Small area estimation (SAE)

Small area estimation is a family of statistical methods for producing usable estimates, a poverty rate, a disease prevalence, an unemployment figure, for geographic areas or demographic subgroups where the available survey sample is too small, or nonexistent, to produce a direct estimate with acceptable precision.

Rather than relying only on the small or absent sample in the target area, these models "borrow strength" from related areas or from auxiliary data available everywhere, administrative records, census counts, satellite imagery, to predict the value in the small area indirectly. multilevel regression and poststratification is one widely used small area estimation technique, particularly in polling; statistical agencies also use model-based composite estimators for the same purpose. This differs from simply computing a raw average from a tiny sample, which would swing wildly from one area to the next.

Small area estimation lets agencies publish figures, such as county-level poverty rates used to allocate federal funding, in areas the underlying survey, an American Community Survey sample or a national poll, did not collect enough responses in to measure directly, often anchored to census tract geography. The pitfall is that because these estimates come from a model rather than a direct count, they can be systematically off wherever the model's assumptions do not hold, particularly in atypical small areas.