Thread by @gelliottmorris

#NEW from me: The Economist is now publishing state-level estimates of 2020 vote intention based on my analysis of our our polling data from YouGov. Let me explain how this model works & what it tell us about what might go down in November. THREAD economist.com/united-states/…

The problem: National polls can mislead us about what might happen in POTUS elections. That's b/c there can be large gaps between how left- or right- leaning competitive states are compared to the country as a whole—& the electoral college elects the POTUS, not the popular vote.

State-level polls are also few and far between these days, and many still haven't fixed the problems that caused SOME of them to misfire in 2016 (no education weights being the biggest issues). We think that a lack of weights for political lean could be a similar issue this year.

So it doesn't make a whole lot of sense to report on these national polls as if they're that predictive of the EC.

Luckily, we have data from YouGov (which wasn't involved in the model-building for this project) that we can use to turn the national polls into state-level ones!

The solution: A statistical technique called “multilevel regression and poststratification” (MRP) that takes raw polling data, figures out how likely different demographic groups are to vote for Dems or Reps, and estimates state-level opinion given the demographics of each state.

(We've used MRP before, FYI:

(1) In a story about how high turnout helps Democrats: economist.com/graphic-detail…

(2) About the Senate being biased in favor of Donald Trump's acquittal on impeachable offenses economist.com/graphic-detail…)

As on April 28th, this is what our MRP model said about the election **if it were held today**:

Joe Biden is favored over Donald Trump in states worth a majority of votes in the electoral college victory. Notable that TX, AZ and GA are all competitive. National vote of Biden +6.

Our MRP model ensures that the state-level estimates are representative of the electorate by weighting on:

- Gender
- Age
- Race
- Education
- Marital status
- Religion
and
- 2016 turnout and vote choice

[[Hang in there, methods talk is almost over now]]

The model works by creating predictions for each of these 380,000+ groups of American eligible voters. We know how likely they are to turn out to vote, and also how much more Democratic- and Republican- leaning each of them are compared to the 2016 election.

So to figure out _why_ Biden is performing better than Clinton in certain states, we can look at shifts among all demographic groups.

We see that Biden is doing better w/
- White voters (both with and w/o a college degree)
- Boomers

& worse with
- Non-whites
- Young people

These shifts have improved Biden's standing across the board, but especially in states with more working-class white voters. In the crucial Midwestern states, Biden is out-polling Clinton's 2016 result by about 5 points.

Perhaps counter-intuitively, however, this has not broken Donald Trump's advantage in the electoral college relative to the popular vote. While Biden is currently favored to win the election, the president would be the favorite if he closed the popular vote gap to 3pts or less.

This is because we expect turnout for non-whites to increase versus 2016 more than for whites, but most e.g. Hispanics are not located in competitive states, pushing the nationwide Democratic popular vote margin up more the increase in the electorally decisive states.

By using MRP, we:

- Have moved beyond relying on national polls (or, even worse, low-quality state-level ones) to infer what **is currently happening** in the electoral college
- Can dive into voter behavior among 380k demographic subgroups.

A technical note: similar to polls (or any other model, really), MRP is only an approximation of reality. As such, our estimates come with margins of error. Depending on the number of people sampled in each state, these margins of error can range between 5 and 20 points.

(Oh, I've also done something that I haven't seen pollsters do elsewhere, which is that our margins of error take into account the uncertainty in both the likely voter filter and vote intention model. That partially explains why they are wider than the MOEs you'll see elsewhere.)

So you shouldn't refer to our estimates as the end-all be-all for election handicapping. Instead, we think we have produced better-than-replacement-level estimates of state-level public opinion when we don't have high-quality state polls to benchmark off of.

A final clarification: this is not an election forecast. Think of the MRP model as producing 51 state-level polls from one national sample. Just as the national poll is only a snapshot in time, so too is the MRP model limited to the present day and the present day only.

Expect more on this throughout the election year (including how we're going to use MRP estimates to improve our election forecasts beyond what other outlets can produce) but for now just read my piece outlining our findings:

economist.com/united-states/…

Final thoughts: Several people provided feedback on the formal Bayesian modeling work here, but none more than @cwarshaw. Thanks also to the many people on the Stan discourse forums who provided feedback on fitting sparse MRP models with MCMC.