|
Abstract:
|
The American Community Survey (ACS) publishes annual poverty estimates for large counties. In response to the demand for poverty estimates for U.S. counties and school districts, the Census Bureau's Small Area Income and Poverty Estimates (SAIPE) program produces poverty estimates for all U.S. counties and school districts from area-level Fay-Herriot models. This paper focuses on estimating county-level poverty rates for the year 2014, using the unit-level logistic mixed effects model. We fit an unweighted multilevel logistic regression (MLR) model with demographic predictors and state- and county-level random effects. To account for the design of the ACS survey, we consider models with weights scaled based not only on the person sample size (PSS), but also on the housing unit sample size (HUSS). Given the individual demographic characteristics, we estimate the predicted probability that an individual is in poverty. An aggregation within county is made to generate the corresponding county-level poverty rates using the U.S. Census Bureau postcensal population estimates. In comparing ACS direct estimates, SAIPE estimates, and the estimates from the MLR models, the distributions of all three estimates are similar for large counties. For small size counties, the distributions of the estimates produced via MLR models tend to have smaller ranges compared with the distribution of ACS direct estimates. MLR models with weights scaled based on the person sample size almost always yield estimates with smaller mean absolute differences with the ACS.
|
