3 ms·
In 2016 the polls were arguably "correct" in that, though they incorrectly predicted a Clinton win, the results were still mostly within the margin of error, i.
by ineedasername 6y ago
In 2016 the polls were arguably "correct" in that, though they incorrectly predicted a Clinton win, the results were still mostly within the margin of error, i.e., the polls weren't really predicting one candidate or another because Clinton's lead was too small to be statistically significant.
This time, we might end up in the counter-intuitive territory where the poll were "wrong" even if they pointed in the direction of the winner, because the results are outside the margin of error.
- AnthonyMouse 6y ago> In 2016 the polls were arguably "correct" in that, though they incorrectly predicted a Clinton win, the results were still mostly within the margin of error, i.e., the polls weren't really predicting one candidate or another because Clinton's lead was too small to be statistically significant. The problem is that Presidential elections are always pretty close. It has been almost a half century since any candidate has gotten more than 60% of the popular vote. If the result is always within the margin of error, how useful does that make the polls?
- Swizec 6y ago> If the result is always within the margin of error, how useful does that make the polls? From the perspective of ad clicks and news readership, very.
- Karrot_Kream 6y ago> If the result is always within the margin of error, how useful does that make the polls? That's what Gelman's post is about. There's something wrong with polling methodology and statisticians' models, and it's time to see what.
- dragonwriter 6y ago> That's what Gelman's post is about. There's something wrong with polling methodology and statisticians' models, and it's time to see what. The problem is that it's very hard without more data than can be gathered, and it may be a moving target. This same fact (or, rather, highly likely conclusion of substantial, correlated polling error indicating a systemic error) was, remember, widely observed in the wake of 2016 and the common hypothesis (for which fixes were implemented in 2020) was that the problem was pollsters not weighting samples for education.
- ineedasername 6y agoSomewhat: no individual poll should be considered all that useful, but the overall direction of the polls should be a better indicator as it's less likely to occur by chance that all of the polls are off in the same direction. But that's not what we've seen: Something is wrong with the methodology here, perhaps some form of sampling bias. Perhaps Trump is attracting supporters who typically didn't vote before, so polls' that use things like "likely voters" aren't performing well. I suspect it's still more complicated than that though. I'd like to see more polls after the election that ask people "who did you vote for". If those polls are skewed against actual results in the same way as pre-election polls, that might point in the direction of the problem.
- Balgair 6y agoI mean sure, but if the best you can do is get +/- 3% that's not good enough anymore. This US election cycle was ~$50B over all the races. 3% doesn't cut it at those ad spends. Something more precise is needed and the funds are more than there to develop it.