Site Meter

Tuesday, October 19, 2004

Bush 47.31 % Kerry 46.43 % from Meta analysis of State Polls

This time I did some calculations myself. I took the state polls from 2.004k.com . For each state I averaged polls from October (if available). I used only the most recent poll by each polling agency. If there were no polls in a state in October, I used the most recent poll in September.

I learned that there are an absurd number of polls, even in states where the outcome is not in doubt.

Then I took the census estimates of population in 2003 and calculated a population weighted average
Bush 47.31 % Kerry 46.43 % so Bus-Kerry = 0.88%

This strikes me as closer than the average national poll in October.

If one assumes that the only difference between the polls and the actual election is sampling error, the estimates have an absurdly low standard error. I don't have data on the sample sizes of the polls, but even assuming each poll has a small sample, the standard error of the population weighted average is tiny.

Assuming each poll has a sample size of 600 (or that the square root of the inverse of the average of the inverse of the sample size squared is 600)
gives a standard error for the estimate of Bush-Kerry of 0.57%. Assuming a sample size of just 400 for each poll gives a standard error of 0.70%. Finally assuming a sample size of 1000 for each poll gives a standard error of 0.44%. 1000 is large compared to many samples for state polls and almost all likely voter samples.

The race is so close that even this absurdly low standard error based on assuming all error is sampling error implies that Bush's lead is not quite statistically significant.

Of course given the pattern that undecided voters tend to vote for the challenger, Kerry is clearly ahead.

I am interested in testing whether the average of state polls gives a better result for Kerry. It looks that way to me, but I might not be considering the facts that the state polls are, on average, slightly older than the national polls and that 2.004k.com uses polls of registered voters if available.

Anyway, I crunched some numbers and got Bush 47.31 % Kerry 46.43 %


No comments: