Intention to treat is a rule for analyzing the results of a clinical trial: every participant is counted in the group they were randomly assigned to, regardless of whether they actually took the treatment, switched groups, dropped out, or broke the study rules in any way. The idea sounds almost stubbornly simple, but it exists to prevent a specific and serious problem. Without it, researchers can unintentionally (or intentionally) skew results by shuffling people between groups after the fact. The principle is considered the default standard for randomized trials, yet the reality of how it gets applied is far messier than the textbook definition suggests.
What Problem Does It Solve
When researchers randomly assign participants to a treatment group or a control group, that randomization is doing something powerful: it balances out all the differences between people, both the ones you can measure and the ones you cannot, so that any difference in outcomes between the groups can be attributed to the treatment. The moment you start removing people from the analysis after randomization, you risk destroying that balance. Maybe the sickest participants stopped taking the drug because of side effects. Maybe the healthiest ones in the control group sought the treatment elsewhere. If you exclude those people or move them to a different group, you are no longer comparing two balanced groups. You are comparing two cherry-picked subsets, and your results could easily be wrong.
Intention-to-treat analysis prevents this by keeping everyone where randomization put them, no matter what happens next. A participant assigned to the drug group who never swallowed a single pill still counts as part of the drug group. A participant in the placebo group who somehow obtained the real drug still counts as placebo. This preserves the prognostic balance that randomization created and allows researchers to draw unbiased conclusions about the treatment’s effectiveness.1PubMed Central. Understanding the Intention-to-treat Principle in Randomized Controlled Trials
The trade-off is that an intention-to-treat analysis tends to give a conservative estimate of a treatment’s effect. If a bunch of people in the treatment group did not actually take the treatment, their outcomes will dilute the measured benefit. The analysis is answering the question “what happens when you assign this treatment?” rather than “what happens when someone actually takes it?” Those are different questions, and the distinction matters a great deal.2PubMed Central. Intention-to-treat concept: A review
The Real-World Question Versus the Ideal-Conditions Question
Intention to treat is not the only way to analyze a trial. Two common alternatives exist, and each answers a different question about the treatment.
A per-protocol analysis includes only the participants who followed the study rules as written: they took the right dose, showed up for all their appointments, and did not violate the protocol in any major way. This approach tries to answer whether the treatment works under ideal conditions, when everything goes according to plan. The problem is that the people who follow protocols perfectly are often systematically different from the people who do not. They may be healthier, more motivated, or less burdened by side effects. By restricting the analysis to compliant participants, you lose the protection of randomization and risk overstating how well the treatment works.3PubMed Central. Intention-to-treat versus as-treated versus per-protocol approaches to analysis
An as-treated analysis groups people by what they actually received rather than what they were assigned. If someone randomized to placebo ended up getting the drug, they move to the drug group for this analysis. This approach can be useful in certain situations, but it shares the same fundamental weakness as per-protocol: it breaks the randomization and opens the door to bias.
Intention to treat answers the pragmatic question, the one that matters most for real clinical decisions: if a doctor prescribes this treatment to a typical patient, what outcome should they expect? Some patients will not fill the prescription. Some will stop taking it because of side effects. Some will take it inconsistently. All of those messy realities are baked into the intention-to-treat estimate, which is exactly what makes it useful for guiding real-world practice.4Journal of Clinical Epidemiology. Intention to treat and per protocol analyses: differences and similarities The per-protocol analysis, by contrast, answers a more idealized question: does the treatment work when everything goes perfectly? Both questions are legitimate, but they serve different purposes.5PubMed. Intention to treat and per protocol analysis in clinical trials
The Missing Data Headache
The intention-to-treat principle sounds clean in theory: analyze everyone as randomized. In practice, it runs headfirst into a brutal problem. People drop out of trials. They move away, lose interest, get too sick to continue, or die. If someone’s outcome data does not exist, you cannot just pretend it does. Yet the principle says you must include them. This tension between “include everyone” and “we do not have everyone’s data” is one of the most contested issues in trial methodology.
A naive approach is to simply analyze the data you have and ignore the people who dropped out. This is sometimes called a complete-case analysis. It can work if dropouts are random, meaning the people who left the study are not systematically different from those who stayed. But that assumption is often wrong. If sicker participants are more likely to drop out, and dropout rates differ between the treatment and control groups, the remaining data gives a distorted picture of the treatment’s effect.6PubMed Central. Differential dropout and bias in randomised controlled trials: when it matters and when it may not
Researchers have developed several strategies for dealing with this. One of the oldest is “last observation carried forward,” where the participant’s last recorded measurement is treated as their final outcome. If someone dropped out at week four of a twelve-week study, their week-four measurement stands in for their week-twelve result. This is easy to implement but often misleading, especially in diseases that get worse over time. If the treatment was supposed to slow decline, carrying forward an early measurement can make the treatment look better than it actually was.
More sophisticated techniques include multiple imputation, where missing values are filled in many times using statistical models, and mixed models that use all available data without explicitly filling in gaps. These methods generally perform better, but none of them is foolproof. When the reason people dropped out is related to their outcomes in ways you cannot measure, no statistical technique can fully correct for the bias. Experts recommend using a primary analysis under reasonable assumptions and then running sensitivity analyses to see whether the conclusions change when those assumptions are varied.7PubMed Central. Strategy for intention to treat analysis in randomised trials with missing outcome data The honest truth is that there is no single strategy that works well for every combination of dropout rates, dropout reasons, and disease trajectories.8PubMed. Intention-to-treat: methods for dealing with missing values in clinical trials of progressively deteriorating diseases
This means that even when a trial claims to have used intention-to-treat analysis, the details of how missing data were handled can dramatically affect the results. Two trials can both call themselves “intention to treat” and reach different conclusions partly because they dealt with dropouts differently. If you are reading a trial report and the dropout rate is high, it is worth paying close attention to how the missing data were managed and whether the researchers tested alternative approaches.9PubMed Central. ‘Intention-to-treat’ meets ‘missing data’: implications of alternate strategies for analyzing clinical trials data
Modified Intention to Treat and the Labeling Problem
A growing number of trials report using something called “modified intention to treat,” often abbreviated mITT. The idea is to exclude a small, well-defined group of participants after randomization, usually for reasons that seem reasonable: they never received even one dose of the study drug, they were later found not to meet the eligibility criteria, or they had no follow-up data whatsoever. It sounds like a sensible compromise between strict intention to treat and the realities of running a trial. The problem is that “modified intention to treat” has no agreed-upon definition, and researchers use the label to justify a wide range of exclusions.
A systematic review of trials using the term found that the definitions were irregular and often arbitrary. Receiving at least one dose of the drug was the most common condition for inclusion, but many trials also required certain lab results, follow-up measurements, or other criteria. Roughly four in ten trials fixed one condition for their modified analysis, another four in ten used two conditions, and a smaller group stacked three or more. In most cases, the researchers did not offer adequate explanations for why they used a modified approach instead of a strict one.10Cochrane Database of Systematic Reviews. Modified Intention to Treat: frequency, definition and implication for clinical trials
A separate systematic review examining published trials found that about 80% of trials using a modified approach made post-randomization exclusions, and more than half reported two or more types of deviation from a true intention-to-treat analysis.11BMJ. Modified intention to treat reporting in randomised controlled trials: systematic review This matters because the label “modified intention to treat” gives readers a false sense that the analysis is close to the gold standard, when in reality the modifications can be substantial enough to reintroduce the very biases that intention to treat was designed to prevent. The estimand being targeted by such approaches, and the assumptions needed for them to be unbiased, are often unclear.12PubMed Central. Using modified intention-to-treat as a principal stratum estimator for failure to initiate treatment
When you encounter a trial that says “modified intention to treat,” the instinct should be to look for the specifics. Which participants were excluded, and why? Did the exclusions affect the groups equally, or could they have tilted the balance? A trial that excludes only participants who were randomized in error and never received any intervention is doing something very different from a trial that excludes everyone who missed a follow-up visit.
How Well Do Published Trials Actually Follow the Principle
Surveys of the published literature paint a somewhat discouraging picture. A landmark survey of randomized trials in major journals found that about half mentioned intention-to-treat analysis, but of those, a meaningful fraction did not actually analyze all randomized participants as allocated. Among trials that did appear to follow the approach, only about a third explicitly stated that all participants were analyzed as randomized. Three-quarters of the trials had at least some missing outcome data, and the methods used to handle it were generally inadequate.13BMJ. What is meant by intention to treat analysis? Survey of published randomised controlled trials
More recent systematic work has confirmed that deviations from intention to treat remain common across randomized trials and even carry over into systematic reviews that synthesize those trials.14PubMed. Deviation from intention to treat in randomized trials and systematic reviews: a systematic review A meta-epidemiological study of trial analyses found that more than half of trials classified as using a modified approach reported multiple types of deviation, and the majority were not explicitly labeling their approach as “modified” at all. They simply described an analysis that deviated from true intention to treat while using the term as if it were fully intact.15PubMed Central. Deviation from intention to treat analysis in randomised trials and treatment effect estimates: meta-epidemiological study
The takeaway for anyone reading medical research is that “intention to treat” in the methods section is not a guarantee. It is worth checking whether the trial actually included all randomized participants, how many dropped out, and what was done about the missing data. The label alone tells you the researchers’ stated philosophy, not necessarily what the numbers reflect.
When Intention to Treat Can Backfire
The conservative bias of intention-to-treat analysis is generally considered a feature, not a bug, because it means the analysis is unlikely to declare a treatment effective when it is not. But there are two important situations where this built-in conservatism becomes a liability.
The first is in non-inferiority trials. These are trials designed to show that a new treatment is “not worse” than an existing one by more than a small, pre-specified margin. Here, the logic of intention to treat flips. If people in the study do not follow their assigned treatment, the two groups can end up looking more similar than they should, because non-adherence dilutes the real difference between them. In a standard superiority trial, that dilution makes it harder to declare the new treatment better, which is conservative and safe. In a non-inferiority trial, that same dilution makes it easier to declare the new treatment “not worse,” which could lead to approving an inferior treatment. Researchers have shown that intention-to-treat analysis in non-inferiority trials can increase the chance of incorrectly concluding that a new treatment is acceptable when it is actually worse than the standard.16BMJ. Non-adherence in non-inferiority trials: pitfalls and recommendations When non-adherence patterns differ between the two arms, the risk of accepting an inferior treatment or missing a genuinely effective one grows further.17Statistics in Medicine. Assessing efficacy in non-inferiority trials with non-adherence to interventions: Are intention-to-treat and per-protocol analyses fit for purpose?
The second situation involves safety trials, where the goal is to detect whether a treatment causes harm. If intention to treat dilutes the measured effect toward zero, a harmful treatment could appear safe simply because many participants assigned to it did not actually take it. Failing to find a toxic effect in an intention-to-treat analysis does not guarantee the treatment is safe.18PubMed Central. Beyond the intention-to-treat in comparative effectiveness research For these types of questions, researchers need to supplement the standard analysis with methods that better reflect what happens to people who actually receive the treatment.
Methods That Go Beyond Intention to Treat
When intention to treat gives a diluted or potentially misleading answer because of non-adherence, researchers turn to more targeted statistical approaches. One family of methods uses the concept of instrumental variables. The idea is to use the original randomization assignment as a kind of natural experiment: since assignment was random, it can serve as a tool to estimate the treatment’s effect among people who would actually comply with whatever group they were placed in. This produces an estimate called the complier average causal effect, which answers the question “what is the treatment effect in the subgroup of people who would follow their assigned treatment?”
These methods can avoid some of the bias introduced by per-protocol analysis while still getting closer to the effect of actually receiving treatment. In simulation studies, the instrumental variable approach has been shown to be unbiased for this complier effect when compliance is binary (people either take the treatment or they do not), while the per-protocol approach is unbiased only under specific, somewhat narrow conditions related to the pattern of non-compliance.19arXiv. Estimating the Complier Average Causal Effect in Randomised Controlled Trials with Non-Compliance: A Comparative Simulation Study of the Instrumental Variables and Per-Protocol Analyses Inverse probability weighting is another technique: participants are weighted in the analysis according to the probability that they would have adhered, which attempts to reconstruct what the randomized comparison would have looked like if everyone had followed the protocol.20PubMed Central. Estimating treatment effects in randomised controlled trials with non-compliance: a simulation study
These methods are increasingly recommended as supplements to intention-to-treat and per-protocol analyses, particularly in non-inferiority trials and safety evaluations. They are not replacements for the intention-to-treat principle but rather tools for answering the questions that intention to treat was never designed to handle.
What This Means If You Are Reading a Trial
If you are a patient trying to understand a study about a treatment you are considering, or a clinician evaluating evidence, the intention-to-treat result is generally the one that best reflects what will happen in your practice. It accounts for the fact that some people will not stick with treatment, and it gives you a realistic picture of the benefit you can expect from prescribing or starting a therapy. But it pays to look beyond the label.
A few things are worth checking when you read a trial report. First, how many participants were randomized versus how many actually contributed outcome data? A large gap between those numbers means substantial missing data, and the handling of that data shapes the results. Second, does the trial say “intention to treat” or “modified intention to treat,” and if modified, who was excluded and why? Third, if the trial is a non-inferiority design, be aware that the intention-to-treat result may actually overstate the case for the new treatment being acceptable. In that context, the per-protocol result and any supplementary causal analyses carry extra weight.
The intention-to-treat principle remains one of the most important safeguards in clinical research. It is not perfect, and anyone who tells you it eliminates all bias is oversimplifying. But the alternatives, when used carelessly, are worse. The principle’s strength is not that it gives a precise answer about what a drug does inside the body. Its strength is that it gives an honest answer about what happens when a treatment is offered to real people in a real healthcare system, with all the messiness that implies.21PubMed. Intention-to-treat analysis in clinical trials: principles and practical importance
Dropout Bias and Why Even Good Intentions Fall Short
One subtlety that catches people off guard is that intention to treat does not protect you from all forms of bias, even when you genuinely include every participant. The protection comes from randomization, and randomization balances the groups at the moment of assignment. After that moment, things can go wrong in ways that no analysis strategy can fully repair.
Consider a trial where sicker participants in the treatment group drop out because of side effects, while sicker participants in the control group drop out because they seek the real treatment elsewhere. Both groups lose their sickest members, but for different reasons and possibly in different proportions. The remaining data looks like two healthy-ish groups with similar outcomes, but the true treatment effect could be quite different from what the numbers show. The validity of a complete-case intention-to-treat analysis requires the untestable assumption that dropout is not related to outcomes, and when that assumption fails, even the well-intentioned analysis can be biased.22NIH Pragmatic Trials Collaboratory. Missing Data and Intention-to-Treat Analyses
This is why the quality of a trial’s effort to minimize dropout matters as much as the analysis technique chosen after the fact. Shorter follow-up periods, simpler protocols, better participant support, and designs embedded in routine clinical care all reduce the chance that missing data will compromise the results. The best intention-to-treat analysis is the one that does not have to wrestle with large amounts of missing data in the first place.