Handling missing responses in patient-reported outcome data

Describe the missingness first

Missing data in patient-reported outcomes come in two shapes. Item-level missingness is a participant skipping one or two questions inside an instrument. Unit-level missingness is a participant skipping an entire questionnaire or a whole follow-up visit. They call for different handling, and a manuscript should report both counts separately.

Before choosing any method, tabulate how much is missing and check whether it relates to observed characteristics. Participants with worse baseline scores who stop completing follow-ups are a common pattern in clinical cohorts. When the probability of missingness depends on things you measured, the standard assumption is missing at random, and methods that use those measured variables can recover much of the lost information. When it plausibly depends on the unobserved outcome itself, no method fully fixes it, and the honest response is a sensitivity analysis.

Item-level gaps: use the instrument's rule

Most validated instruments specify how to handle skipped items. Some prorate the total when a minimum share of items is answered. Some declare the score missing if any item is skipped. PROMIS provides pattern-based scoring for partial responses. Apply the published rule, report how many respondents it affected, and resist the temptation to invent a more lenient one.

Questionnaire-level gaps: the defensible options

Multiple imputation creates several completed datasets by drawing plausible values conditioned on the observed data, analyzes each, and pools the results so the uncertainty from imputation is carried into the confidence intervals. It is widely accepted in clinical journals when the imputation model includes the variables that predict missingness and the outcome. Reporting the number of imputations, the variables in the model, and the software is expected.

Likelihood-based methods such as mixed-effects models for repeated measures use all available observations without filling anything in, and they are valid under the same missing-at-random assumption. For longitudinal PRO data with intermittent missing visits, this is often the simplest defensible route.

Sensitivity analyses test how much the conclusion depends on the assumption. Common versions compare the primary result against complete cases, against a pattern-mixture model that assumes dropouts had worse outcomes, or against tipping-point analysis that finds how strong the departure from the assumption would need to be to change the conclusion.

What reviewers reject

Complete-case analysis, meaning dropping anyone with a missing value, is not wrong in itself, but it shrinks the sample and biases the estimate whenever the dropped participants differ from the retained ones. Reviewers usually ask for it to be a sensitivity analysis rather than the primary analysis.

Single imputation with the mean or the last observation carried forward understates uncertainty because the imputed values are treated as if they were observed. Both methods have been discouraged in methodological guidance for over a decade, and regulators have moved away from last observation carried forward in trials.

• Report item-level and unit-level missingness separately, with counts.
• State the assumed missingness mechanism and why it is plausible.
• Name the method, the software, and for multiple imputation the number of imputations and the model variables.
• Include at least one sensitivity analysis.