Here’s a summary performance, starting from 2004, of a meta-analytic approach based purely on available polls. The consistent finding in three elections: simple is best.
Click on the years to see earlier postings and news clips. Numbers in [square brackets] indicate the 68% confidence interval for the Democratic-side outcome.
Senate polls – 50.5 D/I, 49.5 R (50% chance of Democratic takeover).
Final outcome – 51 D/I, 49 R.
House polls – 234.5 D, 200.5 R, [232,237].
Final outcome – 233 D, 202 R.
2008 (results pending a few remaining results):
Final single-day snapshot – Obama 352 EV, McCain 186 EV. My prediction: Obama 364 EV, McCain 174 EV, [352, 367]. Popular vote 53.0% to 46.0%, [52.5,53.5]%.
Outcome – Obama 365 EV, McCain 173 EV. Popular vote 52.7% to 45.9%.
Senate polls – 58 D/I, 42 R (16% chance of a 59-41 split).
Outcome – 58 D/I, 41 R, MN recount pending.
House polls – 257 D, 178 R, [254,260].
Outcome – 257 D, 178 R.
Where comparisons exist, meta-analysis by straightforward reduction of poll data outperformed econometric predictive models, poll aggregators (electoral-vote.com, the 3BlueDudes average, and FiveThirtyEight), and the InTrade electronic market. Overall, the results support the conclusion that the collective “market” of pollster methods contains sufficient information to converge extremely well on final outcomes. Adjustments based on undecided-voter assignment, voter demographics, and pollster reliability have the effect of reducing accuracy.
I’m about to fade back into academic mode, which means reduced posting. I hope to write up this whole project for a regular peer-reviewed publication. In the meantime, for one last bit of bloggy goodness, including some pretty, um, robust comments, let’s look at the mailbag.