Target trial emulation for addiction research—A useful method but not a panacea

Target trial emulation makes causal research using observational data more robust and transparent and presents an opportunity to increase the rigour and relevance of research in addiction. However, it is not without its limitations, and there are avenues of research where it is not appropriate or useful. Randomised controlled trials (RCTs) are the gold standard for generating causal evidence. However, there is a growing understanding that randomised studies are not always possible, especially in public health [1], and in those cases alternate approaches are needed to produce high-quality causal evidence [2, 3]. One such method is target trial emulation (TTE) [4]. TTE is a framework that aims to design observational studies through the lens of replicating the ideal RCT for the question of interest [4]. First, the ideal RCT—called the ‘target trial’—is completely specified, which involves outlining the eligibility criteria, treatment strategies, assignment procedures, follow-up period, outcome(s), causal contrasts and data analysis plan, as would be done when designing an RCT. The observational study is then designed with the goal of ‘emulating’ the target trial. This approach requires researchers to critically consider and articulate the potential sources of bias that might prevent study findings from being interpreted causally and specify study design and methods that would optimally minimise those risks of bias [5]. TTE presents an opportunity to strengthen evidence on the comparative effectiveness and safety of substance use and related interventions, where RCTs would otherwise not be possible because of financial, ethical or other constraints. For example, in the context of treatment for opioid dependence, there exist evidence gaps relating to the comparative effectiveness of opioid agonists (methadone and buprenorphine) on rare events (e.g. mortality, perinatal and neonatal health outcomes). These gaps could be addressed with the larger cohort sizes and event counts that are provided by population-level administrative datasets [2, 6, 7]. Application of the RCT framework and related nomenclature in TTE studies has the advantage that they are more likely to be well understood—and therefore more widely accepted—among clinical audiences. Even so, TTE are not a panacea for producing causal evidence from observational data and have notable limitations in the context of addiction research. In many cases, the available observational data is not capable of answering the causal questions being posed of it. This is especially true when attempting to use observational data, which has already been collected, to answer a different question to the one for which the study was designed [8, 9]. TTE studies are unlikely to be feasible or valid if the necessary data to define eligibility, exposure or confounders at baseline are missing or poorly measured [4]. For example, observational studies of drug use cannot fully emulate placebo-controlled trials, because they lack randomisation, blinding and true control conditions [10]. More broadly, some causal questions cannot be answered using observational data, because they relate to exposures or treatments that are not available in the real world (e.g. novel pharmaceuticals). Fine-grained data, such as information on adherence to the treatment protocol, which is now commonly collected in RCTs, also may not be available in observational data, especially when using data from surveys and other discrete measurements. TTE can only be used when it is reasonable to consider the question in terms of a ‘target trial’. That is, it must be possible to frame the research question in some form of a RCT, which typically involves discrete comparisons, for example, between specified treatment groups and/or control. In addiction research, there are often scenarios where the causal contrast of interest relates to a continuous exposure and the marginal effect (the slope or rate of change). For example, investigating the effects of blood alcohol content (a continuous exposure) does not have a straightforward trial analogue, at least not without potential loss of power and nuance in the relationship. It is possible to consider such exposures in more simple terms, such as a many-group RCT with participants randomised to one of a large number of different doses [11], or a crossover trial, but doing so only approximately answers the research question, because assignment to all possible values of blood alcohol is not practically feasible. The validity of TTEs can also be constrained by the difficulty of identifying an appropriate comparison group, because treatment initiation is strongly influenced by clinical severity, prior substance use trajectories and social vulnerability, leading to substantial risks of confounding by indication [8]. TTE studies also have somewhat dubious interpretation when examining causal effects of things that are not ‘intervenable’, even in theory [12]. For example, the question of whether familial history of substance use alters the risk of drug-related harm is not one that can be answered via a RCT. This does not mean that it cannot be on the causal pathway of drug-related harm, but such effects cannot easily lead to meaningful change via public health implementation [13]. Several of the sources of biases to which TTE studies are subject, they share in common with other observational studies. Of particular note is immortal time bias, which occurs when, by design choice, there is a period of follow-up time where the study outcome cannot occur (e.g. from baseline to receipt of treatment). Specific methods have been created to handle this bias [14], however, limitations remain, particularly when exposure is ascertained based on discrete measurements (e.g. annual surveys) rather than administrative or linked data. Confounding is another common cause of bias common to both TTE and observational studies, requiring the measurement and appropriate adjustment of all confounders of the treatment–outcome relationship on occasion, in a time-varying nature. It is worth noting that observational causal evidence can still be valid and informative, even when a target trial cannot be perfectly specified. Consider the question of whether delaying the initiation of drug use until adulthood reduces later drug-related harm. In this case, it is not possible to define a hypothetical intervention with a well-specified assignment mechanism, and any estimated effect would depend heavily on the underlying mechanism that led to that change in exposure [13]. For example, delaying use by restricting supply might have very different downstream consequences to delaying use through shifts in social norms or attitudes to drug use. In this scenario, a TTE may offer limited value over a well-designed observational study. However, this does not invalidate the causal insights that may emerge from such a study. Rather, it means that, although an observational study may demonstrate a causal effect of delaying drug use on related harm, the real-world impact of any specific intervention aiming to delay drug use will depend on the mechanism by which the delay is achieved and result in varied outcomes. TTE offers a powerful opportunity for strengthening causal inference in addiction research through observational data, but its applicability is limited to research questions that can be reasonably framed as hypothetical RCTs. Many important causal questions involve continuous exposures, non-intervenable variables and nuanced mechanisms of action that do not lend themselves to trial-like comparisons. In these cases, a well-designed observational study—grounded in strong theoretical reasoning and transparent assumptions—may provide more appropriate and interpretable evidence. Rather than treating TTE as a universal solution, researchers should view it as one tool among many in the broader causal inference toolkit, choosing methods that best align with the nature of the question, the structure of the data and the feasibility of hypothetical interventions. Philip James Clare: Conceptualization (equal); writing—original draft (lead); writing—review and editing (equal). Chrianna Bharat: Conceptualization (equal); writing—original draft (supporting); writing—review and editing (equal). None.