In this presentation, I will discuss how to formulate and choose an estimand, beyond the marginal intention to treat effect, from the perspective of a decision maker and drug developer. I will specifically consider randomized clinical trials with time-to-event outcomes. I will emphasize that a careful articulation of a practically useful research question should either reflect decision making at this point in time or future drug development. A common feature of estimands that are practically useful is that they correspond to possibly hypothetical but well-defined interventions in identifiable (sub)populations. To illustrate my points, I will consider examples from clinical trials involving competing, recurrent and intercurrent events.

The ICH E9(R1) addendum on estimands and sensitivity analysis in clinical trials calls for clarity and precision when describing the clinical question of interest. It defines the estimand as a population-level summary of “what the outcomes would be in the same patients under different treatment conditions being compared”. Thus, while not explicitly using the term “causal”, both the framework and language used in ICH E9(R1) are aligned with causal reasoning.

In randomized clinical trials with a time to event endpoint, the hazard ratio is still the most common effect measure. Post-randomization (i.e., intercurrent) events are often addressed through censoring without explicitly discussing or stating the underlying clinical question of interest. Alternative summary measures, especially on a probability scale or time scale, are rarely considered in clinical trials despite being seemingly easier to interpret and potentially more meaningful to patients and practitioners.

In this talk we will present the status of ongoing discussions among a working group of statisticians from different pharmaceutical companies on estimands for clinical trials with time-to-event data. In detail, we will discuss what key clinically meaningful questions of interest are when measuring the effect of an intervention through a time-to-event endpoint. We will reflect on the interpretation of various summary measures, the role of causality when defining an estimand in a clinical trial, and on how the choice of the estimand affects the design of a trial with a time-to-event endpoint. We will also elaborate on the practicalities of summarizing the effect of treatment through a single number in a time to event setting and discuss separating testing and estimation.

The publication of the ICH E9 addendum on estimands and sensitivity analysis moved the importance of the estimand as the exactly defined target of inference into the focus of regulatory scientific advice and assessment of clinical trials. For time-to-event endpoints, awareness for the need to account for post-randomisation events such as discontinuation or change of treatment existed before the ICH E9 addendum and is reflected in regulatory recommendations for study design and censoring rules. However, the estimand framework draws attention to the need for a precise definition of the underlying question prior to aligning design and analysis accordingly, and to the need to differentiate between intercurrent events and missing data. As it appears questionable that all estimands of interest can be appropriately addressed by the still most commonly used approach of adapting censoring rules, alternatives may need to be advocated. However, regulatory experience with alternatives is still limited.

The hazard ratio (HR) estimated based on the Cox model is still the most commonly used summary measure for the treatment effect in studies with a time-to-event endpoint. However, it has been increasingly criticized due to its lack of causal interpretation. Furthermore, deviations from non-proportional hazards in certain therapeutic areas have aroused additional interest in alternative effect measures. From a regulatory point of view, an effect measure should not only inform benefit-risk decisions for approval but provide relevant information for labelling across the patient population that can be easily interpreted by patients and prescribers. The HR has never been the only effect measure to support regulatory decisions and labelling, but several measures are used to provide the overall picture. Still, stronger emphasis on less commonly used alternative summary measures allowing a causal interpretation may be useful.