Estimating lifetime medical costs from censored claims data

[[abstract]]Objectives: Estimates of expected lifetime survivals and lifetime costs for cohort with specific conditions are usually needed in cost-effectiveness analysis. However, the survival data of followed-up patients were often censored with high rates and observed expenditures were incomplete. It is desirable to develop reliable and robust methods for extrapolating survival and cost functions beyond the follow-up. Methods: We propose using a semi-parametric extrapolation method to replace parametric survival models for estimating lifetime survival rates. We extrapolate the lifetime monthly mean costs using a weighted average of mean expenditures of patients in their final years and months prior to their final years. The weights are functions of hazards which can be estimated from the extrapolated lifetime survival rates. The expected lifetime cost can be estimated by summing the product of the estimated survival probabilities and monthly mean costs. Results: We evaluate performance of the proposed approach using simulated data and empirical data. For demonstration, we use population-based claims data from the Taiwan National Health Insurance to establish cohorts of ischemic stroke and intracerebral hemorrhage and estimate the lifetime direct medical costs of first-ever stroke patients. We found that life expectancy of patients diagnosed with intracerebral hemorrhage and ischemic stroke is about the same of 9 years since the onset of stroke. The expected lifetime direct medical costs are also about the same amount of US$ 35,000 for both cohorts. Conclusions: We demonstrated the proposed semi-parametric method of survival extrapolation performed well using simulated data and empirical data. We also showed in the simulation that even perfectly fitted parametric model may not be accurate for long-term extrapolation. Our estimates of lifetime direct medical costs for first-ever stroke patients can have implication in public health policy planning and clinical decision making