Comparison of Methods for Clustered Data Analysis in a Non-Ideal Situation: Results from an Evaluation of Predictors of Yellow Fever Vaccine Refusal in the Global TravEpiNet (GTEN) Consortium

Abstract

Not accounting for clustering in data from multiple centers might yield biased estimates and their standard errors, potentially leading to incorrect inferences. We fit 15 different models with different correlation structures and with/without adjustment for small clusters, including unadjusted logistic regression, Population-averaged models (Generalized Estimating Equations), Cluster-specific models (linear and non-linear with random intercept) and Survey data analysis methods to study the association of variables with the probability of declining yellow fever vaccine among patients seeking pre-travel health consultations at 18 US practices in the Global TravEpiNet Consortium from 1 January, 2009, to 6 June, 2012. Results varied by the method chosen. Generally, when the odds ratio estimates were similar, adjusting for clustering and the small number of clinics increased the standard errors. We chose the random intercept model with the Morel, Bokossa and Neerchal (MBN) adjustment to be the most preferable method for the GTEN dataset since this was one of the more conservative models that accounted for clustering, small sample sizes and also the random effect due to site. Investigators should not ignore clustering and consider the appropriate adjustments necessary for their studies.