In this presentation, I will start with an overview on common estimands used in causal analyses of observational (aka ‘real world’) data, their motivation, and the key structural assumptions underlying typical analytical methods. Applied examples range from estimating the (side)effects of therapies or the effectiveness of cancer screening using electronic health data, to evaluating lifestyle recommendation aimed at preventing childhood obesity. In most of these examples, the relevant exposure or treatment is not a binary point variable but a sustained or time-varying exposure/treatment, and the outcome is longitudinal or a time-to-event. This means the estimand should be carefully formulated in view of what sustained, time-varying or possibly adaptive treatment strategies we wish to compare. While the key assumptions, required for most methods, of sufficient measured time-varying confounder information and sufficient overlap between observed and targeted strategies often appear hard to defend, approaches to strengthen their plausibility have been developed and are increasingly applied.

Secondly, I will discuss possible implications and lessons-to-be-learned from observational causal inference for the design and analyses of RCTs with typical intercurrent events as addressed by the ICH E9 Addendum. The latter can often be seen to alter either the intended meaning of treatment or of the outcome, so that relevant estimands must be formulated. This will often take us back to dynamic or adaptive treatment strategies. However, sometimes, notions of direct effects are invoked, and I will discuss why these are more problematic to interpret. In this context, and with view to drug development, I will explain how so-called separable effects might be an interesting alternative. Particular attention will be paid to how different estimands, in observational studies or RCTs, inform different decision makers, such as individuals, physicians or public health authorities. A further key difference, compared to most sources of observational data, is that in RCTs (and to some extent in studies using real-world data) we have more influence on what information will be available for the analysis, and thus can ensure that a rich set of confounders is measured as well as other useful information enabling sensitivity analyses.

Finally, I will come full circle and address the role of formulating a target trial, i.e. a somewhat idealized hypothetical trial, in the analysis of observational / real-world data. Target trial emulation is an important principle especially for eliciting actionable estimands in complex longitudinal data situations, but also for avoiding self-inflicted biases in the analysis of observational data. Thus, on the one hand, the analysis of RCTs with intercurrent events looks to learn from approaches to causal inference developed for observational data; on the other hand, the causal analysis of observational data can be much strengthened by adhering to certain design principles of randomized trials.

The presentation will focus on general principles, examples and interpretation more than on technical details.

When choosing estimands and estimators in randomized clinical trials, caution is warranted, as intercurrent events, such as, due to patients who switch treatment after disease progression, are often extreme. Statistical analyses may then easily lure one into making large implicit extrapolations, which often go unnoticed. This is problematic as it can lead to significant bias, large variance and invalid inference. We will illustrate this problem for different estimators, e.g. imputation and weighting methods, using real case studies. Moreover, we show that estimands used to handle intercurrent events are often too ambitious and cannot be inferred from the data without relying on very strong assumptions. In the second part of the talk, we will discuss different solutions in terms of estimands, estimators and trial design.

Causal thinking and related inference methods are gaining increasing prominence in global drug development in light of the recently published ICH E9(R1) guideline on estimands and sensitivity analysis (2019) and the FDA draft guideline on covariate adjustment (2021). These guidelines refer to terminology, concepts and methods from the causal inference literature, such as potential outcomes, principal stratification, non-collapsibility and standardization.

In this talk we build upon these recent developments by examining how causal inference provides a convenient mathematical language and tools to formally establish causal relationships between, e.g., drug and effect. We believe that causal inference methods can be used in many drug development settings, including those outlined in the two guidelines above, but also for the use of external control data and understanding cause and effect in pharmacometric and pharmacovigilance applications. We illustrate the importance and potential impact of causal inference by presenting two case studies. In the first case study, we will discuss how trial external data can be leveraged using the target trial and estimand frameworks for an oncology trial. In addition, the use of a principal stratum estimand to answer a clinical meaningful question in multiple sclerosis will be discussed.