Design of Multiple Binary Outcome Studies with Intentionally Missing Data
We discuss the design and analysis of studies involving multiple binary outcomes in which only a subset of these outcomes can be measured on each individual. Such studies with 'intentionally missing data' may arise due to practical or economic constraints; several examples from toxicology serve as illustrations. A global test statistic based on generalized estimating equations is presented and evaluated under a variety of missingness patterns and correlation structures. Extensions of the global test statistic to allow for clustered data are also described. The relative efficiency of the global test statistic with missing data relative to that for complete data is investigated, both under a common dose effect alternative and when exposure has differential effects on the multiple endpoints. The implications of these efficiency calculations on study design are explored, and several recommendations are provided
Williams, P., & Ryan, L. (1996). Design of Multiple Binary Outcome Studies with Intentionally Missing Data. Biometrics, 52(4), 1498-1514.