The comparison and extension of weighted-least-squares and generalized estimating equations approaches for dependent kappa coefficients

• Main Authors: Jung-Feng Wang, 王榮蜂
• Other Authors: Miao-Yu Tsai
• Format: Others
• Language: zh-TW
• Published: 2009
• Online Access: http://ndltd.ncl.edu.tw/handle/86884922673855432398

碩士 === 國立彰化師範大學 === 統計資訊研究所 === 97 === Many studies in medical, industry and business fields commonly use questionnaires and diagnoses reported by different raters or instruments to evaluate the agreement of multiple ratings. The most popular indices of agreement for a categorical response are the kappa measures, Cohen’s kappa for binary outcomes and weighted kappa for ordinal outcomes (Cohen 1960, 1968). However, when raters assess the same observation on two or more occasions, these ratings are dependent, and, therefore one must take into account the correlation between the kappa estimates for making inference. In this thesis, we focus on estimating kappa measures on inter and intra-agreement using the weighted least squares (WLS) approach for dependent categorical data and testing the equality of the dependent kappa coefficients by using PROC CATMOD in SAS. We modified the SAS code in Barnhart and Williamson (2002) to analyze dependent agreement data with more than two raters and to use different numbers of categories. In addition, it will be compared with the results using generalized estimating equation (GEE) for Cohen’s kappa and weighted kappa. Simulation studies are conducted to evaluate the performance under WLS and GEE approaches. Four applications of repeated ratings by multiple raters are illustrated, including similarity in myopic status for twin pairs with respect to zygosity, conformity between an optometric device and ophthomologist in diagnoses of high myopia, test-retest radiographic examinations by seven endodontists, and comparison between a computerized planimetry method and direct visual assessment for assessing cervical ectopy. The results reveal that the performance of WLS is better than that of GEE.