Robust correlates of county-level growth in the United States

Abstract

Higgins et al. (2006), report several statistically significant partial correlates with US per capita income growth. However, Levine and Renelt (1992) demonstrate that such correlations are hardly ever robust to changing the combination of conditioning variables included. We ask, whether the same is true for the variables identified as important by Higgins et al. Using the extreme bounds analysis of Levine and Renelt, we find that the majority of the partial correlations can be accepted as robust. The variables associated with those partial correlations stand solidly as variables of interest for future studies of US growth.

Acknowledgement

We thank Jordan Rappaport for kindly sharing with us some of his data and computer codes and for providing helpful suggestions throughout the study. All errors are ours.

Notes

¹ Levine and Renelt (1992) find that, using an international sample, very few variables are robust correlates with growth. Sala-i-Martin et al. (2004) introduce an alternative Bayesian sensitivity analysis. Their analysis is motivated by the belief that Levine and Renelt's ‘test is too strong for any variable to pass: any one regression model (no matter how well or poorly fitting) carries a veto’ (p 814). In contrast, we conclude that the majority of variables identified as important by Higgins et al. (2006) are not ‘vetoed’ by the Levine and Renelt test.

² Statistically insignificant coefficient estimates are discarded from the analysis; including them would make for an unreasonably demanding test. An insignificant coefficient estimate of a different sign than extreme bounds of like signs is, rather than a contradiction of those bounds, merely a tentative acceptance of the null of zero partial correlation.

³ This type of convergence is known as β-convergence and is necessary but not sufficient for σ-convergence, i.e. for a narrowing of the income distribution over time. Young et al. (2008) find that, over the same 1970 to 1998 time period, statistically significant σ-divergence actually occurred.

⁴ Our conditioning variables include a dummy variable that takes the value of 1 if the county includes a college or university with enrollment of 10 000 or more and accounts for at least 5% of the total population. In Higgins et al. (2006) the inclusion did not render the bachelor+ coefficient estimate insignificant.

⁵ Levine (2005) provides an overview of the empirical findings, as well as the theoretical literature motivating the studies.