Missing not at random values, often referred to as MNAR (“Missing Not At Random”), occur when the probability of missingness depends on the unobserved value itself, even after accounting for observed information. An example is a symptom value that is not collected because a severe deterioration makes further participation in the study difficult. This missingness mechanism cannot be demonstrated from the available data alone and therefore requires justified assumptions.
Why MNAR is not testable
For a missing value, the very information that could explain whether its absence is related to its magnitude is not available. Different MNAR models can fit the same observed data yet lead to different treatment effects. A statistical test therefore cannot determine whether the mechanism is present or which MNAR scenario applies.
The assessment starts with the data collection pathways: time of discontinuation, documented reason, prior endpoint values, concomitant therapy, and other clinical events may indicate informative dropout. They do not provide a direct observation of the missing value, but they make assumptions about its plausible trajectory more transparent. For confirmatory trials, the EMA requires a presentation of the reasons, patterns, and potential consequences of missing data.
Models and sensitivity scenarios
MNAR is often not formulated as the sole primary assumption, but is examined through scenarios that deviate from a MAR-based primary analysis. Pattern-mixture approaches describe the endpoint trajectory separately by missingness patterns; a delta adjustment shifts unobserved values by a clinically justified amount. Reference-based imputation can also express an MNAR-like assumption about the course after treatment discontinuation.
Which approach is appropriate depends on the estimand. First, it must be defined whether treatment discontinuation, a change in concomitant medication, or death changes the clinical question. Only then can it be determined which data are actually missing and how an assumption about them enters the estimation. A sensitivity analysis without reference to this objective may be a plausible computational variant, but it does not necessarily answer the same clinical question.
Distinction from MAR and from missing-data
MAR is the counterpart to MNAR: under MAR, conditional on the observed variables included in the model, missingness does not additionally depend on the missing endpoint. An MMRM or standard imputation can be used under this assumption. MNAR, by contrast, requires an additional specification for the unobserved part of the trajectory that cannot be verified from the data.
The existing missing-data entry covers all missing information and the mechanisms MCAR, MAR, and MNAR. Missing not at random values do not describe the amount of missing observations, but a specific dependency of their occurrence. Even a small proportion of missing values can be MNAR-relevant if its reason is closely linked to a condition that is important for the endpoint.
MNAR should also not be equated with a single reason for discontinuation. Two patients may both stop treatment due to lack of efficacy and still have different uncollected values. A scenario must therefore explain which deviation is assumed for which patient group and which data available after the event constrain this assumption.
Relevance for clinical trials
MNAR risks do not only emerge during the final analysis. The protocol may stipulate follow-up after the end of therapy, the recording of reasons for discontinuation, and the documentation of subsequent therapies, allowing the assumption of an informative dropout to be verified against concrete data. During the data review, missing values should be broken down by treatment group, visit, and event pathway, rather than merely being reported as an overall rate.
Full-service CROs such as Mediconomics support the assignment of data collection discontinuations and intercurrent events to the estimand, the planning of MNAR-oriented scenarios, and the programming of sensitivity analyses. Biostatistics, the clinical project team, and medical writing can substantiate why the assumed deviations are relevant for the indication and endpoint and how they contextualize the primary finding.
Each scenario should be assigned to a clearly defined analysis population.
Frequently Asked Questions (FAQ)
Does a high proportion of treatment discontinuations prove that the data are MNAR?
No. Treatment discontinuations can increase the risk of informative missingness, but they do not prove the missingness mechanism. What matters are the subsequent study status, the data actually collected after discontinuation, and the clinical plausibility of an association with the unobserved endpoint.
Is MNAR synonymous with data manipulation?
No. MNAR describes a statistical dependency that can arise even with correctly collected data due to disease severity, intolerance, or care settings. The term makes no statement about the integrity of study staff or patients.
Does every study have to use an MNAR primary analysis?
No. The primary analysis should match the estimand and the most plausible prespecified assumption. If MNAR is a relevant alternative explanation, a sensitivity analysis should show how strongly the conclusion depends on that assumption.
Regulatory References
- ICH E9(R1), addendum on estimands and sensitivity analyses – requires explicit handling of missing data for the estimand.
- EMA/CPMP/EWP/1776/99 Revision 1, Guideline on Missing Data in Confirmatory Clinical Trials – defines MCAR, MAR, and MNAR.
- EMA/CHMP/205/95 Rev.6, Guideline on the Clinical Evaluation of Anticancer Medicinal Products – highlights informative censoring and sensitivity analyses.
