探討二元資料組內相關不同估計方法之效率

Abstract

Binary data occur frequently in research of epidemiology, sociology, biology and so forth, such as studies of familial diseases and repeated measurement studies and like versa. In these conditions, there is a correlation between these observations. The intraclass correlation coefficient (ICC) is a quantitative measure of the resemblance among observations within clusters. At least 20 different point estimators of the ICC have been proposed in various areas of application. However, techniques for binary data have been less developed. And the efficiency and application of these techniques are differ from each other. The main purposes of this research aimed to estimate the intraclass correlation coefficient and to evaluate the efficiency of the three ICC methods for binary data proposed by Ridout et al. (1999). Base on common correlation model, computer simulation is applied to compare these three methods in this research. All the simulations in this article are implemented by using the program language of Visual Basic. According to the changes of the three parameters (i.e. intraclass correlation coefficient ρ, the probability of success π, and number of clusters k), the performances of each method are discussed under different parameter combinations. Simulation results indicated that the coverage of confidence interval obtained from ANOVA method performed well only when ρ is small. However, it is not stable under other parameter combinations. On the other hand, the coverage of confidence interval based on the Pearson and the FC estimates for ρ, showed highly stable under any combinations of parameter. In general, the coverage levels of the three methods are low when k is small. Even k is increasing, the coverage rates perform not ideally when the ρ is extreme (ρ is close to 0 or 1). Thus, research requires these three methods to apply in certain particular situations, such as sample size is small or extremely high or low intraclass correlation, the correlation coefficients and confidence intervals estimated by these methods should not be credible.