130 likes | 257 Views
Combining propensity score and regression approaches in economic evaluation – simulation and case study. Presented at the Population Health - Methods and Challenges Conference, Birmingham 2012 Noémi Kreif. Acknowledgements.
E N D
Combining propensity score and regression approaches in economic evaluation – simulation and case study Presented at the Population Health - Methods and Challenges Conference, Birmingham 2012 NoémiKreif
Acknowledgements Co-authors: Richard Grieve (LSHTM), RosalbaRadice (LSHTM), Susan Gruber (Harvard School of Public Health) Jasjeet S. Sekhon (UC Berkeley), Collaborators: Roland Ramsahai, Zia Sadique, Rhian Daniel, James Carpenter Data: Kathy Rowan, David Harrison (ICNARC) Funders: Economic and Social Research Council
Context Cost-effectiveness analysis (CEA) evaluates health interventions Many circumstances only information from non randomised studies (NRS) Non-random selection into treatment Statistical methods in NRS need to address selection bias Focus on selection on observed characteristics
Context and aim Statistical methods to adjust for observed confounding: • Modelling endpoint (costs, effects): regression methods • Modelling assignment mechanism: propensity score (PS) methods Assumptions rarely assessed in CEA (Kreif et al., 2012) Combined methods: • Regression + PS weights (doubly robust methods) • Matching followed by regression (regression-adjusted matching) Aim: to compare the performance of individual vs. combined methods in the context of CEA
Single methods • Regression • Assume endpoint models correct • Propensity score matching • Assume PS correct • Inverse probability of treatment weighting (IPTW) • - As above + weights stable Combined methods • Doubly robust (DR) methods • Weighted regression, inverse PS weights • “DR”: unbiased if either endpoint or PS model correct (Robins et al., 1994) • Regression-adjusted matching • Regression on matched data (Ho et al., 2007) • Reduce bias due to remaining imbalance (Abadie and Imbens, 2011)
Simulation overview Concern in case study: • Regression and PS misspecified • Unstable weights Individual methods biased under misspecification Compare to combined methods: DR method: weighted regression - Can reduce bias and increase precision compared to IPTW - Can be biased and inefficient due to unstable weights (Kang and Schafer, 2008) Regression-adjusted PS matching - Regression adjustment reduces residual bias vs. matching alone - Balance after matching to reduce sensitivity to endpoint model misspecification - Limited simulation evidence
Simulation setup Treatment assignment: Cost and QALY generated with bivariate copula (Trivedi and Zimmer, 2007) • normal QALY (Y), gamma cost (C) • correlation parameter 0.4. True treatment effects: ΔC=6000, ΔE=0.4, INB=2000 (WTP=£ 20,000)
Misspecification Common functional form misspecification for PS and regression: (Kang and Schafer, 2008) Instead of latent variable Z, we observe X e.g. X2 = Z2/(1+Z1)+10 PS model: Logistic regression includes X1, X2, X3, X4 excludes nonlinear term Regression models: QALYs: GLM - identity link, normal error, X1, X2, X3, X4 Costs: GLM - log link, Gamma error, X1, X2, X3, X4
Simulation results (1) Stable weights, moderate misspecification: • When all correct: regression is best • Regression-adjusted matching and DR increases precision and reduces bias compared to PS methods alone • Even when PS is correct • Even when regression is misspecified (2) Unstable weights, moderate misspecification: • Correctly specified: IPTW high RMSE • DR improved on IPTW • Even with misspecified regression • Dual misspecification: regression and regression-adjusted matching better than DR
Simulation results(3): Unstable weights, major misspecification of both PS and regression
Case study DrotAApharmaceutical: patients with severe sepsis, 3 to 5 organs failing Published PS (Rowan et al., 2008), including nonlinearities Unstable PS weights Regression models developed for cost and QALY endpoints -> possibly misspecified PS and endpoints Report incremental net benefit (INB)
Case study results Bootstrapped CIs; conditional on estimated PS / matched data
Summary Combined methods promising in CEA context Extreme weights challenge for DR, regression-adjusted matching robust Rely on assumption of no unobserved confounding Next steps: Challenge of bivariate endpoint when choosing potential confounders in regression and PS Data adaptive methods: reduce reliance on parametric model specification • Genetic Matching – automated matching method (Sekhon and Grieve, 2011) • Targeted Maximum Likelihood Estimation – a flexible DR method (Gruber and van derLaan, 2010)