Adam Pearce github twitter email rss

Election Forecast Correlations

538 and the Economist have both released detailed data from their election forecasts, listing how each state votes in 40,000 simulations of the presidential election. To understand some unusual scenarios from the 538 model, like every state voting for Biden but New Jersey, Andrew Gelman examined the correlation in the Trump vote share between pairs of several states.

I was curious what the whole universe of pairwise correlations looked like; you can click on a grid cell below to see voting patterns in two states in more detail along with the electoral maps from individual simulation scenarios. In the 538 model, 242 pairs of states are negatively correlated!

Mousing around the edges of the scatter plot pulls out more unusual scenarios, like Trump losing everywhere but CA-HI-VT.

On the correlation matrices, it appears that both models have identified similar groups of states. A scatter plot shows this directly:

Outside of the CA-DC-VT-WA and LA-MS-ND-KY clusters, where the 538 correlation dips below 0, the models are mostly aligned. Glancing over the outliers, it looks like the Economist might not have an equivalent to 538’s regional regression that groups states in the same geographic region together; the Economist has HI at 0.2 correlation with WA & OR while 538 has it around 0.7.

Stepping back from the funky correlation charts, comparing the Trump vote share state by state clearly shows a bigger difference between the models: the Economist model considers really surprising outcomes, like Trump decisively winning CA, less likely than the 538 model.

I haven’t followed the extensive discussion around election modeling closely enough to have a strong opinion on what all of this means, but it does look like the 538 model is allowing for the possibility of a broad realignment in politics–something you’d want to incorporate when modeling 2024 today, but not plausible for an election next week with sixty million ballots already cast.

Only 5,000 scenarios are shown on the scatter plots; the scenarios are a snapshot from 2020-10-25 and not updated (looking at the correlations over time might be interesting though!). The rendered electoral college scenarios ignore the possibility of NE or ME spliting their votes.

The correlations matrix orders states by clustering on 538’s correlations. Sorting using the Economist correlations splits up the negative 538 correlations.

chart code