General Discussion
In reply to the discussion: This! This! This all day long. Without comment [View all]BzaDem
(11,142 posts)"Inferential statistics has margins of error to account for the factors which make sampling inaccurate. And sampling involves ascertaining that the sample is representative of the group."
The reported margin of error in a poll takes into account sampling error. It does not take into account modeling error.
The correct population for a political poll is the universe of actual voters on election day. A pre-election poll is conducted before the election (by definition). Because of this, the universe of actual voters is unknowable with certainty. Whenever you are asking human beings to make a prediction about whether or not they will take a certain action (such as voting), many responses will be inaccurate. And nothing from statistics says (or possibly could say) that such inaccuracy in predicting the future is uncorrelated from poll to poll. (In fact, we know that when we ask people who they DID vote for in the past, without any prediction, a small percentage will consistently claim they voted for the winner despite not having voted for the winner. This has been shown over and over.)
Furthermore, even if we discount the problems inherent in asking humans to predict their own future actions, getting an actual representative sample is extremely difficult when 90% of people you contact don't answer their phones. Pollsters use weighting techniques to try to massage their extremely non-representative data into a representative sample, but such weighting necessarily involves making assumptions about the electorate's composition that could be wrong. And the idea that such modeling error is independent from poll to poll is ludicrous. (It isn't even taken into account when calculating the margin of error, because there isn't even a quantifiable way to take it into account prior to the election.)
In this election, the polls slightly under-represented non-college educated whites, even after correcting for other factors. This naturally had a disproportionate effect in states with large numbers of non-college-educated whites, which unfortunately for us included critical swing states.
(Similarly, in 2012, the average of national polls had Obama at +0.7%. He won by 3.7%. This overperformance was even larger than Trump's overperformance. The proability of such a disparity due to sampling error alone, after combining all pre-election polls, is very low. But such error is utterly commonplace, and that is because of error inherent in modeling a future electorate.)
Edit history
![](du4img/smicon-reply-new.gif)