SC01: A Practical Guide to Estimand Implementation

Since the release of ICH E9(R1), the implementation of the estimand framework has received considerable attention. Despite this well-worn topic, based on presentations at recent statistical conferences and publications in the statistical literature, there is still considerable misunderstanding, differences of opinions, and outright misconceptions about estimands--from how to define them to creating consistent study designs and data collection instruments through interpretation of final results and ultimately labeling. This presentation will also highlight how some efforts have been unnecessarily complicated and confusing, thereby alienating clinical and other nonstatistical collaborators in the clinical drug development process. This course will emphasize first principles and build from there to create clear and actionable considerations for creating meaningful estimands in clinical drug development. The first principles start with two basic definitions and two obvious supposition. First, science is about understanding true cause-and-effect relationships in nature. Second, statistical science is about inferring what is likely to be true--true in the sense of causal effects to support scientific endeavors. The first supposition is so obvious it actually goes unnoticed and therefore needs to be stated: Clinical drug development is about answering the question, "Does this treatment cause that outcome?", where the outcome may be an efficacy outcome or a safety outcome. The second supposition is that the estimand must first and foremost be clinically meaningful. With this foundation, the estimand framework will be presented through a series of lectures and examples. Briefly, the course will start with cause and effect--what it means and how randomization plays an important (or even essential) role. But randomization is not enough; it is necessary but not sufficient. There also needs to be complete data on the randomized experimental units. It is this latter requirement that creates consternation in the definition of the estimand, the estimator, and finally the interpretation of the estimate. A brief history of cause and effect and the introduction of the intent-to-treat (ITT) principle in the late 1950s and early 1960s will be reviewed. This lays the foundation for ICH E9 and the oft-used ITT or treatment policy analysis that is so prevalent (perhaps overused) but not always consistent with the historical perspective or its initial intent. The course will then highlight how ICH E9(R1) is an attempt to clarify the definition of the treatment effect when there is incomplete data on the randomized study treatment. The four attributes will be examined carefully, but none more carefully than the definition of the study treatment. This will be called the estimand-defined study treatment (EDST). With the EDST in hand, many issues related to intercurrent events and the five strategies discussed in ICH E9(R1)--treatment policy, hypothetical, composite, while-on-treatment, and principal stratification--become much more clear and simple. All these strategies are in one way or another a mechanism for creating the complete data necessary for making causal inference, but each carries a different interpretation. The tripartite estimand approach (TEA) will also be discussed in detail as another strategy not covered in ICH E9(R1) but should be given strong consideration in certain situations. Concepts from causal inference and the use of potential outcomes will underpin much of the course. Examples across many disease states will be given--oncology, immunology, diabetes, neurodegeneration, etc. The ultimate learning objective is to help statistical practitioners simplify and clarify the estimand framework so they understand what needs to be done and can engage clinicians and other nonstatistical colleagues in a meaningful way.