Apparent drug overdose spike in 2025 driven by flawed modeling

Last June, several mainstream media outlets reported a surge in U.S. drug overdose deaths in early 2025 that was based on data from the U.S. Centers for Disease Control and Prevention (CDC). 

A new Northwestern University study has found that surge was not real, and the apparent spike was a statistical modeling artifact, not a true reversal in overdose trends. In fact, the study found overdose deaths have continued to decline following a peak in August 2023, marking the longest sustained decrease in more than four decades. 

The findings directly address speculation that the CDC made data reporting errors or intentionally mischaracterized or concealed public health data due to political pressure or interference, said lead author Lori Ann Post.

"Many people think CDC drug overdose data are being cooked, but they're not," said Post, director of the Buehler Center for Health Policy and Economics at Northwestern University Feinberg School of Medicine. "We can trust them because they're scientists trying to do the best job they can with difficult circumstances. There was no clear incentive for any administration to inflate these numbers. This was not politics."

The study was published today (April 8) in the American Journal of Public Health.

What happened?

To prepare the June 2025 report, CDC scientists used data from prior years (2022 and 2023) when there was a massive acceleration of drug overdose deaths, Post said. Because of that, provisional federal data suggested overdose deaths had risen sharply in January 2025, prompting national concern and media coverage. 

"That spike made headlines at the time, but it didn't reflect reality," Post said "What we found was a mismatch between predictive models and a rapidly changing epidemic. CDC scientists did the best job they could with fewer people, more constraints and more people watching them."

Subsequent revisions to federal estimates corrected the overestimate and confirmed that overdose deaths had continued to fall, Post said. 

The U.S. relies on provisional mortality estimates from the National Center for Health Statistics, which uses statistical models to account for reporting delays in death investigations. These models performed well for years but struggled when overdose deaths began declining after a prolonged period of rapid growth, Post said. 

The study authors emphasize the episode highlights a broader issue: surveillance systems are most vulnerable during turning points, when trends shift direction.

Why it matters

Accurate data are essential for public health response. Misinterpreting trends can misdirect policy decisions, undermine public trust and distort resource allocation, Post said. 

Despite this episode, the researchers stress that federal mortality data remain the most reliable near real-time source for tracking overdose death.

Why the models broke

The overdose crisis has been reshaped by fentanyl, a synthetic opioid that spread unevenly across the country and drove rapid increases in deaths through 2022 and 2023.

Models trained on that explosive growth were later applied to a period of decline. The result:

• Overestimation of deaths in early 2024 and 2025
• False signals of a national "spike"
• Confusion among policymakers, researchers and the public

Looking ahead

The authors call for greater transparency in federal data systems, including advance notice of methodological changes and clear documentation of revisions.

"When the numbers change, people notice," Post said. "We need to make sure they understand why."

More about the study

The scientists used Northwestern's national, regional and state dashboard, the OD Pulse, which is designed to track drug overdose deaths between January 1999 to April 2025, with unprecedented timeliness and precision. They analyzed national overdose death data from 2015 through early 2025, comparing observed counts with multiple versions of federal predictive estimates.

The paper is titled, "The 2025 Drug Overdose Spike That Wasn't: Neither Politics nor Data Errors Explain the Anomaly." Co-authors include researchers from University of California, San Francisco, University of Maryland, Baltimore, and University of North Carolina at Chapel Hill.