The Geometry of Informativeness: A Bregman-Centroid Diagnostic for Partial Pooling in Multisite Trials
Abstract
Two multisite trials can have the same informativeness yet differ in how reliably individual site effects can be estimated. We show that the informativeness index averages site-specific reliabilities on the logit scale and is a Bregman centroid under Bernoulli-Kullback-Leibler divergence. This explains the geometric mean of sampling variances and yields a measure of profile dispersion. Under stated conditions, we relate profile dispersion to estimation error and show when balanced allocation maximizes informativeness at fixed total enrollment. Simulations show that estimator performance depends on the reliability profile and whether the target is individual effects, ranks, or their distribution. An online credit recovery trial illustrates the diagnostics and their conditional posterior uncertainty. We recommend reporting informativeness with arithmetic mean reliability and profile dispersion. The accompanying R package poolinfo provides tools for computing the diagnostics, posterior summaries, and design calculations.
Related
- Consistent Spatial Clustering of Complex Objects via Bregman Geometry
- A positive solution to the Lp projection centroid conjecture
- Breaking the Fourth Wall: The Planar Centroid Banach--Mazur Diameter Is Strictly Less than $4$
- Relax and Follow: L0-Path Computation with L0-Bregman Relaxations
- Task-Oriented Quantization for Quadratic Scheduling: Centroid Water-Filling and Power-Diagram Encoders
BibTeX
@misc{lee2026geometry,
title = {{The Geometry of Informativeness: A Bregman-Centroid Diagnostic for Partial Pooling in Multisite Trials}},
author = {JoonHo Lee},
howpublished = {arXiv},
year = {2026},
month = oct,
doi = {10.48550/arxiv.2610.24398},
eprint = {2610.24398},
archivePrefix = {arXiv},
url = {https://arxiv.org/abs/2610.24398},
}
RIS
TY - UNPB
TI - The Geometry of Informativeness: A Bregman-Centroid Diagnostic for Partial Pooling in Multisite Trials
AU - JoonHo Lee
PY - 2026
DA - 2026/10/17
PB - arXiv
DO - 10.48550/arxiv.2610.24398
UR - https://arxiv.org/abs/2610.24398
AB - Two multisite trials can have the same informativeness yet differ in how reliably individual site effects can be estimated. We show that the informativeness index averages site-specific reliabilities on the logit scale and is a Bregman centroid under Bernoulli-Kullback-Leibler divergence. This explains the geometric mean of sampling variances and yields a measure of profile dispersion. Under stated conditions, we relate profile dispersion to estimation error and show when balanced allocation maximizes informativeness at fixed total enrollment. Simulations show that estimator performance depends on the reliability profile and whether the target is individual effects, ranks, or their distribution. An online credit recovery trial illustrates the diagnostics and their conditional posterior uncertainty. We recommend reporting informativeness with arithmetic mean reliability and profile dispersion. The accompanying R package poolinfo provides tools for computing the diagnostics, posterior summaries, and design calculations.
ER -