Authors: Miguel A. Hernán (1.CAUSALab, Department of Epidemiology, Harvard T.H. Chan School of Public Health, Boston, MA; 2.Department of Biostatistics, Harvard T.H. Chan School of Public Health, Boston, MA), Issa J. Dahabreh (1.CAUSALab, Department of Epidemiology, Harvard T.H. Chan School of Public Health, Boston, MA; 2.Department of Biostatistics, Harvard T.H. Chan School of Public Health, Boston, MA; 3.Richard A. and Susan F. Smith Center for Outcomes Research in Cardiology, Beth Israel Deaconess Medical Center, Boston, MA), Barbra A. Dickerman (1.CAUSALab, Department of Epidemiology, Harvard T.H. Chan School of Public Health, Boston, MA), Sonja A. Swanson (4.Department of Epidemiology, University of Pittsburgh School of Public Health, Pittsburgh, PA)
Categories: Article
Source: Annals of internal medicine
Authors: Miguel A. Hernán, Issa J. Dahabreh, Barbra A. Dickerman, Sonja A. Swanson
When randomized trials are not available to answer a causal question about the comparative effectiveness or safety of interventions, causal inferences are drawn using observational data. A helpful 2-step framework for causal inference from observational data is 1) specifying the protocol of the hypothetical randomized pragmatic trial that would answer the causal question of interest (the target trial), and 2) using the observational data to attempt to emulate that trial. The target trial framework can improve the quality of observational analyses by preventing some common biases. In this article, we discuss the utility and scope of applications of the framework. We clarify that target trial emulation resolves problems related to incorrect design but not those related to data limitations. We also describe some settings in which adopting this approach is advantageous to generate effect estimates that can close the gaps that randomized trials have not filled. In these settings, the target trial framework helps reduce the ambiguity of causal questions.
Randomized trials are often not available to answer a causal question about the comparative effectiveness or safety of interventions. In such cases, causal inferences are drawn based on observational data. The use of observational data to explicitly emulate a hypothetical randomized trial—the target trial—has received much attention from researchers in recent years (1). Antecedents to the idea of the target trial emerged in statistics and econometrics in the middle of the 20th century (2–5), and concepts akin to “target trial emulation” were the basis of subsequent causal inference discussions (6–9) until its general formalization by Robins in 1986 (10).
The increasing popularity of target trial emulation as an approach to causal inference from observational data raises questions about the utility and scope of this approach. Two key questions are “Why is it helpful?” and “When is it helpful?”
In this article, we answer both questions. To do so, we propose an updated (11) structure for target trial emulations, clarify what problems target trial emulation does and does not solve, and describe some settings where adopting this approach is advantageous. We start with a brief review of the reasons that randomized trials are a deservedly trusted tool for causal inference in health research.
Randomized trials are trusted to guide decisions about treatment strategies (or interventions or policies or regimens) for 3 They ask a well-defined causal question, describe a plan to collect data, and incorporate a mechanism (randomized assignment of treatment) to answer the causal question using the collected data without the need for unverifiable assumptions, provided that the data were collected as planned. The random assignment of treatment strategies to study participants is the defining characteristic of randomized trials.
Well-conducted randomized trials ask well-defined causal questions because their protocols specify the eligibility criteria for the study participants, the treatment strategies that will be compared, the method of treatment assignment (for example, double-blinded), the outcomes of interest, the start and end of follow-up, and the causal contrast (for example, intention-to-treat effect). These 6 components of the trial protocol articulate an unambiguous question about a causal effect, which is referred to as the causal estimand. A seventh component of the protocol is the description of the statistical analysis, which can be referred to as the estimation procedure or estimator of the causal effect.
In many randomized trials, the causal estimand involves an intention-to-treat contrast (a contrast of the outcome distribution under assignment to different treatment strategies). The corresponding estimator is based on an intention-to-treat analysis—that is, a comparison of the outcome distribution between groups that were assigned to different treatment strategies, regardless of whether the participants actually received the assigned treatment. This intention-to-treat effect can therefore be referred to as the effect of assignment. An intention-to-treat analysis (estimator) is expected to yield an unbiased estimate of the effect of assignment (causal estimand) because, under random assignment of the treatment strategies, groups of participants assigned to different strategies are expected to be comparable at the time of assignment. Therefore, in the absence of losses to follow-up, postassignment differences in outcomes between groups can be attributed to their different treatment assignments rather than to different characteristics of participants in each group.
The effect of assignment is not the only relevant causal contrast. When some trial participants do not adhere to the treatment strategy that they were assigned to, we may want to consider additional causal contrasts, such as the per protocol effect (the effect of receiving the assigned treatment strategies as indicated in the trial’s protocol). Unlike for the intention-to-treat effect, unadjusted estimates of per protocol effects (often obtained by excluding participants when they stop following their assigned strategy) may not be valid because the receipt—as opposed to the assignment—of treatment is not randomized. Therefore, estimating per protocol effects without bias requires analyses whose validity relies on assumptions that are not justified by randomization alone. Valid estimation of the per protocol effect may require that baseline and time-varying confounders be measured and appropriately adjusted for (12). Regardless of whether the causal contrast is the intention-to-treat effect or the per protocol effect, assumptions are also needed if some participants are lost to follow-up. If competing events are present (for example, death when the outcome is some other event), they need to be incorporated in the specification of the causal contrast, whose valid estimation may require additional assumptions (13, 14).
Leaving aside nonadherence, losses to follow-up, and competing events, the main limitation of randomized trials is that they may not exist because they are impractical, unethical, or untimely or because the sheer number of causal questions of interest is much greater than the number of randomized trials that can be conducted. Often the only available human data are observational. A natural approach is then to use the observational data to emulate a hypothetical randomized trial—the target trial—that would quantify the causal effect of interest (11).
As discussed earlier, randomized trials are trusted for causal inference because they ask a well-defined causal question and describe how to generate the data to answer it, as described in their protocol, and because they incorporate randomized assignment to the treatment strategies under study. Observational studies, by definition, do not incorporate randomized assignment, but there is no reason they should not ask well-defined causal questions. An overemphasis on the lack of randomization has sometimes caused investigators to neglect the importance of a well-defined causal question in observational studies. In fact, the lack of an explicitly stated, well-defined question is a common problem of causal inference from observational data, which leads to inappropriate data analyses and estimates that are hard to interpret.
For example, consider 2 noteworthy failures of observational causal inference in the late 20th century. Observational studies reported apparently strong protective effects of hormone therapy on coronary heart disease and of statins on cancer, but randomized trials did not confirm these benefits (15, 16). Subsequent investigations revealed that these failures were not explained by limitations of the observational data but by errors in how the data were used (17, 18). In fact, the discrepancies between the observational studies and the randomized trials disappeared or were greatly attenuated in reanalyses that explicitly specified the causal estimand via the specification of a target trial protocol and emulated the target trial using the observational data. That is, a prerequisite for answering the causal question was translating it into an estimand that included a contrast of well-defined interventions.
The table in the Figure summarizes the key elements of the specification of the protocol of a target trial and its emulation using existing observational data, either from observational studies or from other data sources. The table reorganizes and expands on previously proposed items for the target trial framework (11), which were themselves an expansion of previous proposals (19).
In this updated format, the Specification column is organized into 3 1) the elements that define the causal estimand; 2) the assumptions made to identify the causal estimand in the target trial; and 3) the estimator (data analysis) that, under the identifying assumptions and any additional modeling assumptions (also listed here), would yield a valid estimate of the causal estimand if the target trial were to be conducted. The identifying assumptions are, informally, the assumptions that are needed to quantify (or identify) the effect even if the sample size were infinite. For example, if the target trial with randomized assignment were to be conducted, the identifying assumption of no baseline confounding, used for the intention-to-treat effect, would be expected to hold by design. However, as described earlier, assumptions about potential selection bias due to loss to follow-up or competing events would generally be needed for the intention-to-treat effect, even with random assignment at baseline. Also, assumptions about baseline and time-varying confounding are generally required to identify the per protocol effect in the target trial.
It is important to note that the target trial is not an idealized randomized trial that the investigators would design and conduct in the absence of constraints. Rather, the target trial is a randomized trial that can be reasonably emulated with the available observational data. For example, when outcome ascertainment occurs as part of routine clinical care in the observational data, the target trial is specified as a trial in which the outcomes of participants are ascertained in the course of routine care (for example, cancer diagnosis after symptoms or screening) rather than in an experimental setting (for example, cancer diagnosis after periodic magnetic resonance imaging). That is, the available data impose constraints on what can and cannot be emulated such that the target trial is typically a pragmatic trial. An implication is that investigators specifying a target trial need to have sufficient knowledge about the data that will be used to emulate it. Specification and emulation are not independent A trial that cannot possibly be emulated cannot be the target of an actual observational analysis.
The Emulation column in the table outlines how each element of the causal estimand will be mapped to existing data (or measured if new data are collected), the identifying assumptions, and the data analysis under those assumptions plus any additional modeling assumptions.
The emulation begins with the mapping of eligibility criteria, treatment strategies, assignments, outcomes, and follow-up to the available data items. The process usually is to construct a data set that looks approximately as if the target trial had actually been conducted. In the observational data, individuals may not have been selected according to a set of eligibility criteria, but by mapping each eligibility criterion to observational data items, we can build a data set as if individuals had been selected according to those eligibility criteria. In the observational data, treatment may not have been administered according to the treatment strategies specified by the target trial, but by mapping the components of the treatment strategies to observational data items, we can build a data set as if some individuals had been following the treatment strategies. In the observational data, the investigators do not assign treatment strategies, but by mapping the concept of assignment to observational data items (for example, prescription of a treatment), we can classify individuals as if they had been assigned to that treatment strategy.
The quality of the emulation depends on how close the mapped data items are to the data that would have been generated by conducting the target trial. In the Emulation column in the table, the mapping procedures for the elements of the causal estimand are described and any required deviations are noted. For example, if an eligibility criterion can only be partly operationalized, then the Emulation column must explain that, and sensitivity analyses can be conducted to explore the impact of different possible mappings. Also, as mentioned earlier, specification and emulation are not independent processes. For example, if one of the eligibility criteria cannot be mapped to the existing data at all, the Specification column should be modified by removing that criterion.
Even if the mapping to the elements of the causal estimand were perfect, a fundamental challenge for the emulation There is no randomization in the observational data. In fact, the randomized assignment is the only necessary difference between a target trial and an analysis of observational data that emulates it. Emulating a target trial with observational data requires assumptions that, if they held true, would make up for the lack of randomized assignment and support the identification of some causal effects of interest as if randomized assignment had actually occurred. The table provides an example of emulation in which the key identifying assumption for the intention-to-treat effect is the assumption that the baseline confounders are known and measured; this is also referred to as conditional exchangeability between groups assigned to different treatment strategies (which roughly corresponds to the terms selection on observables and no omitted variable bias used in other disciplines). However, the framework can be used with other identification strategies, such as instrumental variable estimation (20, 21), difference-in-differences (22, 23), or proximal causal inference (24). Also, similar assumptions are generally required in both the target trial and its emulation to identify the per protocol effect and to deal with potential selection bias due to loss to follow-up or competing events. Sensitivity analyses can be conducted to explore the impact of different identifying assumptions.
By explicitly specifying eligibility criteria, treatment strategies, assignments, outcomes, follow-up, and causal contrast, investigators can articulate precise causal estimands without being versed in technical language involving counterfactuals (25, 26). By explicitly specifying the assumptions and data analysis of the target trial, investigators can conduct analyses of the observational data that naturally parallel those of a randomized trial. That is, the target trial framework facilitates both the articulation of the question, especially when the question involves treatment strategies that are sustained over a period of follow-up, and the implementation of sound procedures to answer it (26).
As a result, an explicit target trial emulation helps to eliminate design biases arising from decisions made by the investigators when designing their analyses of observational data. Common design biases, such as those involving selection and immortal time (27, 28), are prevented because by specifying the protocol of the target trial, the study population is defined by eligibility criteria that are met at the time of assignment to a treatment strategy rather than at some earlier or later time. For example, an analysis in which one treatment group is restricted to people who never started treatment during follow-up will be viewed as suspect because it can introduce selection bias. Similarly, an analysis in which one treatment group is restricted to people who started treatment at some point during follow-up will be viewed as suspect because it can introduce immortal time bias. The aforementioned examples of the effects of hormone therapy on coronary heart disease and of statins on cancer illustrate the design biases that can arise in practice when eligibility determination and assignment to treatment strategies are not synchronized.
An explicit target trial emulation can help avoid design biases but does not eliminate data biases arising from data limitations, such as measurement error and insufficient information to adjust for confounding and selection bias from losses to follow-up. A target trial emulation that relies on adjustment for baseline confounders will fail when important confounders are unmeasured or mismeasured in the observational data. A goal of target trial emulation is preventing design biases so that investigators can focus their attention on data limitations (such as unmeasured confounding) and on how those limitations could be mitigated in future studies.
Therefore, the target trial framework is a structured procedure to operationalize good practices for study design, data analysis, and reporting. Of course, these good practices can be appropriately implemented in observational research without explicit reference to a target trial, as investigators have done in the past and continue to do now. Yet these good practices are not universally incorporated in observational research in the health and social sciences. Many investigations continue to be affected by design biases that often go unnoticed and result in misleading analyses. The flaws tend to be more severe when it is necessary to account for time-varying treatments and confounders, such as when the causal question involves treatment strategies that are sustained over a period of follow-up.
The target trial framework can then be viewed as methodological scaffolding for investigators, especially those who are not expected to be experts in causal inference methods and who may have trouble translating the technical literature into good research practices. It is important to note that the target trial framework is not a statistical method or a particular estimation procedure but a unifying approach for asking and attempting to answer causal questions using observational data. From a statistical standpoint, target trial emulation provides few or no methodological innovations because most components of the framework have been described previously.
To understand when the target trial framework is helpful for causal inference from observational data, consider 2 (i) the treatment strategies in the causal estimand are based on sufficiently well-defined interventions, and (ii) the investigators can map the components of the target trial protocol to existing or newly generated data. Under these 2 conditions, analyses of observational data for causal inference can typically be recast as the emulation of a target trial. Explicitly specifying and emulating the target trial clarifies the causal question, helps avoid errors in the design of observational analyses, and facilitates transparent reporting of any compromises that investigators may need to accept to accommodate data limitations. As such, when conditions (i) and (ii) hold, in our experience, the target trial framework is especially helpful.
What if it is uncertain whether condition (i) holds? As an example, suppose there is reasonable debate among subject matter experts over whether a particular intervention is sufficiently well defined (for example, the effect of social media). In such cases, the target trial framework facilitates concrete discussion about this disagreement. We can also consider more extreme examples in which the causal estimand is based on ill-defined interventions that may be too vague for articulation of a hypothetical experiment (such as the effect of loneliness). For these types of questions, it may be argued that meaningful quantitative causal inference is not possible, regardless of the chosen identification strategy and estimator (29, 30). However, engaging with investigators who pose such a question and asking them to articulate the causal estimand by specifying a target trial protocol may lead to targeting better-defined interventions.
What about when condition (ii) does not hold? Suppose we would like to study the causal effect of tax reform on the next generation’s life expectancy in the United States. This causal question might be expressed as a thought experiment that compares the outcomes in 2 one in which the tax reform was approved by the U.S. Congress in, for example, March 2025, and another one, identical to the first one until March 2025, in which the tax reform was not approved and the tax policy remained unchanged for another 30 years. Even if this counterfactual contrast were taken to be sufficiently well defined, and thus condition (i) was thought to hold, no data set contains information on life expectancy over several decades for 2 or more “United States” under different tax policies because there is only 1 United States. Therefore, no observational data set can be used to emulate a hypothetical experiment in which we observe different “United States” under different tax policies over the same period.
Consider 3 possible approaches to answering causal questions like this one for which condition (ii) does not hold. One possibility is to compare subsets of the United States (for example, states or counties) that are subject to different tax policies over the same period. However, data from subsets of the population may not suffice to make inferences about the effect of scaling up the policy change to the entire population, particularly when scaling up the intervention changes the system (or the intervention itself) in fundamental ways. If the new tax policy were actually implemented in the United States in 2025, another possibility would be to combine information from countries other than the United States to estimate the outcome of a “synthetic United States” under no policy change and, a generation later, compare it with the outcome of the actual United States. This approach requires very strong assumptions about stability and comparability across countries (31). Further, these 2 approaches may require waiting many decades to obtain the effect estimate. A third approach is building a mathematical model of the system’s structure to simulate the outcomes under different policies. This approach is often used in public health, economics, and other disciplines for questions that involve system-wide interventions, long-term horizons, or other components of the question that cannot be directly mapped to data. In these cases, identifying the causal effect typically requires unverifiable modeling assumptions about the mechanisms that govern the system (32).
Put simply, target trial emulation is not possible when the corresponding randomized trial is not possible, not even in principle. First, when the treatment strategies are not based on sufficiently well-defined interventions, neither a randomized trial nor its observational emulation is feasible because the causal question is too vague (29, 30). More technically, the counterfactual contrast is ill defined. Second, when the components of the target trial protocol cannot be mapped to data because no suitable data exist or can ever exist, neither a randomized trial nor its observational emulation is feasible because the causal question is too complex. That is, the counterfactual contrast may be well defined, but the causal question involves a thought experiment that is not tethered to the real world (for example, what if the United States could be subject to different tax systems in parallel universes), and thus an observational emulation is infeasible.
The limitations of feasible randomized trials may also be present in their emulations. In particular, the inferences from a randomized trial conducted in a particular study population at a particular time may not apply to another population with different characteristics or to the same population at a different time (33). For example, a trial that provides a valid estimate of the effect of assignment to a COVID-19 vaccine among people previously unexposed to the virus may not appropriately quantify the effect in previously exposed people (different effect modifiers at baseline); in a community under lockdown (different interference patterns); or after changes in behavior, perhaps triggered by the trial itself (different environment). Neither a randomized trial nor its observational emulation can produce inferences that are automatically transportable across times, places, and populations.
Explicitly emulating a specified target trial when analyzing observational data prevents design-induced biases and makes assumptions more transparent. However, the most impactful contribution of the target trial framework is reducing the ambiguity of causal questions because we cannot provide valid answers if we do not know what the question is.