Online Program

Post-Stratification in Contexts Where Strata Population Counts Are Unavailable (304177)

Keywords: post-stratification; small area estimation; multilevel analysis; hierarchical analysis

Post-stratification (PS) is a useful technique for implementing small area estimation of multilevel outcomes and covariates. With PS, estimates made on strata of level 2 units are aggregated to level 1 units using population-weighted averages of strata estimates. Often, population data for PS are only partially available due to many level 2 strata. It is also unclear how to perform hypothesis testing to identify significantly different level 1 units from the overall estimate, accounting for PS and multiple comparisons. For example, a multilevel spatial logistic regression of childhood overweight/obesity (OO) provided county-level estimated OO rates for 256 strata of child-level variables (4 race/ethnicity [r/e], 2 gender, 8 age, and 4 parental education groups). Census population counts are only available for 64 r/e, gender, and age strata. We demonstrate a method utilizing Census and public data to obtain estimates of population prevalence for all strata. We discuss minimal assumptions required and show a bootstrap technique to test if county-level OO rates are significantly lower/higher than the national average, controlling for PS uncertainty and overall false discovery rate.