LVQ與多變數反覆加權法於測驗效度檢驗影響

Abstract

本研究目的探討學習向量量化網路（learning vector quantization， LVQ）與多變數反覆加權法（Raking）的權重校正對於測驗效度檢驗結果之影響。楊志堅、蔡良庭、楊志強（2009）以及Tsai與Yang（2012）指出，利用驗證性因素分析（confirmatory factor analysis， CFA）進行量表編製的效度檢驗時，若忽略權重的計算，將導致參數估計偏誤（bias）。然而上述相關的研究，都侷限於單一變數的分群權重插補。然實際的測驗調查中，背景變數往往同時考量兩個或兩個以上（例如：性別、種族），而導致整份測驗中的分群設計需同時考量多個變數。因此，本研究利用數值模擬方式，在考量兩個變數的分群時，探討LVQ權重校正及多變數反覆加權法對於測驗效度檢驗的影響。本研究設計兩個背景變數，每個背景變數各包含兩個分群，合計共有四個不同的群體。另外的實驗設計包含：取樣數大小、樣本的遺失比例、四個分群的異質性及取樣不均勻程度。研究顯示，當分群間的異質性愈大、且因遺失樣本而導致嚴重的取樣不均勻時，利用多變數反覆加權法的權重校正，將導致參數估算的正確率降低，而相對的，LVQ的權重校正則能提供穩定且正確的測驗效度檢驗結果。The study proposes learning vector quantization networks (LVQ) and Raking approaches of sample weightings on measurement validity. Ignoring sampling weights can lead to severe bias in parameter estimation of confirmatory factor analysis (CFA) (Tsai & Yang， 2012; Yang， Tsai， & Yang， 2009). This study extends Yang et al. (2009) to include multiple weighting factors (e.g.， gender and ethnic group) simultaneously in large validity research. The study design incorporated two auxiliary variables. Each auxiliary variable included two categories resulting in a total of four groups. Experimental factors， including missing proportions， sampling sizes， disproportionate， and heterogeneousity of groups， are designed to examine performances of LVQ weighting adjustment. Results show that accuracies and stabilities of LVQ are much better than the raking method as disproportionate sampling is severe， sampling size is 2，000， heterogeneousity of groups is 0.4， and missing rate is 20%. When a survey research has several auxiliary variables， LVQ weighting adjustment can remove the nonresponse bias when using CFA model to conduct measurement validity.