2011 HSR&D National Meeting Abstract
1088 — Latent Variable Models for “Quality of Care” in Multi-Center Cluster Randomized Trials
Zhao X (Center for Health Equity Research and Promotion, VA Pittsburgh Healthcare System), Stone RA
(Center for Health Equity Research and Promotion, VA Pittsburgh Healthcare System), Ye F
(University of Pittsburgh School of Education), Fine MJ
(Center for Health Equity Research and Promotion, VA Pittsburgh Healthcare System)
In clinical trials, multiple endpoints for treatment efficacy often are obtained and data may be collected hierarchically. Statistical analyses become very challenging for such multidimensional hierarchical data. The objective of this study is to implement a latent variable approach to assess an intervention effect with regard to multiple binary outcomes in a study with three-level hierarchical data.
This latent variable modeling approach incorporates the correlation structure into one latent (hypothetical) construct (“quality of care”) and simultaneously regresses the latent construct on the intervention. Random effects model the hierarchical structure. Bayesian estimation is implemented in WinBUGS. We illustrate the approach in a 32-site cluster randomized trial, the Emergency Department Community Acquired Pneumonia (EDCAP) study, which was designed to compare the effectiveness of three guideline implementation strategies on quality of care of patients with CAP in the ED. The intervention was randomized at site (ED) level, with 8 low-intensity, 12 moderate-intensity, and 12 high-intensity intervention sites. The outcomes were binary indicator measures of four recommended processes of outpatient care (oxygen assessment, first dose of antibiotics in the ED, treatment with compliant antibiotics, and compliant antibiotic therapy upon discharge). These outcomes were collected at the patient level and clustered at the provider and site levels. We conducted simulation studies to check the computational implementation.
High-intensity intervention sites had significantly higher mean quality of outpatient care than sites with low- or moderate-intensity interventions; low- and moderate-intensity sites were similar. Two processes of care (first dose of antibiotics in the ED and compliant antibiotic therapy upon discharge) had relatively higher power to discriminate between sites. A few “outlier” sites were identified using the estimated latent score. Simulation study results demonstrate the accuracy of Bayesian estimates.
This method allows assessment of overall intervention effects with respect to multiple outcomes, quantifies relationships between outcomes, identifies those outcomes that are most informative regarding the latent trait, and provides a summary measure of the “quality of care” for each site.
Latent variable models provide a comprehensive way to analyze multivariate hierarchical data. A single quality of care index can be estimated from multiple indicators.