Evaluating the Small-Sample Bias of the Delete-a-Group Jackknife for Model Analyses

Abstract

The delete-a-group jackknife can be effectively used when estimating the variances of statistics based on a large sample. The theory supporting its use is asymptotic, however. Consequently, analysts have questioned its effectiveness when estimating parameters for a small domain computed using only a fraction of the large sample at hand. We investigate this issue empirically by focusing on heavily poststratified estimators for a population mean and a simple regression coefficient, where the poststratification takes place at the full-sample level. Samples are chosen using differentially-weighted Poisson sampling. The bias and stability of a delete-a-group jackknife employing either 15 or 30 replicates are evaluated and compared with the behavior of linearization variance estimators