03 · What You Need to Know
How Target Trial Emulation Works and What Makes It Credible
What Is the Basic Idea Behind Target Trial Emulation?
Many observational analyses begin with an existing dataset and search for associations between recorded exposures and outcomes.
Target trial emulation reverses that sequence. The researcher first defines the causal experiment that would ideally answer the question, then determines whether the observational records can support an approximation of that experiment.
For example, rather than asking whether frequent AI users obtain higher writing scores, the researcher might ask what would happen if eligible students followed one specified feedback strategy instead of another beginning at the same point in the semester.
This distinction matters because the observational association between actual users and nonusers may not represent the effect of assigning or implementing either strategy.
Hernán and Robins (2016) describe target trial emulation as a way to make causal questions explicit and avoid design problems that can arise when observational analyses are organized retrospectively around recorded treatment histories.
What Is the Difference Between a Target Trial and Its Emulation?
The target trial is the hypothetical randomized experiment that defines the intended causal comparison.
The emulation is the observational study designed to approximate that trial using available records.
The distinction is consequential. A target trial might assign students randomly to two feedback strategies, whereas the observational emulation must account for the fact that students selected or received those strategies through nonrandom processes.
The emulation should identify which components of the ideal trial can be reproduced and which cannot.
Calling an observational analysis a target trial emulation does not establish that the randomized trial's inferential advantages have been recreated.
What Components Must Be Specified?
| Protocol component |
Target trial question |
Observational implementation |
| Eligibility criteria |
Who could enter the trial? |
Identify eligible participants using information available at entry. |
| Treatment strategies |
What interventions would be compared? |
Define corresponding observable strategies or exposure histories. |
| Assignment |
How would treatment be randomized? |
Identify actual treatment selection and relevant confounders. |
| Time zero |
When would assignment and follow-up begin? |
Align eligibility, treatment initiation, and follow-up. |
| Follow-up |
For how long would outcomes be observed? |
Define the corresponding observation period and censoring rules. |
| Outcome |
What result would be measured? |
Use a valid, consistently defined recorded outcome. |
| Causal estimand |
Which treatment effect is targeted? |
Specify the effect that the observational analysis seeks to identify. |
| Analysis |
How would the trial estimate the effect? |
Use methods appropriate to the observational assignment and follow-up processes. |
These components should form a coherent protocol. The researcher should not define treatment using information from the end of follow-up while pretending that treatment status was already known at the beginning.
Why Is Time Zero So Important?
Time zero is the point at which eligibility, treatment assignment or initiation, and the beginning of follow-up are aligned.
Suppose the intended comparison begins at semester enrollment. Students should be eligible for the relevant strategies at that point, and their follow-up should begin according to the same temporal definition.
If AI users begin follow-up only when they first access the system but nonusers begin at enrollment, the groups may have different opportunities to experience the outcome.
Likewise, defining exposure using future behavior can create selection problems.
Hernán and colleagues (2016) explain how explicit target trial specification can prevent immortal time bias and other errors arising from inconsistent temporal definitions.
What Is Immortal Time Bias?
Immortal time bias arises when the classification or analysis of exposure incorrectly handles a period during which participants must remain event-free or under observation to qualify for a particular group.
For example, suppose sustained AI users are defined as students who complete ten consecutive weeks of platform use.
Students must remain enrolled and observed long enough to satisfy that definition. Those who withdraw earlier cannot qualify as sustained users.
If researchers compare these future-defined users with other students from semester start without accounting for this structure, the resulting association may be biased.
The problem is not solved simply by adding baseline covariates to a regression model.
Appropriate treatment definitions, time alignment, and analytical methods are needed.
How Are Treatment Strategies Defined?
Treatment strategies should describe meaningful intervention alternatives rather than vague exposure categories.
For example, offering access to an AI feedback platform is different from requiring weekly feedback sessions.
The former concerns access or assignment. The latter concerns sustained adherence to a specified instructional strategy.
Researchers should define when treatment begins, what participants are expected to do, whether deviations are allowed, and when the strategy ends.
Strategies may be static, such as a fixed number of feedback sessions, or dynamic, such as additional feedback triggered by specified student performance criteria.
However, dynamic strategies require sufficiently detailed longitudinal data to determine when their rules apply.
What Is the Difference Between Treatment Initiation and Sustained Treatment?
Treatment initiation concerns beginning an intervention at a defined point.
Sustained treatment concerns following an intervention strategy over time.
For example, initiating AI feedback during the first week of a course is not equivalent to completing a structured feedback activity every week.
The second question introduces adherence and potentially time-varying confounding.
Students' academic confidence, workload, and prior feedback experiences may influence continued use while also affecting later writing performance.
Researchers should not estimate the effect of sustained use merely by comparing participants who ultimately completed many sessions with those who did not.
What Is the Role of the Causal Estimand?
The estimand specifies the causal quantity the study seeks to estimate.
For a binary intervention, an average treatment effect may be represented as:
However, the meaning of these potential outcomes depends on how the treatment strategies are defined.
The effect of offering AI access differs from the effect of sustained weekly use.
Similarly, an effect among all eligible students differs from an effect among students with particular characteristics.
Researchers should define the estimand before selecting the statistical model, not infer the research question from whichever coefficient the dataset produces.
What Assumptions Are Required for Causal Interpretation?
Target trial emulation does not eliminate the assumptions needed for observational causal inference.
Conditional exchangeability requires adequate control of relevant confounding under the chosen strategy.
Positivity requires that the relevant treatment alternatives be possible across covariate histories in the target population.
Consistency requires that observed outcomes correspond appropriately to the potential outcomes under the treatment strategies actually followed.
Additional assumptions may concern censoring, interference, outcome measurement, and the statistical models used.
Hernán and Robins (2020) discuss these identification conditions and their importance for causal estimation.
A clearly written target trial protocol improves transparency but does not demonstrate that the assumptions hold.
How Does Confounding Affect the Emulation?
Unlike randomized assignment, observational treatment selection may depend on participant characteristics related to outcomes.
Students who begin using AI feedback may differ from nonusers in prior writing ability, motivation, digital access, or instructional support.
These differences can distort the intended causal comparison.
Researchers should identify plausible confounders using substantive knowledge and measure them at appropriate times.
Matching, weighting, standardization, or regression may address selected measured confounding under suitable assumptions.
However, statistical adjustment alone does not establish causation, and unmeasured confounding may remain.
What Is the Clone-Censor-Weight Approach?
Clone-censor-weight methods can be useful when participants are initially compatible with more than one treatment strategy, particularly when strategies include a grace period for treatment initiation.
In a simplified implementation, eligible participants are represented by copies assigned analytically to different strategies. Each copy is censored when its observed treatment history becomes inconsistent with its assigned strategy.
Appropriate inverse probability weighting may then address selection introduced by this artificial censoring under relevant assumptions.
For example, a study might compare starting an intervention within a defined enrollment window with not starting during that window.
However, cloning does not create new independent participants or remove unmeasured confounding. Artificial censoring and weighting require careful specification, and uncertainty estimation must account for the data structure.
The method is not required for every target trial emulation.
What Is Sequential Trial Emulation?
In some settings, participants may become eligible for treatment initiation at several time points.
Rather than define one baseline comparison, researchers may emulate a sequence of trials beginning whenever eligibility conditions are satisfied.
For example, students might become eligible for an additional feedback intervention at several assessment points during an academic program.
The same student may contribute to multiple emulated trials, provided the design and analysis handle repeated eligibility appropriately.
This approach can use longitudinal information efficiently, but it introduces dependencies and requires consistent definitions of eligibility, treatment, and follow-up.
It should be used when the underlying decision process warrants it, not merely because repeated records are available.
How Is Target Trial Emulation Different From Ordinary Propensity-Score Matching?
Propensity-score matching is one possible analytical strategy for addressing measured differences between exposure groups.
Target trial emulation is a broader design framework.
It specifies the causal question, eligibility, strategies, time zero, follow-up, outcome, estimand, and analysis before choosing a particular estimator.
A target trial emulation may use propensity-score methods, but matching alone does not ensure appropriate temporal alignment or prevent immortal time bias.
Conversely, a well-designed emulation may use methods other than matching.
The design framework and statistical estimator should therefore be distinguished.
Does Target Trial Emulation Require Large Administrative Databases?
No. The framework can be applied to different observational data sources when they contain the information needed to approximate the target trial.
However, large longitudinal databases are often useful because they may record eligibility, exposure changes, covariates, and outcomes over time.
Dataset size alone does not determine adequacy.
A large dataset may lack information about important confounders or the actual intervention received.
A smaller, carefully collected longitudinal dataset may provide more relevant measurements for a narrowly defined causal question.
Researchers should evaluate data suitability against the target protocol rather than assume that more records necessarily produce stronger causal evidence.
Can Educational Technology Studies Use Target Trial Emulation?
Yes, when the research question concerns a sufficiently well-defined intervention or strategy and the available observational data support the required design components.
Potential applications include comparing specified learning support strategies, evaluating intervention initiation, or investigating sustained participation in structured educational programs.
However, educational technology exposure is often difficult to define.
Platform logins, time spent online, and number of interactions may not represent equivalent forms of meaningful learning engagement.
Students may also modify their use in response to changing academic performance.
These features make intervention definition and time-varying confounding particularly important.
When Is Target Trial Emulation Not Appropriate?
The framework may be unnecessary when the research objective is descriptive, interpretive, or exploratory rather than causal.
For example, estimating the prevalence of AI use among university teachers does not require specifying a hypothetical randomized intervention.
Likewise, investigating teachers' experiences of AI governance may be better addressed through an appropriate qualitative design.
Even for causal questions, emulation may be infeasible when the intervention cannot be defined meaningfully or when the dataset lacks essential information about timing, confounding, or outcomes.
The framework should clarify the limits of available evidence, not be used as a label that makes an otherwise unsuitable dataset appear causal.
What Are the Most Common Mistakes in Target Trial Emulation?
One common mistake is defining treatment groups using information that becomes available only after follow-up begins.
Another is aligning time zero differently across exposure groups.
Researchers may also define vague interventions, overlook time-varying confounding, or assume that propensity-score adjustment eliminates every source of bias.
Some studies specify a target trial but do not explain how the observational data approximate each protocol component.
Others interpret an observational treatment-receipt effect as though it were the effect of randomized assignment.
These problems should be identified through explicit protocol specification and critical assessment of departures from the hypothetical trial.
How Should a Target Trial Emulation Be Reported?
Researchers should present the target trial protocol and the corresponding observational implementation clearly.
A table comparing the hypothetical and emulated components may be useful.
The methodology should specify eligibility, treatment strategies, time zero, follow-up, outcomes, estimands, and analysis.
It should also explain how treatment selection, confounding, censoring, missing data, and positivity were addressed.
Where assumptions cannot be verified, researchers should state their substantive justification and consider relevant sensitivity analyses.
The conclusion should distinguish what the emulation was designed to estimate from what the available data can credibly establish.
Watch Out
Do not define treatment groups using future behavior and then treat those groups as if they were assigned at baseline. Misaligned eligibility, exposure classification, and follow-up can create serious bias that ordinary statistical adjustment may not resolve.
