Site Meter

Saturday, October 23, 2004

Bush 47.8% Kerry 46.3 % standard error of difference about 0.6%
is the result of averaging state level polls taken after October 10 (if available other wise most recent poll available) and then taking a population weighted average over states. Notice that polls from October 1 through 10 (see below) change the averages essentially not at all and reduce the sampling standard error very little.

Data at or as of 6 pm EST October 22.

A population weighted average state averages of all available polls in October (or latest in September if none in October) except for multiple observations from tracking polls gives
Bush 47.9% Kerry 46.3 % with a sampling standard error of the difference of about 0.53%

using only polls from October 1 through 10 (or latest in September if none available) gives Bush 47.8% Kerry 46.3 %.

Using only polls from September gives Bush 47.9% Kerry 46% with a sampling standard error of about 0.4%.

The amazing absense of change in the weighted average state average poll is in part due to the fact that there is some overlap of polls. In some states no polls have been taken in October so the state "October average" is the most recent poll from September. The overlap of the aiming for after 10/10 average and the aiming for before 10/10 average is modest. Basically the absence of change in the average is a bit surprising even under the assumption that there has been 0 true change in opinions. Certainly there is no sign of the fluctuation one sees in the averages national polls.

For example the standard error due to sampling of the change of the difference from September to so far in October would be root(0.53 squared + 0.4 squared) or slightly less than 0.7%. The observed change is Bush-Kerry dropped by 0.3 %

The reason that including more polls has such a small effect on the standard error is that about half of it is due to sampling error in the "average" of one poll each for Texas and New York. People in these large uncontested states are polled much less often than would be ideal in order to get a precise estimate of national opinion. The "standard error" i, as reported by pollsters due only to sampling under the assumption that the sample is unbiased. It should be independent in each poll. Needless to say, I don't believe this is the only source of error. Part of the point of this little project (if any) is to test for bias with an average with a very small stated standard error.

No comments: