The intention-to-treat principle is a rule for analyzing clinical trials: once a participant is randomly assigned to a group, their outcome counts toward that group’s results no matter what happens afterward. If someone stops taking the drug, switches to the other treatment, or drops out early, they still belong to their original group in the final analysis. The principle sounds almost stubborn in its simplicity, but it exists to solve a real and serious problem with how we draw conclusions from experiments on people.
Why Counting Everyone Matters
Randomization is the single most powerful tool clinical trials have for producing trustworthy results. When you randomly assign thousands of people to receive either a new drug or a placebo, the two groups start out roughly equal in every way that matters: age, underlying health, genetics, motivation, lifestyle. Any differences in outcomes can then be attributed to the treatment rather than to something else about the people in each group.
The intention-to-treat principle exists to protect that balance. If you start removing people from the analysis after randomization, you risk undoing the very thing randomization accomplished. The participants who drop out, stop taking the drug, or violate the study rules are rarely a random cross-section of the group. They tend to be sicker, or experiencing side effects, or less motivated. Removing them creates groups that no longer mirror each other, and the comparison becomes unreliable.1PubMed Central. Understanding the Intention-to-treat Principle in Randomized Controlled Trials
In an intention-to-treat analysis, every randomized participant stays in the dataset. Noncompliance, protocol deviations, and withdrawals are all ignored for the purpose of group assignment. The analysis asks a straightforward question: among everyone assigned to receive the treatment, how did things turn out compared to everyone assigned to the control?2PubMed Central. Intention-to-treat concept: A review
What ITT Actually Estimates
There is an underappreciated distinction between what intention-to-treat analysis measures and what most people assume it measures. ITT does not tell you how well a treatment works when taken correctly. It tells you how well a treatment strategy works when offered to patients in a real-world-like setting, where some people will not take it as directed. That is a different question, and for most practical decisions, it is arguably the more useful one.
Think of it this way. A doctor deciding whether to prescribe a new blood pressure medication does not just need to know whether the drug lowers blood pressure when taken perfectly. The doctor also needs to know what happens when that prescription goes out into the world, where some patients will forget doses, others will stop because of side effects, and a few will never fill the prescription at all. ITT captures that full picture. It reflects the effect of assigning a treatment, not the effect of receiving it flawlessly.
This distinction matters because the estimate ITT produces is generally conservative. When people in the treatment group do not actually take the treatment, their outcomes look more like the control group’s outcomes, which narrows the apparent gap between groups. The real biological effect of the drug might be larger than what the ITT analysis shows.2PubMed Central. Intention-to-treat concept: A review
How Per-Protocol Analysis Differs
The main alternative to intention-to-treat is per-protocol analysis. Instead of including everyone regardless of what happened, per-protocol analysis restricts the dataset to participants who followed the study rules: they met all inclusion criteria, received the assigned treatment as planned, and had outcome data available. The appeal is obvious. If you want to know whether a drug works when actually taken, why not just look at the people who took it?3PubMed Central. Intention-to-treat versus as-treated versus per-protocol approaches to analysis
The problem is that the people who adhered to the protocol are a self-selected subset. Participants who completed the full course of treatment tend to differ from those who dropped out in ways that affect the outcome. They may have been healthier at baseline, more motivated, or experiencing fewer side effects. By keeping only the “good” participants, per-protocol analysis can make a treatment look better than it actually is, or in rarer cases, worse. The groups are no longer balanced the way randomization intended them to be.4PubMed. Intention to treat and per protocol analyses: differences and similarities
When everyone in a trial follows the protocol perfectly, the distinction vanishes. ITT, per-protocol, and as-treated analyses all produce identical results because there is no divergence between what was assigned and what was received. The analyses only part ways when real-world messiness intrudes, when participants do not comply, switch treatments, or leave the study.5PubMed Central. Intent-to-Treat vs. Non-Intent-to-Treat Analyses under Treatment Non-Adherence in Mental Health Randomized Trials
A striking example of how much the choice of analysis can matter comes from trials of anti-TNF-alpha therapies in rheumatoid arthritis. In one trial, a modified intention-to-treat analysis produced an odds ratio of 2.0 for cancer in the treatment arm versus placebo. But when the same data were analyzed per-protocol, the odds ratio flipped to 0.85, suggesting a slight protective effect. The raw numbers of cancer cases were the same; only the denominators shifted because of who was included or excluded. The analytical framework changed the headline finding entirely.6PLoS ONE. Cancer Risk of Anti-TNF-α at Recommended Doses in Adult Rheumatoid Arthritis: A Meta-Analysis with Intention to Treat and per Protocol Analyses
When ITT Is Not the Safe Choice
Most discussions of ITT emphasize that it is the conservative approach, the one that is more likely to underestimate a real treatment effect than to overestimate one. That framing holds for superiority trials, which are designed to show that a new treatment is better than an existing one. If ITT dilutes the effect, and the treatment still looks superior, you can be fairly confident the result is real.
But not all trials are superiority trials. Non-inferiority trials ask a different question: is this new treatment at least as good as the standard? Here, the conservative bias of ITT becomes a liability rather than a strength. By blurring the difference between groups, ITT can make two treatments look more similar than they really are, which is exactly what a non-inferiority trial is trying to demonstrate. In that context, ITT’s tendency to dilute differences actually makes it easier to declare non-inferiority, even when the new treatment is genuinely worse.7PubMed. Choosing the analysis population in non-inferiority studies: per protocol or intent-to-treat
Simulation studies have confirmed that when non-adherence rates differ between the two arms of a non-inferiority trial, both ITT and per-protocol analyses can produce meaningfully biased estimates, inflating the chance of wrongly declaring one treatment non-inferior to another.8PubMed. Assessing efficacy in non-inferiority trials with non-adherence to interventions: Are intention-to-treat and per-protocol analyses fit for purpose? Regulatory agencies generally recommend reporting both ITT and per-protocol results for non-inferiority trials, precisely because neither one alone is reliably trustworthy in that setting.
The “Modified” ITT Problem
If you read clinical trial reports regularly, you will encounter the term “modified intention-to-treat” or mITT. This refers to an analysis that starts with the intention-to-treat framework but excludes some participants after randomization, usually for specific pre-defined reasons like never receiving a single dose of the study drug, or being found ineligible after enrollment. The term sounds reasonable, and in some cases the exclusions are genuinely defensible.
The concern is that “modified ITT” has no standard definition. Different trials define it differently, and the label can obscure how many participants were removed and why. A systematic review of direct oral anticoagulant trials found that in trials for atrial fibrillation and blood clot treatment, the proportion of participants excluded after randomization was low, around one percent. But in surgical and medical thromboprophylaxis trials that also claimed to use an ITT or modified ITT approach, nearly thirty percent of participants were excluded after randomization.9BMJ Evidence-Based Medicine. Use of modified intention-to-treat analysis in studies of direct oral anticoagulants and risk of selection bias: a systematic review
Excluding almost a third of randomized participants while still calling the analysis “intention-to-treat” is misleading at best. The label gives an impression of rigor that the actual analysis may not support.
A cross-sectional study comparing trials that used standard ITT, modified ITT, and no ITT found that mITT trials were about three and a half times more likely to report post-randomization exclusions than standard ITT trials. More troublingly, mITT trials were strongly associated with industry sponsorship and with author conflicts of interest.10PubMed Central. Modified versus standard intention-to-treat reporting: are there differences in methodological quality, sponsorship, and findings in randomized trials? A cross-sectional study That does not prove the analyses were biased, but it raises questions about whether the “modified” label sometimes serves as cover for selectively trimming the dataset in ways that favor the tested product.
In at least some therapeutic areas, the bias concern may be smaller than feared. A comparison of ITT and mITT results in rheumatoid arthritis trials found no significant difference in the estimated treatment effect between the two approaches.11Journal of Clinical Epidemiology. What Is the Intention-to-Treat Principle? But that finding applies to a specific set of trials with specific exclusion patterns. It would be a mistake to generalize it to all uses of mITT across all fields.
What Happens to Missing Data
Here is the practical headache with ITT: if you insist on analyzing every randomized participant, you need outcome data for every randomized participant. But people drop out. They move away, stop responding to follow-up calls, or die from unrelated causes. You cannot measure a blood pressure reading or a tumor response in someone who has vanished from the study.
A systematic review of methods literature found widespread inconsistency in how ITT and missing outcome data are discussed. Out of sixty-six articles examining the topic, more than a third defined ITT without addressing missing data at all, leaving a fundamental tension unresolved: ITT requires everyone’s data, but some data will always be missing.12PLoS ONE. Inconsistent Definitions for Intention-To-Treat in Relation to Missing Outcome Data: Systematic Review of the Methods Literature
Researchers handle this in different ways. In a trial of cocaine dependence treatment, three strategies were compared. One analyzed data only up to the point each participant dropped out. Another assumed outcomes for the full intended duration, including after dropout. A third, more sophisticated approach kept the full sample and full duration but also modeled the effect of staying in treatment as a separate variable.13PubMed Central. Intention-to-treat meets missing data: implications of alternate strategies for analyzing clinical trials data Each strategy can lead to different conclusions, even though all of them could be described as “intention-to-treat.”
Common approaches to filling in missing data include last observation carried forward (using the participant’s last recorded measurement as a stand-in for all later time points) and multiple imputation (using statistical models to estimate what the missing values probably were). Neither is perfect, and the choice of missing-data method can influence results as much as the choice between ITT and per-protocol analysis. When you read that a trial used ITT, it is worth asking how they dealt with the participants whose outcomes were never measured.
Statistical Methods That Go Beyond ITT
Because ITT captures the effect of offering a treatment rather than taking it, researchers have developed methods that try to estimate what the treatment effect would be among the people who actually complied. One of the most widely discussed is the complier average causal effect, or CACE. This approach uses the random assignment itself as a kind of natural experiment: since assignment was random, any relationship between assignment and outcome that passes through actual treatment receipt can be attributed to the treatment.14American Journal of Epidemiology. Estimation of Treatment Efficacy With Complier Average Causal Effects (CACE) in a Randomized Stepped Wedge Trial
CACE and related instrumental variable approaches do not require you to exclude non-compliers from the dataset. Instead, they use statistical techniques to separate the treatment effect among people who would comply from the noise introduced by people who would not. The result is an estimate that is closer to the true biological effect of the treatment while still respecting the randomized design. These methods have been extended beyond simple two-arm trials to more complex designs, though they require assumptions that may not always hold.15PubMed. Intention to treat, per protocol, as treated and instrumental variable estimators given non-compliance and effect heterogeneity
ITT Outside of Randomized Trials
The intention-to-treat principle originated in randomized controlled trials, but its logic has migrated into observational research through an approach called target trial emulation. The idea is to take a large observational dataset, such as electronic health records or insurance claims data, and structure the analysis as if it were a randomized trial. This includes defining a “time zero” (analogous to randomization), specifying treatment strategies, and analyzing outcomes according to the treatment assigned at time zero rather than the treatment actually received over time.16PubMed. Target trial emulation for causal inference using observational data in neurology
Target trial emulation has gained traction across medicine because many important clinical questions cannot be addressed with randomized trials for ethical or logistical reasons. Comparing two medications that are already in widespread use, or studying long-term outcomes over decades, often requires working with the data that already exist. By explicitly emulating the structure of a trial, researchers can reduce the kinds of bias that historically plagued observational studies, particularly the bias that comes from analyzing by treatment received rather than treatment initiated.17The Journal of Clinical Psychiatry. Target Trial Emulation: A Concept Simply Explained
This represents a notable expansion of the ITT principle’s influence. What started as a bookkeeping rule for randomized trials has become a design philosophy for an entire category of research that once operated under very different assumptions.
How to Read a Trial Report With ITT in Mind
When you encounter a clinical trial in the news or in a journal abstract, a few questions are worth asking. First, did the trial use ITT, modified ITT, or per-protocol analysis? If mITT, how many participants were excluded after randomization and why? An exclusion rate of one or two percent is usually not worrying. An exclusion rate pushing toward thirty percent is a red flag.
Second, what kind of trial was it? If the trial is testing whether a new treatment is superior to a control, ITT is generally the more conservative and trustworthy framework. If the trial is testing non-inferiority, ITT alone may not be enough, and you should look for whether per-protocol results were also reported and whether they agreed.
Third, how was missing data handled? A trial claiming full ITT but analyzing only participants with complete data has quietly abandoned the principle. The method used to account for dropouts matters, and reputable trials will describe it explicitly. If the dropout rate is high and the method of handling missing data is not described, the results deserve extra skepticism regardless of what label the authors apply.
Fourth, consider the compliance rate. If almost everyone in both groups took the assigned treatment as directed, the choice between ITT and per-protocol analysis barely matters. The stakes of the analytical choice scale with the amount of real-world messiness in the trial. A trial with high dropout or frequent treatment switching is exactly the trial where ITT earns its keep, and also exactly the trial where the gap between ITT and per-protocol estimates will be widest.