What would it take to turn redblue states bluered?

How swings in party preference and turnout for key demographic groups could affect the 2024 presidential election.

By G. Elliott Morris, Amina Brown, and Aaron Bycoffe

Footnotes:
*Shifts in vote margin among nonwhite voters in this scenario include the following swings in each racial group: Asian American and Pacific Islander, R+2; Black, R+4; Hispanic, R+5; and Other (non-Hispanic), R+2. These numbers add up to an overall swing of R+3 among all nonwhite voters, and reflect trends in recent polls that show Biden losing more ground with Hispanic and Black voters than other racial groups.
**All racial groups other than Hispanic include only non-Hispanic voters.

Methodology: The Swing-O-Matic combines state-by-state demographic characteristics and population sizes with estimated voter turnout to predict the outcomes of the 2024 Electoral College and national popular vote, using three values to predict voting behavior of various demographic groups: (1) the number of people who share each set of demographic and geographic traits, (2) the probability that a person voted in the 2020 election and (3) the probability that they voted for Joe Biden, Donald Trump or an independent candidate for the 2020 presidency, if they were a voter.

To get state demographic characteristics and population sizes for adult U.S. citizens, we used the 2018-22 American Community Survey, downloaded from IPUMS at the University of Minnesota. For each state where respondents reside, IPUMS reports the population counts for each group of demographic variables, including citizenship status, age (which we combined into age categories), educational attainment, race, ethnicity, sex and total household income. (We imputed any missing data on total household income using a procedure called multiple imputation with chained equations, which predicts missing values based on age category, race and education attainment.)

To estimate the share of adult citizens in each demographic group that voted in 2020, we used voter-validated, post-election polling data from the 2020 Cooperative Election Study. The CES includes polling and voting data from 51,550 respondents, including their demographic traits, whether they voted in the general election and who they voted for. We used the CES data to train two statistical models: The first is a Bayesian multilevel logistic regression model that predicts the probability that someone voted in 2020 as a function of their demographic and geographic profile. This model accounts for the demographics mentioned above (age, race/ethnicity, education, sex, income and state of residence) as well as other auxiliary information, such as the median income in each state and the share of voting-eligible population that turned out in 2016. We repeat this process with a second model (Bayesian multilevel categorical regression model) to predict 2024 vote choice, using only the CES data from 35,549 validated (registered) voters who cast votes in 2020.

We then multiply these three values (population size, estimated 2020 turnout probability and predicted 2020 voting behavior) together to estimate the number of voters for Biden, Trump or another candidate in each demographic and geographic group. Then, we calculated the sum of that estimated number of voters to get an estimated total number of popular votes for each candidate at the state and national level. To make sure the estimates from our model match reality, we force the aggregated vote estimates to match the actual election results in each state using a method called the logit shift. We produced estimates for each congressional district in Maine and Nebraska, which split their electoral votes between the winner of the statewide and district-level popular vote, based on how much more favorable each district was to Biden or Trump than the nation as a whole in 2020.

The initial estimates in the Swing-O-Matic’s starting scenario reflect our best guess at the Electoral College and popular vote outcomes we would see in 2024 if every voter matches their predicted behavior from the 2020 election, accounting for changes in each state’s demographics and population, and the number of electoral votes awarded to each state, after the 2020 U.S. decennial census. Of course, that’s not likely to happen, which is where you, dear reader, come in. When you swing a certain group up 5 points for Democrats or down 5 points in turnout, for example, we adjust the Swing-O-Matic’s predicted group-level probabilities so they match the settings from your simulation to reflect the resulting outcomes in each state. Adjustments to multiple groups are run iteratively and additively, such that if any variable is adjusted, then further adjustments would be applied on top of that newly adjusted value.