Causal Inference
Formal causal inference in experimental N-of-1 trials Marco Piccininni* Marco Piccininni Stefan Konigorski Mats Julius Stensrud
Recent interest in precision medicine has challenged the relevance of average treatment effects estimated from randomized control trials (RCTs). Experimental N-of-1 trials aim at assessing the response to an intervention in a single individual.
N-of-1 trials are often characterized by a multiple cross-over design, multiple measurements of the outcome, random allocation of the treatment sequence, and blinding. Results are generally analyzed using standard statistical methods and claims that causal inference is not needed for experimental N-of-1 trials can be found in the literature, where confusion exists around concepts such as time-varying confounding and the guarantees given by randomization.
The aim of this work is to ground experimental N-of-1 trials in a formal potential outcomes framework for causal inference.
This allows us to define a conditional average treatment effect (CATE) that represents a natural target in the N-of-1 setting, termed the U-CATE. We discuss the assumptions needed for its identification and estimation under different data generation mechanisms. We consider settings in which carryover effects, trends over time and outcome autocorrelation are present, with emphasis on scenarios of particular relevance. We show how the simple mean difference may be a valid estimator of the U-CATE when interventions have effects limited in time and address symptoms of a stable medical condition. More complex scenarios require the g-formula to identify the U-CATE. We clarify the role of randomization in experimental N-of-1 trials, the link between RCTs and N-of-1 trials, and more generally, formalize intuitive claims from the N-of-1 trial literature. While the idea behind N-of-1 trials might seem simple, only causal thinking and approaching the problem rigorously can unveil the complex methodological aspects of this design.