What population trends can and cannot tell you

There is a version of this site’s argument that would be easy to write and would be wrong: “prescribing fell, deaths rose, therefore the guideline caused the deaths.” We do not make that claim, and it is worth explaining why, because the same discipline applies to claims we would like to be true.

What is actually observed

Opioid prescribing fell substantially after 2016. Opioid-involved overdose deaths rose from 42,435 in 2016 to a peak of 82,851 in 2022, then fell to 55,007 in 2024 and a provisional 45,008 in 2025.

What that supports

One inference, and it is a strong one: whatever was driving mortality across that period was substantially not the volume of prescribing, because the volume was moving the other way. That is a claim about what the driver was not, and it is well supported.

What it does not support

That the guideline caused the increase. Too much else changed at once — illicitly manufactured fentanyl entering the supply at scale, then becoming near-universal in it; the pandemic; changes in treatment access; changes in the population at risk. Any of those could dominate, and national aggregates cannot separate them.

It also does not support the reverse claim, that prescribing reduction was working and something else got in the way. Same problem, same data.

The 2024 fall makes the point

Deaths fell by roughly a third in two years. Naloxone distribution, changes in the composition of the illicit supply, expanded treatment, and a shrinking susceptible population are all proposed, and it is genuinely unresolved. Anyone who can confidently attribute a 27,000-death annual change to one cause could also have predicted it, and nobody did.

What to ask for instead

• Individual-level data rather than aggregates, where it exists — the tapering cohort studies are far more informative than the national series precisely because of this.
• A comparison group. Which is why pilot design insists on one.
• A dose-response relationship. The finding that faster tapers are associated with worse outcomes is more persuasive than the association itself, because gradients are harder to manufacture through confounding.
• Explicit statements of what could not be measured. The 2014 CDC evaluation of the Florida crackdown said outright that its data “did not permit any assessment of potential unintended consequences… such as reduction of access to pain medication for legitimate prescribing indications.” That sentence is worth more than most findings.

More: after the 2016 guideline.