91828 Chap.8 Intention-to-Treat and Analysis Choices
Intention-to-Treat and Analysis Choices
Intention-to-treat analysis preserves the comparison created by random allocation, but it does not recover outcomes that were never observed. This chapter separates assignment, adherence, crossover and missing data, then compares ITT, per-protocol, as-treated and complete-case sets by the question each answers.
Attrition is appraised through its rate, reasons, relationship with outcome, handling assumptions and sensitivity analyses.
What this chapter covers
- 01
Intention-to-treat principle
- 02
Per-protocol analysis
- 03
As-treated analysis
- 04
Missing outcome mechanisms
- 05
Imputation assumptions
- 06
Sensitivity analysis
Worked example · free
Appraise differential missing outcomes
- +1Missingness is 14% in intervention and 4% in control, so it is material and differential.
- +1Burden and lack of improvement may be related to the unobserved outcome.
- +1Complete-case analysis could favour intervention if poor outcomes are selectively absent.
- +1Seek prespecified handling, outcome recovery attempts and sensitivity analyses.
- +1Reduce confidence if plausible missing-outcome assumptions materially change the conclusion.
Key terms
- Intention-to-Treat
- Intention-to-treat analyses participants according to their original random assignment.
- Per-protocol
- Per-protocol analysis includes participants who meet predefined adherence or protocol criteria.
- Complete-case Analysis
- Complete-case analysis uses only participants with observed data for the variables analysed.
- Estimand
- An estimand defines the treatment effect, population, endpoint and handling of intercurrent events being targeted.
Intention-to-Treat and Analysis Choices FAQ
Does ITT include people who never took treatment?
Yes, if they were randomised, they remain in their assigned group for an ITT effect of assignment. Their behaviour is part of real implementation. Missing outcomes still require a transparent method and assumptions; ITT is not automatic imputation.
Why can per-protocol analysis be biased?
Adherence occurs after randomisation and can relate to prognosis, adverse effects, motivation or response. Removing non-adherent participants creates groups that no longer retain random balance. Per-protocol findings can be informative but answer a different and more assumption-dependent question.
Are equal attrition percentages reassuring?
They are useful but not decisive. Reasons, timing and relationship with outcome may differ even when percentages match. Compare the characteristics of missing participants, the recovery effort, analysis assumptions and whether conclusions remain stable under plausible alternatives.
What does modified ITT mean?
There is no single definition. It may exclude participants without treatment, baseline data or post-baseline outcomes. Reconstruct the exact rule and whether it uses information after randomisation. Those exclusions can break group comparability and change the estimand even when the report retains an ITT label.
Can multiple imputation recover the missing truth?
No method observes outcomes that were not measured. Multiple imputation uses a model and variables to generate plausible values under assumptions, often missing at random. Appraise those assumptions, included predictors and sensitivity analyses that allow informative missingness rather than treating the technique as automatic repair.
Why examine both ITT and per-protocol in non-inferiority trials?
Non-adherence and crossover can make groups appear similar, which may favour a non-inferiority conclusion. ITT and per-protocol analyses have different vulnerabilities in this setting. Concordant results can strengthen interpretation when margins, exclusions and missing-data handling were prespecified and credible.
Assessment move
Create a participant-flow table for every trial before interpreting its effect. Use columns for randomised, received assigned treatment, crossed over, discontinued, supplied the primary outcome and entered each analysis set. Keep those events separate because loss of treatment contrast and loss of outcome information create different problems.
Calculate missing-outcome percentages by arm, then read the reasons and timing rather than stopping at the totals. Ask whether burden, adverse events, lack of improvement or access to outside treatment relates to the unobserved endpoint. Rewrite the authors' analysis labels in plain language.
Intention-to-treat means comparison by original allocation, per-protocol applies adherence rules, as-treated regroups by treatment received and complete-case excludes missing outcomes. Modified ITT has no stable meaning, so list every exclusion. Identify the estimand each set approaches and the post-randomisation selection it risks. For missing-data methods, write the assumption in words before judging the technique.
Compare complete-case, multiple-imputation and model-based estimates, noting when several analyses share the same missing-at-random premise. Design sensitivity scenarios that make missing intervention outcomes worse or missing control outcomes better within plausible limits. Report estimates side by side and decide whether magnitude, precision or clinical meaning changes.
In superiority trials, keep ITT central to the effect of assignment. For non-inferiority reasoning, examine both ITT and per-protocol because deviations can bias toward similarity. Handle safety populations separately when exposure, rather than assignment, defines the question, but label that change clearly. For AT1, connect differential attrition to a mechanism rather than using a percentage threshold alone.
For AT2, specify outcome-recovery procedures, analysis population and sensitivity plan before data collection. End by tracing one participant who crosses over and one whose outcome is missing through every analysis set. If their placement and consequence are explicit, the estimate remains attached to the question it answers.
