Feature engineering and preprocessing · insight-1107
Bracket a suspected MNAR column with a delta adjustment instead of arguing about MAR versus MNAR
- Evidence type
- documented fact
- Difficulty
- hard
- Roles
- Data scientist, Data analyst
- Industries
- healthcare, insurance, public-sector
Summary: situation, action, reason, boundary
Situation
A decision-critical column, such as stated income or a survey score, may be missing because of its own value, and the team debates the mechanism from missingness plots.
Action
Fit the main analysis under MAR, then shift imputed values by a small set of domain-chosen offsets delta and report how the estimate and the decision move; say plainly if the decision flips.
Reason
MAR and MNAR cannot be told apart from observed data, so honest reporting shows how far the answer moves under plausible departures.
Boundary
Inferential or decision-critical analyses with a non-trivial missing share.
Symptom or risk
The conclusion rests on an assumption the observed data cannot check, while a single imputation looks authoritative and hides the range of plausible answers.
Exceptions
Tiny missing shares, or purely predictive work scored on a clean holdout, where slice checks matter more.
Trade-offs
Offsets are judgments to be agreed with domain experts, and they add reporting weight.
Pitfalls
Choosing delta to deliver the preferred answer, or running the sensitivity analysis with no plausible scenario behind it.
Lessons and concepts
Claims and their sources
Flexible Imputation of Missing Data section 3.8 says adaptations for nonignorable missingness rest on unverifiable assumptions and describes delta adjustment of imputed values as a sensitivity analysis.