Comparision Of Statistics Methods For Analysis Of Multistage Clustered Randomized Data

Objectives: To compare the performance of three statistical methods, namely probit regression, generalized estimating equations and probit-normal model, for analysis of two-stage clustered randomized data and the performance of logistics regression, generalized estimating equations and logistics-normal model, for analysis of three-stage clustered randomized data.Method: Simulating two-stage clustered randomized data according to a pupil based somatic study to compare the bias of estimates and coverage of 95% confidence intervals under different values of the regression coefficient parameters and intraclass correlation coefficient. On the basis of two-stage simulating result, establishing unbalanced three-stage clustered randomized model to compare the bias of estimates and coverage of 95% confidence intervals under different values of the regression coefficient parameters and variances of random effects.Main Result: In the two-stage clustered randomized data analyzing, probit-normal model gives the least bias and the best coverage under each parameter combination. GEE gives a consistent estimation, its statistical ability .is inferior to probit-normal model but superior to probit regression. The values of coefficient parameters don't impact the result of analysis. In the three-stage clustered randomized data analyzing, logistic-normal model can give a precise estimation, but the coverage of school effect is poorer than it of class effect. The coverage of GEE on class effect is credible and logistics can not give a credible estimation on cluster effects. The true value of the estimating parameter influences the result of estimation. The bigger the true value is, the smaller the bias will be. The impact of the bias base on the different values of variances is not the same in these three models.