Scoring Learned Health-State Value Sets by Decision Survival Under Preference Drift
A machine-learning paper from the School of Settled Matters and the School of Continuous Improvement treats a health-state value set as a model fitted once to a valuation cohort that cannot later be re-asked, and scores it by decision survival: the share of 400 simulated cost-utility comparisons whose recommendation holds when the population's preferences drift. Across five simulated valuation waves, the model with the lowest held-out error kept the fewest decisions (75.2% at wave 4, against 86.1% for a twenty-parameter additive model), and a decision-weighted, worst-segment training objective kept 90.4%, an advantage that shrank to under a point when the later preferences had no counterpart in the original cohort.
| Authors | Anselm Threlkeld , Mirela Hanke |
|---|---|
| School | School of Settled Matters |
| Output type | Research paper |
| Published | |
| DOI | 10.5555/slop.eymop7 |
| Pages | 6 |
| Version | 1.0 |
| Licence | CC BY 4.0 |
Cites
Cite as
@misc{slop_eymop7,
author = {Anselm Threlkeld and Mirela Hanke},
title = {The Cohort Cannot Be Re-Asked: Scoring Learned Health-State Value Sets by Decision Survival Under Preference Drift},
year = {2026},
publisher = {Slop University},
doi = {10.5555/slop.eymop7},
url = {https://slop.university/outputs/slop-paper-value-sets-fitted-eymop7/},
version = {1.0},
note = {Research paper},
}