Target trial emulation: teaching epidemiology and beyond

Most important advances in epidemiologic methods in recent decades have been coupled with increased complexity both statistically and in the assumptions required for unbiased estimation. As much as g-methods, quasi-experimental methods, and other causal inference tools have improved the quality of observational research, understanding these methods well enough to apply them properly requires a substantial time investment. Trial emulation is different in this respect because the knowledge required to understand and apply it is already part of the traditional epidemiology curriculum. García-Albéniz et al. are not applying new methodologic principles per se: they are carefully applying design principles from trials to observational data where they have sometimes not been applied judiciously. In order to teach trial emulation to epidemiologists, introductory epidemiology courses require little augmentation. No new vocabulary or mathematical notation is needed beyond that already included in the teaching of trials. In fact, the inclusion of trial emulation creates a logical transition from the teaching of trials to observational research [9].

One nice consequence of teaching causal inference principles via trial emulation is that it clarifies immediately why some analytic practices are problematic. Consider how García-Albéniz et al. describe the two common but biased approaches they conduct alongside their preferred analysis. Without the trial emulation framework, the rationale behind these biased approaches can appear to be logical. However, if one were taught from the beginning to think in terms of trial emulation, their flaws could quickly be identified. For example, when examining the cumulative incidence curves of each analysis, it is clear that only the trial emulation analyses would be able to reproduce the crossing curves expected in a screening trial [10, 11]. In addition to avoiding biases, the framework also reveals opportunities for using available data more efficiently, although this involves more methodologic complexity that may be best taught in an intermediate or advanced course.

It seems clear that trial emulation should be a core component to how causal inference gets taught in epidemiology curriculum. The heuristic questions that remain are when and to whom it should be taught, how much emphasis should it be given, and whether potential negative consequences exist if trial emulation becomes commonplace. Educators need to weigh whether the trial emulation merits displacing other core material in epidemiology courses; alternatively, it is possible that its inclusion streamlines or better conveys the fundamentals of observational epidemiology that are already taught in introductory classes and better connects those fundamentals to the statistical analyses taught in advanced classes. Educators who teach epidemiology to non-epidemiologists (e.g., policy-makers; medical researchers from other disciplines) may also see their students’ comprehension improving, as trial emulation seems more readily accessible than teaching causal diagrams or counterfactual notation while still being able to convey the same foundational principles.

Beyond improving teaching and specific research applications, embracing a trial emulation framework can provide opportunities for epidemiologists to be more creative with the scientific process. (It will be important, of course, to formalize and assess any such broadened methods.) Here we briefly describe two such examples.

First, consider the setting from García-Albéniz et al. in which they emulate a target trial of screening colonoscopy knowing that actual randomized trials are on-going but will not be published for years to come. If their study and studies emulating related target trials in other data all demonstrate similar long-term benefits of colonoscopies, would we necessarily need to wait for a decade of follow-up time to accrue in the actual trials to gain confidence in the effects of colonoscopy? Even after a shorter follow-up time from the actual trials, it is possible to see whether the short-term causal effects estimated in observational studies closely mirror what is seen at that point in the trials. If they indeed map closely onto one another, it may serve as an indication that our observational study is on the right track and make us more confident that we can make colonoscopy recommendations much sooner than waiting for the full results of the trial. If they do not map closely onto one another, it may motivated us to re-examine the means by which we emulated random assignment (e.g., what unmeasured confounding could remain?). However, even if random assignment is successfully emulated, it is important to note that trial emulation will not guarantee comparability with real trials. Differences between populations being studied or other reasons the target trial protocol is not exactly aligned with the real trial may also cause estimates to differ [12]. By understanding when and why results agree, however, we can combine the powers of observational studies and actual on-going trials to support more timely conclusions, and continue to refine those conclusions as the trials continue.

Second, consider the rather common setting of studying some treatment’s effect on a rare outcome, such as the effect of statin medications on cardiovascular disease and death. To conduct a trial in a population of all indicated patients with a suitable sample size may not be feasible; instead, trialists may conduct a trial in a high-risk population (so that the outcome is not expected to be so rare) or they may study a surrogate outcome such as cholesterol levels. Neither is necessarily answering the questions most pressing to policy-makers, clinicians, or patients. With an observational study we could emulate the trial as conducted to see whether our estimates coincide with the trial. Armed with this “validation” of our observational study results, we can modify the target trial in our observational analysis to estimate other quantities such as effects in a more heterogeneous population or effects on endpoints of interest rather than surrogate endpoints. That is, we can combine results from actual randomized trials that did not target exactly the ideal question with the powers of observational studies to support more robust conclusions beyond those produced by a trial or an observational study alone.