A result calculated from available measurements may describe only the units that remained observable. To judge its meaning, find out what is missing and why. The problem is not only a smaller sample: the observed units may differ systematically from those whose outcomes are unavailable.
Identify the missing part of the experiment
Cochrane’s guidance on randomised trials assesses missing outcome data through the reasons for missingness and their relationship to the unobserved outcome. It cautions that the proportion missing alone cannot provide a universal boundary between harmless and consequential missingness.Cochrane Handbook — Assessing risk of bias in a randomized trial (opens in a new tab)
The same basic reading question arises in laboratory evidence: did a unit never enter the study, did a measurement fail, or was an obtained value excluded? These gaps occur at different stages and can affect different conclusions.
| Stage | Useful record |
|---|---|
| Before assignment | Eligibility and inclusion counts |
| After assignment | Units lost or unavailable in each condition |
| At measurement | Outcome-specific failures and reasons |
| At analysis | Exclusions and handling of missing values |
Do not treat a missing value as a measured zero. Zero is an outcome on a defined scale; missing means the relevant outcome is unknown or unavailable. Coding both identically introduces an assumption into the analysis.
Explore how the same observed data allow different totals
Consider an original hypothetical group of twenty units with a binary response outcome. Sixteen outcomes are observed, and eight of those units respond. The observed-case response proportion is 8/16 = 50%.
If none of the four units with missing outcomes responded, the full-group proportion would be 8/20 = 40%. If all four responded, it would be 12/20 = 60%. Those are extreme possibilities consistent with the stated binary data, not recovered observations.
| Assumption about four missing units | Full-group proportion |
|---|---|
| No responders | 8/20 = 40% |
| Two responders | 10/20 = 50% |
| Four responders | 12/20 = 60% |
This range is a simple scenario bound, not a confidence interval. A comparison with another group would also require its observed and missing outcomes. Filling the gaps with one convenient assumption does not make that assumption factual.
Ask why outcomes became unavailable
An accidental file loss unrelated to the measurements presents a different concern from a measurement becoming unavailable because the biological response disrupted the readout. The second process may preferentially remove particular outcomes.
The recorded reason is useful but may be incomplete. A label such as sample lost does not explain whether loss occurred before exposure, during measurement or after results were visible. Those details help assess how missingness might relate to the comparison.
Equal missing percentages in two groups do not prove equal consequences. Missing units could have different likely outcomes in each group. Conversely, unequal percentages are a prompt for investigation rather than a numerical estimate of the bias by themselves.
Read how the analysis handles what is unknown
Cochrane notes that an intention-to-treat label can coexist with exclusions of missing outcomes and that methods for accounting for missing data rely on assumptions. The label alone does not establish that missing-data bias has been removed.Cochrane Handbook — Assessing risk of bias in a randomized trial (opens in a new tab)
Look for the actual handling: complete-case analysis, an imputation model, another model-based approach or sensitivity analyses. Each should be explained in relation to the design and the information available about missingness.
A sensitivity analysis asks whether a conclusion changes under plausible alternatives. Agreement across a narrow set of convenient assumptions is less informative than examining the alternatives that the missingness process makes relevant.
An evidence summary should state the observed counts, the missing counts, the reasons given and the dependence of the conclusion on assumptions. This makes an incomplete dataset interpretable without pretending that its gaps have disappeared.
Sources and further detail
- Cochrane Handbook — Assessing risk of bias in a randomized trial (opens in a new tab)
Section 8.5 read for mechanisms, absence of a universal missingness cutoff, analysis labels and assumptions. Clinical-trial scope is stated. The twenty-unit example is original and its 40–60% range is a scenario bound, not an estimated confidence interval.
Sources checked 19 September 2026. Worked examples are illustrative unless a supplied report is explicitly identified. This article has not undergone independent scientific peer review.