Clustered experimental designs are used in primary prevention trials. Their use is often essential, since the treatments may be defined and applied to a unit of intervention (such as a community, a hospital cohort, a class, a work site, or a physician's practice) that consists of an intact group of individuals. The designs are also practical, since primary prevention often requires promulgation of information (using mass media) and education (in classes), or behavioral modification. Such trials require proper analysis to account for the clustering. Typically, when appropriate analyses are compared between a clustered randomized trial and a completely randomized trial with the same number of elementary units (i.e., patients or workers), a clustered randomized trial will have lower power and produce larger standard errors of estimates. As a result, cluster randomized trials require large data collection efforts, while providing disappointing precision of intervention effects, relative to the level of effort. This research will develop methods for achieving gains in power by reducing the intra-class (cluster) correlation through control of covariables measured on subjects in clusters. The gains will result in tighter confidence intervals in existing trials, and make possible the design of smaller studies without loss of power. Specifically, this research will: 1. develop and extend analysis methods that connect sampling to mixed models and to unequal probability sampling and two stage cluster sampling; 2. extend these results to cluster randomized experimental design with unequal numbers of secondary sampling unit measures on clusters; and 3. develop and extend these results to settings where the variance parameters are estimated from the data. The results will be applied to analyze two large scale trials and applied to account for increases in power achieved in the design of new trials.
| Li, Wenjun; Stanek 3rd, Edward J; Singer, Julio M (2012) Design-based random permutation models with auxiliary information(ΒΆ) Statistics (Ber) 46:663-671 |
