Conference Agenda

In this talk I will discuss approaches for structuring and conducting analyses that combine information from multiple and diverse sources to address questions about the effectiveness of medical interventions. By considering two practical examples — (1) extending inferences from a clinical trial to a “real-world" population and (2) augmenting a clinical trial with external data to improve efficiency in estimating treatment effects — I argue that combining information from multiple sources to asnwer causal questions requires novel study designs, causal assumptions, and statistical methods.

Background

Conducting a properly-powered randomised controlled trial is not feasible in certain settings, due to small populations and a lack of trial-eligible patients, for life-threatening and severely debilitating conditions with high unmet need, and for ethical reasons. Regulators recognize that externally controlled clinical trials may be required in special circumstances, and marketing authorisation applications featuring externally controlled trials are increasing. The reliance of payers and health technology assessment bodies on such research designs is also growing. For instance, in the absence of head-to-head comparisons between all relevant comparators, “unanchored” indirect comparisons are often required in therapeutic areas with a rapidly evolving treatment landscape.

Methods

Various statistical methods have been proposed to adjust for imbalances in baseline covariates between the clinical trial and the external control. The most widely-used methodologies are singly robust propensity score-based weighting and outcome modelling-based approaches. Alternative weighting methods based on entropy balancing will be presented, which directly enforce covariate balance and are generally more stable, precise and robust to model misspecification than the standard “modelling” approaches to weighting. The presentation will introduce doubly robust estimators that augment the entropy balancing approaches by fitting a model for the conditional outcome expectation, then combining the predictions of the outcome model with the entropy balancing weights. The methods are evaluated in a simulation study and their application illustrated in an example analysis.

Results and Conclusions

The presentation integrates parallel developments in the areas of indirect treatment comparisons (meta-analysis) and causal inference, under a unified framework for target estimands and covariate adjustment. Decision-makers have expressed a preference for doubly robust estimation approaches that can consistently estimate treatment effects as long as either a propensity score model or an outcome model is correct, but not necessarily both. Augmented entropy balancing-based estimators that are doubly robust and more bias-robust than commonly used approaches, as demonstrated by simulations involving binary outcomes, are presented.

The recent 21st Century Cures Act propagates innovations to accelerate the discovery, development, and delivery of 21st century cures. It includes the broader application of Bayesian statistics and the use of evidence from clinical expertise. An example of the latter is the use of trial-external (or historical) data, which promises more efficient or ethical trial designs. Draft guidance FDA “Considerations for the Design and Conduct of Externally Controlled Trials for Drug and Biological Products” emphasized that sponsors must include patient-level data in the market-application. Though there are considerable literature discussing Bayesian methods to include summary-level external control data in design and analysis, the literature handling individual-level external control data is still small. We provided a robust meta-analytic framework when individual patient-level data is available. analyze individual patient data from various sources. We suggest a two-step approach. In the first step, the data from each source is analyzed in isolation, resulting in an estimate (and standard error) for the main parameter of the target population, taking account of the covariate information from each patient. In the second step, these adjusted data are then analyzed using robust Bayesian hierarchical model. The utility of the method is illustrated using simulation and a case study from Alopecia Arrieta area. The talk will further reflect on the regulatory discussions.