Brookings Trade Forum 2004 (2004) 24-35
[Access article in PDF]
Comments and Discussion
[Return to Article]
Over the years I have enjoyed reading Martin Ravallion's papers, and this one is no exception. It highlights some fundamental issues inherent to the concept of income inequality in a transparent way and provides a convincing explanation of the conflict between the two sides of the globalization debate. Since I agree pretty much with the content of the author's paper, I propose to raise a few issues extending his analysis. I have four comments: the first two are relatively minor points whereas the last two are substantive and go to the heart of measuring income inequality.
First, Ravallion correctly points out that "churning," that is, two individuals swapping places in the income distribution so that one person gains and another loses, would not be seen to have any impact on the inequality measure whatsoever. Yet he remarks that "one should not be surprised if the losers in the process are unhappy about the outcome and that this fuels criticisms of the policies that led to it." But the issue is deeper: any churning, assuming a convex utility function (reflecting declining marginal utility of income), will lead to a net reduction of utility. The loss of utility to the loser will be greater than the gain in utility to the gainer. Hence if enough churning takes place, it could be potentially destabilizing from a societal standpoint.
A second minor comment is that class conflicts could result from vertical inequality. For example, a structural adjustment and trade liberalization program could lead to higher food prices in a developing country, benefiting farmers who are net sellers of food, while agricultural workers (the landless) would be negatively affected by the reform.
Now I come to my more substantive comments. Ravallion is concerned about the robustness of population-weighted inequality series and the need for statistical [End Page 24] caution in inferring that inequality is falling when weighting people equally. I believe that this is a crucial issue that needs to be extended not just to the underlying data but also to the implicit and explicit assumptions and methodologies used in deriving world inequality measures. The fundamental question that needs to be asked is, how sensitive is the Gini coefficient (or any other inequality measure) to measurement errors and assumptions used in deriving it? Instead of reporting one unique scalar value for the Gini coefficient, could one derive a range of values depending on Bayesian and non-Bayesian estimates of the likely effects of measurement errors and underlying methodologies used to derive the worldwide income distribution? One advantage of this procedure would be that it would force the analyst to identify and confront the key assumptions and measurement errors to which the Gini coefficient is sensitive.
Let me illustrate with the help of some examples. In his derivation of the change in inequality in the worldwide income distribution, Sala-i-Martin made a number of assumptions.1 For example, he left out the former Soviet Republics, Bulgaria, and Yugoslavia—countries that all underwent large increases in inequality in the 1990s. He also derived within-country distributions from quintile distributions, assuming zero variance within quintiles. This use of sparse and fragmentary data led Milanovic, paraphrasing Winston Churchill, to claim that "never was so much calculated with so little."2
The question I am raising is how far can one go in estimating the likely effects of such assumptions on the real Gini? It should certainly be possible to infer how much the Gini coefficient would be underestimated by ignoring the intraquintile variance (using six points of an income distribution rather than the whole distribution).
When it comes to measurement errors, we can distinguish between sampling errors—to the extent that much of the information comes from household surveys—and nonsampling errors. The latter clearly dwarf the former. One example given in Ravallion's paper is the systematic bias in computing growth rates for China: the long-term annual per capita growth rate is likely to have been overestimated by 1 to 2 percent...