Stephen Ruberg, Yongming Qu, Sean Yiu, and Martin Linder.
Causal Inference in Clinical Trials: To understand the potential of the practical application of causal inference methods in drug development and be able to apply these to real world problems or clinical trials.
This second webinar in this two-part series is aimed at illustrating real practical applications in drug development using case studies of how such ideas can provide valuable understanding of the effects of treatments in the presence of intercurrent events or where effects may be mediated by intermediate factors.
Estimating Treatment Effects in Patients Who Adhere to Treatment
Stephen Ruberg (Analytix thinking) / Yongming Qu (Eli Lilly)
The estimation of treatment effects has traditionally been based on the value of randomization and the causal inference it confers. However, causal inference from randomized controlled trials requires that all patients be analyzed as randomized AND, importantly, that all patients be followed for the duration of the trial and the primary outcome measured. Since many large or long-term trials involve patients who discontinue the study or discontinue their study treatment, this approach – often called intent-to-treat (ITT) – actually becomes an estimate of the effect of initiating (or being assigned) a treatment and NOT the effect of actually taking the treatment, which we call the direct treatment effect. An alternative approach is to censor the data from the time of treatment deviation and impute the resulting missing values (e.g., a hypothetical strategy). This approach uses all randomized patients but requires strong assumption on the potential outcome after the deviation away from the randomized treatment. While ICH-E9 recommended the ITT approach in general (or at least the use of all randomized patients in the analysis), ICH-E9(R1) has opened the door to other possible estimands and strategies for estimating a treatment effect. One such alternative is the direct treatment effect in patients (principal stratum) who actually would take/adhere to a treatment (Adherers Average Causal Effect – AdACE). This lecture will be divided into two parts: the first will motivate why such an estimand is of major importance, and the second will provide technical details on its estimation using causal inference methods. Examples will be given to highlight the methods, the code needed, and the interpretation of such the AdACE estimate.
Comparative safety analysis of time-varying exposures in post marketing observational studies
Sean Yiu (Roche)
Health authorities often mandate license holders of approved treatments to conduct post marketing observational studies to sufficiently assess long-term risk of important safety events, e.g. malignancies, since randomized clinical trials are typically too short and underpowered to detect treatment effects on such events. Furthermore, comparative safety analysis of newly approved versus other already approved treatments may be requested as part of the post marketing requirement. However, performing comparative safety analysis of long-term observational studies where treatment assignment is based on clinical practice is challenging and not well established in the regulatory setting, particularly when treatment switching (from control to active and vice versa) is anticipated to be frequent and often occurs prior to safety events of interest. Using a case study for OCREVUS, which is an approved treatment for adult patients with relapsing or primary progressive forms of multiple sclerosis, I will describe one specific post marketing requirement from the FDA on comparative safety analysis, the challenges of performing such analyses in the presence of multiple treatment switching, and highlight severe limitations of conventional methods based on time fixed treatments. I will then describe how established methodology for drawing causal inferences for the effects of time-varying exposures in the presence of time-dependent confounding, e.g. marginal structural Cox models, can address limitations of the conventional methods, and provide feedback from the FDA on the use of causal inference methodology in this observational setting.
Mediation analysis for a cardiovascular outcome trial
Martin Linder (Novo Nordisk)
There is a growing interest in statistical analyses that can answer questions concerning how a drug may affect an outcome via intermediate variables (mediators). The LEADER trial is an example. The trial showed a beneficial effect of the drug liraglutide on cardiovascular outcome in people with type 2 diabetes and high cardiovascular risk. Key opinion leaders as well as regulatory agencies asked whether the effect on cardiovascular outcome could be explained by previously known effects of liraglutide on blood glucose levels or body weight. The question is best answered within the framework of causal inference which provides methods for statistical analysis but also clarifies the assumptions necessary for a meaningful interpretation of the results.
In this presentation, we will consider some selected methods for causal mediation analysis that will be applied to the LEADER data. The methods include an approach developed jointly with experts from academia which specifically handles the case where the outcome is a time-to-event variable and the mediator is repeatedly measured.
