US Election 2024 - Poll aggregator review - Princeton Election Consortium

Who are Princeton Election Consortium

One of the older electoral aggregators - they’ve been going since 2004 - PEC is associated with Princeton University and hosted by them. And despite being named a “consortium”, they appear to be pretty much a one-man-plus grad student show: it has been run from the beginning, by a professor of biostatistics: Sam Wang.

Their new logo is a little US bug, which I assume is a wry nod to the fact that shortly before the election Professor Wang claimed on television that if Donald Trump got more than 240 Electoral votes (that being still a long way short of the magic 270), he would eat a bug.

Donald Trump of course gained this this (he got 304) and so Professor Wang went ahead and ate a cricket on CNN. In the linked article, he explains all this, and tries to go into how he managed get things so very wrong. I think he slightly dances around the core issue though - and in the below - I’ll be a bit more blunt. The PEC method does not account for correlation of error. It essentially assumes independence across US states. And so it is massively overconfident in its probabilistic predictions for the General Election result.

To be fair, PEC do not actually display a probability estimate for the general election. Instead, they show the “Presidential Meta-Margin” which can be thought of (roughly) as the number of expected electoral votes above or below 270. But their model does produce probabilities nonetheless: they show a histogram of all the possible electoral college outcomes and their estimated probabilities.

Here it is (snapped on 29 August 2024).

In order to derive the probability of a (Trump) win all one has to do is to add up all the bars up to 270. And in order to get a Harris win, you just take this number away from one.

And this is what I have done to get the comparison to other pollsters (or, more precisely, I’ve used a Python script to get the numbers from their Github, and asked it to run backwards through their commit history to get a time series).

PEC’s line is quite something.