Article
Inferring Time-Varying Treatment Effects from Observational Data: Lessons for Clinical Trials
Search Medline for
Authors
Published: | February 26, 2021 |
---|
Outline
Text
The field of causal inference deals with approaches for investigating the effects of (possibly hypothetical) interventions from observational data or imperfect trials. Logically, the first step of such an investigation is to define the target of inference, aka causal estimand. In this presentation, I will focus on the fact that in biomedical and epidemiological research most treatments or exposures are time-varying. Hence relevant target interventions will often need to be time-varying, too. This also applies to many RCTs, in particular when faced with intercurrent events, and essentially this is where the ICH E9 estimands addendum comes into play. I will give an overview of and discuss the ensuing issues, as well as ways to address them.
It is relatively easy to decide on a target of inference for a point-treatment/exposure, e.g. as a simple contrast like the average causal effect or the causal risk ratio, with further refinements to subgroup effects or comparison of survival curves. In case of time-varying treatments/exposures, specifying the target of inference is typically more challenging. Even if we simply wished to compare hypothetical interventions like, say, “always-treat” versus “never-treat”, the fact that over time patients will not comply with these two options, possibly for good reasons, must be taken into account. So, while ideally, the choice of target of inference should be dictated by the research question or the decision problem at hand, we may often wish to allow for what is actually feasible in practice. The translation of the research question into a target of inference, or estimand, is a key issue, and should be an explicit part of any investigation (experimental or observational).
I will argue and illustrate that in many situations, including those addressed by the ICH E9 estimands addendum, it makes sense to consider dynamic or adaptive treatment strategies. This has long been recognized for epidemiological studies with time-varying exposures: see Robins [1] and the many follow-up papers since, e.g. Dawid & Didelez [2]; also see, for instance, the monograph by Chakraborty & Moodie [3] on optimal adaptive treatments. The renewed interest has been due to the debate around the ICH E9 addendum on estimands, which has opened up the analysis of clinical trials to causal inference approaches originally aimed at observational data. On the observational studies side, interestingly, there is a recent push to use the “target trial” principles for analyzing observational data [4]; these principles can be regarded as a systematic guide to formulating sensible causal estimands, and then carrying out the appropriate inference. Both developments are two sides of the same coin: an increasing awareness of the general need to be explicit about the target of inference, to design the study towards this aim, and to use suitable models and methods. This ensures a principled and coherent statistical analysis with practically useful results, where avoidable biases are avoided while plausibly minimizing biases that are not avoidable.
The authors declare that they have no competing interests.
The authors declare that an ethics committee vote is not required.
References
- 1.
- Robins J. A new approach to causal inference in mortality studies with sustained exposure periods – application to control of the healthy worker effect. Mathematical Modelling. 1986;7(9–12):1393-1512
- 2.
- Dawid AP, Didelez V. Identifying the consequences of dynamic treatment strategies: A decision theoretic overview. Statistics Surveys. 2010;4:184-231.
- 3.
- Chakraborty B, Moodie E. Statistical Methods for Dynamic Treatment Regimes: Reinforcement Learning, Causal Inference and Personalized Medicine. Springer; 2013.
- 4.
- Cain LE, Saag MS, Petersen M, May MT, Ingle SM, et al. Using observational data to emulate a randomized trial of dynamic treatment-switching strategies: an application to antiretroviral therapy. Int J Epidemiol. 2016 Dec 1;45(6):2038-2049. DOI: 10.1093/ije/dyv295