Why does income grow faster in some countries than others? A common empirical approach in recent growth analysis has been to adopt an 'agnostic' approach and let the data do the talking (i.e. weak priors). But a new paper by Antonio Ciccone from Barcelona's ICREA-Universitat Pompeu Fabra, and ECB economist Marek Jarocinski, question this approach. In their new ECB working paper no. 852, Determinants of economic growth: will data tell? (PDF), they show that small differences in the comparative income data can have a substantial effect on the outcomes:
As many potential explanatory variables have been suggested, these agnostic empirical approaches inevitably need to start out with a long list of variables. We show that, as a result, the growth determinants emerging from these approaches turn out to be sensitive to seemingly minor variations in international income estimates across datasets. This is because strong conclusions are drawn from small differences in the R2 of different growth regressions. Small changes in the relative fit of different models—due to Penn World Table income data revisions or methodological differences between the PWT and the World Bank income data for example—can therefore lead to substantial changes regarding growth determinants.
Their analysis clearly shows that agnostic growth regressions can be sensitive even to small data revisions. They suggest that "the available income data may be too imperfect for agnostic empirical analysis". So what to do? Stronger (and fewer) priors:
At the same time, we find that the sensitivity of growth determinants to income differences across data revisions and datasets falls considerably when priors regarding potential growth determinants become stronger. That is, the data appears good enough to differentiate among a limited number of hypotheses. Empirical models of the typical size in the literature, for example, tend to point to the same growth determinants using different versions of the PWT or the World Bank income data.
Researchers who want to continue giving equal a priori weight to all potential growth determinants in the literature, should consider shrinkage priors, explicitly incorporating priors about measurement error in the income data, or implementing Zellner’s (2002) adjustment for data quality.