After posting my reaction to the new critique of my Mariel paper yesterday, a few friends and many other people contacted me to ask if I had done any additional statistical work to back up my claim that the key message in the Clemens-Hunt paper was, as I put in the title of my blog post, “fake news.” I obviously had, but chose to write a post that summarized my take in a way that would be easiest to explain to a broad audience. Yesterday I presented some simple graphs showing that wages for low-skill workers fell after Mariel even when blacks are excluded from the sample. The point of the exercise is easy to grasp, but the samples, as I emphasized, are very small.
So several people asked me exactly the same question: What would happen to the regression results in the key Table 5 of my paper if one were to redo the entire analysis using the age- and race-adjusted wage of workers? This approach has the huge advantage that we do not need to cut down on sample size. I can still use my original sample (which was small to begin with) and simply include a variable in the wage regressions indicating whether a particular worker was black or white. We can then use the regression results to calculate the average age- and race-adjusted wage in a particular city in a particular year and see if there is any Mariel effect in those trends. Here is what those regression coefficients look like. (Here are the programs for those who want to replicate; all I’ve done is add a “black” indicator variable to the individual-level regressions):
Image may be NSFW.
Clik here to view.
In rough terms, the regression coefficients give the percent wage difference between Miami and the placebo cities at some time after 1980 relative to what the difference was prior to 1980. So for example, the -0.117 statistic in the last column of the first row tells us that the low-skill wage in Miami relative to all other cities fell by 11.7 percent between 1977-1979 and 1981-1983.
It’s pretty obvious that the regression coefficients–which account for the changing black share of the workforce in Miami and elsewhere–still show a significant wage drop in Miami relative to any placebo one cares to pick. And this is true both in the March CPS data as well as in the ORG. So the key inference from the key regression table in my paper is unchanged. Something happened to Miami’s low-skill wage after 1980.
Several people also asked me how I could be so sure that there is no relation between the very strange increase in the fraction of black workers in Miami and the wage drop exhibited by Miami’s low-skill workforce between 1980 and 1985. Because black workers tend to have lower wages (even after adjusting for education), a higher fraction of blacks in the sample would mechanically reduce the average wage of the population. So it is certainly possible that the wage drop could be attributed to the change in sample composition.
It is trivially easy to show that this cannot possibly be the explanation by simply looking at the year-by-year data in either the March CPS or ORG. Let’s look at the March CPS first. The figure below shows the trend in the age-adjusted wage used in my original paper–which includes blacks–and plots it alongside the trend in the black share of the workforce.
Image may be NSFW.
Clik here to view.
To emphasize my point, I’ve shaded in the period 1979 through 1983. It is obvious that nothing whatsoever happened to the black share of the workforce (as measured by the March CPS) in this particular period. But it is also obvious that this is the period where the average wage of Miami’s low-skill workers fell most. In short, it is impossible to explain that steep wage drop in terms of a rising black share. And this leads to an obvious inference: the Clemens-Hunt argument is not consistent with the timing of the increase in the black share and the drop in the average low-skill wage.
(A geeky point about the March CPS graph. The March CPS data in a particular year gives earnings in the previous calendar year. So, for example, the 1980 wage data comes from the 1981 survey. To make sure everything is consistently timed, I’ve lined up the graph so that the 1980 data for both earnings and percent black come from the 1981 survey).
The ORG data in this next graph is equally striking. Again, the average low-skill wage in Miami (including blacks) fell dramatically between 1980 and 1984 while the black share rose slightly and then declined slightly over the period–ending up pretty much at the same place it started. So how could the change in the black share possibly account for the drop in the average low-skill wage? It can’t.
Image may be NSFW.
Clik here to view.
Finally, several people wanted me to opine on where things stand and where we go from here. Well, let me give credit where credit is due. Clemens and Hunt discovered a really weird thing about the racial composition of Miami’s low-skill workforce as measured by the March CPS, with a somewhat similar trend in the ORG data. This is something that future work must take into account. I think we would all agree that the ideal exercise is to track the average person over time to see what happens as a result of the Mariel shock–and we definitely don’t want that “average” to change as a result of changes in sample composition.
It would not surprise me if the weird pattern in the black share of the low-skill workforce as measured by the March surveys is the result of a data glitch or imputation problem that lies undetected in the vaults of the BLS or IPUMS offices. But I also suspect that the less weird ORG pattern of a gradually increasing black share (although with ups and downs through 1987) is not something we should altogether dismiss. This trend may contain valuable information. Could it be that, for reasons maybe related to Mariel or maybe not, the Miami of the 1980s increasingly became a place that did not reward whatever it is that low-skill whites bring into the workplace? And that is something worth investigating.
Image may be NSFW.
Clik here to view.
