Do GLP-1 Drugs Lower Blood Cancer Risk? Why the Answer Changed With the Comparator

The best answer is no—not from this study. In adults with type 2 diabetes, GLP-1 receptor agonist use was not associated with lower incidence of any of seven studied hematologic malignancies when compared with DPP-4 inhibitors. Lower incidence of non-Hodgkin lymphoma, myeloproliferative neoplasms, and Hodgkin lymphoma appeared only when GLP-1 drugs were compared with SGLT2 inhibitors. That comparator-dependent pattern is an observational signal, not evidence that GLP-1 medicines prevent blood cancer.

The same GLP-1 exposure produced different associations against two active comparators—a warning that who forms the comparison group can shape an observational result.

Quick answer: what did the study find?

Shah and colleagues found no consistent reduction in GLP-1 blood cancer risk across two active comparators. Their open-access Blood Cancer Journal correspondence, published October 12, 2026, reported:

• Against DPP-4 inhibitors: none of the seven hematologic malignancy outcomes differed statistically.
• Against SGLT2 inhibitors: GLP-1 receptor agonist users had lower observed incidence of non-Hodgkin lymphoma (NHL), myeloproliferative neoplasms (MPN), and Hodgkin lymphoma.
• Across both comparisons: there was no consistent increase in any studied hematologic malignancy, but that is not the same as proving zero risk.
• Causal conclusion: the study cannot show that GLP-1 medicines prevent cancer, caused the differences, or should be chosen over another diabetes medicine to alter cancer risk.

The useful lesson is methodological: an observational association that appears against one reasonable comparator but disappears against another may reflect differences between the patient groups, prescribing patterns, or other unmeasured factors.

How was the GLP-1 blood cancer risk study designed?

This was a retrospective electronic-health-record study, not a randomized trial. The researchers used the global TriNetX federated network, which contains more than 100 million de-identified records.

Eligible participants were adults aged 18 or older with type 2 diabetes who initiated a qualifying GLP-1 receptor agonist between January 1, 2015, and March 31, 2021. The qualifying medicines were semaglutide, dulaglutide, liraglutide, and exenatide.

To strengthen exposure identification, the study required:

• a type 2 diabetes diagnosis within one year before the first prescription;
• another prescription for the same medication class; and
• another type 2 diabetes diagnosis between one month and one year after the index date.

The authors then ran two separate 1:1 propensity-score-matched analyses:

1. GLP-1 receptor agonist initiators versus DPP-4 inhibitor users: 68,949 matched pairs.
2. GLP-1 receptor agonist initiators versus SGLT2 inhibitor users: 71,391 matched pairs.

Matching used recorded demographics, comorbidities, medications, laboratory values, and selected blood-cancer risk factors, including HIV, tobacco exposure, prior irradiation, H. pylori infection, and autoimmune or inflammatory disorders. Standardized differences for matched covariates were below 0.10.

Outcomes were counted from 180 days through five years after treatment initiation. That 180-day lag helps reduce the chance that an already-developing cancer detected immediately after treatment initiation is attributed to the new medicine, but it does not remove all reverse causation or detection bias.

The seven outcomes were NHL, Hodgkin lymphoma, myeloid leukemia, lymphoid leukemia, plasma cell neoplasms, myelodysplastic syndromes, and MPN. The supplement shows that these outcomes were defined with ICD-10-CM codes rather than pathology review.

What happened versus DPP-4 inhibitors?

No studied hematologic malignancy had a statistically significant lower or higher incidence in the GLP-1 group versus the DPP-4 inhibitor group.

| Outcome | Events: GLP-1RA vs DPP-4i | Hazard ratio (95% CI) | Interpretation | |---|---:|---:|---| | Non-Hodgkin lymphoma | 208 vs 203 | 0.99 (0.82–1.20) | No statistically significant difference | | Myelodysplastic syndromes | 105 vs 121 | 0.84 (0.65–1.09) | No statistically significant difference | | Plasma cell neoplasms | 140 vs 148 | 0.92 (0.73–1.15) | No statistically significant difference | | Myeloid leukemia | 98 vs 82 | 1.16 (0.86–1.55) | No statistically significant difference | | Lymphoid leukemia | 104 vs 94 | 1.07 (0.81–1.41) | No statistically significant difference | | Hodgkin lymphoma | 34 vs 34 | 0.97 (0.60–1.56) | No statistically significant difference | | Myeloproliferative neoplasms | 298 vs 288 | 1.01 (0.86–1.18) | No statistically significant difference |

A hazard ratio of 1.00 would indicate no difference in the modeled event rate over time. Every confidence interval in this comparison crossed 1.00. These estimates therefore do not establish either benefit or harm.

What happened versus SGLT2 inhibitors?

Three lower-incidence associations appeared against SGLT2 inhibitors, while four outcomes did not differ statistically.

| Outcome | Events: GLP-1RA vs SGLT2i | Hazard ratio (95% CI) | Interpretation | |---|---:|---:|---| | Non-Hodgkin lymphoma | 200 vs 241 | 0.82 (0.68–0.99) | Lower observed incidence; borderline confidence limit | | Myeloproliferative neoplasms | 300 vs 379 | 0.78 (0.67–0.91) | Lower observed incidence | | Hodgkin lymphoma | 19 vs 36 | 0.52 (0.30–0.91) | Lower observed incidence; few events and wide interval | | Myeloid leukemia | 85 vs 101 | 0.83 (0.63–1.11) | No statistically significant difference | | Plasma cell neoplasms | 120 vs 132 | 0.90 (0.70–1.15) | No statistically significant difference | | Lymphoid leukemia | 97 vs 107 | 0.90 (0.68–1.18) | No statistically significant difference | | Myelodysplastic syndromes | 100 vs 83 | 1.20 (0.89–1.60) | No statistically significant difference |

The NHL estimate was HR 0.82 with a 95% confidence interval that ended at 0.99. The Hodgkin lymphoma estimate looked larger in relative terms, but it was based on only 19 versus 36 events and had a wide confidence interval. Those details matter because apparently dramatic relative differences can still be statistically fragile when outcomes are rare.

The authors also explored NHL subtypes. Follicular lymphoma, non-follicular lymphoma, mature T/NK-cell lymphoma, and other specified or unspecified NHL did not differ statistically in either comparator analysis.

Why did the comparator change the apparent answer?

An active comparator defines the counterfactual question: lower incidence compared with whom? A GLP-1 user is not being compared with a generic person. The estimate depends on the characteristics of people who received the alternative treatment.

DPP-4 and SGLT2 inhibitors were reasonable active comparators because, like GLP-1 receptor agonists, they were commonly used as non-insulin second-line therapies during the study period. This can reduce some confounding compared with using untreated nonusers. It does not make the groups interchangeable.

The authors noted different prescribing patterns:

• DPP-4 inhibitors may be favored for older or frailer people and when avoiding hypoglycemia or controlling medication cost matters.
• SGLT2 inhibitors may be favored for people with chronic kidney disease, heart failure, or established cardiovascular disease.
• GLP-1 receptor agonists produce more weight loss on average than either comparator class. Because obesity is itself associated with NHL risk, weight change could contribute to an observed difference without proving a direct anticancer action of GLP-1 receptor signaling.

These are examples of confounding by indication and treatment channeling: clinical features influence which drug a person receives, and some of those same features may also relate to cancer detection or risk.

Propensity matching balances recorded factors. It cannot balance variables that were missing, incompletely measured, miscoded, or never entered into the model.

What can propensity-score matching fix—and what can it miss?

Matching can improve balance on recorded baseline variables, but it cannot reproduce random assignment. In the supplement, substantial pre-matching differences became small after matching. That is useful. It is not a causal guarantee.

Important remaining limits include:

• Residual and unmeasured confounding: cancer risk, treatment choice, healthcare access, surveillance intensity, socioeconomic factors, and disease severity may not be fully captured.
• Incomplete BMI and HbA1c data: BMI was available for roughly 65%–72% of matched patients and HbA1c for roughly 80%–82%. The authors did not use multiple imputation.
• Changing treatment eras: GLP-1, DPP-4, and SGLT2 medicines entered practice at different times and were adopted under evolving guidelines.
• Detection bias: groups may have differed in healthcare contact or testing, affecting when a malignancy code appeared.
• Exposure misclassification: TriNetX recorded prescriptions, not dispensing, administration, adherence, dose, or continuous use. A repeat prescription within one year strengthened the definition but did not prove exposure.
• Outcome validation limits: cancers were identified through diagnosis codes without histologic confirmation. The supplement lists the ICD-10-CM definitions, but code presence is not the same as central pathology review.
• Relatively few events: this was especially important for Hodgkin lymphoma and for exploratory NHL subtypes.
• Multiple outcomes without multiplicity adjustment: testing many outcomes increases the chance that some nominally significant findings appear by chance. The authors did not adjust for multiple comparisons.

Together, these limitations make the comparator-dependent results more useful as a signal to investigate than as proof of cancer prevention.

Does this study show GLP-1 drugs prevent blood cancer?

No. The study estimated associations between treatment records and later diagnosis codes; it did not test a cancer-prevention intervention.

To establish prevention, researchers would need evidence capable of separating a medication effect from differences in patient selection, weight change, follow-up, competing risks, and cancer ascertainment. Ideally, that would include prospectively specified cancer outcomes, validated diagnoses, detailed longitudinal exposure, adequate latency, and designs with stronger causal identification.

The biological ideas discussed by the authors—reduced inflammation, altered immune function, or effects on lymphocyte proliferation—remain hypotheses in this context. A plausible mechanism does not rescue an inconsistent epidemiologic association.

How is this different from randomized cancer-safety evidence?

Randomized trials answer a different question and usually provide better protection against baseline confounding, but existing GLP-1 trials were generally not designed to detect individual blood cancers.

The paper notes that meta-analyses of randomized controlled trials have generally found no increased overall cancer risk with GLP-1 receptor agonists. That is reassuring as a broad safety signal, but it does not prove protection from hematologic malignancy. Rare cancers, long latency, limited follow-up, and non-cancer primary endpoints constrain what those trials can resolve.

Observational databases can add scale and longer real-world follow-up. Randomization adds stronger comparability. Neither should be made to answer more than its design allows.

Why prior GLP-1 cancer studies do not settle this question

Earlier observational results depended on different populations and comparison groups, so their hazard ratios are not interchangeable with the Shah study.

• In people with obesity, Levy and colleagues reported lower lymphoid and hematopoietic cancer incidence among GLP-1 receptor agonist users versus matched controls: HR 0.69 (95% CI 0.60–0.80).
• Ashruf and colleagues reported lower incidence versus insulin for NHL, HR 0.42 (0.30–0.58); myelodysplastic syndromes, HR 0.19 (0.11–0.35); and MPN, HR 0.50 (0.41–0.61). The same study found no statistically significant differences when GLP-1 drugs were compared with metformin.
• A Japanese claims analysis reported lower hematologic-malignancy incidence, including lymphoma, among SGLT2 inhibitor users versus non-SGLT2 antidiabetic agents: lymphoma HR 0.71 (0.57–0.89).

Insulin users may have longer-standing or more advanced diabetes. Metformin is often first-line therapy and may continue across treatment stages. Nonusers can differ from treated patients in many ways. These comparator choices change the clinical question—and the confounding structure.

For another example of how active comparators shape interpretation, see our review of a GLP-1 and tuberculosis-risk observational study. Our guide to evaluating peptide claims online explains why human data, causal design, and endpoint validation matter more than a dramatic headline.

Bottom line

The Shah study did not show that GLP-1 drugs lower blood cancer risk. It found no lower incidence for any studied hematologic malignancy versus DPP-4 inhibitors, while lower NHL, MPN, and Hodgkin lymphoma incidence appeared versus SGLT2 inhibitors.

That disagreement is the result worth remembering. It suggests that the association is sensitive to comparator selection and may reflect who received each drug class, how they were followed, or other unmeasured differences. The findings do not justify starting, stopping, switching, or choosing a GLP-1 medicine to change cancer risk.

Frequently asked questions

Did GLP-1 drugs lower blood cancer risk in this study?

Not consistently. No studied malignancy differed versus DPP-4 inhibitors. NHL, MPN, and Hodgkin lymphoma incidence was lower versus SGLT2 inhibitors, but those comparator-dependent observational associations cannot establish prevention.

Was the Shah study randomized?

No. It was a retrospective TriNetX electronic-health-record analysis using 1:1 propensity-score matching. Matching balanced recorded factors but could not remove all residual or unmeasured confounding.

Why were outcomes counted from 180 days after treatment initiation?

The lag reduced the influence of cancers detected very soon after treatment started. It does not prove those cancers began after exposure or eliminate reverse causation and detection bias.

Does a hazard ratio below 1 mean the drug prevented cancer?

No. It means the modeled event rate was lower in one group during follow-up. In a nonrandomized study, the difference may reflect treatment effects, patient differences, measurement, chance, or a mixture of these.

Did the study find an increased blood cancer signal?

It found no consistent increase in the seven evaluated malignancies against either comparator. That is not proof of zero risk, especially for rare outcomes with limited event counts.

Should someone change a GLP-1 medicine because of this paper?

This paper does not support medication changes to alter cancer risk. It was not a treatment-selection trial and did not test screening, dosing, switching, or prevention strategies.

Sources

1. Shah P, Singhal A, Wang TH, et al. Glucagon-like peptide-1 receptor agonists and risk for hematologic malignancies: a multicenter propensity score-matched analysis. Blood Cancer Journal. 2026;16:178. Full text and supplementary material verified through the publisher and PubMed Central record PMC13658783.
2. Ko A, Chang YC, Bahar F, et al. Risk for cancer with glucagon-like peptide-1 receptor agonists and dual agonists: a systematic review and meta-analysis. Annals of Internal Medicine. 2025. doi:10.7326/annals-25-02237.
3. Ashruf OS, Hundal J, Mushtaq A, et al. Hematologic cancers among patients with type 2 diabetes prescribed GLP-1 receptor agonists. JAMA Network Open. 2025;8:e250802. doi:10.1001/jamanetworkopen.2025.0802.
4. Levy S, Attia A, Elshazli RM, et al. Differential effects of GLP-1 receptor agonists on cancer risk in obesity: a nationwide analysis of 1.1 million patients. Cancers. 2024;17:78. doi:10.3390/cancers17010078.

This article is for general education and evidence interpretation only. It is not medical advice and does not recommend starting, stopping, switching, dosing, screening, or using a GLP-1 medicine to alter cancer risk.