Last week when I was hatin' on Ray Fair's election equation, skeptics told me to just wait and see how well regression equations would predict the Olympic medal counts.
I don't know how well Fair's Olympic model did this year. However, two other high profile models delivered a mediocre performance. Both models correctly predicted that developing countries would win a larger share of medals than at previous Olympic games. However, neither model came close to pegging the exact ratios for Athens 2004. See Daniel Gross' latest Moneybox column in Slate for details: Medal Miscount.
Last week some readers accused me of dismissing regression equations out of hand. In fact, I don't. The past is often a good predictor of the future. Fair's models describe interesting relationships. The election equation confirms the intuitive hypothesis that voters prefer incumbents during good economic times. The model is a concise way of describing historical patterns, it's just not an appropriate tool for making precise forecasts about the upcoming election.
If the future is often like the past, why should we doubt that this election will be a lot like previous elections? Well, as a great man once said, things are more like they are now than they've ever been. We know that this election is atypical in many respects. The model will only work if voters assign a historically average weight to economic conditions. This year voters are preoccupied with unusual non-economic issues like the war in Iraq.
Fair's model assumes that voters who feel good about the incumbent's economic record are likely to reelect him. The model only works if voters' sense economic well being is closely related the subset of economic variables specified by the equation: present economic growth and inflation, and history of growth during the incumbent's last term.
These variables may not correlate as closely with the average person's sense of economic well being as they once did. Economic growth is usually associated with job creation and rising standards of living. However, Bush's so-called jobless recovery has created an unusually small number of jobs. The jobs that have been created have been predominantly bad jobs with low pay, poor job security and few benefits. The number of Americans living in poverty has risen for three years in a row. The facts on the ground suggests that voters won't go into the voting booth feeling good about the incumbent and the economy.
To recap, regression equations have their place, but they are more suited to describing major trends over long periods. They aren't necessarily useful for forecasting complex outcomes with great precision. If we had nothing else to go on, we might be justified in relying on a regression model to predict medal shares, or election results. Luckily, we can do better. For example, I suspect that medal results for from the last series of world championships would have predicted the 2004 medal counts better than the regression equations. Likewise, public opinion polls are imperfect, but at least provide a snapshot of political opinion within a known margin of uncertainty. At least there's some direct causal relationship between political opinion today and voting behavior in November.
An aside: Fair fans point out that the equation as currently specified would have predicted 18 out of the last 22 elections. But the model as currently specified doesn't have a track record. The election equation gets re-calibrated after each contest. Fair doesn't say how accurate each incarnation of the equation was in predicting that year's election.
[Addendum: Newmark's Door tabulates the 2004 medal count, the predicted medal count of Bernard and Busse, and the 2000 medal count.]