Premium
Applications: Oversampling Through Households or Other Clusters: Comparisons of Methods for Weighting the Oversampled Elements
Author(s) -
Wells J.
Publication year - 1998
Publication title -
australian and new zealand journal of statistics
Language(s) - English
Resource type - Journals
SCImago Journal Rank - 0.434
H-Index - 41
eISSN - 1467-842X
pISSN - 1369-1473
DOI - 10.1111/1467-842x.00031
Subject(s) - oversampling , weighting , statistics , mathematics , estimator , variance (accounting) , sample (material) , sample size determination , yield (engineering) , sampling (signal processing) , bootstrapping (finance) , population , group (periodic table) , econometrics , computer science , demography , economics , telecommunications , medicine , chemistry , materials science , accounting , chromatography , organic chemistry , detector , sociology , metallurgy , bandwidth (computing) , radiology
If a subgroup of a population is of particular interest in a survey, researchers may wish to increase the yield of this special subgroup by oversampling. One procedure for oversampling through households, or other clusters, is to divide the households into two segments: the main sample and the oversample (for which only members of the special group are eligible). Members of the oversampled special group come from both segments. This paper describes three methods for weighting the members of the special group. The household method treats the segments as strata and weights according to the proportion of households in each segment. The yield method uses weights according to the yield of special‐group members in the two segments. The combined probability method provides a Horvitz‐Thompson estimator using the sum of the probabilities that a person will be selected through either segment. Simulations show that the yield method produces estimates with variance lower than those of the household method. The combined probability method appears to be even more efficient. The difference in precision between the methods is small for estimates from the total sample but the household method can be markedly worse than the other two methods for estimates from the oversampled special group (over 40% greater variance in one scenario). Results from a community sample illustrate the comparisons. Because the household method can be much less efficient it should not be used.