« Two More SUSA Polls (MO & PA) | Main | Rasmussen: 13-Point Lead for Bush in NC »

Saturday, September 11, 2004

To Weight, or Not to Weight

Posted by DavidNYC

One polling-related topic that has come up a lot lately is the issue of weighting by party identification. That is, if a polling outfit does a survey and the sample contains, say, too many Republicans for the given population, should the pollsters adjust the numbers according to what they believe the actual party ratios are? This issue came up most prominently when the LA Times released a poll back in June that was very pro-Kerry, had a seemingly outsize number of Democrats in the sample, and was not weighted to reflect this.

Into this debate charges Alan Reifman, a professor of Human Development and Family Studies at Texas Tech University. He has the most comprehensive essay I've read to date on whether polling firms should weight or not weight. He lucidly presents the arguments on both sides. If you're a serious poll junkie, check it out.

Posted at 12:46 AM in General | Technorati

Comments

I can see the advantages of weighting, but if it's a randomly-selected sample and there is little or no non-response bias, then the % of Republicans vs. Dems should never be too far off from the true distribution. If the percentages ARE too far off (more than 2 SDs), then that probably is a signal that there is something wrong with the way the sample is drawn, in which case the weighted #s would probably still be biased (inaccurate) but better than the unweighted #s. This is why I think they should show the #s BOTH ways -- weighted & unweighted -- just like they do registered & likely voters separately.

Posted by: Jason at September 13, 2004 10:42 AM | Permalink | Edit Comment | Delete Comment

Weighting definitely takes fluctuations out of the poll but has flaws. The biggest problem with weighting is changing demographics. We saw that some in the 2000 election in Florida. Early exit polls showed an Al Gore victory in FL. The problem was that the demographics had changed from 1996 to 2000. The population of the Florida red panhandle had increased as a percentage quicker than blue south Florida. The weighting was faulty toward blue south Florida. Of course, it didn���t help Gore any when Jeb disenfranchised 50,000 registered voters.

Posted by: DFuller at September 13, 2004 10:52 AM | Permalink | Edit Comment | Delete Comment

An argument for weighting is Kerry polls best on Sunday.

I have looked at the Rasmussen polls over the last three weeks. One thing I have noticed is that Kerry always jumps on Monday. That is probably because fundamentalists, which are strong Bush, are in church on Sunday nights. Here is my analysis over the last three weeks:

Kerry polls +4.4% on Sunday vs. Thursday, +1.65% on Monday vs. Friday, and +.1% on Thursday vs. Monday.

Bush polls +4.0% on Wednesday vs. Sunday, +1.2% on Friday vs. Tuesday, +1.05% on Tuesday vs. Saturday, and +.5% on Saturday vs. Wednesday.

Bush's best overall poll for the week is the poll that includes Thursday, Friday, and Saturday. That poll excludes both Wednesday night and Sunday night when fundamentalists are in church.

If you look over the past three weeks:

Th, Fr, and Sa Bush +2.43%
We, Th, and Fr Bush +2.27%
Mo, Tu, and We: Bush +1.90%
Tu, We, and Th: Bush +1.87%
Fr, Sa, and Su: Bush +.97%
Su, Mo, and Tu: Bush +.85%
Sa, Su, and Mo: Bush +.50%

My prediction for Rasmussen tomorrow is Bush +.4%.

Posted by: DFuller at September 13, 2004 01:19 PM | Permalink | Edit Comment | Delete Comment

I did an identical day-of-week analysis for roughly the final six weeks of the 2000 election, using the CNN/USA Today/Gallup daily tracking poll. Here's what I found:

Sat-Sun-Mon Bush ahead by 0.17
Sun-Mon-Tue Gore ahead by 0.17
Mon-Tue-Wed Gore ahead by 0.83
Tue-Wed-Thu Bush ahead by 2.67
Wed-Thu-Fri Bush ahead by 4.00
Thu-Fri-Sat Bush ahead by 5.00
Fri-Sat-Sun Bush ahead by 3.17

I think we both find that W-Th-F and Th-F-Sat are when Bush does best!

Posted by: Alan R. at September 13, 2004 05:42 PM | Permalink | Edit Comment | Delete Comment

Well, since the election is on a TUESDAY, this should be great omen for Kerry!

Posted by: Pepe at September 13, 2004 06:35 PM | Permalink | Edit Comment | Delete Comment

50000? I thought I disenfranchised 50,491 voters.

Posted by: Jeb at October 26, 2004 03:06 PM | Permalink | Edit Comment | Delete Comment