Regression Dissection

I am normally a fan of Kevin A. Hassett’s NR contributions, but I found his most recent column (“The Metamorphosis,” December 21) unconvincing. The graph accompanying the piece appears to show that in states where voters supported Obama, businesses were more likely to exaggerate the number of jobs that had been created by the stimulus. A closer look reveals several problems.

Most important, there is an outlier in the upper-right corner of the graph — a state that hosted more fake jobs per capita than any other, and also voted for Obama at an astounding 90-plus percent. (Presumably, this is the District of Columbia.) Given that the other dots show no obvious pattern, it is reasonable to suspect that the correlation wouldn’t exist were this single data point removed.

Also, the source of the fake-job data — newspaper reports compiled by the Washington Examiner — is not comprehensive and may not be representative. There are undoubtedly many fake jobs that newspapers did not uncover (some major papers probably didn’t even look), and the Examiner may have missed some reports. As I count only 41 dots, there are about ten states for which the Examiner found not a single fake job; Hassett simply leaves these states out of his analysis, introducing another potential problem.

Further, the data are not weighted by population. States with few inhabitants can throw off the correlation as a whole, and states with many inhabitants don’t influence the correlation as much as they should.

Finally, the fake-job data are presented by percentile rank. That is, the state with the highest number is ranked 100, the lowest zero, with the rest spaced evenly between. This can both exaggerate variation (even if there is very little difference between the highest and lowest number, they appear as zero and 100 on the graph) and obscure it (even if there is a huge gap between the 50th and 52nd percentile, they appear very close on the graph).

Stewart Jefferson Evanston, Ill.

Kevin A. Hassett replies: Our charts must meet a very high standard before we will elevate them from the minor leagues to the big leagues (National Review, of course), but we seldom have room to describe every test that we have run. We investigated each of the questions raised in this letter before publication.

The outlier question is a genuine concern, but when we tested the relationship without D.C., it held. We used only 41 states (including D.C.) because those are the only states for which we have data. It may be that the relationship would be weaker if we added data on the other states. It seems unlikely to us, but this is clearly an area worthy of future research.

It is true that we did not population-weight the regression, but the fake-jobs numbers are calculated on a per capita basis. If you do weight by population, the regression is still significant, with a slightly steeper slope. We normalized the data into percentiles because the actual numbers are very small decimals — percentiles make the chart easier to read.

Readers can always contact me at [email protected] and ask for more detail on any chart. We will always provide data, programs, and outputs to skeptics and supporters alike.