Yet more on unemployment and midterm elections

My previous post on unemployment and midterms continues to live, like an undead brooding teenage vampire. Via Twitter, Naadir Jeewa runs some slightly different numbers and comes up with this graph:The odd implication is that an increase in unemployment is associated with improved fortunes for the president's party. However, the R-squared of .02 suggests that there's probably not much going on here. I'm guessing that slope isn't close to statistically significant. And again, there's not much here to suggest that high unemployment will lead to a particularly bad year for Democrats in 2010.

Meanwhile, over at Real Clear Politics, Sean Trende offers a bit of a critique on my trying to draw lessons from so few observations:

We have only had 15 midterm elections since 1950. This is barely data; it’s more of a good collection of anecdotes.

This statement struck me as a bit anti-quantitative at first, but the rest of the article makes it clear that Trende is serious about looking for appropriate hard data to address this question. He goes back a lot further in time than I do. The point that he was making above is that you can come up with a story to explain why the president's party did or did not lose a bunch of seats in any given year:

Is your President pursuing an unpopular war and controversial policies at home (1966, 2006)? Then it probably doesn’t matter that the economy is blazing ahead. Is the President kicking some al Qaeda arse a year after they attacked us, and getting ready to take out a longtime nemesis (2002)? The public is going to be more forgiving of the sluggish growth in real disposable income and rising unemployment. The end result of this is that every election becomes something of an explainable, unique event – in other words, they’re almost all outliers.

There's certainly some truth to this, although this is true of pretty much any dataset. When you're dealing with a small number of observations, everything looks like an outlier. Yet there's nothing particularly wrong with drawing inferences from only 15 observations, or even fewer. There is, for example, a broad acceptance that economic performance affects presidential elections, and those data are only drawn from post-WWII presidential elections. The relationship between economic performance and voting behavior is a bit stronger for presidential elections than it is for midterms, but it either case we shouldn't lose the forest for the trees.