A practical evaluation of finite-sample BOIN calibration and Bayesian smoothing in phase I dose finding
We evaluated an asymmetric loss-based BOIN boundary and a monotone Bernstein estimator. The boundary preserved lookup-table operation but did not improve correct selection uniformly in seven monotone scenarios. It reduced overdose selection relative to BOIN, aBOIN, and gBOINS in six. Its accuracy depended on MTD position and overdose aversion. Bernstein selection reduced accuracy in most scenarios, especially for middle and high MTDs. Reference MCMC showed that importance-sampling error did not explain these losses. Target-based elimination eventually removes a repeatedly sampled MTD when toxicity exceeds the target. An upper-equivalence-bound threshold avoids this asymptotic conflict but not finite-sample elimination. Standard BOIN should remain the default. The loss-based boundary is justified only when the protocol prespecifies overdose avoidance and calibration shows an acceptable tradeoff. The Bernstein estimator is not recommended in its present form.
Authors
- Siyi Chen (ORCID: https://orcid.org/0000-0002-0153-6152)
Institutions
- Louisiana State University Health Sciences Center New Orleans (US)
Publication Details
- Journal
- Journal of Biopharmaceutical Statistics
- Published
- 2026-11-14
- DOI
- https://doi.org/10.1080/10543406.2026.2762006
- Primary Topic
- Statistical Methods in Clinical Trials
- Type
- article
- Field-Weighted Citation Impact
- 0.00