Analysis of dichotomous outcome data for community intervention studies

Abstract

Community intervention trials are becoming increasingly popular as a tool for evaluating the effectiveness of health education and intervention strategies. Typically, units such as households, schools, towns, counties, are randomized to receive either intervention or control, then outcomes are measured on individuals within each of the units of randomization. It is well recognized that the design and analysis of such studies must account for the clustering of subjects within the units of randomization. Furthermore, there are usually both subject level and cluster level covariates that must be considered in the modelling process. While suitable methods are available for continuous outcomes, data analysis is more complicated when dichotomous outcomes are measured on each subject. This paper will compare and contrast several of the available methods that can be applied in such settings, including random effects models, generalized estimating equations and methods based on the calculation of 'design effects', as implemented in the computer package SUDAAN. For completeness, the paper will also compare these methods of analysis with more simplistic approaches based on the summary statistics. All the methods will be applied to a case study based on an adolescent anti-smoking intervention in Australia. The paper concludes with some general discussion and recommendations for routine design and analysis.