Paul Johnson, Steven Durlauf and Jonathan Temple 1129of differences in development levels. The same point holds for variables such as
institutions or the extent of corruption, where the quality of institutions, and the
nature of social norms, may well be partly a function of the level of development.
The standard response in the literature has been to use instrumental variables,
with Frankel and Romer (1999), Hall and Jones (1999) and Acemoglu, Johnson
and Robinson (2001) as three particularly well-known and influential examples.
This approach is potentially informative, and has greater claims to identify causal
effects than much other work. But it is also open to a range of important criticisms
that we will discuss in detail in section 24.6.
Some of these criticisms could be avoided by estimating models in which initial
income appears on the right-hand side. In that case, the empirical model may still
be used to identify the long-run impact of institutions on the level of development.
Recall that in a conditional convergence regression, the explanatory variables are
not necessarily explaining long-run growth, but instead determine the long-run
steady-state position that countries are converging towards. Put differently, imag-
ine that the hypothesis of interest is the effect of institutional quality on long-run
income levels. If we take two countries with the same initial level of income, the
country with better institutions should grow more quickly over a given time inter-
val. This is because it must be further below its steady-state growth path; otherwise,
its better institutions would not be consistent with initial income levels that are
the same.
Hence, growth regressions are often best seen as models of the height of the bal-
anced growth path – that is, as models of long-run level effects. This point is perhaps
most easily understood by considering how the analysis in Mankiwet al.(1992)
would be adapted to include geographical characteristics or measures of institu-
tional quality. In either case, the extension is straightforward, and the growth
regression framework can be retained. These points suggest that some of the argu-
ments usually advanced in favour of levels regressions are potentially misguided,
and that the conditional convergence regression still has much to recommend it.
Based on arguments like these, Bhattacharyya (2004) is an example of a study which
revisits the evidence on institutions and development using a conditional conver-
gence specification, rather than a levels regression. One remaining issue concerns
whether the convergence specification can identify long-run effects with sufficient
precision for the approach to be informative.
24.3.4 Interpreting errors in growth regressions
The development of the relationship between cross-country growth regressions and
neoclassical growth theories in section 24.3.2 illustrates the practice, common in
the literature, of deriving a deterministic growth relationship and then appending
an error term in anad hocway to capture all aspects of the growth process omit-
ted from the model. One problem with this practice is that some types of errors
often have important implications for the asymptotic behavior of the estimator
used in the subsequent empirical analysis. Binder and Pesaran (1999) conduct an
exhaustive study of this question and conclude,inter alia, that if one generalizes
the assumption of a constant rate of technical change so that technical change