日本評価研究5巻1号 The Japanese Journal of Evaluation Studies Vol. 5, No. 1

【研究論文】 定量化された評定データのための分析手法の研究 −個人差による影響を受けない測定モデルの開発− 《PDFで全文ダウンロード　1.3MB》 大久保　智哉　　中川　正宣　　牟田　博光　　前川　眞一 東京工業大学　大学院社会理工学研究科 要　約 　現在、内部評価報告書を、外部有識者などの第三者によってさらに評価するような2次評価がおこなわれるようになってきている。これら2次評価では、各報告書について評価項目ごとに評定者が評定付け（rating）をして、定量化した上で分析・検討されることが多い。しかし、このように定量化された指標は客観的でなければならないにもかかわらず、実際は評定者によって「甘い・厳しい」の個人差があるという指摘がなされている。そこで、本研究では評定データにおいて定量化された指標の客観性を向上させるために、評定の個人差を除去した指標を算出するため因子分析モデルを導入し、より客観性の高い測定モデルを提案した。また、特殊な欠損パタンにおいても因子分析モデルを適用できることを示し、実際場面での評定データの分析の幅が広がることを示した。また、それらの精度等についても検討を加え、有用な測定モデル・デザインについて提案した。 キーワード 評定データ、組織的欠損、評定の個人差、因子分析モデル、欠損パタン

【Article】 Statics Analysis for Rating Data ― Suggesting Statistical Models to Elimiinate the Personal Differences in Ratings ― Tomoya Okubo Masanori Nakagawa Hiromitsu Muta Shin-ichi Mayekawa Tokyo Institute of Technology. Graduate School of Decision Science and Technology Abstract There have been reports suggesting that defferences occur in ratings for each individual. Therefore the grading figures show the merits of the subjects of the evaluations, as well as the influences by the personal defferences in ratings. When it comes to rating data, the true rating scores can be expressed by showing the personal defferences with the factor analysis. However, reality is that ratings often cannot be done alone, and in these cases, systematic missing patterns would be included to the analyses. Under the data in which systematic missing patterns occur, however, the usual factor analysis cannot be utilized because the rating scores cannot be shown on common scales. Under these circumstances, the suggestion made by Allison(1987) could be applied to show the factor scores on common scales. With this in mind, the purpose of this research lies in suggesting effective rating designs and rating models for the rating data which include numbers of systematic missing patterns. As a result of this research, the following two points were proven to be effective:1.) speaking of rating design, increasing the ratio of the common raters, and2.) placing raters who well describe the tendency of the whole as the common raters. Also, as a result of comparing the rating models using AIC, models that acknowledged the personal defferences with both the rating difficulty and the rating discrimiration were proven to be the most appropriate. As shown in this research, efficient rating designs and highly reasonable rating models were examined and taken under consideration. Keywords rating data, systimatic missing, personal differences in ratings, factor analytic model, missing patterns