Long-term cancer risk in users of GLP-1 agonists in Denmark: a nationwide emulated trial

Summary

Background

The long-term cancer safety of glucagon-like peptide-1 receptor agonists (GLP-1RAs) in real-world settings remains unclear, with limited long-term clinical and observational studies. We clarify the long-term cancer risk.

Methods

This register-based nationwide emulated trial includes all Danes initiating treatment with GLP-1RA or dipeptidyl peptidase-4 inhibitors (DPP-4i) 2007–2019, propensity score matched 1:1 on baseline characteristics and followed 10 years. The primary outcome was risk differences for cancer, estimated for long-term sustained use of GLP-1RA vs DPP-4i using g-computation accounting for time-varying patient characteristics. Secondary outcomes included “death without prior cancer” and the composite outcome “death or cancer”. Analyses included sex stratified estimates and Cox hazard ratios (HR).

Findings

After 195,702 person-years 4758 developed cancer. Among sustained users of GLP-1RA, 4·1 (95% CI 0·4–7·2) more patients developed cancer per 100, compared to 100 DPP-4i patients 10-years post-initiation (HR: 1·35 [95% CI 1·05–1·73] 6–10 years post-initiation). The excess cancer risk was 6·6 (95% CI 1·8–10·7) per 100 women and 2·2 (95% CI −2·2 to 6·2) per 100 men. Fewer patients “died without prior cancer” in users of GLP-1RA (per 100 users: −4·9 [95% CI −7·6 to −2·4]). There was no difference in risk of “death or cancer” per 100 users: −1·15 (95% CI −4·9 to 2·5).

Interpretation

Long-term sustained users of GLP-1RA had a small increased risk of cancer; potentially explained by a survival benefit. Residual confounding by body mass index cannot be ruled out.

Funding

The Scientific Committee of the Danish Cancer Society (R354-A20492-23-S3 to LSM).

Research in context.

Evidence before this study

We conducted a comprehensive literature search in PubMed, without language restrictions, from beginning to December 15, 2024. We used the following search terms: (“GLP-1 receptor agonist∗” [Title] OR “GLP1 receptor agonist∗” [Title] OR GLP [Title] OR “Glucagon-like peptide∗” [Title] OR “GLP-1 Analog∗” [Title] OR liraglutide [Title] OR albiglutide [Title] OR dulaglutide [Title] OR exenatide [Title] OR lixisenatide [Title] OR semaglutide [Title]) AND (cancer [Title] OR neoplasm∗ [Title] OR tumor∗ [Title] OR tumour∗ [Title] OR malignan∗ [Title] OR carcinom∗ [Title]) to identify studies examining the relationship between glucagon-like peptide-1 receptor agonists (GLP-1RAs) and risk of any type of cancer. Previous evidence suggests both a potential increased risk of certain cancer types and decreased risks in patients using GLP-1RAs. However, evidence from observational studies has been inconsistent, with most research limited to short-term follow-up periods. Moreover, these studies frequently focus on a ‘intention-to-treat’ analogue approach, limiting causal interpretation and insights into the long-term effects of sustained GLP-1RA use in real-world settings. Consequently, the long-term overall cancer risk among new users of GLP-1RA remains unknown.

Added value of this study

This study contributes to existing research by leveraging an emulated trial framework with advanced statistical methods (e.g. g-computation) and high-quality Danish registries to examine long-term cancer risk associated with sustained GLP-1RA use in a nationwide, real-world setting with up to 10 years of follow-up. In the absence of large, long-term randomized trials and observational studies, this approach provides novel real-world evidence on the long-term cancer safety of sustained GLP-1RA use—knowledge needed for a comprehensive public health assessment of its overall risk-benefit balance.

Implications of all the available evidence

The findings of this study have significant public health implications, indicating that GLP-1RAs, widely used for diabetes and obesity, are generally safe at a population level concerning overall cancer risk. A suggested small increased cancer risk with long-term use of GLP-1RA was possibly linked to extended survival rather than a direct carcinogenic effect of GLP-1RA.

While site-specific cancer risks require further investigation, this study addresses critical gaps in the literature by providing essential evidence on the overall net balance of long-term cancer risk associated with GLP-1RA use from a public health perspective.

Introduction

Glucagon-like peptide-1 receptor agonists (GLP-1RAs) are potent glucose-lowering and weight-reducing agents that are now widely used in the treatment of type 2 diabetes (T2DM) and obesity.¹ Due to their efficacy and the rapidly increasing number of people with T2DM and obesity, it is forecasted that the use of GLP-1RA will rise dramatically globally during the forthcoming decade, e.g. that the total number of GPL-1RA users in US may reach 30 million by 2030 (around 9% of the population).² GLP-1RAs have shown a favourable safety profile³ and post-hoc analyses from the cardiovascular outcome trial of liraglutide (LEADER) provided no indication of a major increase in the risk of total malignant neoplasms with use of GLP-1RA.⁴ However, the LEADER trial exclusively enrolled patients with established cardiovascular conditions or who were considered at high risk for cardiovascular events.⁵ Finally, the patients were followed for maximum of five years which leaves a knowledge gap regarding effects of long-term use. Therefore, these data cannot stand alone in an evaluation of the overall cancer safety with use of GLP-1RAs in real-world clinical practice of unselected patients with varying baseline health statuses. Although no immediate concerns of carcinogenicity have emerged,⁴ some preclinical^6,7 and epidemiological studies8, 9, 10, 11 have shown increased risk of thyroid^6,8, 9, 10, 11 and pancreatic cancer^7,10,11; while others do not support these findings,12, 13, 14 and some studies report a decreased risk for prostate,^4,8,15 lung, and colon cancer.⁸ In clinical and public health contexts the overall cancer risk is relevant to understand the broad, cumulative effects of a drug across various organ systems, particularly for long-term treatment evaluations. Here, we use high-quality administrative and health registries to examine the long-term cancer risk with use of GLP-1RA in a nationwide, real-world setting of patients treated for up to ten years.

Methods

Study design

In this register-based nationwide emulated trial, we applied randomized controlled trial principles to observational data, reducing design-induced biases by explicitly defining eligibility, treatment, assignment, follow-up, outcome, causal contrast of interest, and analysis plan (Table S1).¹⁶ Thus, the present study was based on a pre-prepared protocol of a target trial to estimate the effects of GLP-1 agonists on the risk of overall cancer. We used an active-comparator, new-user design with dipeptidyl peptidase 4 inhibitors (DPP-4i) as active comparator to GLP-1RA. According to clinical guidelines DPP-4i is an alternative to GLP-1RA in second-line therapy for T2DM¹⁷ and does not per se increase the risk of cancer overall.¹⁸

Data sources

Data was retrieved from Danish nationwide health registries; The Danish Cancer Registry¹⁹ (incident cancer cases), The Danish National Prescription Registry (study drugs, co-medications), The Danish National Patient Registry (comorbidities), The Danish Civil Registration System (administering the unique personal identification number) encoding sex and date of birth, address, and vital status for all residents in Denmark), and Statistics Denmark (educational and socioeconomic parameters). Linkage of registry data was secured by the unique personal identification number. (References and further information about the registers are detailed in Table S2).

Emulated trial population

The eligible source population were all citizens in Denmark during 2007–2019. In this nationwide population, we retrieved first time initiators of GLP-1RA vs DPP-4i naïve to both study drugs indicated for diabetes. Patients were excluded if their age at baseline was below 50 years or if they had previous filled prescriptions for the liraglutide under the trade name Saxenda used in treatment of obesity. Inclusion criteria were uninterrupted residence in Denmark within the previous 10 years and no history of cancer. Inclusion and exclusion of emulated trial patients is presented in Fig. 1. According to Danish legislation, ethics approval or informed consent is not required for register-based research.

Fig. 1.

Flowchart of the emulated trial selection for initiators of GLP-1RA and DPP-4i. Abbreviations: GLP1RA, glucagon-like peptide 1 receptor agonists; DPP-4i, dipeptidyl peptidase-4 inhibitors; DK, Denmark; N, number.

Study drugs

Information on use of GLP-1RA and DPP-4i was obtained from The Danish National Prescription Registry, which holds information on all prescriptions filled at all Danish pharmacies including information on e.g. date of dispense and Anatomical Therapeutic Chemical (ATC) codes. ATC codes for the study drugs are presented in Table S3. Based on information from the nationwide Danish prescription data, GLP-1RAs were typically redeemed every three months. The potential carcinogenic mechanisms of GLP-1RAs in cancer development was assumed to involve a role as promoters of latent cancers or precancerous conditions, or as initiators with a short latency period. Hence, to account for a short latency period and diagnostic delays, we applied a 180-day exposure window after each redeemed prescription. Specific reasons why the assumption of a potential promoting effect of GLP-1RA was made are discussed in detail in the Discussion section. Patients were considered adherent to their initial study drug if they did not discontinue or cross over to the other study drug (switch from GLP-1RA to DPP-4i or vice versa). Discontinuation was defined as the absence of any additional redeemed prescriptions for the initial study drug (GLP-1RA or DPP-4i) within the 180-day exposure window (i.e. at the date corresponding to 180 days after redeemed prescription). If a patient redeemed an additional prescription for the initial study drug within the exposure window, 180 days were added from the date of the new filled prescription. Switch was defined as the redemption of a prescription for the study drug other than the one initially initiated (i.e. at the date when a prescription for the other study drug was redeemed).

Outcomes

The primary outcome was cancer overall defined as first diagnosis of any type of cancer. Cancer diagnoses were obtained from The Danish Cancer Registry¹⁹ (further details in Table S2) recorded according to the International Classification of Diseases, Tenth Revision (ICD-10), and the International Classification of Diseases for Oncology for morphology codes (Table S4).¹⁹ The date of cancer diagnosis was from The Danish Cancer Registry which assigns a diagnosis date that reflects the specific and earliest clinically verified diagnosis date.¹⁹ Additionally, secondary outcomes were death without prior cancer was assessed and a composite endpoint of cancer or death. Exploratory outcomes were the 10 major specific types of cancer, and ‘other cancer types’ containing the remaining cancers. ICD-10 codes are detailed in Table S11.

Covariates

Baseline characteristics at time of entry to the emulated trial were selected a priori and included: age, year, sex, other drug use (individual covariates (yes/no) of glucose lowering drugs, statins, aspirin, non-aspirin non-steroidal anti-inflammatory drugs, antihypertensives, other cardiovascular drugs, 5α-reductase inhibitors, and psychotropic drugs), proxies for diabetes severity (number of non-metformin antidiabetic drugs, duration of metformin use, and duration of non-metformin antidiabetic drug use); diagnosis of comorbidity (individual covariates (yes/no) of obesity, ischaemic heart disease, heart failure, cerebrovascular disease, chronic obstructive pulmonary disease, moderate to severe kidney disease, and moderate to severe liver disease), education (basic/high school, vocational, higher education). Income (quintiles 1–5, where 1 represents the 20% with lowest and 5 the 20% with highest income); and region of residence (North, Central, Southern, Capital, and Zealand Region of Denmark).

Comorbidities, concomitant medications, and diabetes medications were additionally used as time-varying covariates (Table S5). Codes for population characteristics are presented in Table S6. There was no missing information in the registries except for data on income, region, or educational level (∼6%). Missing values were handled by exclusion (Fig. 1 and Supplementary Statistical analyses).

Statistical analyses

Initiators of GLP-1RA were matched 1:1 with initiators of DPP-4i based on nearest neighbour propensity score matching without replacement. The propensity score was estimated using a logistic regression model for being a GLP-1RA initiator adjusted for all baseline covariates (section Covariates). Nonlinear effects of continuous covariates were modelled by restricted cubic splines. Patients were followed from initiation of study drugs (GLP-1RA or DPP-4i) until first diagnosis of cancer, emigration, death, end of follow-up (30 December 2019), or at time of switch or discontinuation of treatment (per-protocol analogue approach).

Cancer

We used g-computation to estimate the risk of cancer with sustained use of either GLP-1RA or DPP-4i over a 10-year period following initiation considering death without prior cancer a competing event (Supplementary Statistical analyses).²⁰ The per-protocol analysis approach was used to estimate the influence of long-term sustained use of GLP-1RAs on cancer development, as were all the patients fully adherent to the treatment. To account for potential informative censoring, we included time-varying covariates (section Covariates) into the g-computation. Each patient's covariate history and outcomes during follow-up were split into intervals of 91 days. The evolution of the time-varying covariates and outcomes at each interval were modelled using parametric models adjusted for study drug use and covariate history (Table S7). The interaction between sex and study drug use was investigated by adding the interaction term in the g-computation outcome model for cancer risk.

Cox proportional hazards model was used to estimate cause-specific hazard ratio (HR) of cancer for GLP-1RA vs DPP-4i considering death without prior cancer as a competing event. If the proportional hazards assumption was violated the HR was allowed to be piece-wise constant over time. A cut off was selected to ensure proportional hazards within each time-interval. Finally, sex specific estimates were computed as separate analyses for both g-computation and the Cox model.

Death

When estimating absolute risk differences between treatments, competing events like death can influence the cancer risk. Even without a true treatment–cancer relationship, differing mortality risks between treatment groups may create distorted differences in cancer risk. This is further detailed in ‘Interpretation of the treatment effect on the risk of cancer when death is a competing event’ in the Supplemental Material (page 19). To elucidate if the effect on cancer could be explained by an effect on death²¹ corresponding analyses were performed for the outcome death without prior cancer and for a composite endpoint of cancer or death. Sex specific estimates were also computed.

Cancer types

To elucidate how the overall cancer risk was influenced by the risk of specific cancer types, we assessed the rates for the 10 major cancers, and for the remaining cancers combined in the category ‘other’. Cox proportional hazards models were used to estimate cause-specific HRs of cancer types for GLP-1RA vs DPP-4i. HRs were reported for the entire follow-up (0–10 years). Sex specific estimates were also computed (Table S10).

Additional analyses

An analysis disregarding adherence (intention-to-treat analogue approach) was conducted by not restricting the follow-up by switch or discontinuation of study drug use to assess any potential initiating effects of the study drugs on cancer development that might emerge after treatment cessation. For this analysis, the Aalen-Johansen estimator was used to compute absolute risks.

To further evaluate the issue of cancer occurring years after cessation analyses was conducted restricted to patients who discontinued treatment within 2 or 3 years after initiation. These patients were followed from landmarks of 2 and 3 years (if discontinued before and still at risk at landmark) until cancer, death or censoring (emigration, end of follow-up, re-initiation of either GLP-1RA or DPP-4i, or 7 or 8 years after the landmarks (corresponding to 10 years after initiation)).

An analysis of a negative control outcome (i.e. cataract (ICD-10: H25), which is age-related and not expected to be influenced by the study drugs) was conducted to further evaluate potential bias by unknown factors, such as a survival benefit. This analysis was conducted with g-computation similar to the analysis of the risk of cancer, however, patients with cataract before baseline were additionally excluded.

An analysis restricted to patients with a known type 2 diabetes diagnosis at baseline (assessed according to The Danish National Prescription Registry (ICD-10: E11) was conducted to limit potential influence of off-label use.

Analyses restricted to five years of sustained use, along with subgroup analyses among patients 65 years or older, and among those with prior CVD at baseline was assessed for comparison with previous studies.

Finally, analyses were performed to assess 1) the impact of switching or discontinuation, 2) changing the interval length in the primary analysis, and 3) if death was not considered a competing event but only a cause for censoring the patients follow-up, i.e. a scenario where death was ‘eliminated’.

All additional analyses, including methods, are detailed in Table S8 of the Supplemental Material.

The goodness-of-fit was assessed for all models. All analyses were performed using the statistical software R (version 4·3·2) and g-computation was performed using the R package gfoRmula (Supplementary Statistical analyses).

Ethical approval

In Denmark, ethical approval is not required for studies that are entirely register-based.

Role of the funding source

The Scientific Committee of the Danish Cancer Society (R354-A20492-23-S3) had no role in considering the study design or in the collection, analysis, interpretation of data, writing of the report, or decision to submit the article for publication.

Results

The emulated trial included a total of 39,460 patients (19,730 new users of GLP-1RA and 19,730 new users of DPP-4i) followed for an interquartile range (IQR) of 1·6–8·2 years (Fig. 1), with 2559 sustained users after 8 years and 344 after 10 years (Fig. 2), IQR of sustained treatment: 0·6–3·6 years. During follow-up a total of 4758 patients were diagnosed with cancer; of which 2340 were diagnosed during sustained use (other malignant neoplasms of skin: 21·7%, colorectal: 11·3%, prostate: 9·7%, lung: 9·2%, neoplasms in urinary system: 7·6%, breast: 7·3%, pancreas: 5·2%, lymphatic tissue: 3·9%, corpus uteri: 3·2%, melanoma of skin: 3·2%, and other cancer: 17·7%). Before propensity score matching, baseline characteristics were similar in the two treatment groups and after matching all covariates were well-balanced (Table S9 and Table 1).

Fig. 2.

Absolute risk of cancer with sustained use of GLP-1RA or DPP-4i. The curves were adjusted for time-varying covariates using g-computation.

Table 1.

Baseline characteristics among diabetes patients in the emulated trial after matching.

	After propensity score matchinga
	GLP-1RA	DPP-4i
N	19,730	19,730
Follow-upb (median, IQR), yr	5 (1·6, 8·3)	4·7 (1·7, 8·2)
Characteristics at study entry
Sex
Female	8409 (42·6)	8290 (42·0)
Male	11,321 (57·4)	11,440 (58·0)
Age (median, IQR), yr	62·6 (56·6, 68·5)	62·7 (56·7, 68·9)
Index year, n (%)
2007	92 (0·5)	135 (0·7)
2008	481 (2·3)	492 (2·5)
2009	1006 (5·1)	1033 (5·2)
2010	3369 (17·1)	2560 (13·0)
2011	2783 (14·1)	3025 (15·3)
2012	1988 (10·1)	2092 (10·6)
2013	1203 (6·1)	1309 (6·6)
2014	889 (4·5)	922 (4·7)
2015	974 (4·9)	979 (5·0)
2016	1042 (5·3)	1033 (5·2)
2017	1167 (5·9)	1369 (6·9)
2018	1781 (9·0)	2142 (10·9)
2019	2955 (15·0)	2639 (13·4)
Education, n (%)
Basic or high school	7978 (40·4)	7877 (39·9)
Vocational	8539 (43·3)	8609 (43·6)
Higher education	3213 (16·3)	3244 (16·4)
Income quintiles, n (%)c
1	3989 (20·2)	3903 (19·8)
2	4010 (20·3)	3882 (19·7)
3	3936 (19·9)	3956 (20·1)
4	3915 (19·8)	3977 (20·2)
5	3880 (19·7)	4012 (20·3)
Region, n (%)
North Denmark region	1571 (8·0)	1537 (7·8)
Central Denmark region	4316 (21·9)	4381 (22·2)
Region of Southern Denmark	3896 (19·7)	3815 (19·3)
Capital region of Denmark	5810 (29·4)	5921 (30·0)
Region Zealand	4137 (21)	4076 (20·7)
Comorbidities, n (%)
Ischaemic heart disease	4959 (25·1)	4549 (23·1)
Heart failure	1683 (8·5)	1521 (7·7)
Cerebrovascular disease	1683 (8·5)	1929 (9·8)
Chronic obstructive pulmonary disease	1621 (8·2)	1515 (7·7)
Moderate to severe kidney disease	495 (2·5)	525 (2·7)
Moderate to severe liver disease	94 (0·5)	98 (0·5)
Obesity	6546 (33·2)	5443 (27·6)
Diabetes medication, n (%)
Metformin	18,690 (94·7)	18,901 (95·8)
Bolus insulin	3221 (16·3)	1507 (7·6)
Basal insulin	8406 (42·6)	4583 (23·2)
Sulfonylureas	11,059 (56·1)	10,215 (51·8)
Combinationsd	0 (0·0)	7 (0·0)
Alpha glucosidase inhibitors	458 (2·3)	305 (1·5)
Thiazolidinediones	635 (3·2)	495 (2·5)
SGLT2 inhibitors	1476 (7·5)	1205 (6·1)
Other blood glucose lowering drugs	790 (4·0)	590 (3·0)
Concomitant medications, n (%)
Statins	16,948 (85·9)	16,669 (84·5)
Aspirin	11,310 (57·3)	10,691 (54·2)
Non-aspirin NSAIDs	17,639 (89·4)	17,553 (89·0)
Antihypertensive medication	18,097 (91·7)	17,953 (91·0)
Other cardiovascular drugs	4875 (24·7)	4561 (23·1)
5-alpha reductase inhibitors	437 (2·2)	441 (2·2)
Psychotropic drugs	11,772 (59·7)	11,454 (58·1)
Proxies for diabetes severity
N of non-metformin antidiabetic drugse
n (%)
0	5505 (27·9)	7239 (36·7)
1	6284 (31·8)	7831 (39·7)
2	4802 (24·3)	3197 (16·2)
3	2478 (12·6)	1203 (6·1)
4+	661 (3·4)	260 (1·3)
Duration of metformin usef (median, IQR), yr	6·2 (2·5, 10·1)	5·2 (2·2, 8·8)
Duration of non-metformin antidiabetic drug usef (median, IQR), yr	4·6 (0·0, 11·6)	2·1 (0·0, 8·6)

Abbreviations: DPP-4i, dipeptidyl peptidase 4 inhibitors; GLP-1RA, glucagon-like peptide 1 receptor agonists; NSAIDs, non-steroidal anti-inflammatory drugs; SGLT2, Sodium-glucose co-transporter 2.

^aMatching was based on propensity score from a logistic regression model with being a GLP-1RA initiator as outcome.

^bBased on follow-up without censoring at non-adherence.

^cIncome quintiles 1–5: 1 represents the 20% with lowest and 5 the 20% with highest income.

^dCombinations of oral blood glucose lowering drugs, except combinations with metformin.

^eOther than the respective study drugs.

^fCalculated as time since first filled prescription.

Cancer

The absolute risk of cancer with sustained use of either GLP-1RA or DPP-4i is shown in Fig. 2. After 10-years of sustained use the absolute risk of cancer was 25·5 (95% CI 23·3–27·4) per 100 users of GLP-1RA and 21·4 (95% CI 18·8–24·1) per 100 users of DPP-4i. The risk difference was 4·1 (95% CI 0·4–7·2) per 100 uses of GLP-1RA vs DPP-4i (Table 2). The cause-specific HR was 0·97 (95% CI 0·89–1·06) within 0–6 years and 1·35 (95% CI 1·05–1·73) within 6–10 years.

Table 2.

Effect of sustained use of GLP-1RA vs DPP-4i on the 10-year absolute risk of overall cancer, and secondary the outcomes death without prior cancer, cancer and death as composite outcome.

	All patients	Men	Women
	10 years	10 years	10 years
Cancer overall
GLP-1 RA
Na	19,730	11,321	8409
Eventsb	1364	820	544
Absolute risk (95% CI)c	25·53 (23·26, 27·44)	25·46 (23·13, 28·02)	25·9 (23·3, 29·11)
DPP-4i
Na	19,730	11,440	8290
Eventsb	976	621	355
Absolute risk (95% CI)c	21·42 (18·84, 24·05)	23·3 (20·11, 26·84)	19·31 (15·78, 23·6)
GLP-1 RA vs DPP-4i
Absolute risk difference (95% CI)	4·11 (0·36, 7·16)	2·16 (−2·23, 6·16)	6·58 (1·8, 10·71)
Additional analyses
Death without prior cancer
Eventsb	1433	900	533
GLP-1 RA vs DPP-4i
Absolute risk difference (95% CI)	−4·88 (−7·59, −2·36)	−4·89 (−8·56, −1·41)	−4·96 (−8·83, −1·62)
Composite endpoint (cancer or death)
Eventsb	3773	2341	1437
GLP-1 RA vs DPP-4i
Absolute risk difference (95% CI)	−1·15 (−4·91, 2·48)	−2·94 (−8·26, 1·72)	0·63 (−4·74, 6·01)

The absolute risks with sustained use were estimated using g-computation accounting for time-varying covariates.

^aNumber of patients at baseline.

^bNumber of cancer diagnoses within 10 years.

^cExpected number of cancer diagnoses per 100 users within 10 years.

The absolute risk difference for cancer among users of GLP-1RA vs users of DPP-4i seemed more pronounced among women, than among men after 10 years of sustained use [risk difference among women was 6·6 (95% CI 1·8–10·7) and 2·2 (95% CI −2·2 to 6·2) among men per 100 users] Fig. 2. However, in the cancer-risk model the interaction term between sex and treatment group was not statistically significant (p-value of 0·12).

The HRs were 0·94 (95% CI 0·84–1·05) after 0–6 years and 1·29 (95% CI 0·95–1·75) after 6–10 years among men, among women the HRs were 1·03 (95% CI 0·89–1·19) and 1·49 (95% CI 0·98–2·27).

Death

After 10 years the absolute risk difference of death without prior cancer was lower with sustained use of GLP-1RA vs DPP-4i [−4·9 (95% CI −7·6 to −2·4) per 100 users] (Table 2, Figure S1). No significant risk difference was found for the composite endpoint of cancer or death [risk difference after 10 years of use: −1·2 (95% CI −4·9 to 2·5) per 100 users] (Table 2, Figure S2).

Cancer types

Melanoma of skin, neoplasms in urinary system, prostate, and other cancers showed no association or a suggested maximum decrease by 10% in point estimates, while point estimates suggested a decrease by approximately 20% for lung and colorectal cancer. The point estimate for other malignant neoplasms of skin tended to be increased by approximately 10%, while the remaining cancers showed increase in point estimates of approximately 20% or more. None of the HRs were statistically significant except for corpus uteri cancer (Table S10). HRs for men and women separately showed only minor variations in HRs; however, these were affected by few numbers of events (Table S11).

Additional analyses

Not accounting for adherence (i.e. intention-to-treat analogue approach), the 10-year risk difference for cancer overall was 1·4 (95% CI 0·2–2·6) per 100 users among initiators of GLP-1RA vs DPP-4i. The HR was 0·97 (95% CI 0·9–1·0) after 0–8 years and 1·40 (95% CI 1·1–1·7) after 8–10 years of follow-up, respectively (Table S12, Figure S3).

Among patients who discontinued within the first 2 years, the 8-year post-landmark cancer risk difference was −0·24 (−3·27 to 2·78) per 100 discontinued GLP1-RA users vs discontinued DPP-4i users (Figure S4, Table S13). For those who discontinued within 3 years, the 7-year post-landmark risk difference was −1·26 (−5·93 to 3·42) (Figure S5, Table S14).

Assessment of the negative control outcome cataract showed a 10-year risk difference of 1·22 (95% CI −0·88 to 3·09) per 100 users among sustained users of GLP-1RA vs DPP-4i (Figure S6, Table S15).

Among patients with a known type 2 diabetes diagnosis at baseline, the 10-year risk difference of cancer overall was 4·18 (0·4, 7·19) per 100 users among initiators of GLP-1RA vs DPP-4i (Figure S7, Table S16).

The 5-year absolute risk of cancer overall with sustained use was 11·2 (95% CI 10·4–11·9) per 100 users of GLP-1RA and 10·6 (95% CI 9·9–11·7) per 100 users of DPP-4i. The risk difference was 0·5 (95% CI −0·6 to 1·6) per 100 users. The HR within 5 years was 0·98 (95% CI 0·90–1·08). Among patients 65 years or older at baseline and among patients with prior CVD at baseline, the absolute risk differences were similar to the main analysis (results not shown).

DPP-4i users had a higher risk of switching or discontinuing than GLP-1RA, e.g. obese users of DPP-4i were more likely to switch to GLP-1RA during follow-up (Figure S8).

After changing the interval length in the primary analysis from 91 to 73 days, the results were virtually unchanged [risk of cancer after 5 years with sustained use: 0·5 (95% CI −0·6, 1·6), 10 years: 4·0 (95% CI 0·8–6·8)] (Figure S9). The risk difference for cancer with death eliminated was 3·6 (95% CI −0·6 to 7·1) per 100 users after 10 years of sustained use (Figure S10, and Table S17).

Discussion

In this nationwide emulated trial using high-quality disease and prescription registries, long-term use of GLP-1RA was associated with a small increased risk of cancer overall, compared with use of DPP-4i. A survival advantage among users of GLP-1RA may explain this finding, which was partly supported by a negative control outcome (age-related cataracts), which showed a tendency in line with the results of overall cancer. Although residual confounding by body mass index (BMI) is possible, assessments of specific cancer types showed tendencies for increased rates for cancers not linked to obesity among GLP-1RA users, suggesting that factors beyond BMI, potentially a drug effect, could still play a role. Further, not all cancers showed increased rates among GLP-1RA users, as would be expected if a general survival advantage were to completely explain the increased cancer risk.

Sustained use of GLP-RA for a maximum of 5 years was not associated with an increased incidence of cancer, compared with use of DPP-4i. This is in accordance with findings from the randomized controlled LEADER trial assessing several cancer types in patients assigned to a GLP-1RA (liraglutide) during a maximum follow-up of 5 years, compared to placebo in addition to standard care.⁴ Other RCTs assessing shorter-term use of GLP-1RA (i.e. maximum of 3·8 years)) point to the same conclusion of no increased cancer risk.^12,22,23 RCTs typically include individuals not representative of the general population which limits generalisability to diverse real-world populations. However, our results including all patients treated with GLP-RA in Denmark suggested no modified influence of GLP-1RA on cancer risk among patients with prior CVD or among elderly patients (≥65 years).

The most comprehensive real-world assessment of neoplasms in users of GLP-1RA is a disproportionality analysis of the FDA Adverse Event Reporting System (FAERS) which conclude no association between use of GLP-1RA and cancer risk.⁹ FAERS is based on data from a voluntary reporting system containing adverse events submitted to the U.S. Food and Drug Administration with limitations that have previously been debated.²⁴ Common for the available real-world studies published today is that the follow-up is limited to a maximum of 5 years. A recent real-world study, reported reduced risks of some of the most common types of cancer i.e. prostate, lung, and colon cancer with GLP-1RA use, compared with use of metformin.⁸ Conflicting results have emerged in real-world studies following GLP-1RA users for breast cancer risk during a maximum follow-up of 3·5 years.^15,25, 26, 27 The few available real-world studies assessing associations with less common types of cancer have reported inconsistent findings for thyroid cancer and pancreatic cancer suggesting both an increased risk8, 9, 10^,28 and no elevated risk12, 13, 14 with GLP-1 based therapies. No real-world study or randomized trial has addressed the overall long-term cancer safety with GLP-1RA use. This study offers novel real-world evidence on GLP-1RAs long-term cancer safety, needed for assessing its public health risk-benefit balance.

In the absence of high-quality randomized trials with large sample sizes, this emulated target trial using real-world data and g-computation methods may provide better evidence on the cancer safety of long-term GLP-1RA use than conventional observational studies. The strengths include a causal interpretation of results if the assumptions for causal inference (treatment consistency, positivity, no unrecognized confounding, and correct model specification) are met. We adjusted for patient characteristics, changes in comorbidities, and other diabetes treatments using time-varying information in the g-computation. Additionally, the use of nationwide high-quality registers in Denmark provided comprehensive data on patient characteristics, prescription drug use, medical conditions, sociodemographic variables, and validated cancer diagnoses. All information is linked at individual level through a unique personal identifier (PIN). Moreover, the information is continuously updated,^19,29 and long duration of follow-up was available, which allowed for detailed long-term assessments of patients exposed to GLP-1RA use and their subsequent development of cancer. The Danish Cancer Registry is estimated to be 95%–98% complete, and have a strict validation process of cancer diagnosis thus recognized for its high degree of validity and completeness.³⁰ As we use nationwide real-world data, our findings are expected to have good generalizability and results to be broadly applicable considering the nationwide real-world population. We used an active-comparator new user design to reduce confounding by indication or by unmeasured clinical characteristics, since the comparator drug are recommended according to clinical guidelines for the same clinical purpose as GLP-1RA at similar disease stage of type 2 diabetes.¹⁷ Additionally, we controlled for a large number of patient characteristics at baseline, and several time-varying covariates using g-computation in an observational analogue of a per-protocol analysis in a randomized trial. By use of the per protocol-analogue analysis we accounted for adherence to the study drugs. Accounting for adherence can be crucial in real-world settings where non-adherence is common, especially when the assessment of interest is the long-term drug effects.³¹ In the present study, novel methods were used, making it possible to account for the dynamic nature of adherence and the changing risk profiles of patients over time.³²

In the present study, GLP-1RAs were assumed to act through a potential promoting effect of latent or precancerous conditions, with the expectation that any increased cancer risk would diminish after discontinuation. To challenge this assumption, two additional approaches were applied. An intention-to-treat analogue approach was used to examine potential cancer development, including after discontinuation, when specific lag times are unknown. The intention-to-treat analogue approach disregards whether individuals adhere to their medications and applies no criteria for a minimum exposed duration. Descriptive data showed patients had sustained use of GLP-1RA for an IQR of 0·6–3·6 years during the 10-year follow-up. Findings yielded a less pronounced absolute risk difference for overall cancer in the analysis disregarding adherence. Further, two landmark analyses of exclusively discontinued users showed no evidence of effects on cancer development occurring 7–8 years after a maximum of 2–3 years of GLP-1RA vs DPP-4i use. These findings suggest that short-term GLP-1RA exposure is not associated with delayed cancer onset, supporting the hypothesis that any increased cancer risk linked to GLP-1RA use could act through promoting rather than initiating mechanisms. However, due to low discontinuation rates and limited post-discontinuation follow-up among long-term GLP-1RA users, potential cancer effects after discontinuation could not be reliably assessed in this group.

While the biological mechanisms of GLP-1RAs on cancer development remain incompletely understood,³³ the assumption that GLP-1RAs may act as cancer promoters rather than initiators is grounded in: pre-clinical³⁴ and rodent studies^6,35 suggesting that GLP-1RAs may promote certain cancers via GLP-1 receptor-mediated mechanisms. Some observational studies indicate that any increased cancer risk with GLP-1RA use appears reversible after discontinuation, arguing against potential cancer effects occurring years after exposure.^26,28

Previous studies on other hormone-related drugs, such as hormonal contraceptives and breast cancer, confirm that drugs can accelerate the progression or growth of premalignant or latent disease, with increased risk observed among current or recent users, rather than emerging years later.³⁶ Similar trends are seen for hormone replacement therapy, where increased risk of cancer is observed with longer duration of use, but declines after discontinuation and approaches the risk in non-users approximately after the same amount of time as the duration of use.37, 38, 39, 40 However, given the complexity about potential mechanisms of GLP-1RAs on cancer development, it is not possible to rule out an initiating effect of GLP-1RAs—or the possibility of both initiating and promoting effects.

The overall cancer risk was mainly driven by skin neoplasms and colorectal cancer (both >10%). Prostate (9·7%), lung (9·2%), urinary (7·6%), and breast cancer (7·3%) with frequencies above 7%, while others like pancreatic, lymphatic, corpus uteri, and melanoma were around 5% or less. Effects among the rarest types of cancer would be less clearly expressed in the present study. However, the assessment of the overall cancer risk provides awareness of the public health net-balance according to cancer risk with sustained use of GLP-1RA from a public health perspective.

The survival advantage among users of GLP-1RA could cause a selection of patients whereby more frail patients survive on the long-term when using GLP-1RA compared to users of DPP-4i. This could explain the higher risk of cancer late in follow-up among users of GLP-1RA, compared to the comparator. This assumption seemed to be supported by the assessment of a negative control outcome indicating similar trends as cancer overall. However, results for specific types of cancer did not show a consistent indication for increased rate of different cancers with use of GLP-1RA, which would be expected if a general survival advantage were to completely explain the results. Altogether, we cannot exclude that the small increased cancer risk found in this study could be linked to use of GLP-1RA.

Reverse causality can occur if an undiagnosed cancer influences the initiation of a drug prescribed e.g. for treatment of diabetes. This could lead to a misleading association between the drug and cancer. However, in a situation where diabetes medication is prescribed because of a latent cancer not diagnosed yet, an increase in cancers diagnoses shortly after drug initiation would be observed. This was not found in the current study. In contrast a slight increase in cancers was found years after initiation. This indicates that reverse causation is not likely in the current study. Furthermore, for reverse causality to significantly influence the results, it would need to be more pronounced for GLP-1RA than for DPP-4i, which seems unlikely. Consequently, reverse causation is not likely to explain our findings.

Both baseline and time-varying adjustments of obesity were included in our g-computation model. However, switching from DPP-4i to GLP-1RA was common, especially among obese users. This may affect estimates for long-term GLP-1RA use, as obese patients were not censored upon switching to the same extent in the GLP-1RA group. However, assessments of specific types of cancer showed no consistent pattern in rates for obesity vs non-obesity associated cancers. Indicating that other factors than BMI is likely to play a role, which could be a potential drug effect. Reassuringly, an additional analysis did not indicate a potential influence of off-label GLP-1RA use for weight loss (i.e. an analysis among patients with known type 2 diabetes) as results were the same as for overall cancer.

The influence of different sub-types of GLP-1 analogues were not possible to assess since liraglutide accounted for the vast majority of use in Denmark during the study period.

While the emulated trial framework avoids design-induced biases, it does not eliminate biases arising from data limitations, such as confounding by unobserved covariates.⁴¹ Insufficient information to adjust for confounders (e.g. lack of information on BMI below the threshold of obesity and weight loss) may hinder the full potential of the emulated trial design.⁴² Hence, a target trial emulation that relies on sufficient adjustment may not succeed if this is mismeasured in the observational data (e.g. if assessments of obesity was incomplete), which cannot be excluded in the present study. Future studies on GLP-1RA and cancer risk could further benefit and mitigate these limitations by improving adjustments for obesity and BMI, particularly when examining obesity-related cancers.

Another limitation was lower statistical precision at very long-term durations, however, the estimates for long-term durations still reached statistical significance with a substantial proportion of individuals at risk.

Conclusion

This nationwide emulated trial of patients treated in routine clinical practice suggests a small increased cancer risk with long-term sustained GLP-1RA use, potentially due to a survival benefit among users. Moreover, residual confounding by BMI cannot be excluded despite the advanced statistical methods. Findings were generally reassuring, particularly for short-term treatment. Given the widespread and increasing use of GLP-1RAs for diabetes and obesity, this study provides novel findings for public health evaluations of the overall risk-benefit balance of GLP-1RAs.

Contributors

Mads Gamborg, MSc (Conceptualization; Investigation; Data curation; Methodology; Project administration; Validation; Visualization; Writing—original draft; Writing—review & editing), Mia Klinten Grand, PhD (Data curation; Formal analysis; Methodology; Validation; Visualization; Writing—review & editing), Kathrine Grell, PhD (Data curation; Formal analysis; Methodology; Validation; Visualization; Writing—review & editing), Susanne Rosthøj, PhD (Methodology; Validation; Writing—review & editing), Ulrik Pedersen-Bjergaard, PhD (Validation; Writing—review & editing), Christian Torp-Pedersen, PhD (Methodology; Validation; Writing—review & editing), Lina Steinrud Mørch, PhD (Conceptualization; Funding acquisition; Investigation; Data curation; Methodology; Project administration; Supervision; Validation; Visualization; Writing—review & editing). Mørch is the guarantor. The corresponding authors attest that all listed authors meet authorship criteria and that no others meeting the criteria have been omitted.

Data sharing statement

Data are stored remotely on a secure platform at Statistics Denmark. According to Danish regulations, individual-level data can only be made available for researchers who fulfil legal requirements for access to sensitive data. Please contact Lina Steinrud Mørch (morch@cancer.dk) for further questions about data access.

Declaration of interests

All authors have completed the ICMJE uniform disclosure form at www.icmje.org/disclosure-of-interest/ and declare: MG, MKG, SR, and LSM have nothing to disclose. UPB reports payments from Abbott (educational events), Novo Nordisk, and Sanofi (educational events and expert testimony). CTP reports payments from Bayer and Novo Nordisk (research grant unrelated to the current study).