Princeton Election Consortium


After all the canvassing, donation, and activism, tonight now we observe. Later this afternoon I will publish the Princeton Election Consortium biennial Geek’s Guide to the election. It will detail what I will be watching.

This is also an open thread – please chime in. I may host a Zoom call tonight as well – stay tuned.

Reweighting can be mistaken for herding

There’s a lot of talk about how pollsters are watching each other, and coming up with results that don’t get too far out of line with the average. This is called herding.

In my view herding is somewhat overblown, and not to be confused with expert judgment. It is hard to tell the difference, but I would caution against impugning the craft of how one generates a good survey sample.

Pollsters are faced with the problem of identifying who will vote, and normalizing their sample to match. If some demographic of people is disproportionately likely to answer a survey, more often than they would vote, then pollsters have to weight those responses appropriately. This is a matter of judgment.

As it turns out, re-weighting reduces variability. Here is an example.

Imagine 3 groups of voters, distributed as 45% Democratic-leaning, 10% independent, and Republican (45%). They vote for the D candidate with probability 0.9, 0.5, and 0.1 respectively.

Below in blue is the raw result of 10,000 simulations of samples from this population (my MATLAB script is here). The results are pretty spread out.

In red are the same results, reweighting the sample with prior known proportions of the voter population. Reweighting narrows the apparent result. Pollsters reweight by more variables, and these various weightings (sometimes it’s called raking) will narrow the distribution further. This is just what pollsters do routinely.

The judgment comes from knowing what weights to use. Every pollster has their own set of weights that they use. Poll aggregation collects the crowd wisdom of these pollsters. In the aggregate, they will still be off – that is called systematic error.

How much will the systematic error be this year? In 2016, it was about 1.5 points, leading to a surprise Trump win. It would not be crazy to imagine an error that large in 2024. But in which direction, I don’t know.

The one thing we can say is that this error is much larger than the virtual margin of the presidential race, where that margin is defined in terms of how far the race is from a perfect electoral tie. I calculate this “meta-margin” as Trump +0.3%. That’s much smaller than the error. This has two implications:

  • We don’t know who’s going to win based on polls alone.
  • The final outcome is likely to be less close in one direction or the other.

Later this afternoon I will give evidence for why I believe the error will end up favoring Kamala Harris.

Topics:

We will be happy to hear your thoughts

Leave a reply

Som2ny Network
Logo
Register New Account
Compare items
  • Total (0)
Compare
0
Shopping cart