cross-section dispersion of a measure, usually log per capita output, across
countries. As Durlaufet al. note, whilst many studies use the log variance, other
measures, such as the Gini coefficient or those suggested in Atkinson (1970), may
be preferred. In this measure of convergence, a reduction in the dispersion mea-
sure across countries suggests that they are getting closer together. As in Chapter
22 on exchange rates, an important methodological conclusion of Durlaufet al.is
that nonlinearity (due in this case to endogenous growth models) is likely to be
an important modeling characteristic, which is not well captured in many existing
studies, whether based on cross-section or panel data.
Having considered the question of convergence in Chapter 23, in Chapter 24
Durlaufet al. turn to the details of the methods of growth econometrics. Whilst
concentrating on the methods, they first note some salient facts that inform the
structure of the chapter. Broadly, these are that: vast income disparities exist despite
the general growth in real income; distinct winners and losers have begun to
emerge; for many countries, growth rates have tended to slow, but the dispersion
of growth rates has increased. At the heart of the growth literature is the one-sector
neoclassical growth model, transformed to yield an empirical form in terms of the
growth rate of output per labor unit, such that growth is decomposed into growth
due to technical progress and the gap between initial output per worker and the
steady-state value. Typically, an error is then added to a deterministic equation
derived in this way and this forms the basis of a cross-country regression, usu-
ally augmented with “control” variables that are also thought to influence growth
rates. However, as Durlaufet al. note, there are a number of problems with this
approach; for example, the errors are implicitly assumed to be exchangeable, but
country dependence of the errors violates this assumption; the plethora of selected
control variables leads to a multiplicity of empirical models; and parameter hetero-
geneity. To assess the question of model uncertainty an extreme bounds analysis
(Leamer, 1983) can be carried out, and model averaging as in a Bayesian analysis
can be fruitful. Parameter heterogeneity is related to the Harberger (1987) criti-
cism that questions the inclusion of countries with different characteristics in a
cross-country regression. The key to criticisms of this nature is the meaning of
such regressions: is there a DGP that these regressions can be taken as empirically
parameterizing? The chapter continues by providing,inter alia, an overview of the