This presentation is part of: Oral Concurrent Session B - Health Economics
A MULTI-STATE STOCHASTIC MODEL FOR ASSESSING COVARIATE EFFECTS ON SURVIVAL AND COST
JOSEPH C. GARDINER, PhD1, ZHEHUI LUO, PhD1, CORINA M. SIRBU, PhD1, CATHY J. BRADLEY, PhD, MPA2, and CHARLES W. GIVEN, PhD3. (1) Michigan State University, Department of Epidemiology, East Lansing, MI, (2) Michigan State University, Medicine, East Lansing, MI, (3) Michigan State University, Family Practice, East Lansing, MI
PURPOSE: We describe a longitudinal multi-state regression model to assess covariate effects on patient outcomes and cost, addressing issues of heteroscedasticity, skewness and censoring. METHODS: Event histories of patients are modeled by finite-state Markov processes in continuous time. Cost incur at transitions between health states and in sojourn in health states. These expenditure streams over a specified period are combined to form net present values. First, we model the impact of treatments, patient-specific demographic and clinical characteristics on survival through the transition intensities using a Cox regression model. Second, we use a mixed model for transition and sojourn costs with transition times as random effects and patient characteristics as fixed effects. Censoring of patient costs and outcomes, discounting of costs and benefits are all incorporated into the model. APPLICATION: We used data on 605 incident cases of lung, prostate, colon and breast cancer recruited from Michigan community hospitals during 1994-1997. Charge data were obtained from Medicare claim files which included reimbursements for inpatient and outpatient care, physician services, home health care/skilled nursing facilities, laboratory tests and treatments. A patient’s total charge was used as a proxy for cost. Complete cost was deemed known if a patient had at least 2 years of follow up, or if death occurred within 2 years of diagnosis. Otherwise, a patient’s cost was regarded as incomplete. Approximately 31% of the cost data were censored. Physical function as assessed by the SF-36 at baseline and subsequently at 4 waves was categorized as “good” or “fair”. During follow up at least one transition in physical function occurred in 50.4% of the patients, and 22.7% remained in fair physical function throughout. Our regression model revealed that treatments (surgery, chemotherapy/radiation) cancer site and comorbidity were significantly associated with cost. Transitions were influenced by these same factors as well as a count of symptoms. Mortality was 10%. Survival was best among breast cancer patients and worst among lung cancer patients. CONCLUSIONS: The joint regression model provides a rigorous and flexible statistical approach to assessing the influence of patient variables on both cost and health outcome. While incorporating explanatory variables, the model accommodates heteroscedasticity, skewness and censoring. It also provides a unified framework for inference on summary measures used in cost-effectiveness analysis.
