Kappa統計量於量測一致性之應用與使用限制

Abstract

一致性研究不管在臨床研究或實驗設計中都有其日益著重的影響。而針對資料型態的屬性須選擇適合的一致性量測，若是資料屬離散型包含類別或二元情況，則可採用kappa統計量作為一致性的量測工具。文中探討資料類型為二元情況時，kappa的使用狀況，其抽樣分配的表現及kappa出現不完備之處，如兩種悖論和發生負值的情形。另外，本文也針對kappa的計算提出使用廣義線性模式及貝氏統計的作法，並比較各自適用的情況。最後以牙科根管治療研究中，X光片判讀之一致性研究為範例說明，並採用廣義線性模式方法估計其一致性係數值。

The study about “agreement” has played an important role in clinical studies and experimental designs. Researchers often need to choose a suitable measurement based on the data type. For instance, when dealing with discrete data, kappa is a popular index for evaluating the degree of agreement. In this thesis, we focus on the dichotomous data, and investigate the sampling distribution, performance, and paradoxes of the kappa statistic. When covariates are involved, we use Generalized Estimating Equations to estimate kappa. When negative value of kappa is observed, Bayesian procedure is recommended. Finally, a dental study is considered for illustration.