National Meeting 2007

1049 — Using Mixed Effects Logistic Regression Models for the Analysis of Blood Pressure Control in a Clustered and Longitudinal Trial

Olsen MK (Durham VAMC/Duke University MC) , Oddone EZ (Durham VAMC/Duke University MC), DeLong ER (Duke University MC), Bosworth HB (Durham VAMC/Duke University MC)

Clustered and longitudinal designs are becoming more common in health services research. When the outcome variable is dichotomous, analysis methods for this type of design are computationally challenging. In this study, we describe and compare statistical methods for analyzing trajectories of blood pressure (BP) control from a clustered and longitudinal randomized trial of hypertensive veterans.

The Veterans Study to Improve the Control of Hypertension (VSTITCH) examined study interventions which occurred at two levels (provider and patient). In VSTITCH, patients were nested within providers and followed over time. BP was measured at patients’ primary care visits; therefore, each patient had a varying number of outcome assessments at varying intervals of time. Mixed effects logistic regression models with random effects at the patient and provider level were used to derive expected trajectories of change in BP control. These models were fit using both penalized quasi likelihood (PQL) and numerical integration techniques.

BP was obtained from 571 patients at their primary care visits over 24 months; the median number of visits per patient was 5 (range 1 to 24). These patients were clustered within 17 provider groups; the median number of patients per provider group was 37 (range 10 to 56). In the models, the fixed effects included intervention, time, and their interaction; the random effects included a provider-level intercept and a patient-level random intercept and slope. BP control was approximately 40% at baseline and increased linearly over time. Estimates for the fixed effects terms in the models were similar between the two techniques. Standard errors, however, were smaller for the model fit with PQL, leading to higher rates of statistical significance. The estimated variance components of the random effects were greater with PQL as compared to numerical integration.

Different techniques for mixed effects logistic models produced varying results in analyses of longitudinal BP control among patients clustered within providers.

Health services research studies often involve interventions for both patients and providers and, therefore, require statistical models which incorporate multiple levels of clustering. Different computational approaches should be examined when using mixed effects logistic regression models for this type of study design.