RTI uses cookies to offer you the best experience online. By clicking “accept” on this website, you opt in and you agree to the use of cookies. If you would like to know more about how RTI uses cookies and how to manage them please view our Privacy Policy here. You can “opt out” or change your mind by visiting: http://optout.aboutads.info/. Click “accept” to agree.
To the Editor: We would like to object to the arguments made by Pearce and Vandenbroucke1 because they criticize a strawman. The misrepresentation they refute is that “(Target Trial Emulations [TTE])… assumes any causal analysis (…) has to be framed within this (randomized clinical trial) paradigm, focusing on a single ‘ideal’ study.” There is no such requirement: TTE does not impose restrictions on the number of studies to address a specific causal question, nor on the choice of design to emulate the target trial(s). It does require the specification of the (hypothetical) target trial that would answer the research question, but then researchers are free to apply any design feature to the underlying cohorts, for example, it has been shown that TTE can be used with case–control sampling, instrumental variables, and even postbaseline selection as in test-negative designs.
The authors’ focus on residual confounding betrays another misunderstanding: TTE’s main purpose is not controlling for confounding but rendering the research question clear and easy to communicate without obscuring it by mathematical formalisms. The focus of TTE on the alignment of time zero, eligibility, and exposure assignment helps avoid selection bias, which can be a substantial threat to validity. For example, when emulating a target trial using a test-negative design2 to address confounding due to unmeasured health-seeking behavior, it becomes clear that the target trial would discard participants who do not receive a test during follow-up—making this explicit helps evaluate if such selection introduces bias. More generally, we see no conflict between triangulation and TTE, as illustrated by the common use of negative and positive control outcomes in TTE.3
The authors additionally coin a novel concept of “matched target trial emulation” and criticize it by elaborating on the limitations of matching. Attributing the limitations of matching to the TTE framework is a logical error: TTE imposes no restrictions on methods to emulate randomization; approaches applied in TTEs include the g-formula, inverse probability weighting, or targeted machine learning. One of the advantages of matching, however, is its simplicity. Its first application in TTE allowed generating evidence on the effectiveness of COVID-19 vaccines within a few weeks,4 in a unique moment of human history when a global pandemic obligated to a massive deployment of vaccines with only limited evidence on effectiveness and safety.
In summary, we find that causal inference using observational data can only benefit from the process of explicitly formulating the target trial it aims to emulate. That does not make TTE a “gold standard” but rather a “best practice.”