雙底冊電話調查組合設計權數之應用

Application of Weighting Strategies for the Combination of Data from a Dual frame Telephone Survey

隨著手機的普及與生活形態的改變，以市內電話為主所進行的電話訪問面臨涵蓋率及代表性的問題，進而影響資料品質；因此近年來多採用雙底冊抽樣設計的方式，同時納入手機與市話的訪問對象。然而，雙底冊的抽樣設計會因手機樣本與市話樣本的重疊性或組合差異，使得資料必須以合理的加權方式處理後，才可進行合併分析。
過去國內外的相關研究已嘗試數種不同的資料加權處理程序，希冀對雙底冊電話調查組合獲取較佳之估計加權方式；由於各國影響電話調查成本之因素不一，在既有成本的條件下，國內研究採用的加權方式主要考量不同電話使用族群（包含「唯市話族」、「唯手機族」及「市話與手機者」等三類對象）所占的比例，進行事後分層加權。本研究以兼具全國與縣市代表性的雙底冊電話調查資料，以拔靴法(bootstrapping)重複選取樣本，來評估不同雙底冊電話調查組合加權程序的估計結果。拔靴法的估計獲得偏低的誤差值，顯示兩種程序的穩健性，本文並對於相關之研究限制提出討論。

With the widespread use of cell phones and changes in life style, conventional telephone surveys, which are conducted using mainly household landline numbers, have encountered the problems of coverage and representativeness. As the use of landline numbers cannot reach people who only use cell phones, data quality of telephone sur¬veys is further affected. Dual-frame telephone surveys have been applied in recent years to include both landlines and cell-phone users to reduce undercoverage bias due to the incompleteness of the land¬line frame. With a fixed budget, previous studies have suggested adopting an overlapping sampling design for dual-frame surveys. The weighting procedures and the combination of data from dual frame surveys, which can be seen as unequal weighting methods, therefore, become a challenging issue. In particular, data combination for those who use both landlines and cell-phones in the overlapped sampling domain has drawn researchers’attention. As sample characteristics derived from cell phone and landline samples differ in many aspects, it is important to employ suitable weighting strategies for combining data from dual-frame telephone surveys.
Previous studies have developed various weighting approaches, with the consideration of non-sampling or nonresponse errors. Due to the absence of external population estimates for telephone usage, several weighting strategies are not suitable to be applied to dual-frame telephone surveys in Taiwan. This study aims to examine dif¬ferent procedures for incorporating unequal weights for dual-frame telephone survey using data from the 2020 Adult Smoking Behavior Survey (ASBS). Two types of weighting procedure, which are modi¬fied based on previously developed approaches, are examined. One is to apply weights to landline and cell-phone samples separately before the combination of the dual samples. The other is to combine the dual samples first with the consideration of sample characteristics in the cross-tabulation of age and gender, education, and geographic areas. During the weighting procedures, the proportion of different tele¬phone users, including“landline only”,“cell phone only”and“dual users”, derived from the landline and cell-phone samples, is used to adjust the combined data, and final weights are applied using post-stratification raking. We use bootstrapping to evaluate the robustness of sample estimates, with different sample allocations of the ran¬domly selected resamples from the 2020 ASBS. Auxiliary informa¬tion of labor participation rate and the distributions of ethnicity groups in Taiwan is used as bench mark to assess the bootstrapping estimates.
A total of 26,065 complete landline cases is obtained with a response rate (AAPOR RR1) of 18.3%. For cell-phone sample, a total of 4,299 complete cases is obtained, with a response rate of 10.8%. Consistent with previous studies, the unweighted cell-phone sample included higher proportions of younger respondents and those with a college degree, while more females and the elderly responded to the landline survey. The proportions of different telephone users, i.e. landline-only, dual users, and cell phone only, slightly differed between the two weighting strategies. The bootstrapping results indi¬cated that sample estimates derived from the first weighting strategy performed somewhat better than the second one. Although an optimal weighting approach for the data combination of dual-frame telephone surveys is not available, it is important to obtain relatively low stan¬dard errors for sample estimates. We discuss the limitations of this study and call for more research on this issue.