Published on 23 September 2026

Analysis · Schizophrenia · Pharmacoepidemiology

Antipsychotic nonadherence and mortality in schizophrenia: does the association hold up to scrutiny?

◆ Collection
European Neuropsychopharmacology · 2026; 104: 112745 · Correll et al.
DOI 10.1016/j.euroneuro.2025.112745
PMID 41435593
Scientific 66
Editorial 80

Editorial collections A Study Under the Microscope PEB Practice Change

The essentials

This American retrospective cohort includes 6,417 adults with a schizophrenia diagnosis who initiated or reinitiated an antipsychotic, identified in the Optum Clinformatics Data Mart healthcare claims database. Adherence was measured as the cumulative proportion of days covered by a pharmacy dispensing. During follow-up, 1,094 patients died, 17.1% of the cohort. Compared with patients whose coverage reached at least 80%, those below 20% coverage had an all-cause mortality hazard ratio of 1.80 (95% CI 1.43 to 2.27); patients with intermediate coverage had a hazard ratio of 1.53 (95% CI 1.25 to 1.89). The gradient is clean, the precision adequate, and five sensitivity analyses point the same way. Three caveats frame the reading. The design is observational and does not establish causality, despite the marginal structural model used, yet the body of the source publication reportedly uses causal wording for this same finding in its conclusion. Median follow-up is only 17.3 months, while the headline survival figures are five-year model estimates, and the number needed to treat is even projected to eight years. And the population is not what the word schizophrenia typically evokes: mean age at index was 55.3 years (SD 20.5), and roughly three patients in ten carried a comorbid neurocognitive disorder diagnosis. The work was funded by Johnson & Johnson Innovative Medicine, and five of the six authors are affiliated with that company.

Context

That stopping an antipsychotic raises the risk of relapse has been established for decades. Its relationship to mortality is less settled, and it is rarely expressed in a form usable at the bedside. A patient weighing whether to stop treatment does not hear a hazard ratio. What they hear is an order of magnitude, and, just as importantly, the uncertainty the clinician is willing to show.

The authors start from this gap. Their own abstract states it plainly: the impact of nonadherence on mortality remains unclear, and their conclusion stays at the same cautious register, that improved adherence may enhance survival and should be reinforced early in the disease course. What matters for this reading is that the full text of the article, according to the verified French version of this analysis, uses firmer, causal language in its own conclusion section, stating that suboptimal adherence increased mortality risk, a formulation the observational design does not support. The interest of this study is therefore as much methodological as epidemiological. It forces a distinction that a quick read tends to blur: the association itself, which is solid, and the derived quantities built on top of it, survival curves and numbers needed to treat, which are considerably less so.

The study at a glance

Population
6,417 adults retained from a pool of more than 2.1 million people who received an antipsychotic prescription. The schizophrenia diagnosis required at least two outpatient claims on separate dates, or at least one inpatient claim, with a schizophrenia diagnosis during the 12 months before or at the index date. The index date was antipsychotic initiation or reinitiation, with no antipsychotic prescription and no bipolar disorder diagnosis in the preceding 12 months, and 12 months of continuous plan enrollment before index. Mean age 55.3 years, SD 20.5. Male 54.0%. White 51.0%. At index, 81.7% were prescribed a second-generation oral antipsychotic, 10.4% a first-generation oral agent, 6.9% a first-generation long-acting injectable, 2.7% a second-generation long-acting injectable, and 1.0% clozapine. Mean Quan-Charlson comorbidity index 1.87, SD 2.38. Source: Optum Clinformatics Data Mart, United States, commercial and Medicare Advantage plan enrollees. Main analysis study period: January 1, 2011 to December 31, 2023.
Exposure
Antipsychotics of any class, measured as the cumulative proportion of days covered (PDC) by a pharmacy dispensing. Three levels: high adherence, PDC of at least 0.80; low/moderate adherence, 0.20 to under 0.80; nonadherence, under 0.20. Within six months of initiation, mean cumulative PDC was 0.59 (SD 0.33). By the end of follow-up it was 0.51 (SD 0.35), with 1,993 patients classified as high-adherent, 2,563 as low/moderate-adherent, and 1,861 as nonadherent.
Comparator
The most adherent category, PDC of at least 0.80, served as the reference group.
Outcome
All-cause mortality during follow-up: 1,094 deaths, 17.1%. Median follow-up 17.3 months, interquartile range 6.2 to 36.7 months, range 0.03 to 133.2 months, with about a quarter of patients followed beyond 36 months. Additional analyses: five-year estimated survival, numbers needed to treat at five and eight years, E-values, and five sensitivity analyses.
Design
Retrospective claims-based cohort study, analyzed with a marginal structural model adjusting for baseline and time-varying confounders, inverse probability of treatment weighting and inverse probability of censoring weighting, and a weighted pooled logistic model. Analyses conducted in R, version 4.3.3.
Funding
Johnson & Johnson Innovative Medicine. Five of the six authors are Johnson & Johnson employees, according to their affiliations in the publication, and are reported to hold Johnson & Johnson stock. The first author discloses ties with a very large number of industry sponsors, including Janssen and Johnson & Johnson. Medical writing support is reported to have been funded by the sponsor.

Quality control

Point checked	Verdict
Sample size and number of events	Adequate
Finding6,417 patients and 1,094 deaths give reasonable precision to an all-cause mortality endpoint, visible in the width of the confidence intervals. Overall eight-year survival, accounting for censoring, is estimated at 60%, roughly five deaths per 100 person-years.
Length of follow-up	Short median
FindingMedian follow-up is 17.3 months, interquartile range 6.2 to 36.7 months. Half the cohort is therefore followed for under a year and a half, while the headline figures are five- and eight-year estimates. A quarter of patients do exceed 36 months, though, and the longest observed follow-up reaches 133.2 months, which genuinely feeds the later part of the survival curves.
Handling of time-varying confounders	Appropriate
FindingA marginal structural model is the appropriate method when exposure and confounders influence each other over the course of follow-up. The published list of time-varying covariates includes relapse, hospitalization, emergency visits, intensive care admissions, outpatient visits, lithium, benzodiazepines, antidepressants, cardiovascular and metabolic medications, opioids, polypharmacy, the Quan-Charlson comorbidity index, and psychiatric and somatic comorbidities, each measured over the six months preceding each assessment point.
Analytic weights	Wide range
FindingThe stabilized weights, combining the treatment-probability weight and the censoring weight, had a mean of 1, SD 0.65, and a median of 0.92, which is reassuring. Their range, however, spanned 0.01 to 16.74, even after excluding patients whose weights fell in the extreme percentiles, an exclusion that affected only 10 patients and 224 longitudinal records. A small number of patients therefore carry disproportionate weight in the estimate.
Proportional hazards assumption	Tested
FindingInteraction terms between adherence level and time were not significant (p>0.1), which the authors interpret as support for the proportionality assumption. This test carries the usual limitation: a nonsignificant result is not proof of proportionality.
Exposure measurement	Indirect
FindingProportion of days covered measures pharmacy dispensing, not actual medication intake. The authors acknowledge this and note that coverage may overestimate true adherence. Hospital days were counted as covered, and overlapping dispensings were not double-counted in the primary analysis.
Ascertainment and cause of death	Cause unavailable
FindingDeath was identified from several cross-checked sources: hospital discharge status, coverage termination for death, electronic health records, the national social security death index, public program data, and death notices. The date is known only to the month. The authors acknowledge that some deaths may have been missed and judge that this would bias toward the null. Cause of death is not available in claims data, so no breakdown between suicide, somatic causes, and accidental causes is possible, leaving the mechanism open.
Sensitivity analyses	Consistent
FindingExclusion of the pandemic period, leaving 4,050 patients; an alternative coverage calculation accounting for medication stockpiling; three different cut-points for the adherence categories; adjustment for the initial antipsychotic type; and stratification by age. All point the same direction. The authors specify that no correction for multiple comparisons was applied, as these analyses were exploratory.
Funding and conflicts of interest	Industry
FindingThe work is funded by Johnson & Johnson Innovative Medicine. Five of the six authors are Johnson & Johnson employees, and the text states that the employee authors took part in the study design, analysis, interpretation, writing, and the decision to submit. The first author, an academic, discloses ties with several dozen industry sponsors, including Janssen and Johnson & Johnson, along with research support from Janssen. The disclosure is thorough and detailed, which is to the study’s credit, but independence from the sponsor is not assured.

Results

1.80
All-cause mortality hazard ratio for patients whose treatment coverage stayed below 20%, compared with those whose coverage reached at least 80%. 95% CI 1.43 to 2.27.

Result	Value
Nonadherence versus high adherence	HR 1.80, 95% CI 1.43 to 2.27
ReadingA clean, adequately precise association. An association, not demonstrated causation. Worth flagging, an internal inconsistency reported in the full publication: the discussion is said to refer to an 83% increase in risk, while the published hazard ratio corresponds to an 80% increase. This discrepancy is not explained in the source.
Low/moderate adherence versus high adherence	HR 1.53, 95% CI 1.25 to 1.89
ReadingThe relationship follows a gradient, a classic argument in favor of a real effect, though not proof of one. A patient taking their treatment more than one day in five but fewer than four days in five remains at elevated risk.
E-values	3.00 and 2.43
ReadingAn unmeasured confounder would need to be associated with both adherence level and mortality by a relative risk of at least 3.00 to fully explain away the point estimate for nonadherence, and 2.21 to explain away its lower bound. For low/moderate adherence, the corresponding values are 2.43 and 1.81. The authors judge such confounding unlikely. This argument is solid against ordinary residual confounding; it is considerably weaker against reverse causation, which is not confounding but a reversal of the causal arrow.
Time to survival falling to 80%	1.3 years, 2 years, and 5 years
ReadingRespectively for nonadherence, low/moderate adherence, and high adherence. This is the figure closest to the actual observed follow-up, and so the least dependent on extrapolation, and probably the most usable of the published numbers.
Estimated five-year survival	80%, 71%, and 67%
ReadingRespectively for high adherence, low/moderate adherence, and nonadherence. These are model-predicted survival estimates, not observed five-year proportions: median follow-up is 17.3 months.
Number needed to treat at five years	7.7 and 11.1
Reading7.7 patients, 95% CI 6.6 to 8.7, shifted from nonadherence to high adherence to avoid one additional death; 11.1 patients, 95% CI 9.0 to 13.2, for those starting from low/moderate adherence. These figures assume the association is causal, which the design does not establish.
Number needed to treat at eight years	5.9 and 8.3
Reading95% CI 5.1 to 6.6 and 6.9 to 9.8. Variance was computed using the delta method. The authors themselves truncated predictions at eight years, judging estimates beyond that horizon unreliable because of excessively wide confidence intervals.
Mean coverage at six months and at end of follow-up	0.59, then 0.51
ReadingSD 0.33, then 0.35. Six months after initiation, an average patient is covered only six days out of ten, and only 39.3% reach the high-adherence threshold. By the end of follow-up, only 31.1% remain high-adherent, while 29.0% are nonadherent. The wide dispersion reflects two very different populations more than an average behavior.
Sensitivity analyses, nonadherence	1.59 to 2.32 depending on analysis
ReadingExcluding the pandemic period, HR 1.79 (95% CI 1.32 to 2.42). Alternative coverage calculation, 1.78 (1.42 to 2.24). Adjusting for initial antipsychotic type, 1.84 (1.46 to 2.33). Under age 65, 2.32 (1.51 to 3.56). Age 65 and over, 1.59 (1.23 to 2.05). Complete absence of treatment for at least six consecutive months at end of follow-up, 2.11 (1.68 to 2.64). The association is therefore stronger under age 65, which partially tempers the concern raised by the cohort’s older mean age.

Critical appraisal

Domain	Judgment
Confounding by indication and reverse causation	Serious risk
FindingNonadherence is not an assigned treatment. It is a behavior correlated with illness severity, socioeconomic precarity, substance use, and social isolation, all independently associated with mortality. The baseline table bears this out: nonadherent patients had more substance use disorders, 44.9% versus 33.6% among the highly adherent. Reverse causation compounds this: a patient whose somatic health is deteriorating, who is hospitalized or institutionalized, stops picking up prescriptions, so a drop in coverage can be a consequence of what is about to kill them rather than its cause. The marginal structural model, which adjusts for hospitalizations and comorbidities over time, reduces this bias. It does not remove it, and the E-values do not answer this particular objection.
Extrapolation beyond follow-up	Moderate risk
FindingThe figures most likely to circulate, five-year survival and the number needed to treat at five and eight years, cover horizons well beyond the 17.3-month median follow-up. A quarter of patients do exceed 36 months, and the longest follow-up reaches 133.2 months, so the curves are not purely extrapolated. The authors did deliberately truncate predictions at eight years. Even so, the published confidence intervals reflect sampling uncertainty, not the uncertainty of the model’s assumptions, and the number of patients remaining at risk at the longer horizons is small.
Classification of exposure	Moderate risk
FindingA patient on a long-acting injectable and a patient on an oral formulation do not leave the same dispensing footprint, so identical true adherence can produce different proportions of days covered. This matters little here quantitatively, since 81.7% of patients were on a second-generation oral antipsychotic at index and fewer than 10% on an injectable formulation. A sensitivity analysis adjusting for initial antipsychotic type left results unchanged, but the authors explicitly state that the study does not evaluate adherence patterns by antipsychotic type or dose. No comparison between molecules or formulations can be drawn from this work.
Outcome measurement	Low risk
FindingDeath is an outcome with little room for misclassification, which is the study’s main strength, and its ascertainment cross-checks six sources. Two caveats remain: the authors acknowledge that some under-ascertainment is possible, and the date of death is only known to the month. Cause of death is not available.
Diagnostic validity	Moderate risk
FindingDiagnosis relies on billing codes, never on a clinical interview. In a cohort with a mean age of 55.3 years, where 29.5% of patients also carry a comorbid neurocognitive disorder diagnosis, rising to 36.4% among the highly adherent, the question of overlap with late-onset psychosis or behavioral disturbances associated with a neurodegenerative disorder is a legitimate one. The authors do not address it. This is an editorial observation, not a result of the study, but it bears on how the figure should be read.
External validity	Limited
FindingThe population is covered by commercial insurance and Medicare Advantage plans in the United States, with a mean age of 55.3 years, SD 20.5. The title describes newly treated or reinitiating patients, which, at this mean age and after 12 months without any antipsychotic dispensing, does not describe a first-episode cohort of young patients. Patients without insurance coverage, arguably the most socially disadvantaged, are absent by design, as are those lacking 12 months of continuous enrollment before index. The authors themselves state that the findings cannot be generalized beyond the population represented in this database. Healthcare organization differs from one country to the next, and readers should weigh how far these figures transpose to their own setting.
Priority claim	Fragile
FindingThe authors write, with a qualifying clause, that to their knowledge this is among the first data to directly quantify mortality risk across different adherence levels. Their own discussion nonetheless cites a Korean case-cohort study of 80,581 adults comparing adherence quartiles, which reported an all-cause mortality hazard ratio of 0.86 (95% CI 0.78 to 0.95) for the most adherent quartile. The novelty here lies in the marginal structural model and the three-category breakdown, not in the fact of quantification itself.
Independence	Industry-funded
FindingFunded by Johnson & Johnson Innovative Medicine, five of whose six authors are employees, with disclosed sponsor participation in design, analysis, interpretation, and the decision to submit. The disclosure is thorough and the statistical protocol is described precisely, which limits room for interpretive latitude. Still, the take-home message, reinforce adherence from initiation onward, directly serves the commercial interest of a manufacturer of long-acting injectable formulations, even though the study compares no formulation against another.

Level of evidence

Scientific66

Editorial80

Oxford CEBM level of evidence 2b, a retrospective observational cohort. Confidence is high that a strong, graded association exists between poor treatment coverage and all-cause mortality in this population, and it is reinforced by concordance across five sensitivity analyses. Confidence is moderate that restoring adherence would itself reduce mortality: the numbers needed to treat assume a causality the design does not establish, and the authors’ own abstract states that improved adherence may enhance survival, a more cautious formulation than the one reportedly used in the article’s conclusion section. Confidence is low regarding the five- and eight-year horizons, which extend well beyond the observed median follow-up. It is low regarding numeric transposition to other healthcare settings, and it is absent for any conclusion about a specific molecule or route of administration, which the authors explicitly state they did not evaluate.

What is demonstrated: in this cohort, patients with low treatment coverage die more often, and more precociously, than patients with high coverage, with a clear dose-response gradient. What is suggested, not demonstrated: that improving adherence would itself lower that mortality, and that the five- and eight-year projections describe what these patients will actually experience. What amounts to expert opinion: that the observed gradient reflects a genuine causal pathway from nonadherence to death rather than confounding by underlying severity or reverse causation, a reading the authors favor but that their own design cannot settle.

The colleague test

What an experienced colleague might say about this study in two minutes, between two consultations.

“Nobody disputes the direction of the result, and I like the gradient. What stops me is the gap between what was measured and what is being announced: a median follow-up of seventeen months, and I’m handed a number needed to treat at eight years. A mean age of fifty-five is not the population I most often talk to about stopping treatment, even if the effect is sharper under sixty-five. And I note who paid for it: five authors out of six work for the sponsor.”

Translated for practice: the clinical message is usable, the quantification is not. You can tell a patient that stopping treatment is associated with a worse survival outlook. You cannot give them a precise number attached to that risk.

What you can do with this

Frame the discussion of stopping treatment as a question of overall prognosis, not relapse alone. That is what these data genuinely add, and it is their most solid contribution.
Use the order of magnitude in consultation without quoting the exact figure: the estimated survival gap is real, its translation into individual benefit is not.
Anchor on the time horizon rather than the percentage. The time for estimated survival to fall to 80% runs from five years among highly adherent patients to a little over a year among nonadherent patients, a quantity that stays within the actually observed follow-up window.
Treat the first six months as the window to watch. Mean coverage has already dropped to 0.59 by then, only 39.3% of patients reach the high-adherence threshold, and this is where the later trajectory is largely set.
Do not relax vigilance for the partially adherent patient. Coverage between 20% and 80% still carries an elevated hazard, HR 1.53.
Actively look, in a patient whose coverage is slipping, for what is accompanying that slide: somatic decline, hospitalization, precarity, substance use. These factors themselves carry part of the excess risk, and they are targets for action.
Do not infer from this work that one formulation or molecule is superior to another. The authors explicitly state that adherence by antipsychotic type was not evaluated, and eight patients in ten were on an oral second-generation agent.
Mention the industry funding if these figures are cited in a meeting. It does not disqualify the work, but it should shape how its conclusion is read.
The course of action is set out in the NICE decision tree for schizophrenia in adults.

Frequently asked questions

Does this study prove that adherence reduces mortality?

No. It shows a strong, graded association. The observational design cannot rule out that low coverage is largely a marker of a more compromised clinical and social situation, or that it is sometimes a consequence of a somatic decline already under way rather than its cause. The authors’ own abstract stays conditional, even though the conclusion section of the full article is reported to use causal wording.

Is the number needed to treat usable in consultation?

Only with real caution. It is derived from survival differences projected at five and eight years, while median follow-up is 17.3 months, and it assumes a causal relationship the study does not establish.

Which patients do these figures apply to?

Adults covered by commercial insurance or Medicare Advantage in the United States, mean age 55.3 years, SD 20.5, who are starting or restarting an antipsychotic after 12 months with no dispensing at all. This is not a cohort of young, first-episode patients. A stratified analysis does show a stronger association under age 65, HR 2.32, than at 65 and older, HR 1.59.

Do the results hold up when the assumptions change?

Yes, in the same direction. Five sensitivity analyses, excluding the pandemic period, an alternative coverage calculation, three different cut-points for the adherence categories, adjustment for initial antipsychotic type, and stratification by age, all yield a significant excess risk for nonadherence, ranging from 1.59 to 2.32. The authors specify that no correction for multiple comparisons was applied.

Who funded this study?

Johnson & Johnson Innovative Medicine. Five of the six authors are Johnson & Johnson employees, and are reported to have taken part in the design, analysis, interpretation, writing, and decision to submit. The first author, an academic, discloses ties with several dozen industry sponsors, including Janssen and Johnson & Johnson. Medical writing support is reported to have been funded by the sponsor. This information is clearly disclosed in the publication.

Annotated bibliography

Source study. Correll CU, Bookhart BK, Benson C, Liu Z, Zhao Z, Tang W. Association of antipsychotic nonadherence with all-cause mortality in adults with schizophrenia newly treated or reinitiating antipsychotic medication: A retrospective healthcare claims study. European Neuropsychopharmacology, 2026, volume 104, article 112745. Epub December 22, 2025. DOI 10.1016/j.euroneuro.2025.112745. PMID 41435593. Funded by Johnson & Johnson Innovative Medicine; five of six authors are Johnson & Johnson employees, and the first author discloses ties with numerous industry sponsors, including Janssen and Johnson & Johnson. No supplementary material accompanied the document consulted: only the published abstract was accessible for this English version, and no data appearing solely in the full text or its supplementary material is reproduced here beyond what the French version, verified against the full text, already establishes.

Priority claim to qualify. Chang J, Kim JA, Kim K, Choi S, Kim SM, Nam YY, et al. Association of antipsychotics adherence and cardiovascular disease among newly diagnosed schizophrenia patients: a national cohort among Koreans. Asian Journal of Psychiatry, 2020, volume 52, article 102161. Cited by the source study’s own discussion, which credits it with a comparison of adherence quartiles and an all-cause mortality hazard ratio of 0.86 (95% CI 0.78 to 0.95) for the most adherent quartile. It qualifies the source study’s priority claim. These figures come from the source study’s discussion and have not been independently verified against the Korean publication itself.

Relapse definition. Turkoz I, Daskiran M, Starr HL, Najarian D, Lopena O, Obando C, et al. Comparing relapse rates in real-world patients with schizophrenia who were adequately versus not adequately treated with paliperidone palmitate once-monthly injections before transitioning to once-every-3-months injections. Neuropsychiatric Disease and Treatment, 2022, volume 18, pages 1927 to 1937. Cited in the source study’s supplementary material as the source of the billing codes defining relapse, a time-varying covariate in the model. Relapse is therefore identified through administrative codes, psychiatric hospitalization, suicidal ideation, suicide attempt, hostility, aggressive behavior, incarceration, rather than through clinical assessment. Two events within a 14-day window count as a single relapse. During follow-up, 2,742 patients, 42.7%, had at least one relapse so defined.

Comorbidity classification. Elixhauser A, Steiner C, Kruzikas D. Healthcare Cost and Utilization Project (HCUP) Comorbidity Software. Rockville, Agency for Healthcare Research and Quality, 2015. Source of the somatic comorbidity covariates, cited in the source study’s supplementary material.

Pharmacological note, outside the scope of the study. The baseline characteristics table indicates that 61 patients, 1.0% of the cohort, were receiving clozapine at index. Clozapine prescribing is subject to structured monitoring requirements in most jurisdictions, typically including mandatory hematological monitoring, weekly during an initial period and then monthly, and contraindications that generally include a history of clozapine-induced agranulocytosis or granulocytopenia, bone marrow impairment, uncontrolled epilepsy, and paralytic ileus. Readers should confirm the specific requirements of their own regulatory framework, since these vary between countries. This point falls outside the study’s results, which compare no molecule against another.

Editorial collections

A Study Under the Microscope PEB Practice Change