Planning for dropout and missing data
Missing data are not one problem with one solution. A participant may skip an item, miss the entire follow-up survey, leave the organization, or become ineligible for a later assessment. A record may also be missing because the collection process failed.
SQUIRE asks authors to report missing data as part of improvement results. [1] Plan that reporting before collection. Otherwise, you may know how many rows are in the file without knowing how many people were eligible or why observations are absent.
Keep the participant counts distinct
Hypothetical example. A project invites 30 eligible staff members. Twenty-six complete baseline, 21 complete follow-up, and 19 have usable linked scores at both times. A paired analysis uses 19 pairs. The other counts still describe recruitment and response.
| Missingness situation | What to record |
|---|---|
| One unanswered item | The item and the instrument’s rule for forming a score. |
| Entire follow-up missing | Response status and an approved reason, when known. |
| Responses present but not linkable | The number of unmatched observations and linkage limitations. |
| No assessment opportunity | Why the observation was not applicable, rather than treating it as a failure. |
A follow-up that is not yet due is different from a follow-up that was due but not completed. Keep observation windows in mind before classifying records.
Prevention starts with the design
Use a survey that is no longer than necessary. Make code instructions clear, provide a reasonable response window, and use approved reminders. Check exports early for broken scoring, unexpected blanks, or duplicate identifiers. These are opportunities to fix the process, not reasons to change the outcome after seeing its results.
Do not require unnecessary sensitive questions solely to eliminate blanks. A “prefer not to answer” response communicates something different from a technical failure and must not be scored as an ordinary substantive response.
Do not fill gaps automatically
A missing score is not zero. Replacing it with the group average or copying the baseline score forward makes assumptions about what would have been observed. It does not recover the actual missing response.
Follow the instrument’s missing-item rules and discuss the analysis strategy with your statistical advisor. Advanced methods also depend on assumptions and may not be useful for every small project. A transparent descriptive analysis can be preferable to an elaborate but poorly supported correction.
Consider whether people missing follow-up might differ from those retained. For example, staff who found training less useful might be less likely to respond. This is a possible explanation to evaluate, not a missingness mechanism you can prove from the missing values alone.
Your next decision
Create a response flow record before launch. In the report, give the number contributing to each analysis, explain exclusions, and distinguish known reasons from guesses. See Pre-post surveys: Same instrument, linked responses for linkage and Setting up your project data for coding.