What is intention to treat analysis?

Intention-to-treat analysis compares participants in the groups they were randomly assigned to, regardless of whether they took, completed or switched the assigned treatment. It preserves the prognostic balance created by randomization and estimates the effect of offering a treatment under real-world adherence. It is often summarized as "once randomized, always analyzed."

What is the difference between intention-to-treat and per-protocol analysis?

Intention-to-treat analysis includes every randomized participant in the assigned group, while per-protocol analysis includes only participants who followed the protocol, such as those who took a minimum amount of the treatment. Per-protocol analysis estimates the effect under ideal adherence, but it becomes a nonrandomized comparison because adherence is related to prognosis. Intention-to-treat is the preferred primary analysis in superiority trials, with per-protocol results used as supporting evidence.

Why is intention-to-treat analysis conservative?

Participants who stop the treatment or cross over remain in their original groups, which makes the groups more alike and dilutes the measured difference. In a superiority trial this biases the estimate toward no effect, so a benefit shown by intention-to-treat is unlikely to be exaggerated, at the cost of lower power and more risk of a type II error. In non-inferiority trials the same dilution is not conservative, because it makes it easier to conclude that a new treatment is no worse.

What is a modified intention-to-treat analysis?

A modified intention-to-treat analysis includes randomized participants except for certain exclusions, such as people who never received any study treatment or who have no outcome data. Excluding people who never started treatment keeps the intention-to-treat principle only if that decision could not have been influenced by knowledge of the assigned group. Because the term is used inconsistently, readers should check exactly who was excluded and whether exclusions differed between arms.

What is an as-treated analysis in a clinical trial?

An as-treated analysis groups participants by the treatment they actually received rather than the treatment they were assigned. A control patient who obtained the study drug is analyzed with the treatment group, and an assigned patient who never took it is analyzed with the controls. Because those switches are not random, as-treated analysis breaks randomization and can be confounded in either direction.

Intention-to-Treat Analysis: ITT vs Per-Protocol

Intention-to-treat (ITT) analysis compares trial participants in the groups they were randomized to, whether or not they actually took, finished or switched the assigned treatment. Keeping everyone where randomization put them preserves the prognostic balance between groups, so ITT is the standard primary analysis for superiority trials and estimates what happens when a treatment is offered in real practice. The trade-off is dilution: nonadherence and crossover pull the result toward no difference, which makes ITT conservative. The one-line rule is "once randomized, always analyzed."

ITT vs per-protocol vs as-treated

Analysis	Who is analyzed, and in which group	What it estimates	Main weakness	Usual effect in a superiority trial
Intention-to-treat	Everyone randomized, in the assigned group	Effect of assigning (prescribing) the treatment at the adherence actually achieved	Dilution by nonadherence and crossover; needs outcome data on everyone	Smaller, conservative estimate
Per-protocol	Only participants who followed the protocol, in the assigned group	Effect under ideal adherence (efficacy)	People who are excluded differ from those kept, so it becomes a nonrandomized comparison	Often larger; bias can be severe
As-treated	Everyone, regrouped by the treatment actually received	Effect of receiving the treatment	Breaks randomization completely	Biased in whichever direction the switchers push it
Modified ITT	Everyone randomized minus certain exclusions (for example, never received a dose, or no outcome data)	Close to ITT if the exclusions are unbiased	Definitions vary between trials	Depends on who was excluded

The Cochrane Handbook warns that naive per-protocol analyses, restricted to participants who adhered, can be seriously biased because prognostic factors influence who adheres. That is the core idea: only ITT compares the groups that randomization actually created.

Why ITT matters: adherers are different people

Randomization balances prognosis at the moment of assignment. What happens afterward (stopping a drug because of side effects or worsening illness, crossing over, dropping out) is not random, so removing or regrouping those people reintroduces confounding. Patients who adhere tend to do better than those who don't, even when the pill is a placebo, and even after adjustment for known prognostic factors.

The classic demonstration is the Coronary Drug Project, a placebo-controlled trial of the lipid-lowering drug clofibrate in men after myocardial infarction. Men who took at least 80% of their clofibrate had much lower 5-year mortality than poor adherers (15.0% vs 24.6%), which looks like a drug effect until you check the placebo arm: good adherers to placebo also did better (15.1% vs 28.3%). Adherence marked healthier patients; it did not show that the drug worked.

Stopping treatment, crossing over and dropping out are post-randomization problems, a close cousin of the selection and attrition bias covered in bias types for Step 1.

Worked example

Vignette. A hypothetical trial randomizes 1,000 high-risk patients: 500 to a new drug and 500 to placebo. The outcome is a major cardiac event within 2 years.

Drug arm: 400 take the drug as directed (20 events); 100 stop it early, many because they felt unwell (20 events).
Placebo arm: 400 stay on placebo (48 events); 100 obtain the drug outside the trial and cross over (12 events).

What are the absolute risk reduction (ARR) and number needed to treat (NNT) by ITT, per-protocol and as-treated analysis, and which should be the primary result?

Analysis	Drug group	Comparison group	Event rates	ARR	Relative risk	NNT
ITT	All 500 randomized to drug (40 events)	All 500 randomized to placebo (60 events)	8% vs 12%	4%	0.67	25
Per-protocol	400 adherent (20 events)	400 adherent to placebo (48 events)	5% vs 12%	7%	0.42	about 14
As-treated	500 who took the drug: 400 adherent + 100 crossovers (32 events)	500 who did not: 400 on placebo + 100 who stopped (68 events)	6.4% vs 13.6%	7.2%	0.47	about 14

Answer: By ITT, the event rate is 8% vs 12%, so ARR = 4 percentage points, RR = 0.67 and NNT = 1 / 0.04 = 25. That ITT result is the primary answer.

Reasoning: The 100 patients who stopped the drug had a 20% event rate, worse than either arm's average, because people who stop a drug are often the sicker ones. Per-protocol analysis throws them out of the drug arm, and as-treated analysis moves them into the comparison arm, so both make the drug look nearly twice as effective (ARR about 7% instead of 4%). ITT keeps them where randomization put them and answers the prescriber's real question: treat 25 similar patients for 2 years to prevent one event. If a stem asks which analysis preserves randomization or gives the unbiased effect of offering the drug, the answer is ITT.

Missing outcomes and "modified" ITT

Strict ITT needs an outcome for every randomized participant. That means continuing to follow people after they stop the treatment, not just while they take it.
Complete-case analysis is not strictly ITT. Dropping participants without outcome data (sometimes labeled "modified ITT") can bias the result when whether an outcome is missing depends on the outcome itself. Imputation, from carrying the last observation forward to complex models, rests on assumptions and is unlikely to remove that bias.
Watch differential dropout. Bias is most likely when missing outcomes depend on both the assigned group and the outcome itself.
Modified ITT is a spectrum. ICH E9, the international trial-statistics guideline, says excluding participants who never took a dose can still respect the ITT principle, for example when the decision to start treatment could not have been influenced by knowing the assignment.
The label is not the method. The Cochrane Handbook notes that "intention-to-treat" is used inconsistently in trial reports, so check who was actually analyzed.

Superiority vs non-inferiority trials

Superiority trials ("is A better than B?"): the ITT (full analysis) set is the primary analysis because noncompliers diminish the estimated effect, which avoids over-optimistic claims.
Non-inferiority and equivalence trials ("is A no worse than B?"): ICH E9 warns that the full analysis set is "generally not conservative." Dropouts in both arms tend to look like nonresponders, which makes the arms look alike and biases the result toward showing equivalence, so a non-inferiority claim is more convincing when the per-protocol analysis agrees.
In confirmatory trials ICH E9 says it is usually appropriate to run both analyses; when they agree, confidence in the result increases.

How ITT shows up in USMLE stems

Stem phrase	Analysis being described
"Analyzed in the groups to which they were randomly assigned, regardless of adherence"	Intention-to-treat
"Only participants who took at least 80% of doses were included"	Per-protocol
"Analyzed according to the treatment they actually received"	As-treated
"Excluded participants who never received a dose"	Modified ITT

Classic question patterns: which analysis best preserves randomization when crossover occurs (ITT); which analysis is more likely to overstate benefit (per-protocol); and which analysis is more prone to a false negative, because dilution lowers power (ITT, a type II error risk; see type I vs type II error).

Exam traps

ITT is not a completers analysis. Analyzing only those who finished is a per-protocol-style exclusion and breaks randomization.
"Conservative" applies to superiority trials. In a non-inferiority trial, ITT dilution can hide a real difference; look for both analyses.
ITT cannot rescue a nonrandomized study. It preserves randomization; it cannot create it.
ITT does not fix missing data. Participants lost to follow-up still need outcomes, from continued follow-up or imputation.
Per-protocol is not worthless. It estimates efficacy under full adherence and is a useful supporting analysis, just not the unbiased primary one.
NNT and ARR follow the analysis population. Calculate them from the ITT numbers unless the question says otherwise.

Key takeaways

ITT = analyze everyone in the group they were randomized to, regardless of adherence, dropout or crossover: once randomized, always analyzed.
It preserves randomization and estimates the effect of prescribing the treatment in practice (effectiveness).
Nonadherence dilutes the effect: ITT is conservative in superiority trials and more prone to type II error, but not conservative in non-inferiority trials.
Per-protocol keeps only adherent participants (efficacy, but biased); as-treated regroups by treatment received (randomization lost); modified ITT requires checking who was excluded.
Adherers differ from nonadherers even on placebo, as the Coronary Drug Project showed.
True ITT needs outcomes for everyone, so follow-up must continue after treatment stops.