ABSTRACT
Shoukri1 presents several other examples where the observations are correlated. One very interesting one is from an eye study, where there are two ophthalmologists examining both eyes of a patient for evidence of geographic atrophy, where 0 is scored for absence of the pathology and 1 denotes evidence of atrophy. One would expect the scores from the two eyes to be correlated and this correlation should be taken into account when estimating agreement between the two raters. This study had been analyzed by Oden2, Schouten3, and Donner, Shoukri, Klar, and Baretfay4. The Oden approach is to estimate the agreement between the two raters by kappa, separately for the two eyes, then combine the two kappas by a weighted average, while Schouten also used a weighted kappa to estimate overall agreement. The methodology of Donner et al. is quite different and is based on a generalized correlation model, a generalization of the common correlation model presented in Section 2.8. The main focus of the Donner et al. approach is to compare the two kappas and two see if they are the same for each eye.
