Sample size buys detection, not localisation: an identifiability limit for hippocampal subfield morphometry

Abstract

Automated segmentation has made hippocampal subfield volumes a routine measurement, and studies now report which subfield relates to an outcome rather than whether the hippocampus does. Those reports do not agree with one another, and the standard explanation is insufficient statistical power. We argue that a second limit operates independently of sample size. Using 638 participants from a population-derived adult lifespan cohort, we first show that no individual subfield contributes to a general cognitive factor beyond a single global size component: no coefficient interval excludes zero, the local block carries half a percent of outcome variance, and no model improves out-of-sample prediction over the global factor alone. Because an observed null cannot distinguish an absent effect from an effect the design cannot locate, we then planted effects of known location and size in the measured design and in a whitened copy of it that preserves sample size, dimensionality and effect size while removing only the correlation between subfields. The arms were paired down to the noise vector. Collinearity did not place recovery out of reach; it multiplied the required sample size by a factor of roughly two to three, and the penalty widened as cohorts grew. At the effect sizes this literature reports, neither design reached an adequate recovery rate at any sample size, and coarsening the parcellation rescued neither. The choice of estimator moved recovery further than collinearity did. We provide a calibration surface on which a planned design can be located before data collection.