What the studies found

Study	Population	Design	Principal finding
Agnoli et al., JAMA, 2021	113,618 US adults on stable ≥50 MME/day for 12 months (commercial and Medicare Advantage claims, 2008–2019)	Retrospective cohort	Post-taper: 9.3 vs 5.5 overdose events per 100 person-years (aIRR 1.68, 1.53–1.85); 7.6 vs 3.3 mental health crisis events per 100 person-years (aIRR 2.28, 1.96–2.65). Faster tapers worse.
Fenton et al., JAMA Netw Open, 2022	21,515 tapering events among 19,377 adults	Exposure-crossover cohort	Elevated risk persisted 12–24 months after taper initiation: aIRR 1.57 (1.42–1.74) overdose or withdrawal; 1.52 (1.35–1.71) mental health crisis.
Oliva et al., BMJ, 2020	1,394,102 Veterans Health Administration patients, FY2013–2014	Observational evaluation	Stopping opioids associated with higher risk of overdose or suicide death at every treatment length, rising with duration: HR 1.67 (≤30 days) to 6.77 (>400 days). Risk also elevated after starting.
Metz et al., J Gen Intern Med, 2024	60,040 non-cancer patients across 8 US health systems, 2012–2018	Multi-site cohort	Reductions of 30–<100% over two months vs 1–<15%: opioid overdose OR 1.41 (1.09–1.81); all-cause mortality OR 1.39 (1.16–1.67).
Perez et al., J Gen Intern Med, 2020	1,624 adults on ≥25 MME/day, urban health system, 2008–2012	Retrospective cohort	Taper associated with subsequent termination of care: AOR 4.3 (2.2–8.5).

What this evidence can and cannot establish

Every study above is observational, and every author group says so explicitly. That limitation is not a technicality and it deserves to be stated properly rather than waved at.

• Confounding by indication is the central problem. Patients who get tapered are not a random subset. They are tapered because something prompted it — a concerning result, a deterioration, a new prescriber, an aberrant behavior. Several of those prompts are themselves risk factors for the outcomes being measured. Statistical adjustment reduces this but cannot eliminate it.
• The direction of the residual bias is arguable. If tapering is triggered by risk, some of the observed harm belongs to the trigger. But the Oliva analysis found risk elevated after starting opioids too, which is harder to explain as pure indication bias and suggests transition periods themselves carry risk.
• No randomized trial of involuntary tapering exists, and it is difficult to see how an ethical one could now be designed given these findings.
• “Taper” covers very different things. A slow, collaborative reduction with support is not the same intervention as an abrupt mandated cut, and claims data cannot always distinguish them. The dose-response finding — faster tapers associated with worse outcomes — is the most informative feature here, because a dose-response relationship is harder to produce by confounding alone.

What good practice looks like

A consensus panel published in Mayo Clinic Proceedings set out patient protections for tapering, and it is the most useful practical document in this area. The core of it: tapering is appropriate when benefit no longer exceeds risk, it should be collaborative, it should be slow, it should be paused or reversed when the patient deteriorates, and it should never be a condition of remaining in care.

A separate analysis argued that non-consensual dose reduction mandates are not justified either clinically or ethically. That is an argument rather than a finding, and it is made by authors who also produced the primary cohort evidence above.

CDC’s 2022 position is quoted on the prescribing guidelines page.

Why this shapes the program

It is why mental health crisis events and loss of prescribing clinician are required outcome measures rather than optional ones, and why reduction in prescribing volume is not a success measure. A program that produced tapers and called them a result would be reproducing the harm this evidence describes.

Questions this raises

Does this evidence prove that tapering causes overdose?

No, and no author cited claims it does. All five studies are observational and subject to confounding by indication. What they establish is that the expected safety benefit is not visible and that harm signals appear consistently across independent datasets, with a dose-response gradient. That is enough to change practice under uncertainty. See observational evidence and what to do with it.

What does a well-run taper look like?

Collaborative, slow, initiated because benefit no longer exceeds risk rather than because of a number, paused or reversed if the patient deteriorates, and never a condition of remaining in care. The Mayo Clinic Proceedings consensus panel recommendations are the practical reference. See what a well-run taper looks like.