Use of instrument variables in the presence of heterogeneity and self-selection: An application in breast cancer patients
- Presenter:
Chair: Bruce Stuart; Discussant: Partha Deb Mon June 5, 2006 10:45-12:15 Room 325
Background: Instrumental variables (IV) are widely used in health economics literature to adjust for hidden selection biases in observational studies and estimate causal treatments effects. However, less attention is paid to the proper use of instrumental variables if treatment effects are heterogeneous across subjects and more importantly, if individuals select treatments based on expected idiosyncratic gains or losses from treatments. In the context of this interaction of heterogeneity and self-selection, we highlight the role conventional instrumental variable analysis and alternative approaches using instrumental variables in estimating causal treatment effects on 5-year costs in breast cancer patients. Given the similarities in 5-year mortality rates between breast conserving surgery with radiation therapy (BCSRT) and mastectomy (MST), estimation of treatment effects on costs is important in evaluating the cost-effectiveness of alternative treatments. Data: We use the Center for Medicare and Medicaid Services national claims database of a 5% random sample of all Medicare beneficiaries. The data were collected as part of the Outcomes and Preferences in Older Women Nationwide Survey (OPTIONS) project, and was used by other researchers to study costs of breast cancer patients under alternative treatments. Outcome: Average Treatment Effect (ATE) on 5-year direct costs. Treatments: BCSRT versus MST Alternative Methods: Traditional covariate adjustments, traditional IV analysis, control function approach using ordinary least-squares regressions. Results: We find that covariate adjustments estimates a treatment effect of $12,829 (se=1,639). Using two instruments, the treatment effect is estimated to be $32,136 ($15,005). Using one instrument at a time and also the propensity score as an instrument give estimates of treatment effects ranging from $8,421 ($8,828) to $44, 921 ($21,790), where some of these estimates are significantly different from each other. Employing a newly proposed test, we find evidence of self-selection based on idiosyncratic differences in outcomes between treatments. This implies the treatment effects estimated by traditional IV analysis may not be consistent estimates of ATE. We therefore estimate the marginal treatment effects (MTEs) using the control function approach with the polynomials of propensity scores. However, in this application, we find that it is not possible to estimate the MTE over its full support; hence ATE cannot be estimated with this data. Any attempt to estimate ATE with these data would require extrapolation of the control function beyond the range of observed support for the propensity scores. Consequently, depending on the function form of the control function we get estimates of ATE ranging from -$46,698 ($69,571) to $88,727 ($34,435). Conclusions: Estimation of causal ATE with instrumental variables faces several limitations. Traditional IV analysis do not provide consistent estimates of ATE if heterogeneity in treatment effects exist and patients self-select into treatment based on idiosyncratic effects. Marginal treatment effects can be estimated and used to overcome this limitation of traditional IV analysis. However, one may still fail to estimate an ATE, if the data do not provide full support for the propensity scores. Consequently, development and application of instrument-free semi-parametric methods to estimate causal treatment effects can be extremely valuable.
