Skip to content
Scintria
ToolsGrant writerFunding
HomeLiteratureClinical trialsEvidence LabProtocol packData analysisLearnProjectsLibraryReferencesApplications
Start free
ScintriaStart free

Learn / Biostatistics / Statistics in clinical trials and evidence synthesis

Intention-to-treat, per-protocol and missing data

14 minute read · 6-question quiz

Download this lesson as a colour Word or PDF study guide with the figures and answers. Included in paid plans; reading and quizzes stay free.Sign in

What you will learn

  • Define intention-to-treat (ITT), modified ITT and per-protocol analyses and choose between them

  • Explain how non-adherence and missing data bias a trial and what the estimand framework adds

  • Describe the options for handling missing outcome data and why last-observation-carried-forward is discouraged

Randomisation creates comparable groups, but only if they stay comparable. In practice participants stop taking treatment, switch arms, are found ineligible or vanish. How these people are handled in the analysis decides whether the trial's comparison is still protected by randomisation.

Analysis populations

PopulationWho is analysedStrengthWeakness
Intention-to-treat (ITT)Everyone randomised, in the arm they were assigned, whatever happened nextPreserves randomisation; unbiased for the effect of a treatment policy; the primary analysis in superiority trialsDilutes the effect when many do not take treatment
Modified ITT (mITT)ITT minus a few pre-defined exclusions (never received any dose, found ineligible after randomisation due to a blinded process)PracticalOpen to abuse; define in the protocol
Per-protocol (PP)Only those who followed the protocol adequatelyCloser to the effect of taking the drugBreaks randomisation; compares unlike groups
As-treatedBy the treatment actually receivedDescribes safetyTreatment received is not random

Worked example: why ITT matters

A trial randomises 200 patients per arm. In the new-drug arm, 20 people never take the drug (they are sicker, with severe nausea or poor access). Results at 28 days:

AnalysisNew drugControlDifference
ITT130/200 = 65.0%110/200 = 55.0%10.0 points (95% CI 0.4 to 19.6; p = 0.041)
Per-protocol128/180 = 71.1%110/200 = 55.0%16.1 points

Only 2 of the 20 non-takers were cured (10%), because they were sicker. Removing them makes the drug look far better, because we have taken the sickest patients out of one arm only. ITT keeps them in. It estimates the real-world effect of offering the drug in this population. Do not drop people because they were non-adherent.

The principle: analyse as randomised

'Once randomised, always analysed.' ITT is protected from the bias that arises when the reasons for leaving treatment are linked to prognosis. The price is a diluted estimate; that is acceptable in superiority trials because dilution makes it harder to claim a benefit, but it is anti-conservative in non-inferiority trials, where PP is analysed alongside ITT.

The estimand framework (ICH E9 R1)

Modern regulators ask trial teams to define the estimand: the exact quantity the trial estimates. It has five attributes: the population, the treatment, the outcome variable, the handling of intercurrent events (stopping treatment, switching, death), and the summary measure. Different questions give different estimands. 'What is the effect if everyone takes the drug as directed?' differs from 'What is the effect of prescribing it?' Strategies for intercurrent events include treatment-policy (ITT), hypothetical (as if the event had not happened), composite, while-on-treatment and principal-stratum. State the estimand in the protocol and align the analysis with it.

Missing data

Missing outcome data are inevitable and damaging. The mechanism matters.

  • MCAR (missing completely at random): unrelated to anything. Rare in practice; complete-case analysis is unbiased but loses power.
  • MAR (missing at random): missingness depends on observed data (older patients miss more visits, but within age groups it is random). Handle with multiple imputation or likelihood-based models.
  • MNAR (missing not at random): missingness depends on the unobserved value itself (people with worse outcomes drop out). No analysis fully fixes this; use sensitivity analyses.
AvoidBetterAlwaysDropping people withmissing data(complete-case) when muchis missingLast observation carriedforward (LOCF): assumes nochangeReplacing with the mean:understates variabilityMultiple imputation: fillin several plausiblevalues, analyse each, poolwith Rubin's rulesMixed models for repeatedmeasures (MMRM)Inverse probabilityweightingPrevent missingness:retention plans, visitwindows, phone follow-upReport numbers and reasonsRun sensitivity analysesunder MNAR scenarios
Handling missing data: prevention is better than any statistical cure.

Why LOCF is discouraged

Carrying the last observation forward assumes patients remain unchanged after dropout. In a progressive disease they would have worsened; in a condition that improves with time they would have improved. The method understates uncertainty and can bias in either direction. Modern guidance (e.g., the NRC report, ICH E9 R1) prefers multiple imputation or model-based methods plus sensitivity analyses.

A simple sensitivity analysis: tipping point

Ask how bad the outcomes of the missing participants would have to be (in the treatment arm only) before the conclusion changes. If a plausible scenario overturns it, the result is fragile.

Summary

  • ITT analyses everyone as randomised and is the primary analysis for superiority trials.

  • Per-protocol analysis breaks randomisation and can exaggerate benefit.

  • Define the estimand, including how intercurrent events such as stopping treatment are handled.

  • Missing data are MCAR, MAR or MNAR; use multiple imputation or mixed models, with sensitivity analyses.

  • Avoid LOCF and mean substitution; prevent missingness through trial conduct.

Check your understanding

Question 1
1

A trial analyses participants in the groups they were randomised to, regardless of adherence. This is:

Question 2
2

Per-protocol analysis can exaggerate treatment benefit because:

Question 3
3

Missingness that depends on the unobserved outcome itself is:

Question 4
4

Which approach to missing data is generally preferred to LOCF?

Question 5
5

An ICH E9(R1) estimand includes all EXCEPT:

Question 6
6

Why is ITT potentially anti-conservative in a non-inferiority trial?

Draft study guide. Spotted an error or something unclear? Tell us at hello@scintria.com. This is educational material, not clinical or regulatory advice.

Clusters, non-inferiority, attrition and writing the sectionInterim analyses, data monitoring and multiplicity in trials