Editor’s note: Yesterday’s post by Brookings Institution Fellow Scott Winship prompted a reply by University of Ottawa economist Miles Corak. Scott has kindly agreed to post his reply to Corak in this space.
University of Ottawa economist Miles Corak, who I know and like, has put up a long post on his blog taking issue with my critique of Council of Economic Advisors Chair Alan Krueger’s claims about mobility. Miles and I had a friendly exchange over Twitter subsequent to Krueger’s inequality speech at the Center for American Progress. After thinking about Krueger’s evidence some more, I decided to write up a criticism of his claims, which Reihan was generous enough to post and which will appear in modified form in the next issue of National Review. Miles’s post dismisses my criticisms pretty unfairly in my view, expounding on decades of theoretical and methodological developments in the academic mobility literature and implying that I am unfamiliar with them.
So to set the record straight, here is some more detail on my objection to Krueger’s “Great Gatsby Curve,” which plots ten countries’ inequality and immobility to estimate the mobility “today’s children” will experience. A first set of issues relates to how robust is Krueger’s “best-fitting line” through the ten points in the chart, which is the basis for his projection. Before descending into serious weeds, let me address the most fundamental question here: whether inequality necessarily reduces mobility. I’m actually fairly agnostic on this point, and I don’t find the existing evidence very compelling either way. The absolutists here are those who insist that inequality can’t do anything but reduce mobility.
Because, Miles’s invocation of theory aside, this is an empirical question. If inequality rises only among the top 1 percent, will that affect the mobility of the bottom and the middle? Will rising inequality between the top and the middle hurt the mobility prospects of the bottom? The specific aggregate measure of inequality used in these Great Gatsby charts, the “Gini coefficient,” doesn’t tell us who is far apart from whom, which is an important consideration for assessing the impact of inequality on mobility. My dissertation advisor, Christopher Jencks, and his colleagues have found that inequality between the bottom and the middle did not really increase between the late 1960s and the late 2000s, even though the Gini coefficient did. As I’ll come back to below, research that includes the value of health insurance in “income” finds little increase in the 1990s and 2000s (except at the very top).
Equally important, what reference group matters in assessing inequality? Surely not the whole population of people or households. Rising inequality due to increased immigration swelling the ranks of the poor will not necessarily affect the mobility of kids born in the U.S. at the start of an immigration boom. Nor need rising inequality between the elderly and everyone else, or between parents and non-parents. What would seem most relevant is the amount of inequality among the families of the specific birth cohorts for which mobility is considered, but those estimates are rarely estimated for any country. Jencks et al. show that trends in inequality differ depending on whether you look at households without kids or at households with kids–how much more would it differ if we were to look at households with kids born in a narrow range of years?
OK, into the weeds. I gave a number of reasons in my piece for questioning the robustness of the Krueger claims, but I also noted that every single version of the Great Gatsby chart previously compiled shows a positive relationship between inequality and immobility. That is, I explicitly indicated that my objections generalized to past versions of the chart produced by other researchers and are not specific to Krueger’s in particular. Implicit in this admission is acknowledgement of previous versions of the chart (Miles has put together multiple ones, the latest of which was the basis for Krueger’s).
My point was that it is difficult to do these charts well, in a way that is informative. The immobility estimates are taken from studies that are (usually) independent of one another (but not always). Estimating immobility for each country puts demands on data that are often difficult to meet. Different countries have better or worse data, and in the worst-case scenarios, researchers do not actually observe parental earnings of children and the earnings of the children in adulthood. In that case, they must resort to technical methods such as two-sample instrumental variables, whereby predicted parental earnings is used rather than actual parental earnings. The prediction comes from (yes, Miles) a model that relates the characteristics of adults in an auxiliary data set to their earnings via an equation. Then the equation is used with characteristics of the parents in the main data set to predict their earnings. Predicted parental earnings are then combined with actual earnings of adult children to produce a mobility estimate.
Miles suggests that two-sample instrumental variables is unproblematic, but it relies on some strong assumptions. If the prediction equation does not actually predict earnings that well in the auxiliary sample, it will not predict parental earnings that well in the main data set, and there are likely to be systematic biases. Even if parental earnings are well-predicted, the resulting prediction eliminates the idiosyncratic factors that are precisely the reasons that parent and child earnings are not inextricably linked. Indeed, as Miles admits, the method tends to produce estimates of immobility that are too high. Some researchers compiling Great Gatsby charts (but not all) adjust them downward, but there’s not much basis for knowing how much to adjust them. All of this ignores the thorny details of whether the auxiliary and main data sets represent comparable populations, or whether the prediction equation from the former can be validly used for the latter if they do not. You’ll forgive me, I hope, for not getting into this in my original piece.
Many of the estimates of immobility that are used in Great Gatsby charts have a lot of imprecision around them, which makes the best-fitting line through the points imprecise. Don’t take my word for it, look at the vertical bars in this Great Gatsby chart, which I noted is the one I trust the most (compiled by Anders Bjorklund and Markus Jantti). Furthermore, different versions of these charts take mobility estimates from different studies. Compare how much mobility the United Kingdom has in this version by Jo Blanden (Figure 8) versus the Bjorklund/Jantti one. Which one is right? Who knows. I trust the Bjorklund/Jantti chart primarily because, from personal knowledge, I admire the care with which Jantti approaches his research and because they had the advantage of being able to consider the early attempts to compare the immobility of nations, such as Corak’s, and assess which estimates were truly comparable.
In the Bjorklund/Jantti chart, the relationship between inequality and immobility is driven by three countries; that is, a straight line does not describe the relationship well at all and would badly predict immobility for many of the countries in the chart. This brings up another point, which is that the “best-fitting line” through the points in any of these charts may not predict individual points all that well, which means estimated projections may be off as well. The Krueger chart does not predict Sweden’s or the United Kingdom’s mobility levels very well, for instance. Asking a single variable to offer a good prediction of another is generally asking too much, which is why the vast majority of social science models include a number of predictor variables.
I also raised a bigger issue about the timing of when inequality is measured in these Great Gatsby charts. Typically, because there is less data available from earlier decades, estimates of inequality from the 1980s or later are used. But most estimates of mobility are for children born decades ago, with the 1960s being perhaps the most common period. Inequality can affect mobility in one of two ways: by reducing child opportunities to reach the best positions in life, or by increasing the rewards to getting one of those best positions. Krueger’s speech, like most discussions of the issue, focuses on the former effect. But that means he should be using inequality measures from the 1960s (yes, Miles, all the better if they are multi-year averages). Instead, he uses inequality measures from 1985, when adults born in the 1960s are roughly 20 years old and many are no longer living at home. Miles points out that parents may invest in children with an eye toward future inequality, but parents cannot predict inequality levels twenty years out with much precision. Even if they correctly surmise that the returns to investing in their kids will rise, they don’t know the extent to which other parents will also act on that knowledge. More weeds—I apologize that I need to go through all this.
At any rate, inequality has risen over time in most countries, so if one simply picks a year, such as 1985, in which to measure inequality, the slope of the best-fitting line through the points will differ depending on the year chosen. Again, don’t take my word for it, check out Figures 8(a) and 8(b) from Blanden, where the mobility estimates are exactly the same in both charts but the year in which inequality is measured changes.
More complications arise from inequality measurement. Like immobility estimates, these measures are imprecise–the OECD, World Bank, and Luxembourg Income Study give three different U.S. estimates of the same inequality measure for the same income concept for 2000. Different income concepts—such as after-tax versus pre-tax income, or earnings versus family income—yield different inequality estimates, and it is not obvious which concepts are the most theoretically important. In general, the income concept never includes employee benefits such as health insurance. These issues make a difference. It is unclear whether the Gini coefficient really increased in the U.S. between 1985 and the present if employee benefits are taken into account. A seminal paper shows that wage inequality between the bottom and everyone else stopped growing after the mid-1980s, and an increasing number of papers suggest that inequality did not rise at all in the 1990s or 2000s once health insurance is taken into account (except at the very very top of the income distribution—the top 1 percent).
You can decide whether you think I have a case here and whether I am familiar enough with the existing research to comment. You should also remember the main criticism I had with Krueger’s speech, which was his obfuscation of the fact that the middle class has “shrunk” only because more people are richer-than-middle-class than in the past. And you should remember that the Great Gatsby Curve was not the first attempt by the Administration to show a relationship between inequality and mobility—it was the second attempt after I debunked the first one. Miles is silent on these points, but they speak volumes about how much you should trust his colleague’s mobility claims.