|
Abstract:
|
The Survey of Income and Program Participation, conducted by the U.S. Census Bureau, is a longitudinal survey plagued by severe cases of missing data. So, the cross-sectional Horvitz-Thompson estimator for subdomain totals cannot be computed, much less its variance. One solution to this problem is to prorate partial subdomain estimators from previous cross-sections, in lieu of the missing data. The result is a "backdated" HT subdomain estimator. We use MLE's of transition probabilities (Bishop Fienberg Holland 1975 page 267) to prorate. We compute the MLE's through poststratification (Little 2004, 1993). The variance formula combines large sample approximations and BRR. This approach takes advantage of the survey design and the power of the poststratification. A careful diagnosis is conducted to validate the model.
|
