Wednesday, September 17, 2008

Rasmussen Just Changed the Rules mid game

They will add roughly 1% 2% to McCain minus Obama compared to what the results would have been if they had stuck to the system they always said without qualification that they use. Now they "Generally" weight so sample party affiliation matches the three month average,

However, from now through Election Day, we are making two adjustments to that process. First, due to potentially changes in the political environment, we will now base our targets upon survey interviews conducted over the preceding six weeks, rather than three months. Second, we will update the targets weekly, rather than monthly.

Makes some sense to me. The number of observations in the six week average is huge and party affiliation is not constant. However, I am not thrilled about suddenly changing the method.

The effect ?

For polling data released during the week of September 14-20, 2008, the new targets are 38.7% Democratic, 33.6% Republican, and 27.7% unaffiliated. For the first thirteen days of September, the targets were 39.7% Democrat, 32.1% Republican, and 28.2% unaffiliated. It should be noted that these figures indicate an improvement for Democrats since the beginning of the year.

So they just rejiggered their numbers 1.0 from Dem to Rep (about 0.85 0.8 % for McCain) and 0.5 unaffiliated to rep (about 0.15% for McCain)

Also they Kant spell.

"A one-percentage point decrease in the number of Democrats would half the same impact. "

update: and a similar 0.9 to 1% from Obama so a change in McCain-Obama of about 2% Robert you bozo.

This means that efforts to determine whether Obama has gained statistically significant ground on McCain recently have to add about 2% to Sunday and later released Rasmussen results to make them comparable to Saturday and before.

Unweighted average of the 4 most recent at is Obama + 1 now corrected to Obama +1.5 (just for comparison with the past the new weights are probably better).

Older 4 + then values of trackers (data through 9/10) are (0+4+4+3-1+4+2)/7 = Obama - 2.3

Var of the avg of 4 = roughly (1/3000+1/3000+1/1000+1/1000)/16 = 1/6000

Of the older more like (2/3000+5/1000)/49 = about 1/9000 so sum of vars about
5/18000 so se around 1/60 so z-score 2.28 so the correction to the Rasmussen revision just pushes it over 2 (would be 1.98 without).

Boarderline significant evidence that Obama has gained on McCain.

Now weighting by sample size and using new Rasmussen weights implies exactly even (also if you include recent polls with Barr and Nader which average to zero Research 2000 Obama +3 ARG Obama -3).

1 comment:

Anonymous said...

Check out Rasmussen's electoral college "balance of power" table--all summer Obama was gaining, then from July 22 to Aug 22 they didn't make any changes to the table at all. Starting from Aug 22 they changed the way they weighed the states and since then almost all the changes have been in McCain's favor.