建立樣本流失檢視流程：以臺灣幼兒發展調查資料庫為例

Establish an Inspection Procedure of Sample Attrition: A Case Study of the KIT in Taiwan

一直以來，長期追蹤調查無可避免地會有樣本流失的問題，且多數文獻均指出樣本流失並非完全隨機缺失。不過，若與樣本流失有關的變數和追蹤資料使用者所關注的變數沒有相關，那麼這類隨機缺失的資料不會導致分析結果的偏誤。由於臺灣目前沒有一般化的檢視樣本流失程序可供學術研究參考使用，故本研究以「臺灣幼兒發展調查資料庫」為例，建立中選樣本與受訪者不為同一人時的檢視流程，以期能適用於不同的追蹤調查資料。本研究的分析結果顯示，在控制其他變數的情況下，每波均受訪的樣本與曾經流失過的樣本，在幼兒發展領域中的記憶力、注意力╱執行功能、讀寫萌發等變數上沒有差異，以「臺灣幼兒發展調查資料庫」目前四波的平衡追蹤資料進行分析，應不會有樣本流失偏誤的產生。不過，若檢測樣本流失與關注的變項有關時，建議研究者可考慮以傾向分數倒數加權法進行調整。

Sample attrition is inevitable in long-term tracking surveys and most studies indicate that sample attrition is not missing completely at random. However, if the variables associated with sample attrition are not correlated with the variables of the researcher’s interest, then such data missing at random will not lead to analysis bias. Since there is no generalized inspection procedure of sample attrition for academic research in Taiwan, this study takes the“Kids in Taiwan: National Longitudinal Study of Child Development and Care (KIT)”as an example to establish the examination process when the selected sample and the respondents are not the same and to apply to different kind panel survey data. The findings show that, holding other variables constant, there were no differences between the attrition and non-attrition samples in terms of memory, attention/executive function, and emergent literacy. Based on balanced panel data from the KIT, the analysis results should have no sample attrition bias. If the sample attrition is related to the variables of the researcher’s interest, we suggest that the researcher could adjust it through inverse probability weighting.