What is missing, and from which stage?
A study can be missing a questionnaire response, a laboratory result, an entire follow-up visit, or information about someone who was eligible but never enrolled. Those are different problems. A dataset that looks complete for its enrolled participants can still describe a selected part of the intended population. Conversely, a mostly complete study can have missing information concentrated in the outcome that matters most.
A missing value also needs to be distinguished from a genuine zero, an inapplicable question, or a value below a detection limit. Treating all four as the same code can manufacture a pattern. The meaning of the absence should be understood before choosing a statistical method to handle it.
Why the percentage alone is insufficient
Suppose 10% of participants miss a final assessment. If appointments were lost because of an unrelated scheduling error, the implications may differ from a situation in which people with worsening symptoms were less able to attend. The numerical fraction is identical, but the observed outcomes may represent the full group differently.
Even a small amount of missing information can matter when events are rare or when missingness differs between comparison groups. A larger amount need not produce large bias in every setting, although uncertainty may increase. The relevant questions concern who is missing, which values might be absent, and whether that absence is related to the conclusion being estimated. A universal acceptable percentage cannot answer those questions.
What “missing at random” actually means
The phrase is easy to misread. In statistical usage, missing at random means that, conditional on the observed information included in the model, missingness does not depend on the unobserved values themselves. It does not mean the blanks were scattered arbitrarily through a spreadsheet. Whether the assumption is plausible depends on what was measured and why information was lost.
If age predicts both follow-up attendance and the outcome, including age may help account for that pattern. If attendance also depends on an unrecorded deterioration not explained by the available variables, the assumption may remain questionable. Observed data generally cannot establish that all such dependence has disappeared. The assumption needs a substantive argument, not just the name of a software option.
What happens when incomplete records are dropped?
Analyzing only complete records can be straightforward, but it changes which participants contribute. It may also discard useful information and reduce precision. Whether it biases a particular estimate depends on the missingness process and analysis; it is not automatically valid or automatically wrong.
A useful example is a comparison of two programs where participants struggling with one program are more likely to leave. An analysis of completers might then describe people who could tolerate or benefit from that program, rather than the group originally offered it. Random assignment at the start does not prevent selection introduced afterward. Cochrane's trial-bias guidance treats missing outcomes as a distinct concern for that reason.
Imputation represents uncertainty rather than erasing it
Imputation uses available information and a model to supply plausible values for missing observations. Multiple imputation creates several completed versions of a dataset and combines analyses in a way intended to reflect uncertainty about the missing values. It is not evidence that those values were observed, and it does not make an unsuitable model adequate.
The model needs to fit the analysis and the structure of the data. Repeated measurements, constrained values, and relationships among variables can matter. Simply filling every blank with an average can distort variability and associations. The question is not whether a spreadsheet has been made complete, but whether the procedure supports the intended estimate under defensible assumptions.
How sensitivity analysis changes the conversation
A sensitivity analysis asks what would happen under alternative plausible assumptions. For example, the outcomes of missing participants could be worse than predicted by the primary model. Examining how much worse they would need to be to change the conclusion can make the dependence on missing information more concrete.
This does not identify the true missing values. It shows which conclusions are stable and which rely heavily on a particular account of the unseen data. An analysis that changes substantially under modest alternatives calls for a different interpretation from one that remains similar across a defensible range. The range itself still needs justification; extreme scenarios alone may not illuminate the uncertainty relevant to the study.
What a transparent report includes
A reader benefits from counts by group and time point, known reasons for missingness, the method used, and the assumptions behind it. Reporting can distinguish a participant who withdrew from someone whose sample was unusable. Those details help connect the statistical approach with what occurred during data collection.
Missing information does not make a study worthless. It limits what can be learned under particular assumptions. A careful report makes that limit visible instead of allowing a tidy final table to imply that every enrolled participant contributed the same information.
