摘要:This paper examines the chain of weights, beginning with the basic sampling weights for the respondents. These were then converted to reweights to reduce the bias due to missing quantities. If micro auxiliary variables are available for a gross sample, we suggest taking advantage first of the response propensity weights, and then of the calibrated weights with macro (aggregate) auxiliary variables. We also examined the calibration methodology that starts from the basic weights. Simulated data based on a real survey were used for comparison. The sampling design used was stratified simple random sampling, but the same methodology works for multi-stage sampling as well. Eight indicators were examined and estimated. We found differences in the performance of the reweighting methods. However, the main conclusion was that the response propensity weights are the best starting weights for calibration, since the auxiliary variables can be more completely exploited in this case. We also tested problems of calibration methods, since some weights may lead to unacceptable weights, such as below 1 or even negative.
关键词:reweighting;simulation study;macro vs. micro auxiliary variables;case of negative and other implausible weights