比較入選機率分組與其他加權方法對電話調查樣本的調整：上網率的推估

A Comparison of Propensity Score Sub-classification and other Calibration Methods based on a Telephone Sample to Estimate Internet Usage

傳統的事後分層與多重反覆加權法可以降低抽樣誤差，但在改進推估偏差的效果仍然有限。本研究採用「入選機率調整法（propensity score adjustment, PSA）」，對2008年的電話訪問樣本進行18歲以上臺灣民眾之上網率的推估，進行上述兩種加權方法以及其納入PSA前後之估計效果的比較。在幾種入選機率值（propensity scores, PS）的調整方法中，本研究採用次樣本分組法，因此也同時比較四到十個分組之間估計效果的差異。本研究以同一時段蒐集的「臺灣地區社會變遷基本調查五期四次全球化組」為PSA的參考樣本；（前）行政院研究考核委員會執行的「數位落差調查」所推估的全臺灣民眾的上網率為擬母體黃金標準，亦即PSA與兩種傳統加權之估計效果的比較基準。結果證實納入原調查權數（來自事後分層與多重反覆加權）後，在五個次樣本分組時，上網率的估計值最接近數位落差調查的估計值。比起納入事後分層權數，納入多重反覆加權權數的PSA在估計誤差上相對較小。

While post-stratification and raking calibration methods can reduce sampling errors, they have estimation limitations. The author adopts propensity score adjustment (PSA) to estimate Internet usage based on data collected from a 2008 telephone sample. Comparisons were made among post-stratification, raking, post-stratification PSA, and raking PSA. Stratification was used to produce PS weights and to compare estimated Internet usage for seven sub-classifications. The Taiwanese Social Change Survey conducted during the same period was used as the reference sample for PSA, while official statistics based on a Digital Divide Survey as the benchmark for bias reduction comparisons. Results indicate that (a) Internet usage estimates based on PSA adjusted according to base weight (i.e., survey weight from post-stratification or raking) using five sub-classes were more accurate than other estimates, and (b) bias reduction based on PSA adjustedby raking exceeded that of PSA adjusted by post-stratification.