MODEL ASSISTED DESIGN WEIGHT CALIBRATION BY OUTLYINGNESS

DOI Web Site

Bibliographic Information

Other Title
  • ロバスト回帰を利用した乗率の調整
  • ロバスト カイキ オ リヨウ シタ ジョウリツ ノ チョウセイ

Search this article

Description

In this paper, we propose a new design weight calibration method in sample surveys based on robust regression. Classical survey statistics may adopt the Horvitz-Thompson (HT) estimator to have finite population quantities such as mean and total. Design weights used for the estimator are the inverse of inclusion probabilities. They may cause a problem when there is any extreme value with large design weight.<br>　The proposed calibration method utilizes a regression model explaining the target variable, and then estimates the parameters by any robust regression method which weight each record based on its outlyingness. We chose M-estimators and GM-estimators for evaluating accuracy improvement regarding finite population estimation. The weights derived from robust regression indicate outlyingness of each record. Calibrating design weights with those robust regression weights yields a new concept of survey weights considering both the sample design and outlyingness. We made Monte Carlo simulation with random number datasets and real datasets. The results are favorable and the weight calibration using M-estimators provides more efficient estimates for the finite population mean than that of GM-estimators.

Journal

Keywords

Details 詳細情報について

Report a problem

Back to top