A Simulation Study to Compare Two Classes of Models for Analyzing Clustered Data with Informative Cluster Size

• Activity Number: 253
• Type: Contributed
• Date/Time: Monday, July 30, 2012 : 2:00 PM to 3:50 PM
• Sponsor: Biometrics Section
• Author(s): Ana Maria Iosif*+
• Companies: University of California at Davis
• Keywords: clustered ; informative cluster size ; repeated

Abstract

In many chronic diseases or health conditions, a person may have repeated episodes collected together with a measure of each episode's intensity or severity. Data for an individual can be viewed as a cluster of severities. The size of the cluster is a random variable determined by the number of episodes and might inform about the condition severity. Failure to account for the potential relationship between the responses across the member of a cluster and the size of that cluster might lead to bias inference. Recent literature introduced models that are more appropriate for handling the possible informativeness of the cluster size. In this simulation study we compare two classes of such models. The first one employs a maximum likelihood framework using a shared parameter approach, with parametric distributions modeling the number of episodes. The second class does not specify a parametric distribution for the cluster sizes and uses within-cluster resampling to generate a series of data sets that can be analyzed using standard univariate analyses. We discuss finite sample properties of the estimators in each class in simulation scenarios with sample sizes in a small to medium range.