(C) BoingBoing
This story was originally published by BoingBoing and is unaltered.
 . . . . . . . . . .



The overdose spike that alarmed NPR was a model trained on old data [1]

['Ellsworth Toohey']

Date: 2026-04-14  

In June 2025, federal data suggested US drug overdose deaths had suddenly surged in January, reversing a decline that had been going on for well over a year. NPR covered it. Public health officials were alarmed. Some people wondered if the Trump administration was monkeying with the numbers. According to a new analysis in the American Journal of Public Health, none of that was real. The spike was an artifact of a prediction tool that couldn't handle a downward trend.

The tool is called UCD Predict, and the National Center for Health Statistics uses it to estimate current overdose death counts while the slow work of toxicology reports and death investigations catches up. It had worked fine for about a decade. But when the government retroactively retrained it on data from 2022 and 2023, when deaths were still climbing, and then applied the revised model to a period when deaths were falling, it produced phantom deaths. For January 2024, it overshot the real count by more than 5,000. For December 2024, it somehow predicted fewer deaths than actually occurred, a result the researchers describe as implausible.

A later government revision corrected the numbers and confirmed what had really been going on: overdose deaths have been dropping since August 2023, the longest sustained decline in more than four decades. The researchers aren't calling for UCD Predict to be scrapped, but they want the government to start publishing a change log when it tweaks the model, so journalists and policymakers aren't blindsided by phantom surges.

[END]
---
[1] Url: https://boingboing.net/2026/04/14/the-overdose-spike-that-alarmed-npr-was-a-model-trained-on-old-data.html

Published and (C) by BoingBoing
Content appears here under this condition or license: Creative Commons BY-NC-SA 3.0.

via Magical.Fish Gopher News Feeds:
gopher://magical.fish/1/feeds/news/boingboing/