Monday, October 13, 2008

Analysis of national pollsters

The esteemed author of the blog asked me to give a quick rundown of the pollsters so here is a quick (and almost certainly incomprehensive) summary of the main ones. Starting with the national polls:

Gallup Daily: This is a three day tracker, (i.e. it's daily number is the average of polling taken on the previous three days), it measures registered voters rather than likely voters, (it will probably switch to likely voters after the debates) and does not weigh for party-id*, however it does have a large sample size (At least 1000 adults every day). Despite this large sample size, Gallup tends to have rather erratic individual day’s results so it probably best to look at the weekly average.

Rasmussen: This is also a three-day tracker. But unlike the Gallup daily it measures both likely voters and does weigh for party-id. It has a rather unique way of doing this, Rasmussen records the party-id numbers from the previous six weeks of polling before the party weight are applied, and then uses this average for the next week, so effectively this poll has a 6 week rolling tracker on party-id which updates every Sunday. This seems (at least to me) one of the better ways to keep track of party-id numbers. The tracker itself has a sample size of about 2500. The combination of a large sample and party weighting leads to a steady poll which does not quickly or overly reacts to individual events.

R2000: This tracker is sponsored by Daily Kos. Perhaps coincidentally, perhaps not, it has a significant democratic party leaning due to a party weighting of +9 democratic (that is there is 9% more self described democrats than republican, 35D - 26R). While this might occur if the democrats play the perfect ground game while the republicans have a terrible one, the tracker is more useful for measuring trends due to the large number of independents this poll predicts. It also has a complete breakdown of all demographics and sub-groups which I wholeheartedly approve of. It is also a likely voter poll.

Hotline: This poll should be a neutral tracker, but also one which, unfortunately does not explain it's methodology clearly. It has a small sample size and has been incredibly erratic in the last week, it moved 5 points in one day which should never happen in 3day tracker. (In a three day tracker only one third of the sample is changed every day). To start with Hotline was a registered voter poll but now has changed to a likely voter poll.

Zogby/Reuters: The most recent of the tracker. It has a moderate size about 1200LVs and also gives its results to one decimal place which is helpful. Unfortunately it weighs party-id to a very conservative +2 democratic. This should give it a strong republican leaning (even more leaning/biased than the R2000/Kos poll.

GWU/Battleground: This poll has been plagued by early methodological problems. Hence it difficult to come to any particular conclusion about this poll, although in the past GWU has had a good reputation. It is a sometimes 4 day, sometimes 5 day tracker which used to but no longer weigh for party-id. It has a low sample size for individual day (200LVs) so expect it to be erratic and do not put much weight into individual days. (For further information about the early problems, here is a link)

YouGov/Economist: This is a weekly internet poll. Normally internet polls are very unreliable however YouGov has a pretty good record and reputation in Britain. The poll tends to have a relatively large number of undecideds and voter for other candidates. Hence this poll is a probably a good poll to examine the base support for each candidate but may not tell you where the leaners are going. This is the only poll which measures all adults.

ODC/Fox news: Despite their sponsor this is a relatively neutral poll, although Fox News interpretation of their poll is always amusing. (Like the You_Gov this poll has a large number of undecideds but a fairly average party weighting and moderate sample size. Note the fox poll was one of only two pollsters which had Kerry winning in their final poll in 2004.

Pew Research: This is one of the better pollsters, in 2004 they matched the final result exactly (51% Bush - 47% Kerry) and are widely respected. They do not weight their polls. The poll tends to have a good sample size (especially their final poll) and measures likely voters.

CBS & ABC/WaPost: Both these polls have relatively small sample sizes but both were near the 2004 result. Note the CBS uses RVs but ABC uses LVs

NBC/WSJ: This poll uses registered voter and like other news network this poll has a small sample, but anything poll-related that Chuck Todd associates himself with, deserves to be taken very seriously.

Democracy Corp: Despite the company being connected with the democratic party, this poll tends to lean republican due to conservative weighting on key demographics (age and party-id). However it does have one of the most detailed breakdowns on who is included in the poll together with a large amount of information about how voters view the candidates which makes for interesting reading

LA Times/Bloomberg & Newsweek: Neither of these poll weigh for party-id and became somewhat notorious during the summer for giving the number of democratic voters a huge margin over the number of republican voters and hence giving Obama large double figure leads. I would take both polls with a large pinch of salt.

*One of the most difficult decisions a pollster must make is whether to weigh for party-id, and if so at what levels should the party make-up of the electorate be. The advantage of weighing for party-id is it tends to reduce the variability of the poll, the disadvantage is the possibility of introduce a fixed bias if your numbers for the size of each party are wrong.

No comments: