Purpose: To develop “perceptive” methods for obtaining least-confounded cost-effectiveness estimates from non-randomized studies.
Methods: A review of the statistical/methodological literature was conducted to identify current shortcomings and unresolved issues. Statistical methods were developed to help overcome these and were implemented in the R software language.
Results: The start of the quest for least-confounded effects estimates from non-randomized studies began with the realization that comparisons between non-randomized groups would be confounded but that with perfectly accurate non-linear modeling of the covariates (separately by group) or perfect matching of all covariates - the confounding from the available covariates could be removed. For each point in the covariate space, a local treatment effect estimate could be obtained from differences in the response surfaces or differences in the observed effects of the matched pair closest to that point. Given treatments effects are unlikely to be constant over the covariate space, one may wish to define regions of the covariate space of interest and take appropriate averages.
Given neither was attainable, Rubin (JASA 1979) suggested always doing both in combination – for imperfectly matched pairs, within the pair, model the response surface and take the difference between the surfaces (i.e., adjustment after matching). Variations of this have evolved over the years including the consideration of matching just the distribution of covariates rather than the actual covariate values (i.e., conditioning on the propensity score). Recently, approaches have varied the relative combination of response surface modeling to matching, the type of matching and its implementation by stratification, weighting or pair-matching. Perhaps importantly different in health economic analyses, estimands of interest are functions of both effects and costs such as Net Monetary Benefit (NMB) and here to be least-confounded given the covariates, some combination of perfectly accurate response ”joint surface” modeling (i.e., both expected effects and costs) or exact matching on covariates that predict either effects or costs, would in principle, be required.Conclusion: The literature highlights challenges of flexibly specifying bi-variate probability models for effects and costs, correctly specifying the joint systematic effects of covariates on cost and effects parameters, and conditioning this joint modeling on the pair matches. To assist in this challenging modeling, we developed methods to quantify and display individual contributions of observations directly to estimands of interest (e.g. NMB).
Candidate for the Lee B. Lusted Student Prize Competition