About This Site

Arguments and analysis of the different methodologies used by the various “pollsters”, and discussion about strengths and weaknesses of each method (from 538 Electoral Projections - comments section from readers).

mikelow1885 said… Perhaps Pollster.com’s number needs a slight adjustment, since
every third day the trackers aren’t counted. So my guess the true Pollster.com average would be a 1.5 or 1.6% Obama lead. So it may just be a model based so heavily on state polls will show a dead heat.

Tyrone said… Pollster is a biased site. They use polling averages from all the way back in February.

Their “best fit” lines make no sense and are inaccurate. They make their graphs Obama-friendly.

mostman said… Tyrone - Seriously? Is this because your are a McCain fan and approve more of the volatility of the horrible RCP graph based on AVERAGES. The graph that shows wild McCain upticks every time an outlier poll comes along with McCain polling 6 points higher than usual? Pollster is based on Math. Real Math. Not emotion. There is no bias built in. Go read about how they do their trend lines.
VegnaBlitz said… I’d chalk up the difference in results to 538’s treatment of state polls. Nate’s result heavily correlates to the electoral college result, as you would hope such an average would. But that’s the simple explanation.
SarahLawrenceScott said… Isn’t the reason that 538 shows a closer race than RCP or pollster the regression analysis? In fact, isn’t it kind of the point of 538 that it’s not just averaging polls?
fuzzy said… duh… yours (538) are the most McCain friendly because your projection model is structured to assume the losing candidate will close the gap as the election approaches. At least that’s how I understand it. Your number isn’t supposed to reflect current polling, is it?
DaWolf said… Nate, I reckon yours comes low for a couple of reasons

1)Regression Analysis
2)Higher commitment to state polls
3)You automatically discount the results this far out and make it closer.

SarahLawrenceScott said… This all comes so fast now, all we’ll really be able to tell is the net effect of the VP’s and the conventions. Even the tracking polls often show a few days lag for news to percolate through the population.
mike Rappeport said… The problem is not that 538 uses state polls and RCP and Pollster.com use national polls. The problem is that 45% of 538’s August state polls were by Rasmussen and 15% by Survey USA. As the national poll data shows Rasmussen is consistently about 2% more favorable to McCain than the average of the non-tracking polls. That is probably because the tracking polls use a fresh sample nightly, while the non-tracking polls use the better methodology of call-backs on successive nights. This would tend to disproportionately include stay at home (older) voters. As for the Survey USA polls, they did seven polls in August in states where there was at least one other poll, and in 10 of those polls they were more favorable to McCain than the other polls in the same state (and there was one tie in Virginia). Actually the median (a more stable, less poll dependent, measure) of the last five non-tracking national polls in RCP is Obama +3 and the median of the previous five polls was also 3%. In all likelihood Obama was +3% at the beginning of the month and still is.
SarahLawrenceScott said… glenn-in-colorado: Note that ground game, almost by definition, messes with likely voter screens by pollsters. A common likely voter question is “are you registered to vote?” There is no reason for anyone in a same-day registration state who is under 22 to have registered already, unless they voted in the primaries. So polls for those states that use that question get screwed up. The states in that category are Minnesota, Wisconsin, Maine, Idaho, Wyoming, North Dakota, and New Hampshire.
judas_priest said… One reason Nate’s model shows McCain closer is that it expressly reduces the impact of polls the older they are. Since there hasd been a trend toward McCain, the older, more Obama-friendly polls, lose their impact. If the polls turn around, Nate’s model will likely produce readings showing Obama with a higher support level than the other two.
Herunar said… Pollster sucks. And RCP gives equal weight to any poll over a few weeks, a primitive, inaccurate method. Personally I think Nate’s numbers are a bit too McCain-friendly, but they are much more professional than Pollster and RCP.
SarahLawrenceScott said… judas_priest: It’s even beyond that. The model incorporates the current trend when it smooths the data. That means it will always overshoot a little bit at a turning point.

Presumably, Obama is currently at what is at least a local minimum in his polling, since he’ll get some convention bounce. The 538 model will, initially, discount the start of that bounce as anomalous and counter to the trend, which has been down. Then, after several days, it will start to recognize the change as “real,” and overshoot in the other direction, mistaking a bounce for a secular trend.

Michael said… I think that the current pop vote discrepancy between the 538 trend and the other averages may actually be due to a problem with the 538 update procedure. It seems like there was some initial normalization to national polls, then changes in the trend are dominated by changes in the numerous state polls. This sort of incremental process can easily get a little uncalibrated, in part because of the non-systematic pattern of state polling. If the 538 national pop vote stays out of whack with the national polls, then it needs to be recalibrated, or even better to have some other update algorithm that does a better job of continuously recalibrating.

p.s. For the many who think that the assumption of regression towards 50-50 accounts for the shift, that regression is obviously currently zero, as you can see on the super-tracker.

p.p.s. Sarah Lawrence- You may be right that the shift is due to the trend-estimating algorithm, in which case it will reverse sign if the trend reverses sign.

Sedi said… “That means it will always overshoot a little bit at a turning point.”

Exactly. This is what we are seeing now, since all of the polling evidence from the last couple of days suggests that Obama is starting to tick up again after McCain gaining over a couple week period. While we humans can see it pretty clearly, the model is delayed in its response. It was slow to pick up McCain’s gains and is slow to pick up Obama’s now. That’s also one of the reasons why 538.com has the race so close while other sites that simply reflect poll averages show a clear but small Obama lead. I prefer the polls for judging where we are right now, but the projection model will likely work fairly well for the election (it did quite well in the primaries).

I, for one, think that if Nate is really trying to project the election outcome, then campaign spending (not just fundraising) in a state, both on ads and on the ground game. These things absolutely matter in elections, as polls can and do measure. I wouldn’t know how to incorporate it into the model, but I would have more confidence in it if they were included. I think.