Published on 16 September 2026
Do GLP-1 receptor agonists change the risk of depression in older adults with diabetes?
In brief
Since the European safety signal of July 2023 on suicidal thoughts and thoughts of self-harm, the question of the psychiatric effects of GLP-1 receptor agonists keeps coming back in the consulting room, driven as much by patients’ worries as by the opposite promise of an antidepressant effect. This target trial emulation, run on data from the US federal health insurance programme, compares two cohorts, each matched separately on a one-to-one basis, in people aged 66 and over with type 2 diabetes and no history of depression or any other mood disorder: 14,665 pairs against SGLT2 inhibitors, 13,711 pairs against DPP-4 inhibitors. Against DPP-4 inhibitors, the incidence of depression is lower on a GLP-1 receptor agonist, hazard ratio 0.90, 95% confidence interval 0.82 to 0.98. Against SGLT2 inhibitors, no difference is shown, 1.07, interval 0.98 to 1.18. In other words, no significant excess risk of depression emerges, and the favourable association depends on the comparator chosen. Nothing here supports prescribing for mood.
The context
Two narratives circulate at the same time, and they contradict each other. The first, born of the European pharmacovigilance signal opened on 11 July 2023, fears that these drugs might promote depression and suicidal thoughts. That review was closed in April 2024, with the European pharmacovigilance committee concluding that the available data did not support a causal link with suicidal thoughts or thoughts of self-harm. The regulatory conclusion has not put the question to rest in public debate. The second, carried by the media visibility of the class, credits these drugs with a beneficial effect on mood, even a future indication. The psychiatrist is asked about both, without prescribing these treatments. What is needed, then, is solid data on the risk, and a clear position on the alleged benefit.
The methodological point
What target trial emulation corrects, and what it does not
Target trial emulation means first writing the protocol of the randomised trial one would like to run, then reproducing it as closely as possible in observational data. The approach limits two classic artefacts of pharmacoepidemiology: immortal time, and the inclusion of prevalent users already established on treatment. Using an active comparator, here two different second-line classes, also reduces confounding by indication, since the comparison is between patients for whom a decision had been made to add a treatment. What the approach does not correct is what was not measured. The authors state that socioeconomic status, glycated haemoglobin, body mass index and diabetes severity are not in the database, and that glucose-lowering treatment at baseline, insulin included, serves as a proxy for that severity. With an association this modest, that limitation weighs heavily.
The study at a glance
| Population | |
| Adults aged 66 and over with type 2 diabetes, enrolled in the US federal health insurance programme with at least one year of continuous enrolment in Parts A, B and D. Exclusions: type 1 diabetes, end-stage renal disease or dialysis, a history of depression or another mood disorder, and any antidepressant prescription in the previous year. Other psychiatric histories are not excluded: sleep disorders, anxiety, substance use disorders and schizophrenia appear among the baseline characteristics. | |
| Intervention | |
| Initiation of a GLP-1 receptor agonist between January 2014 and December 2020: exenatide, dulaglutide, liraglutide or semaglutide. | |
| Comparators | |
| Two separate cohorts, with a second-line active comparator: initiation of an SGLT2 inhibitor in the first, of a DPP-4 inhibitor in the second. | |
| Outcome | |
| Incident depression, identified by at least one diagnostic code using the chronic conditions algorithm of the Chronic Conditions Warehouse. Intention-to-treat analysis, follow-up of two years at most and up to 31 December 2020, median follow-up of about 1.6 years. | |
| Design and numbers | |
| Target trial emulation on claims data, new-user active-comparator design, one-to-one propensity score matching, nearest neighbour without replacement and a caliper of 0.05. Cox model for the primary analysis, Fine and Gray model accounting for the competing risk of death as a sensitivity analysis. 14,665 pairs against SGLT2 inhibitors and 13,711 pairs against DPP-4 inhibitors, two cohorts matched separately that cannot be added together. |
Quality control
| Criterion | Judgement |
|---|---|
| Explicit target protocol | Sound |
| FindingThe protocol of the target trial is described component by component and set against its emulation. The new-user active-comparator design limits immortal time and the inclusion of prevalent users already established on treatment. | |
| Two active comparators | Strength |
| FindingTwo second-line classes rather than one. It is this choice that reveals the asymmetry of the result, and that is where its value lies. | |
| Balance between groups | Claimed, table disagrees |
| FindingOne-to-one propensity score matching, with a standardised difference threshold set at 0.1. The authors conclude that balance is good, but the table of baseline characteristics publishes several values beyond that threshold, including 0.380 for obesity, 0.251 for sleep disorders and −0.198 for dementias. The standardised difference column is given only for the comparison with SGLT2 inhibitors. This discrepancy is not explained in the text. | |
| Unmeasured confounders | Major limitation |
| FindingSocioeconomic status, glycated haemoglobin, body mass index and diabetes severity are missing from the database, as the authors acknowledge. An association whose upper bound reaches 0.98 is fragile in the face of this limitation. | |
| Outcome definition | Codes, not clinical diagnosis |
| FindingDepression is identified by an algorithm applied to billing codes, a single code being enough. The authors themselves acknowledge the risk of misclassification. Under-detection is likely, and there is no reason for it to be the same in both arms. | |
| External validity | Narrow |
| FindingPeople aged 66 and over with type 2 diabetes, with no history of depression or any other mood disorder, in a US insurance system. The result cannot be extrapolated to younger people, to obesity without diabetes, or to patients who are already depressed. | |
| Independence | Public funding |
| FindingPublic funding, National Institutes of Health, National Institute of Diabetes and Digestive and Kidney Diseases, grant R01DK133465. The funders state that they had no role in the design, conduct or analysis, nor in the decision to submit. Protocol and statistical code available from the corresponding author, data accessible through the ResDAC resource centre. The open-access author manuscript carries only a partial statement of competing interests, the full forms being held by the journal and not consulted here. | |
The findings
| Outcome | Reported value |
|---|---|
| Against DPP-4 inhibitors | HR 0.90 [0.82; 0.98] |
| ReadingIncidence of 51.4 on a GLP-1 receptor agonist against 57.2 on a DPP-4 inhibitor per 1000 person-years, a difference of −5.78 per 1000 person-years, interval −10.49 to −1.07. Cumulative two-year risk of 7.02% against 7.84%. The model accounting for competing risk gives 0.91, interval 0.83 to 0.99. Statistical significance hangs by a thread. | |
| Against SGLT2 inhibitors | HR 1.07 [0.98; 1.18] |
| ReadingIncidence of 49.0 on a GLP-1 receptor agonist against 45.5 on an SGLT2 inhibitor per 1000 person-years, a difference of 3.48, interval −0.81 to 7.78. Cumulative two-year risk of 6.55% against 6.15%. No difference is shown, but the interval remains compatible with a relative excess risk of up to 18%. The absence of a significant difference is not evidence of an absence of effect, in either direction. | |
| By duration of exposure | Significant interaction, p = 0.005 and p = 0.001 |
| ReadingThe authors report an inverse association between treatment duration and the risk of depression, against SGLT2 inhibitors (p = 0.005) and against DPP-4 inhibitors (p = 0.001). Hazard ratios by duration stratum appear only in a figure that is truncated in the copy consulted, so they are not reproduced here. An analysis by duration of exposure selects the patients who tolerate and continue the treatment. The authors themselves cite healthy user bias, reverse causality and weight loss among the possible explanations. | |
Critical appraisal
| Domain | Judgement |
|---|---|
| Level of evidence of the design | Observational |
| FindingA non-randomised cohort. Target trial emulation is the top of the range for such a design, it does not turn it into a trial. The text speaks of association, and nothing in the design allows one to speak of effect. | |
| Internal consistency of the result | Asymmetric |
| FindingA favourable association against one comparator, none against the other. Two readings remain open: a modest benefit specific to GLP-1 receptor agonists, or a disadvantage of DPP-4 inhibitors. The study cannot decide between them, and the authors acknowledge this. | |
| Fit between claim and evidence | Cautious, except the conclusion |
| FindingThe authors describe the association as modest, discuss the asymmetry between comparators without playing it down, and make their interpretation conditional on confirmation in a randomised trial. The last sentence of their conclusion nevertheless goes beyond their data, by suggesting that the class may have antidepressant effects. | |
| Relevance to psychiatric practice | Indirect |
| FindingNo psychiatric indication is authorised for this class in the European Union. The result changes what is said to the patient, not the prescription. | |
Level of evidence
Oxford level of evidence 2b by design: a non-randomised observational cohort, with an active comparator and propensity score matching. Target trial emulation is the best methodological safeguard available for this type of data, it does not raise the level. Confidence is reasonable regarding the absence of a significant excess risk of depression in this specific population. It is low regarding the existence of a specific benefit, since the association disappears depending on the comparator chosen. It is very low regarding any causal interpretation of the association observed with duration of exposure.
The colleague test
What an experienced colleague would say if you put this study to them in two minutes, between two consultations.
“I can reassure an older patient with diabetes who is worried about a risk of depression on these drugs, and I do not advise against them on that ground. But I do not talk about an antidepressant effect, and I make nothing of the signal on duration of exposure, which smells of healthy user bias.”
What this means in practice: the result is an argument for reassurance and a brake on the hype, not an argument for prescribing. The decision on diabetes treatment stays with the diabetologist, and coordination happens without claiming any benefit on mood.
What you can do with this
- Answer an older patient with diabetes who is worried by the media signal: in this population, no significant excess risk of depression appears over two years against two active comparators.
- Do not advise against a GLP-1 receptor agonist on the ground of a depression risk in a patient who is gaining a metabolic benefit from it.
- Do not present the class as having an antidepressant effect: the observed association is modest, it disappears depending on the comparator, and no psychiatric indication exists.
- Remember that the result concerns people aged 66 and over with diabetes and no history of depression or any other mood disorder. It says nothing about younger people, about patients with obesity but no diabetes, or about those who already have a history of depression.
- Keep up the usual clinical monitoring of mood whenever metabolic treatment is changed, regardless of this result.
Frequently asked questions
Does this study refute the European safety signal of July 2023 on GLP-1 receptor agonists?
It provides a reassuring argument in a specific population and on a specific outcome, incident depression. It does not address suicidal thoughts, which were the subject of the signal. That European review was in any case closed in April 2024, with the conclusion that the available data did not support a causal link. This study adds to that body of evidence, it does not sum it up.
Why does the depression risk with GLP-1 receptor agonists differ between DPP-4 and SGLT2 inhibitors?
Because a hazard ratio is always relative to something. A favourable association against DPP-4 inhibitors and a null one against SGLT2 inhibitors may reflect a modest benefit of GLP-1 receptor agonists, or a disadvantage of the reference class. The study does not allow a choice between the two.
Can GLP-1 receptor agonists be prescribed to treat depression?
No. No psychiatric indication is authorised. The authorised indications concern type 2 diabetes and, for some drugs, the management of overweight and obesity. Marketing authorisation, actual availability on the market and reimbursement are three distinct statuses, and none of the three covers use for mood. The level of evidence available would not justify it either.
Is the stronger association with longer treatment duration credible?
It is possible, but the patients still on treatment after a year are those who tolerate it and are doing better. This type of analysis is structurally exposed to healthy user bias and to reverse causality, as the authors themselves point out.
What should a psychiatrist tell the diabetologist about GLP-1 receptor agonists and mood?
That nothing in this study justifies asking for a GLP-1 receptor agonist to be stopped for a psychiatric reason, or asking for one to be started for a mood-related reason.
Annotated bibliography
Source study. Tang H, Lu Y, Donahoo WT, Westen SC, Chen Y, Bian J, Guo J. Glucagon-Like Peptide-1 Receptor Agonists and Risk for Depression in Older Adults With Type 2 Diabetes: A Target Trial Emulation Study. Annals of Internal Medicine, 2025, 178(3), 315-326. DOI 10.7326/ANNALS-24-01347, PMID 39993315. Target trial emulation on Medicare claims data, two propensity score matched cohorts. The reading was based on the open-access author manuscript, whose title differs from the published version by two words.
European safety signal. European Medicines Agency, statement of 11 July 2023 on the ongoing review of GLP-1 receptor agonists for suicidal thoughts and thoughts of self-harm, ema.europa.eu. The review was closed by the Pharmacovigilance Risk Assessment Committee in April 2024, on the conclusion that the available data do not support a causal link, meeting of 8 to 11 April 2024.
What was consulted. Verification carried out on 13 August 2026 on the full text of the open-access author manuscript (PMC12132804, NIHMSID NIHMS2082136), read from the first page to the last: structured abstract, introduction, methods, table 1 of the target protocol, table 2 of baseline characteristics, table 3 of results, figures 1 to 3, discussion, funding statement, reproducibility statements and the 45 references. Every sample size, incidence, risk difference, hazard ratio and confidence interval published here was taken from this source. The supplementary material (tables S1 to S6) could not be consulted: the available copy contains only the download link. Figure 3, which carries the subgroup analyses, is truncated in it: the hazard ratios by treatment duration stratum were therefore not reproduced. Table 3 of the source also contains a manifestly erroneous confidence interval for the two-year risk difference against DPP-4 inhibitors, a value not reproduced here. Full reference, pagination, digital object identifier and PubMed identifier checked on PubMed. Dates and scope of the European signal checked against the European Medicines Agency statements. The competing interest forms, held by the journal, were not consulted. Article subjected to an independent double reading.