If people lie, you can't poll accurately.
And when people view an election as a game to "win", why not lie to throw off the other side?
It's a basic art of war tactic.
And when people view an election as a game to "win", why not lie to throw off the other side?
It's a basic art of war tactic.
Of course pollsters already know this, so they weight the samples to reflect underlying demographics, but it gets tricky to do this accurately if your surveying response rate is 1% to 2%.
And of course... the forecast modelers know this. Which is what Gelman discusses in his post. Given that they now have 2016, and 2020 as examples of presidential polling errors, modelers need to shift their priors in a number of ways to account for what looks like a persistent undersampling of GOP voters.
That doesn't actually help unless the difference in response rate is explained by the underlying demographics. If it is correlated with preferences on the issues polled even controlling for non-preference demographics, it's not going to be correctable by that means, nor is it going to be measurable so that it can be corrected.
OTOH, if it is a long-term consistent effect tied to partisan presidential voting preferences across election (and not, e.g., tied to the particular transitory trends, or evolving significantly over time) you could, as a poll-based forecaster, correct for the effect without measuring the correlation in preference and response rates by adjusting how you draw forecasts from polling data by correctingnbased on empirical election results and their relation to polls.
In the article, he mentions the simplest and probably most correct way to do this: reduce the confidence these models have in the accuracy of the polling. The top line number he threw out as feeling "good" about was 75% for a Biden win.
As it is, it's becoming difficult to find any value in these models because the events they represent as outliers seem to be consistently occurring. On the other hand, if they adjust to more accurately reflect the uncertainty in the the polls, eventually everything starts looking close to a coin toss - in which case they have no practical value, either.
Propensity to participate in polls having even a slight correlation to preferences on the issues being polled will create a systematic polling error, however, and it's an effect that is difficult to measure because how do you measure the preferences of nonrespondents?
Without coordination a bunch of independent actors are generally predictable, because any random choice they makes averages out to the mean choice. With coordination the pollsters should be capable of finding out about it too.
So I think lying to polls only works while it's still rare enough (in terms of # of elections that it happens in) that the pollsters aren't accounting for it.
Either conservatives have started lying to pollsters about their voting intentions or conservative officials started manipulating the vote totals in their favor.
Take your pick as to which possible explanation is most palatable and/or most probable.
If you don't have a representative sample of the population your poll will be inaccurate.
Do you have any reason to believe that pollsters and other statisticians are so ignorant of their own field not to understand this?
So they think they have a representative sample, but they really don't.
It can be explained simply by pollsters mainly living in large cities, and not really understanding rural Americans.
But that would start costing them elections, so instead they coddle and abet the worst of them.
What do you call 9 people staying quiet when a tenth is being blatantly racist? Ten racists. Take it from me, I saw it in action growing up white in Mississippi.