Target Trial Emulation: A Method for Causal Inference

Target trial emulation is a framework for drawing cause-and-effect conclusions from observational data by explicitly designing the analysis to mimic a hypothetical randomized trial. The core idea is straightforward: before touching any data, researchers write out the full protocol of the randomized trial they wish they could run, then use real-world records to approximate that trial as closely as possible. This two-step process, first articulating the ideal experiment and then emulating it, has become one of the most influential developments in modern epidemiology because it forces researchers to confront design choices that observational studies have historically glossed over.

Why Observational Data Needs a Trial-Shaped Scaffold

Randomized controlled trials remain the gold standard for establishing whether a treatment works. But trials are expensive, slow, and sometimes impossible to conduct. You cannot ethically randomize people to smoke for thirty years, or withhold a therapy already known to be lifesaving. In other cases, a trial might be feasible but impractical: studying a rare cancer subtype or comparing dozens of medication sequences would require enormous sample sizes and decades of follow-up.

Observational data, from electronic health records, insurance claims databases, and disease registries, already exists in staggering volumes. The challenge is that simply analyzing these records with conventional regression methods often produces biased answers. Patients who start a medication differ from those who do not, and the timing of treatment decisions introduces subtle distortions that can make a drug look harmful when it helps, or effective when it does nothing. Target trial emulation addresses this by imposing the structural discipline of a clinical trial onto observational analysis.

The Core Mechanics

The framework asks researchers to specify every component a real trial protocol would include: eligibility criteria, treatment strategies being compared, how patients are assigned to those strategies, what outcomes are measured, when follow-up begins, and how the data will be analyzed. This is not just a checklist exercise. The act of writing out a full protocol reveals ambiguities that looser observational designs leave unresolved. Who exactly counts as a “new user” of a drug? When does the clock start ticking for outcomes? What happens to people who switch treatments midway through?

The single most important design requirement is aligning three things at what researchers call “time zero”: eligibility must be confirmed, treatment must be assigned, and follow-up must begin, all at the same moment. In an actual randomized trial this happens naturally on the day a participant is enrolled and randomized. In observational data, these three events often occur at different times, and failing to synchronize them is one of the most common sources of bias in non-experimental research.¹

How Time Zero Alignment Prevents Common Biases

Consider a study asking whether starting statins after a heart attack reduces the chance of dying within five years. A naive approach might identify everyone who had a heart attack, check who eventually started statins, and compare death rates. The problem is that people who died quickly never had a chance to start statins. By classifying them as “untreated,” the analysis makes the untreated group look worse than it really is, and makes statins look better than they might be. This is called immortal time bias: the treated group is guaranteed a stretch of survival (the time between the heart attack and starting statins) that the comparison group is not.

Target trial emulation prevents this by insisting that treatment assignment happens at the same moment eligibility is confirmed and follow-up begins.² In the statin example, the emulated trial would enroll patients on the day they become eligible (say, the day of hospital discharge after a heart attack) and classify them as “initiators” or “non-initiators” based on what happens at that moment. If someone starts statins three months later, they enter a new emulated trial at that later time point. This approach, sometimes called a sequence of nested trials, removes the artificial survival advantage that plagues simpler designs.

The same alignment logic prevents other distortions. Selection bias creeps in when eligibility depends on future events. Confounding by indication arises when the reasons a patient receives treatment also affect their outcome. By forcing researchers to specify exactly when and how each design element locks into place, the framework makes these problems visible before the analysis begins rather than after the results are published.³

Intention-to-Treat and Per-Protocol Effects

Real randomized trials report two types of results. The intention-to-treat effect compares groups based on the treatment they were initially assigned, regardless of whether they actually took it. The per-protocol effect compares groups based on what they actually did, limited to people who stuck with their assigned treatment. Both are useful for different questions: the first tells you what happens when a policy assigns a treatment, the second tells you what happens when someone actually follows through.

Target trial emulation can estimate both. The intention-to-treat analog is relatively simple: classify people at time zero and follow them regardless of what they do later. The per-protocol analog is trickier, because in observational data, the people who stop or switch treatments are often systematically different from those who continue. To handle this, researchers typically censor people when they deviate from the protocol and then use statistical weighting to adjust for the fact that those who deviated may have been sicker or healthier than those who stayed on track. A study emulating trials of statins and antihypertensives demonstrated this approach using electronic medical records. The intention-to-treat estimate for statins versus no treatment showed a modest mortality reduction, while the per-protocol estimate, restricted to people who stayed on therapy and adjusted for adherence, showed a substantially larger benefit.⁴

Does It Actually Match What Trials Find?

The most important question about any observational method is whether it gets the right answer. A 2025 systematic review and meta-analysis tackled this directly, comparing the results of target trial emulations with those of actual randomized trials asking the same questions. Across 106 emulation-trial pairs, the overall agreement was reasonable: about four in five pairs agreed on the direction and rough size of the effect. When the researchers narrowed the comparison to 62 pairs where the emulation more closely mirrored the trial’s design, agreement climbed further, with a correlation of 0.83 and concordance in roughly seven out of eight pairs.⁵

This tells us something important: the method works better when researchers follow its principles more carefully. Emulations that cut corners on eligibility criteria or time zero alignment were more likely to disagree with trial results. The method is not a magic wand; it is a discipline, and the results depend on how rigorously it is applied.

Where It Has Been Applied

Target trial emulation has spread rapidly across clinical fields. In cardiovascular medicine, a large emulation study examined whether starting statin therapy benefits older adults with type 2 diabetes, a group often underrepresented in randomized trials. Among matched patients aged 75 to 84, statin initiation was associated with a roughly 30% lower risk of cardiovascular events and a 35% lower risk of death, with no meaningful increase in muscle-related side effects or liver problems. The benefits held even among those aged 85 and older.⁶ This kind of finding fills a genuine evidence gap: few trials enroll octogenarians, and clinicians need guidance for the patients sitting in front of them.

In oncology, the framework has been used extensively. A systematic review identified 90 target trial emulation studies in cancer, with registry databases serving as the most common data source and overall survival as the most frequent endpoint.⁷ One study used the U.S. SEER-Medicare database to emulate trials of chemotherapy after surgery for colon cancer and of erlotinib for lung cancer. The emulated populations were older and more representative of real-world patients than those in the original trials, offering a view of treatment effects in people trials typically miss.⁸

In diabetes care, a systematic review of observational studies emulating cardiovascular outcome trials for type 2 diabetes medications found that real-world data results were concordant with trial findings across all drug classes examined.⁹ This concordance is reassuring, and it means clinicians can have some confidence that emulations of diabetes drug effects are producing directionally reliable answers.

Dynamic Treatment Strategies

One of the more powerful extensions of the framework handles questions that static trial designs struggle with. In practice, treatment is rarely a one-time decision. Clinicians adjust doses, switch medications, and add therapies based on how a patient responds. Target trial emulation can model these evolving decisions by comparing “dynamic” treatment strategies: rules that say “give drug A, and if the patient does not respond within eight weeks, switch to drug B.”

A study of corticosteroids for COVID-19 illustrated this well. Rather than comparing “steroids yes” to “steroids no,” the researchers emulated a trial of a dynamic strategy: give six days of corticosteroids if and when a patient meets severe oxygen-deprivation criteria, versus never giving steroids. This more closely reflects how the drug was actually used in hospitals.¹⁰ Similarly, researchers have emulated dynamic strategies for major depressive disorder, comparing sequential approaches like “try one antidepressant, then switch class if needed” against “start with combination therapy.”¹¹

Modeling dynamic strategies typically requires a “cloning” procedure, where each patient is duplicated and assigned to every strategy being compared at time zero, then censored when their actual behavior diverges from the strategy they were cloned into. Inverse probability weighting corrects for the informative nature of this censoring.¹² The mechanics sound intricate, but the payoff is that researchers can answer questions no single randomized trial has asked.

Where the Framework Runs Into Trouble

Target trial emulation does not eliminate every problem with observational data. Its most fundamental limitation is unmeasured confounding. If a variable that affects both treatment choice and outcome is missing from the dataset, no amount of careful design can fully correct for it. Researchers can quantify how strong such unmeasured confounding would need to be to explain away a finding. One study of PCSK9 inhibitors, for example, calculated that an unmeasured confounder would need to increase the risk of death sevenfold and be sevenfold more common among treated patients to account for the observed mortality difference.¹³ That is a high bar, which builds confidence in the result, but it does not prove confounding is absent.

Data quality is another persistent concern. Electronic health records are built for billing and clinical documentation, not for research. Diagnoses may be coded inaccurately, lab values may be missing, and the timing of events can be imprecise. A study combining target trial emulation with a nested case-control design found that even when diagnostic codes for heart attack had 95-96% sensitivity and 98% specificity compared to manual chart review, that small measurement error was enough to introduce substantial bias into the results.¹⁴ The lesson: rigorous methodology cannot compensate for data that misclassifies who actually had the outcome.

A scoping review of published target trial emulations also found inconsistency in how well researchers specified the critical components of the framework. Not all papers clearly defined time zero, explained how they simulated random assignment, or described the comparison strategy.¹⁵ The framework provides structure, but like any tool, its output depends on the skill and thoroughness of the person using it.

Machine Learning Meets Trial Emulation

A newer line of work explores whether machine learning can improve the statistical steps within target trial emulation. Traditional approaches rely on researchers manually selecting which variables to adjust for and specifying the mathematical relationship between those variables and the outcome. Machine learning models, including deep learning architectures like Transformers, can process entire longitudinal patient records and learn complex patterns that a human analyst might miss or never think to model.

A study evaluating bias in heart failure emulations compared traditional methods against a Transformer-based deep learning approach that ingested the full sequence of a patient’s electronic health record before the study start date. The deep learning model aimed to capture confounding structures embedded in the trajectory of a patient’s care history, going beyond the handful of variables a researcher would typically select.¹⁶ This kind of approach is still being evaluated, and it trades interpretability for flexibility. You can see what variables a logistic regression adjusts for; understanding what a Transformer has learned from thousands of clinical notes and lab values is much harder.

On the statistical side, targeted maximum likelihood estimation (TMLE) has become increasingly common as a way to improve robustness. TMLE models both the treatment assignment and the outcome simultaneously, and it gives consistent results as long as at least one of those two models is correct. Studies of conditions ranging from acute pulmonary embolism to heart failure have incorporated TMLE alongside more traditional weighting methods as sensitivity analyses.¹⁷

Regulatory Interest and Real-World Evidence

Regulatory agencies have taken notice. The U.S. Food and Drug Administration’s growing interest in real-world evidence has created a natural opening for target trial emulation, particularly in oncology where trial recruitment can be slow and where post-marketing questions about treatment sequencing are common. A 2025 review in the Journal of Clinical Oncology argued that target trial emulation can supplement randomized trial evidence to inform both regulatory and clinical decision-making in cancer, while acknowledging that understanding its strengths and limitations is essential for integrating real-world evidence responsibly.¹⁸

The appeal for regulators is practical. After a drug is approved, questions multiply: Does it work the same way in older patients? In people with kidney disease? In combination with another drug approved three years later? Running new randomized trials for every such question is neither feasible nor always ethical. Emulating trials from the data that accumulates as millions of patients receive routine care offers a path toward answering these questions faster and more cheaply, provided the methodology is sound.

Software and Reproducibility

As the method has gained traction, dedicated software has followed. The R package TrialEmulation implements the statistical machinery needed to emulate a sequence of target trials using time-to-event data from electronic health records, including estimation of both intention-to-treat and per-protocol effects via marginal structural models.¹⁹ Having standardized software matters for reproducibility: when two research groups use the same package with the same data, they should get the same answer, which has not always been true when teams code the analysis from scratch.

Beyond analysis software, the broader ecosystem of support tools is expanding. A review identified 24 distinct tools spanning the design, implementation, and analysis phases of target trial emulation, with most tools concentrated in the analysis phase. The relative scarcity of tools for the design phase is telling. Getting the protocol right, specifying eligibility and time zero and the treatment strategies to be compared, remains largely a matter of clinical and epidemiological judgment rather than automated computation. That is probably as it should be: the hardest and most consequential decisions in any causal analysis are the ones made before any model is fit.

Generalizing Trial Results to Different Populations

A related but distinct problem is taking the results of an actual randomized trial and figuring out whether they apply to a population that differs from the one enrolled. A trial that recruited mostly white men under 65 may not tell you much about older women or people with multiple chronic conditions. Methods for “transporting” or “generalizing” trial findings to new target populations have developed in parallel with target trial emulation. These typically involve reweighting trial participants so they resemble the target population, or building outcome models that adjust for the differences between the two groups.²⁰

Target trial emulation and generalizability methods are complementary. The emulation framework answers “what would the trial have found if we could run it in this database?” while transportability methods answer “what would the trial have found if it had enrolled a different group of people?” Together they represent a broader shift toward making causal evidence more applicable to the patients clinicians actually treat, rather than the narrow, carefully screened populations that populate most trial enrollment logs.