Informative label missingness occurs when the pattern of missing labels is not random but itself contains information about class structure. Previous work has shown that classifiers can exploit this signal to improve efficiency. A new paper by Farinborz Setoudehtazangi and Geoffrey J. McLachlan asks whether that advantage is stable when the deployment objective changes.
The answer is no. The authors demonstrate 'decision-target reversal': holding the data-generating model and statistical experiments fixed, informative partial classification can be favourable under one loss function and unfavourable under another. For binary Gaussian discrimination, they derive a local cost-sensitive excess-risk expansion and, in one dimension, explicit critical cost ratios at which the first-order risk ordering reverses. They then extend the analysis to multiclass finite-action problems using loss-specific active-boundary curvature, producing a general decision-target frontier.
The required covariance geometry can arise under finite uncertainty-dependent missingness mechanisms with a fixed marginal missing-label proportion. Monte Carlo experiments illustrate the predicted reversals, and a descriptive gastrointestinal-lesion application with agreement-defined label availability shows sensitivity to deployment costs. The authors conclude that favourable informative missingness is a property of the statistical experiment relative to the deployment decision, not of the experiment alone.