Statistical interpretation
Missing data, attrition, and censoring: assumptions behind incomplete outcomes
Missing data are not repaired by choosing a convenient imputation rule. Prevention, continued outcome collection, documented reasons, estimand alignment, and sensitivity analyses determine how much confidence an incomplete dataset can support.
Educational content only. Not medical advice.
Missingness begins as a design and follow-up problem
Outcomes can be missing because visits are missed, participants withdraw, measurement fails, sites close, data are out of range, or collection stops after intervention discontinuation. Preventing avoidable missingness and continuing relevant follow-up preserve more information than any later statistical method. Reports should show how much is missing, when, in which groups, and why.
Attrition can erode the randomized comparison
If loss to follow-up is related to prognosis, assigned group, adverse events, or observed response, participants with outcomes may no longer represent those randomized. Equal attrition percentages do not guarantee equal bias because reasons and unobserved outcomes can differ. Complete-case analysis is valid only under assumptions that are often too strong.
Censoring is structured incomplete time-to-event information
In time-to-event analysis, censoring indicates that the event was not observed beyond a participant's last known time or study boundary. Standard methods commonly assume censoring is non-informative conditional on modeled information. Administrative study-end censoring differs from loss caused by deteriorating health, withdrawal, or competing events, so reasons and timing matter.
Every method carries assumptions that need stress-testing
Multiple imputation, weighting, likelihood models, pattern-mixture models, and other approaches answer questions under assumptions about unseen values. Sensitivity analyses test whether conclusions change under plausible departures. Simple last-observation-carried-forward rules can create implausible certainty. The estimand, intercurrent-event strategy, analysis, and sensitivity checks should form one coherent plan.
Evidence limits
- The true values of missing outcomes are unobserved, so assumptions cannot be verified completely from the data.
- Different missing-data mechanisms and estimands require different analyses.
- Low percentages alone do not prove that missingness is harmless or non-informative.
Sources and further reading
These sources ground the definitions and evidence boundaries on this page. A citation is a route for verification, not an endorsement of a product or personal use.
National Academies Press
The Prevention and Treatment of Missing Data in Clinical Trials
Consensus guidance on preventing missing data, documenting assumptions, and using sensitivity analysis.
Open sourceU.S. Food and Drug Administration / ICH
E9(R1) Statistical Principles for Clinical Trials: Estimands and Sensitivity Analysis
Official framework connecting trial objectives, estimands, intercurrent events, analysis, and interpretation.
Open sourceCochrane Handbook for Systematic Reviews of Interventions
Assessing Risk of Bias in a Randomized Trial
Authoritative methods guidance on randomization, allocation concealment, deviations, missing outcomes, measurement, and selective reporting.
Open sourceCommon questions
Is complete-case analysis unbiased?
Only under restrictive assumptions about why data are missing and how that relates to outcomes and covariates.
What is non-informative censoring?
It means event risk after censoring is adequately represented by participants still observed, conditional on the model.
Can imputation recover the true missing values?
It provides principled estimates under assumptions; it does not reveal the unobserved truth, so sensitivity analysis remains important.
Continue with context
Keep building your evidence-reading skills
Explore the public PeptideSchool research library for more source-backed methods, glossaries, and evidence maps.