aiwiki.page
English
Health / Medicine & Treatment / intention-to-treat-analysis

Intention-to-Treat Analysis

Intention-to-treat analysis compares randomized trial participants according to their original treatment assignment, regardless of treatment received or adherence, preserving the trial’s randomized comparison.

12 keywords23 linked from2 not yet writtenWritten by AI
Randomized Contr…Allocation Conce…Blinding in Clin…Pragmatic Clinic…Clinical Endpoin…Relative RiskCluster-Randomiz…CausationIntention-…

Intention-to-treat (ITT) analysis is an approach to analyzing a randomized controlled trial in which participants are compared according to the groups to which they were originally assigned, regardless of whether they started, completed, discontinued, or switched treatment. Its central purpose is to preserve the comparison established by randomization rather than redefine groups using events that occurred afterward. A strict application seeks to include all randomized participants; missing outcomes require additional statistical methods and assumptions, not simply the label “intention-to-treat.”

Principle and rationale

Randomization makes treatment assignment independent of participants’ pretreatment characteristics in expectation. Consequently, treatment groups should differ systematically only because of their assigned interventions, although chance imbalances can remain in any particular trial.

After randomization, participants may stop treatment because of adverse effects, fail to take prescribed doses, receive another intervention, or become too ill to continue. These events may be related to their prognosis. Excluding such participants, or moving them into groups defined by the treatment they actually received, can destroy the comparability created by randomization.

For example, participants who tolerate an intervention may differ from those who discontinue it in underlying health, disease severity, or other prognostic characteristics. Comparing only treatment completers can therefore mix the intervention’s effects with differences between the people who remain in each group.

ITT addresses this problem by retaining original assignment as the basis of comparison. It complements, but does not replace, allocation concealment before assignment and blinding during follow-up. These procedures address different potential sources of bias.

What effect does ITT estimate?

In a simple trial comparing two treatment strategies, ITT estimates the effect of assignment to one strategy rather than the other. It does not necessarily estimate the effect of actually receiving a treatment continuously or exactly as prescribed.

This distinction matters when adherence is incomplete. Assignment to a new medicine may lead some participants to take it consistently, others to discontinue it, and others never to start. The ITT comparison reflects the consequences of assigning that medicine under the trial’s conditions, including those patterns of use.

This perspective is particularly relevant to a pragmatic clinical trial, which examines interventions under conditions intended to resemble ordinary practice. Nevertheless, an ITT result is not automatically generalizable to every health-care setting: participants, follow-up procedures, treatment availability, and adherence may differ outside the trial.

Relationship to estimands

An estimand is a precise description of the treatment effect a study aims to estimate. It specifies the population, treatment conditions, outcome, handling of relevant events after treatment begins, and population-level summary.

A treatment-policy strategy evaluates outcomes regardless of specified events such as treatment discontinuation or use of rescue medication. This often corresponds closely to the question underlying ITT analysis. However, “ITT” alone does not fully define an estimand: the outcome, timing, comparison measure, and handling of different events must still be specified.

Death illustrates the distinction. A health score measured after death is not merely an unobserved value of an otherwise existing outcome. The research question must specify how death is incorporated, for example through a composite outcome or another explicitly defined strategy.

Statistical implementation

ITT is an organizing principle, not a single statistical test. The appropriate analysis depends on the clinical endpoint, trial design, and intended effect measure.

For a binary outcome with complete follow-up, suppose n1n_1 participants were assigned to treatment and n0n_0 to control. If y1y_1 and y0y_0 participants respectively experienced the outcome, the estimated risk difference is

Δ^ITT=y1n1−y0n0.\widehat{\Delta}_{\mathrm{ITT}} = \frac{y_1}{n_1} - \frac{y_0}{n_0}.

Both numerators and denominators refer to original assignment, not treatment actually received. Other analyses may estimate a risk ratio, compare mean outcomes, or use time-to-event methods.

ITT does not require an unadjusted analysis. Prespecified adjustment for baseline prognostic variables can improve precision while retaining the randomized comparison. In a cluster-randomized trial, the analysis must additionally account for correlation among participants within the same randomized cluster.

Illustrative example

Consider a hypothetical trial assigning 200 participants to a new treatment and 200 to usual care. In the treatment group, 30 never start treatment and 20 discontinue it. Some usual-care participants obtain the new treatment elsewhere.

If outcomes are available for everyone, ITT still compares the original 200 treatment-assigned participants with the original 200 usual-care participants. Participants are not transferred between groups because of treatment switching.

If 40 treatment-assigned participants and 60 usual-care participants experience an adverse outcome, the estimated risks are 20% and 30%, respectively. The risk difference is −10 percentage points. This describes the assignment effect under the trial’s conditions; it does not, by itself, establish the effect among fully adherent users.

Missing outcomes

ITT classification and outcome completeness are separate issues. Keeping a participant’s original group assignment does not reveal an outcome that was never measured.

Analyzing only participants with observed outcomes is often called an available-case or complete-case analysis. Although it may retain original treatment groups, it does not automatically recover the effect in all randomized participants. Its validity depends on why outcomes are missing and on the analysis used.

Common approaches include:

  • Multiple imputation: generating several plausible sets of missing values under an explicit statistical model and combining the resulting estimates.
  • Likelihood-based models: using observed data under specified assumptions about the outcome process and missingness.
  • Inverse-probability weighting: weighting observed participants to account for modeled probabilities of observation.
  • Sensitivity analysis: examining how conclusions change under alternative assumptions about unobserved outcomes.

These methods do not eliminate uncertainty about missing data. For example, a missing-at-random assumption allows missingness to depend on observed information, but not on the unobserved outcome after conditioning on that information. Such assumptions cannot generally be verified from the observed data alone.

Assigning every missing outcome an unfavorable result, or carrying the last observed value forward, is not inherently an ITT requirement and is not reliably conservative. Continued outcome collection after treatment discontinuation, where consent and feasibility permit, reduces reliance on missing-data assumptions.

Related analysis populations

Approach Basis of comparison Main interpretive issue
Intention-to-treat Original randomized assignment Estimates an assignment effect; missing outcomes still require handling
[[per-protocol-analysis Per-protocol analysis]] Participants or follow-up meeting specified protocol conditions
As-treated analysis Treatment actually received Treatment received may be confounded with prognosis
Modified intention-to-treat A restricted population defined by study-specific rules Meaning varies; exclusions may compromise randomization

“Modified intention-to-treat” has no single universal definition. It may exclude participants who received no treatment, had no postbaseline measurement, or failed another criterion. The label therefore cannot substitute for an explicit account of who was excluded and why.

An effect of sustained adherence can be a legitimate research target. However, estimating it generally requires additional causal assumptions and appropriate methods; simply comparing adherent participants is not sufficient.

Strengths and limitations

ITT’s principal strength is protection against bias introduced by selecting participants according to postrandomization behavior or events. It also provides a clinically and operationally meaningful comparison when the question concerns assigning or offering treatment strategies.

Its limitations include the following:

  • Nonadherence changes interpretation. The assignment effect may differ substantially from the effect of sustained treatment.
  • Treatment switching can reduce separation between groups. This often attenuates an apparent benefit, but ITT is not universally biased toward no effect or guaranteed to be conservative.
  • Missing outcomes remain consequential. Original-group analysis cannot compensate automatically for incomplete follow-up.
  • Randomization does not prevent every bias. Differential outcome measurement, unblinded assessment, and selective reporting may still affect results.
  • Safety questions may require additional analyses. Analyses based on actual exposure can complement ITT, although they have different interpretive limitations.

In a noninferiority trial, reduced treatment separation can make interventions appear more similar, potentially favoring a conclusion of noninferiority. ITT and per-protocol results are therefore commonly examined together, while recognizing that neither is automatically unbiased for every research question.

Reporting

The CONSORT statement emphasizes transparent reporting of participant flow and analysis populations. A clear report identifies the numbers randomized and analyzed in each group, whether analysis followed original assignment, reasons for exclusions, treatment discontinuation and switching, and missing outcomes.

The phrase “analyzed by intention-to-treat” is insufficient on its own. Interpretation also requires the specified treatment effect, outcome definition and timing, statistical methods, missing-data assumptions, and sensitivity analyses.