1138 The Methods of Growth Econometrics
In other applications, Brock and Durlauf (2001a) and Masanjala and Papageor-
giou (2005) have employed model averaging to study the reason for the poor
growth performance of sub-Saharan Africa. Both papers identify some important
differences between Africa and the rest of the world in terms of the relevant growth
determinants. The study of Brocket al.(2003) takes this idea further by exploring
the use of growth regressions in the evaluation of policy recommendations. Specif-
ically, the paper assesses the question of whether a policy maker should favor a
reduction of tariffs for sub-Saharan African countries. The analysis assumes that
the policy maker has a specific set of mean-variance preferences with respect to the
effects of a change in current policies. At first glance, the analysis supports a tariff
reduction: it shows that a policy maker with these preferences should be in favor,
unless the policy maker has a strong prior that sub-Saharan African countries obey
a growth process distinct from the rest of the world. In the latter case, there is suf-
ficient uncertainty about the relationship between tariffs and growth for African
countries that a change in trade policy cannot be justified. In statistical terms, this
result is a consequence of the strong prior belief in the possibility of heterogeneity.
The prior implies that the growth experiences of non-African countries will have
little effect on the precision of estimates of marginal effects that are constructed
using data for sub-Saharan Africa.
To date, perhaps the most important critique of the Bayesian approach, at least
as applied to growth data, is that developed in Ciccone and Jarocinski (2007).
Their central point is that agnostic empirical approaches, such as model averag-
ing, appear to be sensitive to modest changes in the data. One of their examples
is based on Sala-i-Martinet al.(2004) and its application of model averaging to
1960–96 growth determinants, using Penn World Table (PWT) version 6.0 data for
output levels and growth rates. When Ciccone and Jarocinski update those results
with the revised data provided in PWT version 6.1, they find that the two versions
of the PWT lead to disagreement on 13 of 25 determinants of 1960–96 growth that
emerge using one version of the data or the other. When they carry out a further
exercise, now using the latest PWT 6.2 data for 1960–96, they again find scope for
considerable disagreement. They illustrate the potential concerns using a Monte
Carlo study, which confirm that the Bayesian approach can be sensitive to data
revisions that are modest by the historical standards of revisions to the PWT. Their
results imply that a priority for future research should be the development of meth-
ods which are less sensitive to small changes in the data, perhaps by reformulating
the priors on model coefficients or explicitly allowing for mismeasured data. In this
respect, it is worth noting that the growth literature has made relatively little use
of the methods for simultaneous model selection and outlier identification that
were developed by Hoeting, Raftery and Madigan (1996).
24.4.2 Parameter heterogeneity
The estimation of linear growth models, at best an approximation to the true
law of motion of an economy, has generated unease about the statistical foun-
dations of the exercise. It is difficult to sustain the claim that the data for very