Estimating functions for evaluating treatment effects in cluster‐randomized longitudinal studies in the presence of drop‐out and non‐compliance

Abstract

We describe methods for analyzing longitudinal binary data from cluster‐randomized trials in which responses are incompletely observed and subjects may not be fully compliant with the prescribed treatment regimen. The method is based on a marginal regression model for the response where parameter estimates are obtained from generalized estimating equations. Estimating equations are also employed to estimate parameters of the missing data process which are used to compute inverse probability weights. A model is specified for the compliance process which facilitates estimating the expectation of the contributions to the estimating function for the response parameters among individuals without compliance data, which occurs when the control treatment involves no intervention. The approach is robust in the sense that semi‐parametric models are used for the response and the missing data processes and robust variance estimates are advocated. The proposed method is shown to perform well in simulation studies, and data from a randomized trial of patients with depression are analyzed for illustration. The Canadian Journal of Statistics 38: 232–255; 2010 © 2010 Statistical Society of Canada

Regression analysis of correlated ordinal data using orthogonalized residuals

“…Like alternating logistic regressions and common implementations of generalized estimating equations, the proposed method requires missing observations (such as person-visits in a longitudinal study) to be completely missing at random or covariate-dependent. Extensions of ORTH to missing at random (MAR) dropout may be possible along the lines of Yi et al (2010) using inverse probability weights under a model for MAR dropout.…”

“…First-order estimating equations using conditional residuals in a model for the association between clustered outcomes can be moderately efficient in relation to second-order equations with less burden computationally (Lipsitz and Fitzmaurice, 1996). In the case of multivariate binary data, alternating logistics regressions (ALR), which specify within-cluster associations in term of pairwise odds ratios (Carey, Zeger, and Diggle, 1993), have been applied in diverse settings, including studies of multilevel geographical clustering of drug (Petronis and Anthony, 2003) and alcohol use (Reboussin et al, 2012), surveillance of occupational illnesses with workplace clustering (Preisser, Arcury, and Quandt, 2003) and clusterrandomized trials with longitudinal data (Yi and Cook, 2002;Yi, Cook, and Chen, 2010). Heagerty and Zeger (1996) defined estimating equations for the associations of correlated ordinal data in an extension of Carey et al's method.…”