[Statistical analysis of community-based studies -- presentation and comparison of possible solutions with reference...

The statistical analysis of community-based trials and of other cluster-randomised trials, requires specific statistical methods. We show the consequences of the application of these models for study results, using data of the German Cardiovascular Prevention Study (GCP) as an example.

Data of 30,285 subjects were analysed, which were collected at the beginning and at the end of the study period. These data had been collected in 7 intervention regions and by national surveys. We grouped data of the national surveys in 7 control clusters to mimick a design typical for cluster-randomised trials. We applied the following statistical models to estimate the effect of the intervention on total cholesterol as well as on systolic blood pressure and the respective confidence intervals: a linear model, a mixed model, and fixed and random effects meta-analyses.

While the estimates and confidence intervals for the intervention effect were similar in mixed model analysis and random effects meta-analysis, results from models incorporating fixed effects only were anti-conservative. The underestimation of variance in models incorporating fixed effects only was especially large in the analysis of systolic blood pressure data, where great heterogeneity between intervention communities was observed. Despite seemingly low intraclass correlation coefficients of 0.0019 for total cholesterol and 0.0166 for systolic blood pressure, respectively, the variance of the intervention effect was increased in the mixed model 2.8fold or 17.1fold, respectively, in comparison to the variance estimated in the linear model. Due to this variance inflation the intervention effect on systolic blood pressure lost statistical significance.

Our results emphasise the importance to account for correlations in community-based trials. Besides the mixed model random effects meta-analysis can be applied as an alternative method.