Friday, May 2, 2008

Just How Wrong are the NC Polls?

Looking back at my previous post There is no Bradley Effect!, I couldn’t help but notice a pronounced pattern: The four states in which Obama outperformed the polls by at least 10% were all states with relatively high African-American populations. Intrigued, I decided to see if this pattern would hold for other high-AA states (including DC). Since RealClearPolitics didn’t have published polling data for a number of these states, I turned to usaelectionpolls.com to fill in the gaps. Here is what I found:

State% AAActual SpreadPoll SpreadDifference
DC5751.53615.5
MS3624.515.49.1
LA3221.8**
SC2828.911.617.3
GA2735.31817.3
MD2724.922.32.6
AL2514.4-1.215.6
NC208.2
DE1910.8-212.8
VA1828.217.710.5
*No polling data was available for Louisiana.

The pattern is remarkable. There seems to be a consistently high error rate in these states, with Obama outperforming the polls by an average of 12.6%. One must conclude that there is some systemic bias in the pollsters’ methods that underestimates the level of AA support for Obama. My guess is that the models used to determine Likely Voters fail to account for unexpectedly high turnout among the AA population.

There is one anomaly in the data: Somehow, the polls predicted Maryland with reasonable accuracy. I have no explanation for why that sole state managed to buck the pattern.

But now comes the fun part – predictions! The current RCP average for NC gives Obama an 8.2% edge over Clinton. But if he outperforms the polls by the expected 12.6%, then his win will actually be 20.8%. You heard it here first!

At least one pundit has opined that Clinton has an outside chance of winning NC. Based on the analysis here, I would advise Clinton supporters to avoid setting expectations for even a single digit loss.

Update (5/5/08): As of the eve of the primary, Obama's edge for NC has gone down to 7.0%, according to RCP. That brings my prediction down to 19.6%, plus or minus the usual margin of error. Let's see what happens!

No comments: