Events and Seminars : 2014 Seminars

MULTIVARIATE MULTILEVEL MODELS FOR HEALTH SERVICES RESEARCH

ALAN M. ZASLAVSKY, PH.D.
Professor, Health Care Policy (Statistics)
Department of Health Care Policy
Harvard Medical School

TUESDAY, MARCH 17, 2015
11:00 a.m.–12:00 p.m., CRB 692

Much of health services research involves assessing quality of care and practice patterns of clinical units such as hospitals, health plans, clinics, and individual physicians using multiple measures.A natural statistical tool for modeling such data is the multilevel (hierarchical) model with multiple outcomes at the level of the unit of interest.Typically this implies a vector outcome at the second level of the model, while the observations at the first level represent observations on individual patients.Complications commonly arise at the patient level because not all measures apply to all patients and because the data may arise from a complex sample.To analyze such structures, we estimate multivariate Fay-Herriott-type models in which the patient-level data for each unit are reduced to a mean and covariance matrix for each unit.Once the level-2 covariance matrix is estimated we may further analyze it using factor analysis, or a Kronecker-structured decomposition suitable for cross-classified data.Bayesian methods can facilitate estimation of uncertainty in these analyses.Examples in this presentation include a multiperiodanalysis of survival after colorectal cancer surgery, a factor analysis of regional variations in service and procedure utilization in Medicare, and an analysis of variation across Medicare health plans in associations of patient characteristics with responses on multiple items of a patient experience survey.