Document resource
WHAT IS ALREADY KNOWN ON THIS TOPIC The question of whether incretin-based therapies increase the risk for thyroid cancer (TC) in humans remains unanswered, with studies showing conflicting results.WHAT THIS STUDY ADDS Initiation of glucagon-like peptide 1 receptor agonists and dipeptidyl-peptidase-4 inhibitors, compared with sodium-glucose cotransporter-2 inhibitors, in older US adults with type 2 diabetes did not appear to increase the 3-year risk of TC.HOW THIS STUDY MIGHT AFFECT RESEARCH, PRACTICE OR POLICY The short-term risk of TC should not affect treatment choice in older adult patients with type 2 diabetes, but the risk with long-term treatment with incretin-based drugs or subtype-specific risks remains unknown.Introduction Incretin-based therapies—glucagon-like peptide 1 receptor agonists (GLP-1RA) and dipeptidyl-peptidase-4 inhibitors (DPP-4i)—are commonly used as second-line antihyperglycemics in patients with type 2 diabetes mellitus (T2DM). 1 Incretins promote glucose-dependent insulin secretion, suppress glucagon secretion, slow gastric emptying, and promote satiety.2The GLP-1 receptor is expressed in normal, premalignant, or malignant thyroid tissues.3 Activation of the GLP-1 receptor has been shown to cause thyroid C-cell hyperplasia and C-cell tumors in preclinical carcinogenicity studies in rodents.4 5 Informed by this finding, the US. Food and Drug Administration (FDA) issued a warning regarding medullary thyroid cancer (TC) for long-acting GLP-1RA.6 Nevertheless, the relevance of these findings in humans has been questioned, as there is a discrepancy in the level of expression and biology of GLP-1 receptors in the thyroid between rodents and primates.3 7Currently, the relationship between incretin-based therapies in humans and TC risk remains uncertain, with data showing conflicting results.8–20 Some studies have reported evidence of no increased risk of TC with the use of GLP-1RA or DPP-4i relative to placebo or other antihyperglycemics,9 10 12 13 16–19 while other studies could not rule out possible increased risk of TC with the use of GLP-1RA compared with other antihyperglycemics.8 11 14 15 20 Notably, the nested case–control study by Bezin et al from France found an increased risk of all TC and medullary TC, particularly after 1–3 years of treatment.8 A recent cohort study by Pasternak et al from Scandinavia and a multisite cohort study by Baxter et al showed no increased risk.17 19The objective of this study was to estimate the comparative effect of the use of incretin-based therapies versus sodium-glucose cotransporter-2 inhibitors (SGLT-2i)—a guideline-recommended treatment alternative—on TC incidence among US older adults with type 2 diabetes.Research design and methods Data source We conducted this population-based cohort study using Medicare Fee-for-Service (FFS) Data from 2008 to 2019. Medicare is a US federal health insurance for anyone aged≥65 years, and others under 65 with certain conditions. This database contains deidentified individual-level, longitudinal information on demographics, diagnoses and procedures, and outpatient prescription dispensations recorded during billing of all healthcare encounters. 21The Medicare data available at UNC comprise a randomly selected 20% sample of Medicare FFS beneficiaries who were enrolled in Medicare Part A (inpatient services), B (physician and outpatient services), and D (prescription drugs) plans for a minimum of one calendar month between 2007 and 2019.Study population The base population for this analysis consisted of all UNC FFS Medicare beneficiaries aged 66 years or older, with at least one prescription dispensing claim for GLP-1RA, DPP-4i, or SGLT-2i between January 1, 2008, and December 31, 2018.We implemented an active-comparator, new user (ACNU) cohort study design to identify three pair-wise new user cohorts comparing: GLP-1RA and SGLT-2i (cohort I); DPP-4i and SGLT-2i (cohort II); and GLP-1RA and DPP-4i (cohort III). To be eligible, patients were required to have at least 12 months of continuous part A, B, and D coverage before the first prescription date. New users were defined as individuals who initiated the drugs of interest or their active comparator after a preceding washout period of at least 12 months without a prescription for the drug classes compared. Subjects were allowed to have other antihyperglycemics during the washout period except the drugs being compared. By enrolling only new users and following subjects from the start of treatment, time-varying hazards, including lag times, can be assessed and described, while preserving the temporality of covariate assessment. Patients who did not refill the same drug class within the days’ supply plus a grace period of 30 days after the first prescription were excluded. Requiring two prescriptions increases the probability that patients actually started therapy. Patients with any cancer diagnosis (except non-melanoma skin cancer) or cancer-related procedures were also excluded from the study. Since the earliest pharmacy data (part D claims) available was from January 1, 2007, the earliest possible first prescription date was January 1, 2008. Additionally, since SGLT-2i were approved in March 2013, the earliest possible first prescription date for cohorts I and II was January 1, 2014.Outcome Our primary outcome was TC. We identified outcomes using a prior published algorithm that has been shown to have a high positive predictive value of 0.91 (95% CI 0.81 to 0.96). 22 To establish a TC diagnosis, we required both a thyroidectomy and at least two separate diagnoses for malignant neoplasm of thyroid gland (International Classification of Disease, 9th Revision (ICD-9) 193 or International Classification of Disease, 10th Revision (ICD-10) codes C73, D09.3, or D44.0) within 90 days after the thyroidectomy. The date of TC diagnosis was assigned at the first TC claim associated with thyroidectomy since a definitive procedure date is preferable to ICD codes, which may vary by aspects of care delivery. Because this algorithm has not been validated in Medicare, as secondary analyses, we implemented three other algorithms for defining TC: (1) claims for any non-surgical TC treatment (chemotherapy, radioiodine, radiation) followed by ≥2 separate diagnoses for malignant neoplasm of thyroid gland within 90 days22; (2) requiring ≥2 diagnoses of TC within 2 months (rather than 90 days) and not requiring a thyroidectomy23; (3) claims for (any non-surgical TC treatment (chemotherapy, radioiodine, radiation) OR a thyroidectomy) followed by ≥2 separate diagnoses for malignant neoplasm of thyroid gland within 90 days.Follow-up Follow-up began 6 months after the second prescription date, allowing for a combined induction and latent period for TC, 24 and continued until the earliest occurrence of a diagnosis of TC or any censoring event.25 While it is impossible to know the induction and latent periods given the available data, we chose a 6-month lag period based on examples from prior cancer incidence studies.24 26 This duration, which we varied in both directions in sensitivity analyses, was to strike a balance between being too short (potentially suffering from reverse causality) and too long (potentially missing some cancer-promoting, late-stage effects). Censoring events were: (1) treatment discontinuation, switching to/augmentation with comparator with person-time and events counted for up to 6 months after the censoring event (latent period); (2) death from any cause (with the exception of an analysis treating death as a competing event); (3) disenrollment in Medicare Part A, B, or D; or (4) end of study (December 31, 2019), whichever came first. The primary analysis of this study employed an ‘as-treated’ approach. Patients were defined as exposed to the initial treatment (IT) until the treatment changed due to discontinuation of the index drug class or switching to or addition of a drug from the comparator drug class. Treatment discontinuation was defined as the absence of a prescription of the cohort drug class within the days of supply plus a 30-day grace period following the last prescription. The same definition was applied to switching. Augmenting with comparator was defined as the first dispensed prescription of the comparator. We also varied the length of the grace period and the lag periods (both initial and after stopping, switching, or augmenting) in sensitivity analyses in both directions (longer and shorter) separately to assess the robustness of the primary analysis results. Patients who developed TC during the initial combined induction and latent periods were excluded from the study.Statistical analysis We examined patient baseline characteristics including demographics and socioeconomic status, calendar time, comorbidities, comedications, health behaviors, and metrics of healthcare utilization during the 12 months prior to the first prescription for each cohort ( table 1). To reduce imbalance in patient characteristics and to control for confounding, we used logistic regression to estimate propensity scores (PS)—the individual probability of initiating GLP-1RA or DPP-4i compared with SGLT-2i (or GLP-1RA vs DPP-4i), conditional on all baseline covariates. To achieve our primary aim of estimating the counterfactual scenario of what would have happened to the initiators of the drug of interest if they had initiated the comparator instead, we estimated the average treatment effect in the treated by reweighting the comparator drug initiators by the PS odds (PS/(1−PS)) (standardized morbidity ratio (SMR) weighting).27 We evaluated the adequacy of covariate balance based on standardized absolute mean differences (SAMD), with a threshold of less than 0.1 indicating satisfactory balance.28 We estimated adjusted 3-year cumulative risk differences of TC (aRDs) with 95% CIs using SMR weighted Kaplan-Meier survival functions, and adjusted HRs and 95% CIs using SMR weighted Cox proportional hazard regression models overall and stratified by calendar year of initiation.29 When estimating the risk of cancer outcomes in older Medicare patients, censoring those who died before hypothetically experiencing the outcome of interest, as commonly done, could introduce bias in the risk estimation.30 31 Therefore, in a secondary analysis, we employed Aalen-Johansen (AJ) estimators to assess the impact of this potential bias on our primary risk estimates. First, we estimated the overall survival function and hazard function for each event type (outcome of interest as well as death) in the SMR weighted population. Next, we computed the AJ estimators by multiplying the hazard function of the outcome of interest at each event time by the overall survival at the previous time point. This approach effectively treats death as a competing event by assigning patients a risk of zero for incident TC after death.32 To address the potential for detection bias, we compared the incidence of potential indicators of screening for TC in all pair-wise comparison cohorts and described any observed differences. All analyses were performed with SAS V.9.4 statistical software and R software (V.4.4.2).Table 1Patient characteristics of incretin-based therapies and comparator initiators*CharacteristicsGLP-1RA versus SGLT-2iDPP-4i versus SGLT-2i§GLP-1RA(n=36 238)†SGLT-2i(n=37 150)Weighted SGLT-2i‡ (n=36 449)DPP-4i(n=84 283)†SGLT-2i(n=21 991)Weighted SGLT-2i‡(n=86 499)Age, mean (SD), years72.2 (5.55)72.6 (5.73)72.1 (5.48)75.3 (7.23)72.2 (5.52)75.3 (7.16)Male, n (%)16 236 (44.8)19 238 (51.8)16 305 (44.7)36 060 (42.8)11 706 (53.2)35 987 (41.6)Race, n (%) Non-Hispanic white29 550 (81.5)29 119 (78.4)30 003 (82.3)63 167 (74.9)17 887 (81.3)63 541 (73.5) Black (or African American)3211 (8.9)2670 (7.2)3112 (8.5)9627 (11.4)1547 (7.0)10 337 (12.0) Asian/Pacific Islander838 (2.3)1908 (5.1)787 (2.2)3930 (4.7)757 (3.4)4567 (5.3) Hispanic998 (2.8)1362 (3.7)970 (2.7)3557 (4.2)664 (3.0)3956 (4.6) American Indian/Alaska Native249 (0.7)181 (0.5)251 (0.7)591 (0.7)108 (0.5)587 (0.7) Other724 (2.0)1057 (2.8)689 (1.9)2273 (2.7)505 (2.3)2350 (2.7) Unknown668 (1.8)853 (2.3)637 (1.7)1138 (1.4)523 (2.4)1161 (1.3)Calendar year of drug initiation, n (%) 20142688 (7.4)2925 (7.9)2965 (8.1)13 261 (15.7)1434 (6.5)12 482 (14.4) 20153273 (9.0)5063 (13.6)3421 (9.4)13 527 (16.0)2739 (12.5)13 998 (16.2) 20164837 (13.3)5773 (15.5)4853 (13.3)15 481 (18.4)3127 (14.2)15 084 (17.4) 20176966 (19.2)7465 (20.1)6864 (18.8)16 435 (19.5)4434 (20.2)16 631 (19.2) 20188787 (24.2)6924 (18.6)8719 (23.9)14 237 (16.9)4284 (19.5)15 955 (18.4) 20199687 (26.7)9000 (24.2)9626 (26.4)11 342 (13.5)5973 (27.2)12 348 (14.3)Diabetic complications, n (%) Nephropathy1416 (3.9)983 (2.6)1420 (3.9)4339 (5.1)485 (2.2)4778 (5.5) Neuropathy14 155 (39.1)11 445 (30.8)14 376 (39.4)25 347 (30.1)6540 (29.7)27 628 (31.9) Retinopathy7342 (20.3)6224 (16.8)7458 (20.5)12 994 (15.4)3448 (15.7)13 692 (15.8) Cataract16 396 (45.2)16 343 (44.0)16 543 (45.4)34 654 (41.1)9691 (44.1)35 061 (40.5)Thyroid disorders, n (%) Hyperthyroidism651 (1.8)615 (1.7)663 (1.8)1671 (2.0)310 (1.4)1720 (2.0) Congenital hypothyroidism14 (0.0)15 (0.0)14 (0.0)48 (0.1)12 (0.1)35 (0.0) Acquired hypothyroidism11 264 (31.1)10 140 (27.3)11 455 (31.4)25 054 (29.7)5894 (26.8)26 442 (30.6) Goiter2139 (5.9)1887 (5.1)2222 (6.1)3851 (4.6)1032 (4.7)4066 (4.7) Thyroiditis490 (1.4)406 (1.1)490 (1.3)767 (0.9)225 (1.0)732 (0.8) Other thyroid disorders983 (2.7)913 (2.5)998 (2.7)2407 (2.9)486 (2.2)2648 (3.1)Cardiovascular comorbidities, n (%) Acute myocardial infarction1010 (2.8)931 (2.5)986 (2.7)3018 (3.6)632 (2.9)3245 (3.8) Congestive heart failure7191 (19.8)5412 (14.6)7317 (20.1)18 748 (22.2)3302 (15.0)20 736 (24.0) Hyperlipidemia32 377 (89.3)33 295 (89.6)32 538 (89.3)73 190 (86.8)19 501 (88.7)74 766 (86.4) Hypertension34 063 (94.0)34 384 (92.6)34 211 (93.9)78 495 (93.1)20 248 (92.1)80 729 (93.3) Ischemic heart disease13 940 (38.5)13 893 (37.4)14 168 (38.9)32 971 (39.1)8403 (38.2)34 002 (39.3)Other comorbidities, n (%) Chronic kidney disease24 798 (68.4)20 358 (54.8)24 877 (68.3)48 676 (57.8)11 901 (54.1)50 357 (58.2) Chronic obstructive pulmonary disease8102 (22.4)6814 (18.3)8299 (22.8)19 529 (23.2)4140 (18.8)22 111 (25.6) Depression8790 (24.3)6450 (17.4)9037 (24.8)18 850 (22.4)3921 (17.8)20 841 (24.1) Obesity18 593 (51.3)13 634 (36.7)18 894 (51.8)25 426 (30.2)8550 (38.9)26 778 (31.0) Pancreatic disorders508 (1.4)706 (1.9)519 (1.4)1596 (1.9)448 (2.0)1677 (1.9)Health behaviors, n (%) Alcohol7567 (20.9)6629 (17.8)7760 (21.3)18 675 (22.2)4092 (18.6)20 786 (24.0) Tobacco8689 (24.0)7265 (19.6)8720 (23.9)18 007 (21.4)4728 (21.5)19 263 (22.3)Comedications, n (%) Metformin23 506 (64.9)29 108 (78.4)23 812 (65.3)58 206 (69.1)16 878 (76.7)60 345 (69.8) ACE inhibitors15 777 (43.5)16 427 (44.2)16 014 (43.9)36 896 (43.8)9924 (45.1)36 949 (42.7) Angiotensin II receptor blockers13 791 (38.1)14 071 (37.9)13 960 (38.3)28 901 (34.3)7868 (35.8)30 778 (35.6) Beta blockers20 034 (55.3)19 364 (52.1)20 134 (55.2)46 718 (55.4)11 430 (52.0)47 905 (55.4) Calcium channel blockers13 312 (36.7)12 910 (34.8)13 149 (36.1)33 159 (39.3)7435 (33.8)33 963 (39.3) Sulfonylureas15 052 (41.5)17 251 (46.4)15 256 (41.9)36 801 (43.7)9230 (42.0)37 646 (43.5) Thiazolidinedione3362 (9.3)3942 (10.6)3475 (9.5)6226 (7.4)2113 (9.6)6549 (7.6) Loop diuretics10 382 (28.6)7178 (19.3)10 607 (29.1)22 980 (27.3)4430 (20.1)25 232 (29.2) Other diuretics1683 (4.6)1152 (3.1)1733 (4.8)3345 (4.0)737 (3.4)3846 (4.4) Statin29 050 (80.2)29 795 (80.2)29 103 (79.8)63 873 (75.8)17 273 (78.5)65 083 (75.2) NSAIDs9970 (27.5)10 572 (28.5)10 185 (27.9)21 535 (25.6)5816 (26.4)22 762 (26.3) Dipeptidyl peptidase 4 (DPP-4) inhibitors11 498 (31.7)14 922 (40.2)11 478 (31.5)––– Antiandrogen75 (0.2)65 (0.2)76 (0.2)164 (0.2)47 (0.2)194 (0.2) Antiestrogen59 (0.2)60 (0.2)61 (0.2)128 (0.2)38 (0.2)123 (0.1) Aromatase inhibitors422 (1.2)310 (0.8)437 (1.2)635 (0.8)181 (0.8)627 (0.7) Estrogen1075 (3.0)851 (2.3)1088 (3.0)1936 (2.3)482 (2.2)1978 (2.3) Progestogen235 (0.6)246 (0.7)240 (0.7)790 (0.9)134 (0.6)939 (1.1) Other hormone antagonistsNTSR16 (0.0)NTSR32 (0.0)NTSR30 (0.0) Long-acting insulin16 275 (44.9)8616 (23.2)16 674 (45.7)16 083 (19.1)5767 (26.2)16 810 (19.4) Short-acting insulin8800 (24.3)4050 (10.9)9076 (24.9)8878 (10.5)3001 (13.6)9473 (11.0) Insulin combinations1830 (5.0)1135 (3.1)1977 (5.4)2211 (2.6)813 (3.7)2147 (2.5)Healthcare utilization, n (%) Influenza shot21 009 (58.0)19 701 (53.0)20 866 (57.2)37 527 (44.5)12 097 (55.0)38 760 (44.8) Number of influenza shots, mean (SD)0.7 (0.59)0.7 (0.61)0.7 (0.60)0.7 (0.61)0.7 (0.60)0.6 (0.60) Number of endocrinology visits, mean (SD)12.3 (8.39)10.9 (7.72)12.4 (8.99)10.5 (8.00)10.6 (7.49)11.0 (16.27) Number of HBA1C tests, mean (SD)3.0 (1.73)2.9 (1.57)3.1 (1.79)2.6 (1.60)2.8 (1.56)2.6 (1.58) Number of lipid tests, mean (SD)1.8 (1.33)1.9 (1.29)1.8 (1.28)1.7 (1.29)1.8 (1.25)1.7 (1.33) Number of outpatient visits, mean (SD)12.3 (8.37)10.9 (7.72)12.4 (8.80)10.5 (8.00)10.7 (7.49)11.0 (8.21) Number of hospital admissions, mean (SD)0.3 (0.81)0.2 (0.69)0.3 (0.84)0.5 (1.10)0.3 (0.72)0.7 (1.76) Number of days in hospital, mean (SD)1.9 (6.07)1.4 (5.45)1.9 (6.09)3.5 (10.09)1.4 (5.46)5.5 (15.52) Number of ED visits, mean (SD)0.9 (1.74)0.6 (1.47)0.9 (1.87)1.1 (2.11)0.7 (1.50)1.5 (3.40) Low-income subsidy in index month11 392 (31.4)11 645 (31.3)11 306 (31.0)33 685 (40.0)5762 (26.2)36 975 (42.7)*The comparisons were defined by use of GLP-1RA/DPP-4i and PS-weighted comparator. Covariates were measured in the 12 months before cohort entry including the index date (100% of new users have the treatment at baseline). Initiation is defined as having no prescriptions of either drug class during the 12 months prior to initiation.†The size of the population for a specific drug differed across cohorts because of the requirement not to have been treated prior to the index date with the comparator drug class (online supplemental figure 1).‡Weighted by standardizing to the distribution in GLP-1RA/DPP-4i initiators by using weights of 1 for GLP-1RA/DPP-4i initiators and the odds of the estimated PS for comparator initiators.§Baseline GLP-1RA users were excluded from the DPP-4i versus. SGLT-2i analysis to achieve covariate balance.ED, emergency department; HbA1c, Hemoglobin A1c; NSAID, Non-Steroidal Anti-Inflammatory Drug; NTSR, numbers too small (<11) to report based on Centers for Medicare & Medicaid Services rules and data use agreement; PS, propensity scores; SGLT-2i, sodium-glucose cotransporter-2 inhibitors.SP110.1136/bmjdrc-2025-005090.supp1Supplementary dataSensitivity analyses To examine the robustness of our primary results to changes in study population and condition definitions, we performed several sensitivity analyses. First, we repeated the primary analysis using IT analysis (no censoring for drug discontinuation, switching, or augmentation). Note that for this analysis, we did not require part D data during follow-up and thus only censored for loss of parts A or B. Second, we repeated the primary analysis not requiring a second prescription for a study drug to enter the cohort. Third, we varied the grace period from 30 days to 15 and 60 days. Fourth, we varied the lag periods (both after initiation and after stopping) separately from 6 months to 0 and 12 months. Fifth, we constructed cohorts using only the first new-user period (ie, patients could only enter the cohort once). Sixth, we repeated the analysis restricting to patients with baseline metformin use. This approach has been shown to improve confounding control and covariate balance by restricting to populations that are using study drugs as second-line therapies following initial metformin use. 33 Seventh, we excluded short-acting GLP-1RA (lixisenatide and exenatide). Eighth, we repeated the analysis excluding patients on levothyroxine for postprocedural hypothyroidism, a history of thyroid nodule or biopsy of the thyroid at baseline. Last, we performed asymmetric trimming of PS (1%, 2.5%, and 5% cut points) to assess the significance of any populations treated contrary to expectation (ie, populations treated with incretin drugs rather than the comparator despite low PS and vice versa) and the effect they have on the overall weighting and the effect measure estimate.34Results Study population Of 435 420 initiators of at least one drug of interest, 252 629 patients met the eligibility criteria for inclusion in our analysis. We identified 73 388 new users of the GLP-1RA versus SGLT-2i, 106 274 new users of DPP-4i versus SGLT-2i, and 191 143 new users of GLP-1RA versus DPP-4i. Among GLP-1RA initiators, 37.0% initiated dulaglutide (n=14 563), 33.3% liraglutide (n=13 117), 13.3% exenatide (n=5231), 5.9% semaglutide (n=2298), 1.1% lixisenatide (n=590), and 1.1% albiglutide (n=427). The crude and weighted baseline covariate distributions of all three cohorts are shown in table 1 and online supplemental table 1. Prior to weighting, new users of GLP-1RA were more likely than SGLT-2i initiators to have diabetes complications, thyroid disorders, and other comorbidities such as codes for obesity, chronic kidney disease, and congestive heart failure. DPP-4i initiators were generally similar to SGLT-2i initiators with regard to baseline comorbidities but more likely than SGLT-2i initiators to have codes for obesity. After PS-weighting, all measured patient characteristics were balanced within each cohort as shown by the SAMDs, evincing that we successfully balanced all measured covariates, controlling for confounding by these covariates.Incretin-based therapies and TC The results of the primary as-treated analysis for TC risk are presented in tables 2 and 3. The median duration of follow-up time on treatment ranged from 0.82 to 1.15 years across the three comparison cohorts. Despite the relatively short median follow-up, 11 048–48 644 patients contributed follow-up time beyond 1.5 years across cohorts, allowing for reliable estimation of longer-term risk under the assumption of proportional hazards. The 3-year crude cumulative incidence (risk) of TC ranged from 1 per 10 000 to 17 per 10 000. The cumulative incidence differences were −16 per 10 000 for GLP-1RAs versus SGLT-2i, −1 per 10 000 for DPP-4i versus SGLT-2i, and −5 per 10 000 for GLP-1RAs versus DPP-4i. After adjusting for confounding, the adjusted cumulative incidences (aRDs) for GLP-1RAs compared with SGLT-2i (aRD = −23 per 10 000) or DPP-4i (aRD = −7 per 10 000) did not suggest GLP-1RAs were associated with an increased risk of TC. DPP-4i was also not associated with an increased risk of TC compared with SGLT-2i (aRD = −2 per 10 000). We present weighted Kaplan-Meier curves for GLP-1RA/DPP-4i and comparators in figure 1. Potential indicators of thyroid screening within 6 months of treatment initiation were slightly more frequent among GLP-1RA/DPP-4i initiators compared with SGLT-2i initiators (GLP-1RA: 39.28% vs SGLT-2i: 36.63%; DPP-4i: 37.86% vs SGLT-2i: 34.80%) and were comparable between GLP-1RA initiators (39.17%) and DPP-4i initiators (38.39%) (online supplemental table 6).Figure 1SMR-weighted Kaplan-Meier plots of thyroid cancer. (A) GLP-1RA versus SGLT-2i cohort. RD −0.23% (95% CI −0.51% to 0.04%). (B) DPP-4i versus SGLT-2i cohort. RD −0.02% (95% CI −0.17% to 0.13%). (C) GLP-1RA versus DPP-4i cohort. RD −0.07% (95% CI −0.11% to −0.03%. SMR weights create a pseudopopulation of the untreated (comparators: SGLT-2i or DPP-4i), which has the same covariate distribution as the treated (GLP-1RA/DPP-4i). Every patient receiving GLP-1RA/DPP-4i has a weight of 1, whereas every patient in the comparator group is weighted by PS/(1−PS). The risks on the y-axis were obtained by an SMR-weighted Cox model (weighting comparator drug initiators by the PS odds (PS/(1−PS))). RD treats comparators as reference, and adjusted RD<0 indicates a lower risk for GLP-1RA/DPP-4i. DPP-4i, dipeptidyl peptidase-4 inhibitor; GLP-1RA, glucagon-like peptide-1 receptor agonist; PS, propensity scores; RD, risk difference; SGLT-2i, sodium-glucose cotransporter-2 inhibitor; SMR, standardized morbidity ratio.Table 23-year crude and adjusted RDs for thyroid cancer associated with incretin-based therapies compared with therapeutic alternative*ComparisonCohortPatients, n3-year crude RD per 10 000 (95% CI)Adjusted 3-year cumulative incidence per 10,000†Adjusted 3-year RD per 10 000 (95% CI)†GLP-1RA versus SGLT-2iGLP-1RA36 238−16 (−30 to 1)NTS‡−23 (−51 to 4)SGLT-2i37 1500 (reference)250 (reference)DPP-4i versus SGLT-2iDPP-4i84 283−0.1 (−11 to 11)6−2 (−17 to 13)SGLT-2i21 9910 (reference)80 (reference)GLP-1RA versus DPP-4iGLP-1RA33 771−5 (−7 to −3)NTS‡−7 (−11 to –3)DPP-4i157 3720 (reference)70 (reference)*Analysis based on as-treated exposure definition; latency period is 180 days; 3-year crude cumulative incidence is not shown because numbers were too small (<11) to report based on Centers for Medicare & Medicaid Services rules and data use agreement.†Adjusted RDs were estimated using propensity scores (PS), including all covariates in table 1 and PS odds weighting of comparator cohort (reference).‡Numbers too small to report per Medicare data use agreement.DPP-4i, dipeptidyl-peptidase-4 inhibitors; GLP-1RA, glucagon-like peptide 1 receptor agonists; RDs, risk differences; SGLT-2i, sodium-glucose cotransporter-2 inhibitors.Table 3Crude and adjusted HRs for TC associated with incretin-based therapies compared with therapeutic alternative*ComparisonCohortPatients, nMedian duration (years) of follow-up (IQR)Person-yearsTC incidence per 100 000 person-yearsCrude HR (95% CI)Adjusted† HR (95% CI)GLP-1RA versus SGLT-2iGLP-1RA36 2380.82 (0.54–1.42)21 081NTS‡0.33 (0.09 to 1.17)0.26 (0.06 to 1.08)SGLT-2i37 1500.86 (0.57–1.55)23 67946.45 (23.19–83.12)1.00 (reference)1.00 (reference)DPP-4i versus SGLT-2iDPP-4i84 2830.99 (0.59–1.74)58 06131.00 (18.37–49.00)1.22 (0.36 to 4.15)1.36 (0.29 to 6.34)SGLT-2i21 9910.82 (0.55–1.40)13 63522.00 (4.54–64.30)1.00 (reference)1.00 (reference)GLP-1RA versus DPP-4iGLP-1RA33 7710.83 (0.55–1.45)19 745NTS‡0.25 (0.03 to 1.83)0.14 (0.02 to 1.10)DPP-4i157 3721.15 (0.66–2.22)111 87230.39 (21.05–42.47)1.00 (reference)1.00 (reference)*Analysis based on as-treated exposure definition; latency period is 180 days; number of TC events is not shown because numbers were too small (<11) to report based on Centers for Medicare & Medicaid Services rules and data use agreement.†Adjusted HRs were estimated using propensity scores (PS), including all covariates in table 1 and PS odds weighting of comparator cohort (reference).‡Numbers too small to report per Medicare data use agreement.DPP-4i, dipeptidyl-peptidase-4 inhibitors; GLP-1RA, glucagon-like peptide 1 receptor agonists; SGLT-2i, sodium-glucose cotransporter-2 inhibitors; TC, thyroid cancer.The results for the secondary analyses based on the three additional algorithms for defining TC (online supplemental tables 2–4) were all aligned with the primary results with effect estimates being null to slightly protective across all comparison cohorts, although imprecise.Sensitivity analyses Overall, the results of the sensitivity analyses were consistent with our primary and secondary analyses ( online supplemental tables 7–15). We observed slightly more pronounced decreased risks in the GLP-1RA versus SGLT-2i comparison using a 60-day grace period (aRD = −21 per 10 000) and a 0-day lag period after the index date (aRD = −21 per 10 000). Similarly, we observed consistently slightly decreased risks in the GLP-1RA versus DPP-4i comparison using 15-day and 60-day grace periods (aRD = −6 per 10 000), restricting to patients on metformin at baseline (aRD = −6 per 10 000), restricting to patients on long-acting GLP-1RA (aRD = −7 per 10 000), accounting for death as a competing event (aRD = −5 per 10 000) and for all trimmed populations (aRD = −7 per 10 000 to −8 per 10 000). All the results of the primary, secondary, and sensitivity analyses are summarized in figure 2.Figure 2Forest plot summary of RDs per 10 000 for primary, secondary, and sensitivity analyses. All analysis based on as-treated exposure definition except initial treatment analysis; latency period is 180 days; RDs were estimated using PS, including all covariates in table 1 and PS odds weighting of comparator cohort (reference). GLP-1RA, glucagon-like peptide-1 receptor agonist; PS, propensity scores; RD, risk difference.Discussion Our study of US Medicare beneficiaries with T2DM suggests that the initiation of incretin-based therapies does not increase the 3-year risk of TC versus SGLT-2 inhibitors, with estimates being null to slightly protective across comparison cohorts. However, the results are too imprecise to exclude a moderate increase in risk. These findings align with prior experimental and non-experimental studies as well as meta-analyses that similarly found no relevant increase in TC risk with the use of incretin-based therapies. 9 10 12 13 16–18 This study contributes to the literature as one of the few US-based investigations of TC risk associated with incretin-based therapies. Unlike most prior studies that report HRs, our analysis provides both crude and adjusted cumulative incidence differences (aRDs), offering absolute measures of risk that are directly relevant for clinical decision-making. Our study accounts for death as a competing risk, ensuring valid and reliable risk estimation. By using a large, routinely collected US administrative claims database, our study compensated for the lack of statistical power of the experimental studies to detect rare cancer events such as TC.Our findings are consistent with recent, well-conducted studies by Pasternak et al and Baxter et al, but offer unique enhancements. Pasternak et al’s Scandinavian study (over 200 000 GLP-1RA users, 4-year mean follow-up) also found no increased TC risk versus SGLT-2i (HR 0.93, 95% CI 0.66 to 1.31).17 Our study adds absolute risk differences (eg, aRD −23 per 10 000), improving clinical interpretability. Similarly, Baxter et al’s multisite study showed no increased risk versus DPP-4i (HR 1.00, 95% CI 0.79 to 1.27) using an intention-to-treat approach.19 We advance this with an as-treated approach to ensure patients were truly exposed, extensive sensitivity analyses, and pairwise comparisons (GLP-1RA/DPP-4i vs SGLT-2i), enhancing robustness and nuance.Our results, however, contrast several studies—a few observational studies,8 14 15 20 a meta-analysis of experimental studies,11 and preclinical studies in rodents that suggested a potential TC link with GLP-1RA use.4–6 Several possible reasons may explain the inconsistency between our results and the findings from the nonexperimental studies by Bezin et al, Liang et al, Dore et al and Brito et al. Bezin et al conducted a nested case–control study using the French health insurance claims database and reported an increased risk of TC with GLP-1RAs (adjusted HR 1.46, 95% CI 1.23 to 1.74).8 However, this study did not employ an active comparator, which may have introduced confounding by indication. The rationale behind choosing an active comparator is to minimize the impact of confounding by indication and other unmeasured patient characteristics (such as healthy initiator bias or frailty).35 Further, Bezin et al observed an increased risk with a 6-month lag period, which would imply a very short combined induction and latency period for GLP-1RA-induced TCs or may be indicative of a detection bias.36 In contrast, we found no increased risk with a 6-month lag period, and our sensitivity analyses using a 12-month lag period also did not indicate an increased risk of TC, although these results were imprecise. Liang et al performed a retrospective cohort study, supplemented with a nested case–control study using administrative databases from commercial health plans in the US and reported increased risk of TC with exenatide versus other antidiabetic drugs (OADs) (adjusted HR 1.46, 95% CI 0.98 to 2.19).14 Similarly, Dore et al also conducted a cohort study using commercial health insurance claims data and reported increased risk of TC with exenatide versus metformin or glyburide (adjusted RR 1.40, 95% CI 0.80 to 2.40). However, these studies did not employ suitable active comparators. Metformin is often prescribed as first-line therapy in T2DM management, with glyburide as an add-on.37 Thus, comparing exenatide, a GLP-1RA, to metformin or glyburide introduces heterogeneity in disease severity and patient characteristics, which may confound the observed association. OADs, on the other hand, comprise all antihyperglycemics except exenatide with differing mechanisms of action, treatment indications, and side effect profiles. Such heterogeneity in the comparator group complicates the interpretation of results. Notably, these prior studies included younger adults aged at least 18 years (median age<65 years). This difference may hint at potential effect measure modification by age in the relationship between GLP-1RA/DPP-4i and TC. Although Brito et al studied older adults, they did not account for induction and latent periods of cancer diagnosis and development, which might explain the observed increased risk with GLP-1RA use.23 Additionally, TC was defined using only diagnosis codes which likely decreased the specificity of the outcome.Strengths Our large observational cohort study used data that are representative of US older adults with T2DM to assess the effect of GLP-1RA/DPP-4i on the incidence of TC. Claims data have many advantages when studying the effects of drugs on cancer outcomes as they include complete longitudinal data on dispensed prescriptions and medical encounters and procedures. Another strength of our study is the use of an ACNU cohort design, which helps to reduce the potential for unmeasured confounding by indication by comparing GLP-1RA/DPP-4i initiators with patients initiating a guideline-recommended clinical alternative. Furthermore, we used rigorous statistical adjustment to minimize remaining imbalances between treatment comparison cohorts. Last, we conducted multiple sensitivity analyses to assess the possibility of residual confounding and obtained results consistent with our primary findings.Limitations Our study has limitations that warrant caution in interpreting the findings. The relatively short median duration on treatment (0.82–1.15 years) due to treatment switching or discontinuation during follow-up limits our ability to detect potential TC events associated with long-term use. It does reflect, however, real use patterns, that is, only a few patients with diabetes take incretin drugs over prolonged periods. In settings with short duration on treatment with individual drugs and high frequency of treatment changes as in our study, it is difficult to definitively attribute cancer outcomes to specific treatments without making strong assumptions about induction and latency periods. Notably, results from a sensitivity analysis using the IT approach led to a median follow-up range of 1.42–2.78 years. The IT approach is, however, not ideal for studying safety outcomes as it disregards treatment changes (switching and discontinuation) during follow-up and thus tends to bias results towards the null. Our reliance on claims data to define TC may have introduced outcome misclassification. To address this, we conducted sensitivity analyses using alternative claims-based algorithms that incorporated combinations of diagnostic codes and non-surgical treatment procedures to derive highly sensitive outcome definitions and others that balanced sensitivity and specificity. Results from these analyses were consistent with the primary analysis, requiring thyroidectomy (informed by the American Thyroid Association’s recommendation of surgery as the first step in treatment for all types of TC), 38 suggesting that potential misclassification did not meaningfully affect our findings. Differential intensity of medical surveillance may have also introduced detection bias. GLP-1RA users may have had more frequent healthcare encounters, particularly in the context of weight management or cardiometabolic monitoring, increasing the chance of incidental TC detection through imaging or laboratory workup. Although we attempted to assess this by comparing indicators of TC screening intensity across groups, residual differences in diagnostic surveillance may persist. Despite extensive covariate balance achieved via SMR weighting, we cannot exclude residual confounding from unmeasured or poorly measured variables. Claims data lack information on important clinical risk factors like thyroid hormone levels, thyroid nodule characteristics, or family history of thyroid malignancy. These unmeasured factors may have influenced treatment decisions, particularly in light of the FDA black box warning for GLP-1RA regarding a potential risk of medullary thyroid carcinoma. Clinicians may have steered patients perceived to be at higher baseline risk for TC based on, for example, a family history, away from initiating GLP-1RA, leading to a lower underlying risk of TC in the GLP-1RA group, potentially masking an adverse effect. Further, our study reflects real-world patterns of GLP-1RA use among older adults, and the results may thus not be generalizable to younger populations, among whom indications, patterns of use, adherence, and risk profiles may differ.39 Our findings largely reflect use of dulaglutide and liraglutide, which together comprised over two-thirds of GLP-1RA initiations. While caution is warranted in generalizing results to recently approved agents like tirzepatide and semaglutide, the consistency in pharmacologic mechanism supports the plausibility of a class effect. Lastly, the lack of cancer histology information in claims data precluded subgroup analyses by TC histological subtype. This limitation is particularly relevant given that preclinical concerns have primarily focused on medullary thyroid carcinoma (MTC), a rare subtype. Even if histology data had been available, however, the number of incident TC cases in our study was too small to permit analyses by subtype. Any potential signal specific to MTC is likely to be obscured by non-medullary TCs. Thus, our findings cannot rule out a subtype-specific effect, particularly for MTC, which remains an important focus for future research using linked datasets with histopathology data.Future research should prioritize longer follow-up times and linkage with cancer registries for definitive cancer details and electronic health records to capture additional clinically relevant covariate data. Also, further investigations with larger cohorts to allow precise estimates of effects on medullary TC would enhance understanding of the potential differential impact of incretin-based therapies across cancer subtypes.Conclusion Our population-based, ACNU cohort study of US older adults with diabetes suggests that the real-world use of incretin-based treatments over a period of 0.82–1.15 years does not appear to increase the risk of TC compared with alternative antihyperglycemic treatments.