Multilevel regression and poststratification (MRP)

Multilevel regression and poststratification estimates public opinion or vote choice within small subgroups, a state, a district, or an age-by-education cell, from a survey that is too small to be reliable in each subgroup on its own, by pooling information across groups.

The method has two stages. First, a multilevel, or hierarchical, regression model predicts each respondent's outcome from demographic and geographic characteristics, letting the model borrow statistical strength across similar groups instead of treating each subgroup in isolation. Second, poststratification reweights the model's predictions for every demographic-geographic cell using known population totals, typically from census data down to the census tract level, then aggregates those weighted predictions into an estimate for each geography. MRP is one specific technique within the broader family of small area estimation methods, and it differs from plain survey weighting, which adjusts a single national estimate rather than modeling and reconstructing many local ones.

MRP lets one national survey produce credible state- or district-level estimates without polling each area separately, and it is widely used in election forecasting and policy polling, often alongside poll aggregation and a likely voter model. Its main limitation is that estimates for very small or atypical cells still depend heavily on the model's assumptions, and overall quality is capped by how accurately the poststratification totals reflect the true population.