Accounting for Underspecification in Statistical Claims of Model Superiority

TL;DR

This work addresses the risk that statistical claims of superiority in medical imaging are undermined by underspecification, i.e., seed-induced performance variance. It extends the Bayesian false-claim framework by adding an additive underspecification variance term, formalized as $SE^2_{AB,\text{underspec.}}$, and demonstrates this across segmentation (Dice Score Coefficient) and classification (AUROC) tasks. Using reproducibility-derived variability estimates, the authors show that even modest seed variability ($\delta \approx 0.01$) substantially raises the evidence required to claim superiority, increasing the probability of false outperformance when comparing models trained with different seeds. The study provides estimated parameter values ($s_{\text{seg}}=0.197$, $s_{\text{clf}}=0.737$) and argues for broader validation practices, including stress-testing across seeds and subgroups, to improve robustness of medical-imaging evaluations.

Abstract

Machine learning methods are increasingly applied in medical imaging, yet many reported improvements lack statistical robustness: recent works have highlighted that small but significant performance gains are highly likely to be false positives. However, these analyses do not take \emph{underspecification} into account -- the fact that models achieving similar validation scores may behave differently on unseen data due to random initialization or training dynamics. Here, we extend a recent statistical framework modeling false outperformance claims to include underspecification as an additional variance component. Our simulations demonstrate that even modest seed variability ($\sim1\%$) substantially increases the evidence required to support superiority claims. Our findings underscore the need for explicit modeling of training variance when validating medical imaging systems.

Figures (1)

• Figure 1: Accounting for underspecification drastically increases the probability of false claims.(Top left) Reproduction of the results of christodoulou2025false for segmentation. (Top right) Simulation of the impact of underspecification on segmentation using $\delta = 0.01$. The dashed line shows the trajectory of false claim probabilities above 0.05 without underspecification. (Bottom left) Reproduction of the results of christodoulou2025false for classification. (Bottom right) Simulation of the impact of underspecification on classification using $\delta = 0.01$.