というNBER論文をアセモグルらが上げているungated版)。論文の原題は「Converging to Converge? A Comment」で、著者はDaron Acemoglu、Carlos Molina(いずれもMIT)。

Kremer et al. (2021) revisit how the global distribution of prosperity and growth have evolved over the last six decades and the role of various factors in shaping these distributions. Building on the convergence framework of Barro (1991) and Barro and Sala-i Martin (1992), they investigate how a country’s GDP per capita today depends on its GDP per capita in the past (unconditional convergence) and whether this relationship is different when conditioning on various determinants or “correlates” of growth (conditional convergence). While the earlier literature concluded that there was unconditional divergence and conditional convergence, Kremer et al. report a trend towards unconditional convergence (meaning that growth in rich countries is no longer faster, and in fact may be slower than, in poor countries).


Kremer et al. is a timely paper revisiting the evolution of convergence cross-country patterns over the last six decades. The authors provide evidence that the lack of convergence that applied early in the sample has now been replaced by modest convergence. They also argue this relationship is driven by convergence in various determinants of economic growth across countries and a flattening of the relationship between these determinants and growth. Although the patterns documented by the authors are intriguing, our reanalysis finds that these results are driven by the lack of country fixed effects controlling for unobserved determinants of GDP per capita across countries. We establish theoretically that failure to include for such potential determinants will create a bias in convergence coefficients towards zero and, equally importantly, the resulting estimates may not have straightforward economic interpretations (for example, they will not correspond to any type of local average of the effects at the country level). The root cause of this bias is simple: when there are permanent differences across countries and each country is close to its steady state, a model that does not include fixed effects can only fit the data by having a convergence coefficient very very close to zero. This point is of more general relevance, since it applies not just to Kremer et al.’s study, but also to the majority of the convergence literature.
Empirically, we show that estimated convergence coefficients (from models that do not include fixed effects) are indeed biased towards zero. Moreover, this bias is time-varying, even though the underlying country-level parameters appear to be constant and stable. The authors’ finding that the relationship between economic growth and its country-level determinants (such as institutions) is flattening is as notable. If true, it might suggest that improving institutions and policies may have become less important for explaining and spearheading growth. It might also have important policy implications. However, our reanalysis finds no evidence of a flattening in the relationship between institutional variables and economic growth. Focusing on democracy, we show that this variable’s impact continues to be precisely estimated and, if anything, a little larger than the beginning of the sample.

*1:cf. これ

*2:cf. これ

*3:cf. これ