Candidate for the Lee B. Lusted Student Prize Competition
Purpose:For cost-effectiveness analyses (CEA) that use observational data the key methodological challenge is to minimize selection bias. Propensity score (Pscore) methods can reduce selection bias due to observable differences between treatment groups; but the true Pscore model is generally unknown. Doubly robust (DR) methods exploit information in the Pscore and the response models, and provide unbiased estimates if either model is correctly specified. These methods hold promise for CEA, where selection bias needs to be minimized for the cost as well as the effectiveness endpoint. DR methods have not been examined before in this context.
Method:One implementation of DR methods is inverse probability of treatment weighting (AIPTW). The simple IPTW estimator weights observed cost and effectiveness endpoints with the inverse of the Pscore, to estimate incremental costs and effectiveness. AIPTW extends this by adjusting the formula with weighted predictions from the regression models of the respective endpoints. If a response model is correctly specified, adding this term can reduce bias. The adjustment also stabilises extreme Pscore weights, which can improve the precision of the IPTW estimator.
To compare the methods in a CEA, we evaluate Drotrecogin alfa activated (DrotAA), a pharmaceutical intervention for critically ill patients with severe sepsis. We use data from a published observational study (n=1,898). Potential confounders were selected a priori (e.g. age, APACHE II severity score). Higher order terms and interaction terms were considered, and regression models for both cost and effectiveness were selected by cross-validation. A two-part model was chosen for the QALY and a generalized linear model with gamma distribution for the costs. To maintain correlation between costs and effects, confidence intervals (CI) were constructed by nonparametric bootstrapping.
Result:The incremental net benefit (INB) (λ=£20,000 per QALY) for DrotAA following IPTW was -£4796 (95% CI: - 23927 to 14969). After applying AIPTW, the estimated INB was £4936. Stabilizing the extreme Pscore weights led to tighter CI (-3867 to 12229).
Conclusion:DR methods avoid relying solely on a correctly specified Pscore or response model, and can lead to different point estimates and narrower CI than IPTW. Recent work shows that DR methods, eg. collaborative targeted maximum likelihood, can minimize bias and be efficient even if neither the Pscore or response models are correct, offering further flexibility in CEA.