Site Meter

Monday, September 20, 2004

Robert Rips Rasmussen Report

The Rasmussen Report discusses data on good days (of the week) for Kerry and Bush. For some utterly mysterious reason, Rasmussenreports reports the average by release day "The data presented at the table to the left is based upon release days. Our data is released as a three-day rolling average, so the data is collected on the three nights prior to the release day (i.e.--results released Thursday use interviews conducted on Monday, Tuesday, and Wednesday nights)."

This means that Rasmussen reports the average of moving averages. This is crazy, Give 16,500 observations per week day, there is no need for a moving average. It is not possible to extract the results by day of polling from the moving average (and uninvertable ma). This not only makes it much more difficult to figure out what the numbers means but also reduces the precision of the following estimate.

On Thursday, Friday and Saturday, Rasmussen reports the average of results for 3 week days. on average of average of moving average, Bush is up 2 2/3 % on such days. On Wednesday and Sunday, Rasmussen reports results from two weekdays and one weekend (respectively Sunday and Saturday). Kery is up on average 0.2%. On Monday and Tuesday, Rasmussen averages over two weekend days and one weekday. Kerry is up on average o.35%. On average to the fourth replacing a weekday with a weekend day helps Kerry slightly more than 0.3o8%. Each day is divided by 3 (in the moving average) so the estimated difference
(Bush-Kerry)_weekday - (Bush-Kerry)_weekday is slightly more than 0.924%. This is not a huge number.

However the standard errors of the release day averages are less than half of a percent (1/7*root(10)). The overlap of subtracted days and further averaging makes ... well the standard error of the 0.924% estimate very hard to calculate but small. If the point estimate extracted from Rasmussen's average of moving averages is correct, then the evidence of a weekday vs weekend effect is not statistically significant at the 5% level.

Let's pretend that Bush-Kerry is 0.9% lower on weekends then on weekdays. There are about 33,000 weekend observations and about 82,5000 weekday observations. For each observation the variance Bush-Kerry is almost exactly 1 (slightly less as don't know and other are worth zero but almost exactly). The variance of the average Bush - Kerry on weekends is 1/33,000. The variance of the average Bush-Kerry on weekdays is 1/82,500. These are two independent samples so the variance of the difference is the sum of the variance 7/165,00 so the standard error is 1/(100root(2.34)) that is 1/(root(2.34)) % which is a little less than 2/3. This means that 0.9 is slightly more than 1.35 standard errors. Evidence that weekdends are better for Kerry than weekdays, but not, according to the agreed convention, strong evidence.

No comments: