ACP, SSRS, Elliott Morris Methodology Fall 2026

These geographic estimates come from the MSU American Communities Project omnibus, fielded by the SSRS Opinion Panel from September 3–21, 2026 (n = 2,155 adults). Morris Predictive Insights turned them into county-level estimates using a multilevel regression and poststratification model.

First, we turned each of 47 survey items into a binary yes/no outcome. For agree/disagree questions, “strongly” or “somewhat” agree counts as yes (1), and for the local-issue and news-source questions, choosing that option counts as yes. Other responses on these questions (such as disagreeing, or as not picking a certain item in a list) counted as a no (and a 0).

We then fit a separate Bayesian multilevel logistic regression for each item in Stan, using the brms package in R. Each model predicts the chance that a respondent answers yes using covariates including their sex, age, race and ethnicity, education, household income, and party identification. The model also sees each respondent’s county, including its typology from the American Communities Project (ACP), its region in the country, the state, as well as the county’s Hispanic and college-graduate shares from the Census Bureau’s American Community Survey (ACS). The model sees main effects for each demographic and geographic term, plus interactions of race and county type, race and sex, race and education, and party with each demographic. The model uses partial pooling informed by group sample sizes, so groups with few respondents borrow strength from similar groups instead of resting on a handful of interviews.

The population poststratification frame is built in a three-stage model from the 2019–2023 ACS five-year microdata (via IPUMS). Adult citizens are counted by Census PUMA (areas of at least 100,000 people) and by sex, age, race, education, and income, with missing answers on income imputed via multiple imputation and chained equations. In a first stage, we predict voter turnout using the Census Bureau’s Current Population Survey from 2024. We then predict vote choice, in a second stage, using the 2024 Cooperative Election Study (CES), whose respondents’ turnout is checked against voter files. Predictions from these models are adjusted to reproduce official 2024 turnout and results in every state and, where applicable, county.

Third, we attached leaned party identification to the post-stratification frame using a model fit to Pew Research Center’s address-recruited surveys: the 2023–24 Religious Landscape Study and the 2020–2026 National Public Opinion Reference Surveys (NPORS). The model uses the correlation between demographics, including professed past vote, at the party level, and projects that onto the post-stratification frame. We then adjust the party predictions for predictable differences in past vote by state and PUMA, thereby adjusting post-hoc the party data to be representative of multiple geographies — a solution we are forced to use given Pew does not
make a version of its NPORS data with geographic identifiers available to researchers. The national party mix is then adjusted to Pew’s 2024–2026 benchmark without changing any cell’s size.

To aggregate up (and sometimes down) from PUMA to county, we use the Census Bureau’s block population file. PUMAs and counties are both made of whole census tracts, so each PUMA’s cells are split among the counties it touches in proportion to 2020 Census voting-age population; the total number of adults doesn’t change. For this survey we drop past vote, which the omnibus didn’t ask, and add noncitizen adults from the same ACS sample, giving each the party mix of non-voting citizens like them in their PUMA.. The result covers all 258.5 million U.S. adults in 3.3 million county-by-demographic cells.

To produce estimates, we predict every cell for each of 500 posterior draws from our MRP model, weight each cell by its population and sum to counties and ACP types. The median across draws is the estimate, and the 5th and 95th percentiles give the 90% range. All models converged (R-hat ≤ 1.01). Summed to the nation, the MRP estimates fall within two standard errors of the weighted survey topline for all 47 items, with a mean absolute difference of 0.5 points. For demographic groups the gap averages 0.7 to 2.0 points, and for ACP types 2.6 points. Dropping the state or party terms moved national estimates by at most 0.2 points and county estimates by about 1 point on average.

As with any MRP analysis, this one comes with caveats about the assumptions needed to predict behaviors among people not sampled. In total 2,277 of 3,143 counties in the U.S. have no respondents. Our model’s estimates therefore are what the model expects for a county with that population, type and state. Additionally, where one PUMA spans several small counties, each county gets the PUMA’s average estimate, which may mask U.S. geographic differences.

ACP-type estimates are far more precise than single counties and are the better unit for comparisons.