Propensity Score and Instrumental Analysis of Mortality Differences Between Medicare HMO Enrollees and FFS Beneficiaries
- Presenter:
Chair: Albert Okunade; Discussant: TBA Mon June 5, 2006 13:45-15:15 Room 235
Rationale: Identification of the causal effect of an intervention is challenging in observational data. Propensity score and instrumental variable analysis has been used in health services applications to reduce confounding, but it is unknown whether these two methods are likely to generate similar results.
Objectives: To adjust for selection bias in estimating the effect of Medicare HMO enrollment on mortality for Medicare beneficiaries with diabetes comparing instrumental variables (IV) and propensity score matching (PS) approaches.
Methodology: Medicare claims data were obtained on a 2% sample (N=59,089) of Medicare beneficiaries with diagnosed diabetes in 1995. Managed care benefits and out-of-pocket premiums were also obtained from Medicare to generate county-level measures of plan generosity. We performed cross-sectional analysis of 59,089 patients. IV and PS using radius matching are used to generate the average effect of HMO enrollment on mortality across the entire sample and average treatment effects within quintiles. A non-parametric bootstrap is performed to calculate confidence intervals for the treatment effects by quintile.
Results: The average treatment effect from a PS matching model and a simple probit model suggested that HMO enrollment is associated with a 4% reduction in mortality for Medicare beneficiaries with diabetes in 1995. The mortality reduction from the IV model was much smaller (0.4%). The quintile-specific treatment effects were also very different been PS matching and IV, and also varied across quintiles. Improving balanced in observed covariates via PS were much less effective in reducing bias in the effect of HMO enrollment on mortality than was statistical correction for unobserved confounding via IV. When such a correction is applied, the mortality of HMO enrollees and FFS beneficiaries are almost equivalent.
Conclusions: Propensity score and instrumental variables analysis provided different treatment effect estimates in this study. Propensity score adjustment for imbalance in observed covariates was not able to address bias due to unobserved covariates in instrumental variables, which generated a null effect of HMO enrollment on mortality. Researchers interested in applying methods for identifying causal effects of treatment must clearly understand the nature of the bias that needs to be addressed.
