Published on 17 September 2026
Contingency Management and Mortality: What a Veteran Cohort Can and Cannot Prove
The essentials
In Veterans Health Administration data, 1,481 patients who received contingency management for a stimulant use disorder were matched to an equal number of patients who did not. At one year, there were 27 deaths in the exposed group versus 46 among matched controls, an adjusted hazard ratio of 0.59. The result holds after excluding deaths attributed to COVID. It does not hold, however, on the more specific endpoints: neither overdose deaths nor stimulant-related deaths reach statistical significance, and patients followed in contingency management had more psychiatric hospitalizations. This pattern points toward an engagement-in-care effect rather than a specific effect on stimulant use itself. The title, abstract and conclusion of the publication speak of an association. The discussion, however, states that large-scale access to the program saves lives and prevents overdose deaths, a claim the data presented do not support.
Context
Stimulant use disorder has no medication approved by the U.S. Food and Drug Administration. Contingency management, which rewards a target behavior, typically biochemically verified abstinence, with a tangible incentive, is the intervention with the strongest evidence base, and the U.S. Department of Health and Human Services recommends it as first-line care. Meanwhile, stimulant overdose deaths keep rising.
What this evidence base has lacked is a hard endpoint. Trials measure short-term use behaviors, useful but limited in scope. No prior work had linked contingency management to real-world mortality, and the authors present theirs as the first to do so. In many health systems, including France, the question looks different: the intervention is available almost nowhere, and the very principle of rewarding a clinical outcome with a material incentive runs against deeply held practice norms.
The study at a glance
| Element | Details |
|---|---|
| Population | Adults with stimulant use disorder, Veterans health care network |
| DetailJuly 2018 through December 2020. 138,280 patients identified at baseline, of whom 92.9% were men per the results text and 92.7% per Table 1 of the publication. In the two matched groups, 94.7% were men, mean age 52.5 years. | |
| Exposure | Receipt of contingency management |
| DetailIdentified through electronic health record note templates, which assumes the practice was documented this way. 1,698 patients received contingency management over the period, 1,481 could be matched. The frequency and duration of the protocol are not described in the publication. | |
| Comparator | 1,481 matched controls drawn from the risk set |
| DetailMatched on facility, race/ethnicity, sex, age within two years, opioid use disorder, hospitalization, and addiction follow-up for stimulants in the preceding year. One-to-one matching with replacement: one control served more than one treated patient, so the 2,962 matched rows correspond to 2,961 distinct individuals. | |
| Endpoints | All-cause mortality at one year as the primary endpoint |
| DetailDeaths captured through the National Death Index, linked to the electronic health record. Exploratory endpoint: time to first psychiatric hospitalization, analyzed with a Fine and Gray model, with death treated as a competing risk. In the mortality model, hospitalization occurring during follow-up is handled as a time-dependent covariate. | |
| Design | Retrospective matched cohort on linked data |
| DetailOxford level of evidence 2b. The authors report following the STROBE guideline. Matching is performed on the risk set, with a thirty-day window after the index date meant to mimic the intention-to-treat analysis of a trial. | |
The findings
| Outcome | Value |
|---|---|
| All-cause mortality at one year | 27 of 1,481 (1.82%) versus 46 of 1,481 (3.11%) |
| ReadingCrude hazard ratio 0.58 (95% CI 0.36 to 0.94), p = 0.026; adjusted 0.59 (95% CI 0.36 to 0.95), p = 0.030, a 41% reduction. The publication’s abstract carries a slightly different value for the same adjusted estimate, 0.60 (95% CI 0.37 to 0.97), or 40%. The body text and Table 3 agree on 0.59, the value retained here. | |
| Overdoses | 10 deaths versus 15, hazard ratio 0.67 (95% CI 0.30 to 1.48), not significant |
| ReadingThe interval comfortably spans no effect. This does not prove there is none: with this number of events, the study lacked the power to settle the question. | |
| Stimulant-related deaths | 7 deaths versus 12, hazard ratio 0.58 (95% CI 0.23 to 1.48), not significant |
| ReadingSame caveat. The point estimate points in the same direction as the primary endpoint, but precision is lacking. | |
| Psychiatric hospitalization | Adjusted hazard ratio 1.48 (95% CI 1.25 to 1.75) |
| ReadingMore hospitalizations among patients followed in contingency management, both for all causes combined (45.9% versus 36.4%) and for psychiatric causes specifically. The most plausible reading, and the one the authors themselves put forward, is that these patients are seen more often, and therefore identified and referred more often. It is also what makes an engagement-in-care effect credible. | |
| Sensitivity analysis excluding COVID | Adjusted hazard ratio 0.61 (95% CI 0.38 to 0.98), p = 0.043 |
| ReadingNo death attributed to COVID in the exposed group, three among the controls. The primary result therefore does not rest on excess mortality from the pandemic period. | |
Critical appraisal
| Domain | Verdict |
|---|---|
| Residual confounding | The critical point |
| FindingReceiving contingency management requires showing up repeatedly for scheduled visits; the publication states neither how often nor for how long. This selects patients able to keep to that schedule, that is, patients whose prognosis is already better, independently of the intervention. No amount of matching on administrative data corrects this bias, because what produces it is not recorded. | |
| Number of events | Low |
| Finding73 deaths in total, 27 versus 46. The upper bound of the confidence interval approaches 1. On the raw comparison of published proportions, moving a single death from one arm to the other is enough to lose significance: Fisher’s exact test moves from p = 0.032 to p = 0.057. The primary model is a stratified Cox model, so this calculation is only an approximation, but it gives the order of magnitude. | |
| Discordance between endpoints | A non-specific signal |
| FindingA treatment for stimulant use disorder that reduced mortality through its effect on consumption should perform better on stimulant-related deaths than on all-cause mortality. The opposite is observed here, in terms of statistical significance. The argument is not decisive, for lack of power, but it points in a direction. | |
| Design and adjustment | Careful |
| FindingMatching on the risk set, hospitalization handled as a time-dependent covariate, a converging sensitivity analysis excluding COVID. The authors argue that the thirty-day window makes the estimate conservative if a true effect exists. The work is carried out with method, within the limits of an observational design. | |
| Transparency | Partial |
| FindingThe project was classified as a quality-improvement initiative by the Veterans Health Administration, with a non-research determination, and was therefore not reviewed by an institutional ethics board. No preregistration, no funding statement and no data availability statement appear in the document reviewed. The authors declare no conflict of interest. The study population was 94.7% male, which limits how far the result generalizes. | |
| How the authors frame it | Cautious, then causal |
| FindingThe title, abstract and conclusion stick to the language of association, and the cause-specific death counts are given in the table, not hidden. The discussion crosses that line: it states that large-scale access to the program saves lives, that it prevents overdose deaths, and that it confirms what had previously only been suspected anecdotally. The first claim is causal and the design does not support it; the second is contradicted by the non-significant result on overdoses. | |
Level of evidence
What is established: in this population, patients who received contingency management died less often within the year. What is not established: that the intervention caused this. Standing between the two is the healthy-adherer effect, the well-documented pattern whereby patients who follow a treatment do better, even when what they are given is a placebo. It is an old finding, it is well documented, and nothing in this design rules it out.
Should the result therefore be dismissed? No. A nationwide cohort built on a hard endpoint, in a condition with no approved medication, with converging sensitivity analyses, is not nothing. It is simply one piece of evidence alongside randomized trials, not a substitute for them.
The colleague test
What an experienced colleague might say if shown this study for two minutes, between two consultations.
“Fewer deaths among patients who show up regularly does not surprise me, and it does not yet tell me the program itself is what saves them. That said, in a condition where I have nothing to prescribe, I will take the signal, and I will call it a signal.”
Translated for practice: this is enough to argue for wider access to a recommended but underused intervention, provided the association is never presented as a demonstrated effect.
What you can take from this
- Keep the hypothesis, not the mechanism, as settled: engagement in care could account for part of the result. The study did not test this directly, and in its adjusted model, the number of addiction visits in the preceding year was not associated with mortality in either direction.
- When relaying the result, say association, never a mortality reduction achieved by the treatment. The distinction is one word, and it changes everything.
- Do not present the lack of statistical significance on overdoses as evidence of no effect on overdoses. The study lacked the power to settle that endpoint.
- Know that contingency management remains available in only a limited number of health systems, France among them, and that rewarding a clinical outcome with a material incentive raises ethical and legal questions that remain unresolved there.
- Read this work alongside the randomized trials in the field, which remain the foundation. A cohort study contributes a hard endpoint; it does not contribute comparability.
Frequently asked questions
Does a hazard ratio of 0.59 mean the program saves lives?
It means the exposed group died less often. Moving from that to a causal effect of the program requires ruling out that the two groups already differed before the intervention, which this study cannot do.
Why were there more psychiatric hospitalizations in the group that received contingency management?
The simplest explanation, and the one the authors themselves put forward, is that a patient seen often in outpatient care is a patient whose decompensations get noticed. A hospitalization can be a sign that care is working, not that it is failing.
Does this generalize beyond the Veterans Health Administration, for instance to France?
The delivery model itself does not: contingency management is available in only a small number of health systems worldwide, France among them. The broader argument, that a recommended intervention should be made available where it is currently absent, does generalize. As for the hypothesis of an engagement effect through closer clinical contact, it remains untested, not a finding of this study.
What would it take to settle the question?
A randomized trial with follow-up long enough and a sample large enough to observe deaths. Given the number of events required, this would not be an easy trial to run.
Annotated bibliography
Source study. Coughlin LN, Tomlinson DC, Zhang L, Kim HM, Frost MC, Khazanov G, McKay JR, DePhilippis D, Lin LA. Contingency management for stimulant use disorder and association with mortality: a cohort study. The American Journal of Psychiatry, 2025;182(11):1016-1023. Published online 10 September 2025. DOI 10.1176/appi.ajp.20250053, PMID 40926572. Verification was carried out against the author’s manuscript deposited in PubMed Central (PMC12872285, NIHMS2135837), not against the publisher’s final typeset version: pagination and exact wording may differ after production, and the internal discrepancy between the abstract and the body text on the adjusted hazard ratio cannot be arbitrated on this document. A supplementary appendix, cited by the publication for diagnostic codes, covariate definitions and sensitivity-analysis detail, was not consulted, and nothing that appears only there is reported here.
Editorial collections
Tags
This analysis is intended for healthcare professionals. It does not replace individual clinical judgment or current guidelines, and it does not constitute therapeutic advice.
Psychiatry Evidence Base, evidence-based psychiatry, explained with rigor.