Intention-to-Treat Analysis

Intention-to-Treat Analysis: Concept Architecture

Orientation and learning roadmap

This page follows the randomised groups from assignment to analysis and interpretation, including non-adherence, missing outcomes, and an illustrative event-risk calculation. It then connects the assignment-based result to a treatment-policy question in economic evaluation and distinguishes it from analyses based on treatment actually received.

Assignment defines the comparison

The analysis groups are fixed by the intervention to which participants were originally randomised, even when some decline treatment, stop it, switch treatments, or receive additional care. Keeping those groups preserves the protection that random allocation provides against systematic baseline differences in expectation, provided follow-up and outcome ascertainment do not introduce important bias.

Let $Z=1$ or $Z=0$ denote assignment to the intervention or comparator, and let $Y$ denote an outcome over a specified period. For an additive outcome, a simple assignment-based effect is the difference in group means, $E(Y\mid Z=1)-E(Y\mid Z=0)$; for a binary adverse event, the group means are risks. This contrast describes the effect of offering or assigning the strategies under the trial's actual patterns of adherence and subsequent care, not necessarily the biological effect of taking a treatment continuously.

Conduct and reporting workflow

The trial protocol should define the target population, assigned strategies, outcomes, time horizon, analysis population, and handling of events after randomisation before the result is examined. Reports should show how many participants were randomised, received the intended intervention, contributed outcome data, and were analysed in each assigned group.

Baseline covariates may improve precision when adjustment is prespecified and appropriately modelled, while the randomised assignment still determines each participant's group. Analyses that discard participants because they did not adhere or did not receive treatment should be labelled and justified separately. The phrase “modified intention-to-treat” can obscure such exclusions, so authors should describe exactly who was included and why.

Worked example: adherence and event risk

Imagine 200 patients randomised to a treatment offer and 200 to usual care for one year, with an adverse event fully observed for everyone. Suppose 40 assigned to the offer do not take the treatment and some in usual care later obtain it; these changes do not alter the assigned groups used for the main comparison.

The assignment-based risk difference is $0.15-0.20=-0.05$, an observed five-percentage-point lower event risk in the offer group. Its relative risk is $0.15/0.20=0.75$, but the risk difference and relative risk express different scales and should be labelled separately. These descriptive counts alone do not supply a prespecified confidence interval or establish why the effect arose.

Intercurrent events and missing outcomes

Stopping treatment, switching, rescue medication, and death can alter what outcome is meaningful and how the trial question should be framed. An estimand using a treatment-policy strategy generally evaluates the outcome regardless of specified intercurrent events, but some outcomes cannot be observed after events such as death and require a different explicitly stated strategy.

Remaining in the assigned group does not solve loss to follow-up: if outcomes are missing differentially or depend on unobserved prognosis, an analysis of available cases may be biased despite its ITT label. Plans should specify how missing outcomes are addressed, what assumptions make the chosen method credible, and sensitivity analyses for plausible departures. In the worked example, the complete observation assumption was stated explicitly so the denominators have a clear meaning.

What the result does and does not estimate

The assignment effect reflects the intervention as delivered in that trial, including its uptake, discontinuation, crossovers, and permitted subsequent care. It may be diluted relative to an effect under full adherence, but it can directly inform a policy that offers the intervention under similar implementation conditions.

Per-protocol analyses address different treatment-adherence questions and need appropriate control of selection and time-varying confounding; comparing actual recipients without such control is generally not equivalent. Instrumental variable analysis may use random assignment as an instrument for treatment receipt, but it requires additional relevance, exclusion, independence, and effect-interpretation assumptions. Neither an ITT analysis nor an IV estimate should be relabelled as the other.

Use in health economic evaluation

An evaluation alongside a randomised trial can compare costs and QALYs by assigned group when the decision is whether to offer the strategies under comparable real-world uptake. The analysis must still define perspective, resource-use collection, time horizon, missing-cost and missing-utility methods, and uncertainty in incremental outcomes.

If an economic model instead assumes universal treatment receipt or scales up an intervention with different adherence, directly inserting the trial's assignment effect can misrepresent the decision context. The analyst should explain how trial uptake and post-randomisation care relate to the proposed implementation, and separate extrapolation assumptions from the randomised evidence. An assignment-based clinical contrast alone does not establish cost effectiveness.

Source basis and editorial checks

The sources below distinguish the ITT principle from the exact estimand, analysis population, and handling of missing data. They also support reporting actual randomisation and analysis counts rather than relying on an unqualified label.

• International Council for Harmonisation (2019), ICH E9(R1), Addendum on Estimands and Sensitivity Analysis in Clinical Trials. The guideline separates trial objectives, intercurrent-event strategies, estimation, and sensitivity analysis.
• CONSORT Group (2025), CONSORT 2025 statement. The reporting checklist asks who is included in each analysis, which group they are analysed in, and how participant flow is documented.
• Hollis and Campbell (1999), “What is meant by intention to treat analysis?”, BMJ, 319:670–674. This survey documents ambiguity in reported ITT practice.