Storks Bring Babies

Jonah Goldberg’s post last night stimulated me to take a closer look at Larry Bartels’s widely-discussed regression analysis that purports to prove that income inequality has risen more under Republican presidents than Democratic presidents over the period 1948 – 2005 because of consistent partisan differences in policy.

I spent some time going through the data. I’ll start with the headline of my conclusions. Bartels’s thesis is primarily the statistical artifact of the combination of two very simple observations: (A) there has been a higher proportion of Republican presidential years during the period 1980 – 2005 than the period 1948 – 1979 (64% vs. 48%), and (B) starting in 1979, US income inequality began to rise dramatically after a post-WW II period of relative wage equality. Accepting these two statements of fact does not imply accepting Bartels’s thesis that A caused B, and his academic paper on the subject provides no compelling evidence of causality.

I started by going back to the raw data from the Census (Historical Income Tables; Table F1, 2005 CPI-U-RS adjusted dollars). I assigned presidential party by year and cross-checked that I had a consistent dataset by tying to Bartels’s analysis that income growth at the 95th percentile averaged 2.1% per year from 1948 – 2001 (from his 2004 academic paper on the topic). I then created a very simple metric for the purpose of exploratory analysis: income growth rate at the 95th percentile of income for a given year – income growth rate at the 20th percentile of income for that year. For example, if incomes grew 2.9% for families at the 95th percentile of income in 1976 and they grew 1.6% for families at the 20th percentile of income that year, then this metric for 1976 = 2.9% – 1.6% = 1.3%. The higher this number, the more “the rich are getting richer than everyone else”.

Bartels’s factual claim is, in essence, that this metric (and various more complicated measures of inequality) grew more under Republican presidents than Democratic presidents. In fact, he “lags” the effect by one year, so that for example Jimmy Carter gets credit for changes in income inequality in 1981. When I recreate his analysis on this simplified data for 1948 – 2005, I get exactly the kind of effect that he sees: the simple inequality index averaged 0.8% under Republican presidents (lagged one year) and -0.2% under Democratic presidents (lagged one year). The difference was therefore 0.8% – (-0.2%) = 1.0%.

Now a lag of a year isn’t crazy, but I decided to test this for robustness. It’s amazing, but if you try the effect lagged by two years, the Republican vs. Democratic difference goes from 1.0% to 0.1%, or statistically indistinguishable from zero difference. Almost the same thing is true if you assume no lag (the difference drops from 1.0% to 0.3%). While there are lots of economic policy effects that are most evident after about a year (Bartels cites a couple of papers that describe these), this ought to be a clue that there is a problem with the analysis. Unless there is some kind of magic voodoo effect presidents can have on inequality which is not evident within a year, becomes evident after a year and then can no longer be seen after two years, this is an indication of a possible spurious correlation. Once you know what to look for, you can find Bartels’s admission of the issue down in Footnote 7 of his paper. It seems to me that this ought to be a central analytical issue that he needs to resolve. In this footnote, Bartels says of the 0-year and 2-year lagged regressions that “the resulting regression models fit the data less well…” Ah. More on this later.

Bartels claims that the relationship between (lagged) presidential party and inequality change is so consistent that it would be amazing coincidence if it were not causal. Really? Suppose if instead of presidential party for every year you substitute a variable called “Was the US at war?” (using Korea, Viet Nam 65 – 73 and the current Iraq war starting in 2003 as the war years). If you do this, you then get an effect about 70% as large as the one Bartels found (which by the way is much more robust to changing the lag time). I can get effects 60% – 70% as large as those Bartels found using variables like “Was the president from the Midwest?”, “Was the president from the South?” and “Did the president’s last name start with the letters M-Z?”. Did these factors cause changes in inequality?

The key underlying driver of inequality in this data is that inequality started to rise a lot after 1979. It dominates the relationship between any variable and inequality. Anything that is correlated with time will show the same pattern that Bartels finds. Such as, for example, a political realignment that led to a higher proportion of Republican presidents after 1979.

Bartels is of course aware of this, and claims that presidential party is correlated with changes in inequality within periods of lower and higher structural inequality. But think about this; there were 11 presidents between 1948 – 2005. The guy has a total of 11 independent data points. As an example of what this means for the analysis, consider the entire period until 1960. If you consider the actual years when they were each president, Truman’s change in this simplified inequality metric was -2.1% (he was great for the little guy) and Eisenhower’s was 0.6% (he was pretty good for rich guys). With a one year lag, these numbers go to Truman: -1.1%, Eisenhower: 1.0%. With a 2-year lag the ordering reverses to Truman 0.1%, Eisenhower -0.6%. This is because a 2-year lag moves 1954 from Eisenhower to Truman, and 1954 was a great year to be a rich family – the inequality metric went up 7.8%, and who gets credit for that one year determines who was “better” for inequality. I’m sure that Eisenhower’s economic policies had something to do with this, and I’m sure that Truman’s actions during the years he was president had some impact. But lots of stuff happened in America in 1954. The country was coming out of the Korean War, just to name one thing that was fairly important. Of course, Bartels knows that it was Esienhower’s polices that drove these results in 1954 (though not 1953 – Bartels knows that was caused by Truman). Perhaps surprisingly, the simplified inequality metric in 1955, the next year of Eisenhower’s administration, had its second-biggest biggest one-year drop, going down 8.5%. Bartels is reading tea leaves.

Now, it happens that I think increasing wealth inequality is a big political problem for the US. I’ve written a long article on this for National Review. A combination of structural factors independent of presidential policies has driven inequality growth in the US since the 1970s. First, US domestic production of oil peaked in 1971, oil imports doubled between 1970 and 1975, and OPEC was able to simultaneously drive large price increases. This oil shock was directly regressive, but also tended to disproportionately hammer those industries that were the source of high-wage union jobs. Second, in 1970 “Non-Distributive Services” (i.e., finance, professional services, health care and so on) first became a larger part of the private economy than goods producing industries. The shift to services, especially as leveraged by the increasing technology-dependence of the economy, tended to enhance the prospects of the cognitive elite at the expense of traditional industrial workers. Third, the combination of changes in cultural mores, social programs and the ongoing transition to a service economy began to disassemble the traditional family. The social capital transmitted by intact families became a more and more relevant source of competitive advantage to those individuals raised in functioning homes as this family structure became less universal. Fourth, the foreign-born percentage of the US population, which had reached its historical minimum in 1970, began to rise rapidly as massive immigration resumed after a multi-decade hiatus. This both increased inequality by introducing a large low-income group to the population, and also by intensifying wage competition at the lower-skill end of the income scale.

The objective international conditions that allowed Truman (D), Eisenhower ®, Kennedy (D) and Johnson (D) to preside over an economy with relative wage equality no longer exist. The interaction between: (1) the fact that Republicans achieved a realignment based on a number issues, some related to the economy and some not, as these trends hit the economy with resulting structural increases in inequality, and (2) the fact that the Republican approach to meeting the challenge of increased international competition over this period involved deregulation of various kinds, which has probably had a tendency to increase inequality vs. what it would otherwise have been, is a lot more complex than Bartels is willing to accept. If we tried to recreate the policies that early post-WW II America followed in a misguided attempt to recreate the economy of 1955, we would fail because we face much more severe international competition today than we did then. Trying to pretend otherwise is selling snake oil.

( cross-posted at The Corner )