As expected, the media narrative is all about the headline numbers. According to my calculations, Buttigieg got 26.5 percent, Sanders got 24.4 percent, Warren scored 17.6 percent, and Biden 16.1 percent. (I use population-weighted averages of the final alignment vote share with 86 percent of the precincts reporting.) FiveThirtyEight has dramatically updated its forecast based on the results from Iowa. The odds for Sanders are 1 in 2, and just 1 in 6 for Biden. The sharp jumps in the final forecasts are a sign of model instability. These estimates are not very credible — they are relying too hard on the Iowa bounce. The source of the instability is, as usual, overfitting. While it is good to try to model path-dependence, you should be extra careful of explosive instability in your model.
Meanwhile, the New York Times Upshot tells us that the contamination can be ignored:
There is no reason to believe that Mr. Sanders or Mr. Buttigieg did materially better in the contaminated precincts than they did elsewhere, either over all or controlling for their demographic characteristics.
Um, no. There is good reason to believe that random errors would, in probability, hurt the candidate whose support is more concentrated in a small number of densely populated districts, as is the case with Sanders relative to Buttigieg. This is so simply because the errors would vanish on average if the spatial distribution of the candidates’ performance were the same; if they were different, the one with the more balanced portfolio would be advantaged by random labeling errors. The next figure displays support for both candidates in the bottom and top quintiles by population.
Much more important than the headline numbers is what the cross-section of the Iowa counties tells us about the support base of the candidates. This is what I tried to do in the previous dispatch. Now that we have more data I thought I’d update the estimates. The main results reported previously stand. Indeed, they have strengthened. This is as it should be — more data means more power, and more power means bigger coefficients.
|Table 1. Spearman’s correlations (N=99 Iowa counties, 86% reporting).|
|College Graduation Rate||Population||Per Capita Income||Net Migration Rate||Overdose||Trump|
|Estimates in bold are significant at the 5 percent level.|
Buttigieg turns out to be even more of an empty vessel than reported earlier; even his correlation with the Trump swing falls into insignificance. Biden turns out to enjoy even stronger support in the high school-educated working class. Conversely, Warren’s support turns out to be even more concentrated in the college-educated professional class. Biden’s support is strongly correlated with the Trump swing; while that for Warren and Sanders is strongly anti-correlated. This does not mean that Biden is more viable — it depends on your working diagnosis. It is not clear whether the Dems need someone like Trump (in the sense of sharing the spatial support base) or an anti-Trump. It all depends on why you think Trump won in the first place. As argued in the previous dispatch, if you think that Trump is in the White House because large parts of the country are in serious trouble, Sanders is the man to defeat him. If you think that only someone with strong support in the white working class can oust Trump, Biden’s your man.
As you probably know at this point, Columbia has unlawfully fired the Policy Tensor as a PhD candidate in history. This means that I have to leave the country by tomorrow. But don’t worry, I’ll be back stateside soon enough. Happy to report that I have been appointed as a Visiting Scholar at the University of Arizona to work with Professor Noam Chomsky. Very excited to work with one of the greatest scholars alive on a fascinating question — the origins of the Boas-Chomsky Universal. Meanwhile, I am looking for a more permanent position. I am interested in any position where I can do serious work unmolested. Please get in touch if you have any leads. Thanks in advance.