THE PLANK JUNE 27, 2008
-
Read Later
READ LATERAvailable only to subscribers. SUBSCRIBE TODAY
-
Listen
ARTICLE AUDIO
- Font Size
We seem to have gotten past some point of no return where we have a half-dozen polls or so to look at almost every day:
A lot of these polls are good news/bad news for both candidates. In Rasmussen's poll of Mississippi,
for instance, Barack Obama hasn't improved his numbers from the May
edition of the poll, which had also showed him trailing by a 50-44
margin. On the other hand, that result had looked to me like an outlier
before, and now can probably be taken more credibly. While Mississippi
remains a longshot for Obama, keep an eye on Louisiana, which has
similar demographics but much bluer party identification figures.
The two polls of Texas show the race tightening, but probably not enough to make the race interesting. There's a good rundown here of the pros and cons of Obama investing resources in Texas. The only thing I'd add is that there is far more room at the margins
for the Democrats to make up ground with registration among Latino
voters than among African-Americans. If Tejanos vote in anything
resembling the same proportion that they constitute of Texas's
citizenry, the state could be quite competitive.
But it's the SurveyUSA result in Ohio that I want to focus on. Obama leads by 2 here, but had been ahead by 9 in SurveyUSA's may poll of the state.
That previous poll had shown a heavily Democratic sample -- 52 percent
Democrat, 28 percent Republican, 18 percent independent -- and had
triggered a lot of discussion about whether pollsters should be
weighting their results by party ID. SurveyUSA does not do so --
although if it had applied the May distribution of party IDs to this
poll, it would have shown Obama ahead by 10-11 points rather than by 2.
Conversely, if SurveyUSA had applied the June party ID distribution to
its May poll, that poll would have shown a dead heat rather than Obama
ahead by 9.
I do not mean to be a fair-weather fan on the idea of weighting by party ID. As I implied the other day,
I suspect that pollsters are facing something of a trade-off between
volatility and potentially introducing bias. Weighting by party ID will
almost certainly reduce noise, and perhaps make it easier to perceive
trendlines -- but if the pollster's guesses about party ID are wrong,
they may be reducing the turbulence but landing at the wrong airport.
I'll say this: if a pollster doesn't know what it's doing, I think it
should be letting the numbers speak for themselves. On the other hand,
if the pollster has a robust and thoughtful method for weighting by
party ID, it might be worth the trade-off. It is interesting that,
taking two of our three highest-rated pollsters, Scott Rasmussen is a firm believer that you ought to weight by party ID, and Jay Leve at SurveyUSA is a firm believer that you ought not.
--Nate Silver
5 comments
Weighting raw results to reflect demographics such as party ID is very valuable, but only as a tool for validating the raw results, that is, to ensure that there are no systematic biases being introduced by the sampling methodology. They are not, however, a substitute for proper sampling which would make the weighted results irrelevant. The reason is that demographics don't vote, people do. If a sample is a fair sample of the voting population, such a poll should be more accurate than a weighted analysis.
- roidubouloi
June 27, 2008 at 6:13pm
To clarify a bit, if you have a good sample, then you will not only have a good representation broken down by party affiliation, but a fair representation of such ambiguous groups as "people who never got around to changing their affiliation," "people who are only registered to a party because their parents were," etc. for which it is not possible to adjust the raw results. Adjusting for party affiliation adds some information, but washes out other information. If the sampling is sound, then there is no information to add by adjustment, but information is still lost. Hence, the adjusted results end up less reliable. But that assumes the soundness of the sample, something not easy to establish because it is hard to know what subtle correlations might exist with the sampling methodology (the famous example being the phone polls that predicted "Dewey Beats Truman"). Thus, adjusting results for various demographic cross-sections is a good way to validate the sampling methodology and hence the reliability of the poll. It is unwise, however, to rely on adjusted results when you find significant differences. Better to re-think the sampling technique, unless you are out of time or money.
- roidubouloi
June 27, 2008 at 6:35pm
Nate - I can't tell you how pleased I am that you're blogging here for the election, or even that you're blogging at all on politics. I always enjoyed your baseball related stuff and the political stuff is top notch as well.
- brianlitz3
June 27, 2008 at 6:36pm
Glad to see you Roi, God knows I can never really make head nor tails of polls without your take.
- Wandreycer1
June 27, 2008 at 8:36pm
Nate,
I really enjoy your posts as well. I became a fan of 538 recently.
I live in Louisiana. While Bill Clinton did carry LA twice, and once by 12 points, It seems the state has become more conservative since Katrina. Our Democratic base in New Orleans was dispersed across the country. That's a big reason Jindal had a cakewalk in the Governor's race. He still would have had a very good chance, but the Katrina mess really hurt our Democratic governor's reputation. Before Katrina, she was fairly popular
Losing a large chunk of the New Orleans population has really hurt the Democratic chances here.
Best,
Webb
- webbhaymaker
June 28, 2008 at 5:50pm