Advancing Statistical Methods: Innovations at ISPOR 2024
Enhancing Treatment Comparisons: Introducing 2SMAIC for Improved Precision
Population-adjusted indirect comparison (PAIC) methods allow for comparison of two treatments that have not been evaluated in direct head-to-head clinical trials and adjust for between-trial imbalances in effect modifiers that may bias treatment effect estimates. These comparisons are increasingly used to support market access activities for new interventions. Matching-adjusted indirect comparisons (MAICs) have historically been the most common PAIC method; however, the standard MAIC methodology is susceptible to imprecision when covariate overlap between the trials being compared is poor or when sample sizes are small.1, 2
To improve precision, at no cost to bias, a novel extension of the MAIC methodology has been developed and was discussed at ISPOR International 2024. This method, termed the two-stage MAIC (2SMAIC), includes an additional logistic model for treatment assignment in the index trial fitted to the individual patient data (IPD) which estimates inverse probability of treatment weights. These weights are combined with the odds weights derived from the standard MAIC and aim to balance covariates both between studies and between the treatment arms of the index trial.1
One is likely to question the necessity of this additional step given that randomisation should result in balance between covariates in the index trial. However, randomisation only guarantees balance in large samples, and there may still be finite-sample imbalance due to chance, particularly in trials with small sample sizes. The additional step allows for correction of random finite-sample imbalances and can result in increased precision, efficiency and power compared with standard MAIC. These precision and efficiency gains do diminish as covariate overlap between trials decreases, and 2SMAIC is not appropriate for unanchored comparisons.1 Overall, there are few disadvantages to conducting a 2SMAIC during a standard MAIC workstream; however, careful explication justifying the use of the method and further validation are likely to be required for health technology assessment bodies to accept the method.
Unlocking Causal Insights: Bridging HEOR and Traditional Economics
In addition to methodological developments within the standard health economics and outcomes research (HEOR) space, this year’s ISPOR saw increased crossover between HEOR and traditional economic theory, including a session exploring areas for potential collaboration between ISPOR and the American Society of Health Economists (ASHEcon). During this session, panellists identified the instrumental variable (IV) methodology, which is commonly employed in econometric studies, as a promising methodology in the HEOR space. A session dedicated to discussing the IV approach was held one day later.3
While multivariable regression and propensity score matching (PSM) methods are commonly used to control for measurable confounders, real-world evidence studies often include unmeasurable confounders which may lead to biased estimates. The IV method provides a possibility to account for these unmeasured confounders in the estimation of treatment effects through the creation of pseudo-randomised treatment assignment (for further information see here). However, the choice of IV must fulfil three strong assumptions. Specifically, the IV must:4
- Strongly predict the treatment received
- Be independent of unmeasured confounders
- Only affect the outcome through influencing the treatment received
While the use of IVs is a promising method to establish causal effects, the identification of an IV that meets the above assumptions is difficult, and IV analysis requires larger sample sizes to achieve sufficient power than other methods. The panelists therefore argued in favor of a “triangulation approach” in which a single study is analysed using all three approaches (multivariable regression, PSM and IV analysis) to increase confidence in the findings of causality within RWE studies.4 With the ever increasing importance of RWE, this is a valuable approach to incorporate within the analysis toolkit and will be of particular relevance for studies where unmeasured confounding is of significant concern for analysis validity.