Bayesian Hierarchical Model For The Study Of Clustered Data With Cluster Level Sources Of Measurement

Download or read book Bayesian Hierarchical Model for the Study of Clustered Data with Cluster-level Sources of Measurement written by Maria Esther Perez Trejo and published by . This book was released on 2015 with total page pages. Available in PDF, EPUB and Kindle. Book excerpt: "Cluster randomized trials are an extensively used tool in health research; their main pitfall arises from within cluster correlation of responses, which prevents the use of traditional methods of design and analysis for individual randomized trials. This issue is overcome at the design stage by adjusting the sample size according to the magnitude of the intraclass correlation coefficient, and at the analysis stage by modeling clustered responses through linear and generalized linear mixed models that include random intercepts accounting for the variability due to clustering. However, there is a disadvantage that is usually overlooked when studying cluster trials, that is, when the measurement of the outcomes is naturally clustered, known as clustered measurement. For outcomes susceptible to non-random error, clustering among individuals measured by the same observer will occur if some observers tend to measure systematically differently than others. A particularly troublesome situation arises when the measurements are clustered within the same groups used as units of randomization, which we call "double clustering". The effect of both measurement issues is to increase the intraclass correlation coefficient due to the introduction of a second source of variability (due to observer) at a cluster level. None of these issues can be addressed at the design stage, and when a second set of random intercepts accounting for variation due to observer is introduced in the models at the analysis stage, the variances at cluster level cannot be estimated separately under double clustering. The Promotion of Breastfeeding Intervention Trial (PROBIT) revealed that double clustering is likely to be present in the measurement of several of its outcomes and, in fact this cluster randomized trial motivated the present research. A strategy to deal with these issues is to conduct a second set of measurements that allows us to obtain separate variance estimates and, in turn, to improve the estimation of the treatment effects; such measurements were taken in PROBIT for other reasons. We present a Bayesian hierarchical model for clustered-measured, and more emphatically for double-clustered outcomes, in which audited measurements are available. Due to the complex nature of the resulting posterior distributions, estimation is carried out through Markov Chain Monte Carlo (MCMC) methods. We conducted a simulation study under different conditions reflecting double clustering/clustered measurement severity for continuous and binary responses, to assess the performance of the treatment effect estimators in terms of bias, precision and coverage. Based on the results from this study, we conduct a comparison between double clustering and clustered measurement situations and ultimately we provide practical recommendations that allow us to design optimal cluster trials when clustered measurement or double clustering are likely to be present in the responses. We illustrate the applicability of our methods with continuous and binary variables from PROBIT." --