S17.2: Properties of statistical tests to compare incidence rates in cluster-randomised trials
✍ Scribed by Martina Kron; Josef Hoegel
- Book ID
- 101712321
- Publisher
- John Wiley and Sons
- Year
- 2004
- Tongue
- English
- Weight
- 76 KB
- Volume
- 46
- Category
- Article
- ISSN
- 0323-3847
No coin nor oath required. For personal study only.
✦ Synopsis
In cluster-randomised trials, hypothesis testing methods allowing for correlation of subjects' outcomes within the same cluster are necessary to display the structure of the data adequately. Until now many procedures are proposed and evaluated for continuous and binary outcome variables (Donner, Klar, 2000). Only a few proposals for the comparison of incidence density rates between treatment groups exist whereby a comparison of their properties is lacking. Statistical tests proposed in the literature (see e.g. (Donner, Klar, 2000)), like the t-test, a ratio estimator approach, and a statistic based on a generalized estimating equations approach, will be compared with the unadjusted chi-square test and proposals for adjusted versions of the chi-square test. The test statistics will be adjusted by different estimators of the intraclass correlation. Three types of estimators will be used: an analysis of variance estimator, a Pearson estimator and an estimator derived from a proposal for analysing clustered Poisson data (Rao, Scott, 1999). The statistical tests are compared in a simulation study concerning their level of significance and their power whereby bivariate Poisson distributed variables are generated as described by Paul and Ho (Paul, Ho, 1989). The simulation study identifies powerful test statistics. Limitations, like anticonservatism or lack of power, of some statistics under certain circumstances are shown.