Quasi-Least Square (QLS) and Transition Model Approach for Clustered Longitudinal Binary Data with Multiple Sources of Correlation
Clustered binary longitudinal data often involve more than two sources of correlation. Human papillomavirus (HPV) International transmission (HIT) study is an example of clustered longitudinal data where correlation arises in three ways: correlation between gender within couple, temporal correlation among repeated measurements on the same genotype over time, and directionless clustering due to the multiple genotypes tested within an individual. Current statistical approaches that are used to identify risk factors does not appropriately take into account multiple sources of correlation within a cluster. In this type of data, we are often also interested in modeling persistence of the outcome, but there is no standardized method to calculate persistence, and account for multiple sources of correlation at the same time. For HPV study, given the scientific definition of 12-month persistence, second order transition model provides us with a useful tool. Additionally, we implement Quasi-Least squares (QLS) approach in order to adjust for multiple sources of correlation, and present a combined method of transition model and QLS approach for such data. Using this approach, full specification of the joint multivariate binary distribution is avoided by using conditional argument to handle the temporal correlation and QLS to account for the two sources of correlation. A simulation study is performed for exchangeable correlation structures to compare the proposed approach with commonly used Generalized estimating Equations (GEE). The results demonstrate good coverage probabilities for regression parameter estimates. We also take into account clustering effect when performing imputation missing values, compare with imputation strategies ignoring clustering effect. Using random intercept model with multiple clusters produced increased coverage probabilities for regression parameters. The proposed analytic and imputation model is applied to the HIT data to identify modifiable risk factors for 12-month persistence in couple.^
Chang, Mihyun, "Quasi-Least Square (QLS) and Transition Model Approach for Clustered Longitudinal Binary Data with Multiple Sources of Correlation" (2017). Texas Medical Center Dissertations (via ProQuest). AAI10268770.