Both traditional polling and our modeling approach rest on a central assumption: voters with similar demographic characteristics are sufficiently comparable for the purposes of estimation. This assumption is never perfectly true, but it becomes more plausible as the population is divided into increasingly granular groups. Our modeling approach allows us to segment the population into 13,824 distinct profiles — a level of granularity that cannot be achieved through traditional weighting.
Traditional polling generally weights respondents so the sample matches known characteristics of the population, such as age, race, education, gender, and geography. This helps correct observable imbalances in who was interviewed, but the estimate still depends heavily on the respondents observed within each subgroup. When a group contains only a small number of respondents, its estimate can be unstable or too imprecise to report. As a result, weighting alone generally cannot support reliable estimates for very small demographic groups or geographic areas without an exceptionally large sample.
Our methodology combines predictive modeling with poststratification. Instead of treating each demographic and geographic group as a separate miniature poll, the model learns relationships across the full sample. Information from similar voters and similar places helps estimate groups that may have few respondents in any one survey. For example, evidence from Hispanic women in their twenties contributes to estimates for similar voters across the country, while geographic and political context allows those estimates to differ from place to place. This makes it possible to produce estimates by state, district, county, demographic group, and combinations of those characteristics at a level of granularity that would not normally be reliable through weighting alone.
A second modeling stage then stabilizes these estimates over time, connecting each new survey with the evidence accumulated from earlier surveys rather than treating each week in isolation — described in full in Modeling Change Over Time, below.
We also collect responses through multiple channels, including online panels, SMS, email, and our internal panels. This broadens the range of voters we can reach and reduces dependence on any single method of survey recruitment. It does not eliminate selection bias by itself, but it gives the model a broader and more varied evidence base.
The result is a system designed to produce stable, detailed, and frequently updated estimates for populations that conventional weighted polling often cannot measure reliably on its own.
