ABSTRACT
Objective
To examine the effect of second‐line treatments (insulin secretagogues, thiazolidinediones, glucagon‐like peptide‐1 (GLP‐1) receptor agonists, dipeptidyl peptidase‐4 (DPP‐4) inhibitors or sodium‐glucose co‐transporter 2 (SGLT‐2) inhibitors) on time to insulin initiation among patients with type 2 diabetes.
Research Design and Methods
We conducted a retrospective cohort study using the Clinical Practice Research Datalink (CPRD) Aurum. Initiation of metformin monotherapy (1998–2021) defined base cohort entry, and initiation of second‐line treatment (2013–2021) defined study cohort entry (time zero). After propensity score trimming, we applied inverse probability of treatment weighted Cox models to estimate the association between second‐line agents and time to insulin initiation. The secondary outcome was time to treatment modification, defined as the addition of or switch to another antidiabetic drug class.
Results
Our analytic cohort included 64 404 patients; 44% initiated DPP‐4 inhibitors, 36% insulin secretagogues, 18% SGLT‐2 inhibitors, 1% thiazolidinediones and 3% GLP‐1 receptor agonists. Over a mean follow‐up of 2.9 years, initiation of TZDs, DPP‐4 inhibitors, or SGLT‐2 inhibitors was associated with a lower risk of insulin initiation than initiation of insulin secretagogues (HR [95% CI]: 0.59 [0.43–0.83], 0.75 [0.69–0.80] and 0.62 [0.56–0.68], respectively). For treatment modification, the risk was higher among TZD and DPP‐4 inhibitor initiators, and slightly lower among SGLT‐2 inhibitor initiators and GLP‐1 receptor agonist initiators than among insulin secretagogue initiators.
Conclusion
This study provides real‐world evidence that TZDs, DPP‐4 inhibitors and SGLT‐2 inhibitors may offer an advantage over insulin secretagogues in delaying insulin initiation after metformin monotherapy.
Keywords: insulin initiation, population‐based cohort study, second‐line antidiabetic medication, type 2 diabetes, UK
Twitter Summary
In a large T2D cohort, thiazolidinediones, DPP‐4 inhibitors or SGLT‐2 inhibitors as second‐line therapy delayed insulin initiation versus insulin secretagogues.
1. Introduction
Type 2 diabetes accounts for over 90% of diabetes cases and is currently the 8th leading cause of disease burden [1]. Medication regimens are tailored to maintain glycemic control and preventing microvascular and macrovascular complications. Over the past two decades, several new classes of glucose‐lowering agents have emerged as recommended second‐line therapies, including glucagon‐like peptide‐1 (GLP‐1) receptor agonists, dipeptidyl peptidase‐4 (DPP‐4) inhibitors and sodium–glucose cotransporter‐2 (SGLT‐2) inhibitors [2, 3, 4]. When these medications fail to achieve glycemic targets, basal insulin is recommended [3]. Insulin initiation reflects the progressive beta‐cell function decline [5] and is associated with significant patient burden, including daily injections, higher treatment costs, hypoglycemia and potential weight gain, which may negatively affect patients' quality of life [6]. Therefore, delaying insulin initiation should be considered when evaluating second‐line treatment options.
A series of meta‐analyses of randomized controlled trials [7, 8] and real‐world observational studies [9, 10] have shown that most glucose‐lowering drug classes reduced HbA1c when added to metformin, although the magnitude of reduction differed across classes. However, these studies primarily focused on short‐term glycemic outcomes or evaluated time to insulin initiation in first‐line therapy settings. Few studies [11, 12, 13, 14, 15, 16, 17] have examined their impact on delaying insulin initiation when used as second‐line treatments and were limited to within‐class comparisons [11, 14] or used heterogeneous outcome definitions [12, 13]. To date, no study has simultaneously compared all available second‐line treatments in relation to time to insulin initiation.
Our study aimed to examine the effect of second‐line antidiabetic medications (insulin secretagogues, TZDs, GLP‐1 receptor agonists, DPP‐4 inhibitors and SGLT‐2 inhibitors) on (1) time to insulin initiation and (2) time to treatment modification by applying a target trial emulation framework [18] using data from a contemporary population‐based cohort in the United Kingdom (UK).
2. Materials and Methods
2.1. Target Trial
We sought to emulate a hypothetical target trial that would enroll adults aged ≥ 18 years with type 2 diabetes who initiated metformin monotherapy as first‐line therapy and required an additional antidiabetic drug as second‐line therapy. In brief, eligible participants would be randomly assigned to initiate insulin secretagogues, TZDs, GLP‐1 receptor agonists, DPP‐4 inhibitors or SGLT‐2 inhibitors as second‐line therapy. Follow‐up would begin at treatment assignment. The primary outcome would be time to initiation of any insulin therapy, and the secondary outcome would be time to switching to or adding another antidiabetic drug class. Additional details on the target trial protocol and the corresponding components of its emulation using observational data are provided in Table S1.
2.2. Emulating Target Trial Using Observational Study
We emulated the target trial using data from Clinical Practice Research Datalink (CPRD) Aurum [19], a population‐based clinical database containing information on demographics, medical diagnoses, prescriptions, lifestyle variables (e.g., smoking), laboratory results and clinical measures for over 19 million patients seen in UK general practices, where most patients with type 2 diabetes receive their routine care. In CPRD Aurum, medical diagnoses are captured using SNOMED CT (UK edition), Read Version 2 and EMIS Web‐specific codes while prescriptions were coded based on Dictionary of Medicines and Devices. CPRD Aurum covers around 13% of the English population and is highly representative with respect to demographics.
The study protocol was approved by the CPRD's Independent Scientific Advisory Committee (protocol number: 21_000450) and the Research Ethics Board of the Jewish General Hospital in Montreal, Canada (reference number: 2022–2957).
2.3. Study Population
We first constructed a base cohort of all patients who received a metformin prescription between April 1, 1998 and March 31, 2021, with the first metformin prescription defining base cohort entry. To identify adults with newly treated type 2 diabetes on metformin monotherapy, we excluded patients aged < 18 years, with < 1 year of history in the database, with a prescription for any antidiabetic agents or insulin prior to the base cohort entry and those who received antidiabetic drug prescriptions other than metformin on the date of base cohort entry. We also excluded patients with gestational diabetes diagnosed 1 year before cohort entry and those with end‐stage renal disease on haemodialysis (a contraindication for many of the antidiabetic agents of interest in this study).
From the base cohort, we established a study cohort comprising patients who initiated insulin secretagogues, TZDs, GLP‐1 receptor agonists, DPP‐4 inhibitors, or SGLT‐2 inhibitors between January 1 2013 and March 31 2021. Only treatments with dosage information indicated for type 2 diabetes were included. For example, we excluded those receiving liraglutide at a dose of 3 mg for obesity treatment (subcutaneous, daily). The study period began in 2013 to coincide with the introduction of SGLT‐2 inhibitors in the UK, ensuring availability of all drug classes. The first prescription of a new antidiabetic class during this period defined study cohort entry, defined as time zero of our study.
We excluded patients who had initiated any of these antidiabetic agents before study cohort entry because we aimed to examine them as second‐line medications. We also excluded patients who initiated more than one of the second‐line treatments on the same day. Patients were followed until insulin initiation, censoring due to death, end of CPRD practice registration, or last data collection, or the end of the study period (March 31 2021), whichever occurred first.
2.4. Treatment Strategies
We used an approach that was analogous to an intention‐to‐treat approach in which the exposure was defined by the second‐line treatment received at the study cohort entry day and person‐time was classified into one of five mutually exclusive categories: (1) insulin secretagogues (with or without use of metformin); (2) TZDs (with or without use of metformin); (3) DPP‐4 inhibitors (with or without use of metformin); (4) GLP‐1 receptor agonists (with or without use of metformin); (5) SGLT‐2 inhibitors (with or without use of metformin).
We also adopted a per‐protocol approach in which the exposure was defined by the second‐line treatment received at the study cohort entry day and censored at the time when they discontinued their current second‐line antidiabetic agents, allowing a 30‐day grace period beyond the estimated treatment duration or when they initiated a new class of antidiabetics (other than insulin). Per‐protocol approach was only applied to primary outcome analysis.
2.5. Outcomes
The primary outcome was time to initiation of any insulin therapy since study cohort entry. The secondary outcome was time to treatment modification, defined as initiation of a new antidiabetic class different from the index class initiated at study cohort entry, either as a switch or an add‐on. The date of first relevant prescription defined the event date.
2.6. Potential Confounders
Several potential confounders were measured at study cohort entry, including demographic characteristics (age, sex), year of cohort entry, prescription drug use (angiotensin‐converting enzyme inhibitors/angiotensin receptor blockers, statins, beta blockers, calcium channel blockers; during the year prior to cohort entry). Systolic and diastolic blood pressure (last reading before the study cohort entry and recorded in the previous year), HbA1c (most recent measure recorded in the previous year), lifestyle variables (alcohol use disorder, smoking and body mass index (BMI); most recent measure recorded in the previous 5 years). Since renal function is an important factor in treatment decisions, we calculated the estimated glomerular filtration rate (eGFR) using the most recent recorded serum creatinine measurement from the previous year. In addition, some clinical conditions (dementia, hypoglycaemia, hyperlipidaemia, hypertension) were also important factors for treatment decisions. To assess the severity of diabetes, we measured duration of treated diabetes (time since first metformin prescription to study cohort entry), microvascular (neuropathy, retinopathy, nephropathy) and macrovascular (myocardial infarction, stroke, peripheral arterial disease, heart failure) complications associated with diabetes. We also measured the number of general practice visits in the year before cohort entry as proxies for overall health.
2.7. Statistical Analysis
We described baseline characteristics at cohort entry by second‐line antidiabetic drug class. Categorical variables were summarized as counts (percentages) and continuous variables as means (standard deviations). Covariate imbalance relative to insulin secretagogues was assessed using absolute standardized differences, with values > 0.1 considered meaningful [20]. Crude incidence rates and 95% CIs were calculated assuming Poisson distributions.
For the primary outcome, we compared time to insulin initiation across second‐line treatments, assuming that all patients were eligible for each option. Our target estimands were four pairwise average treatment effects comparing each alternative treatment with insulin secretagogues as the reference group [21]. For example, we compared time to insulin initiation had all patients received DPP‐4 inhibitors versus insulin secretagogues after metformin monotherapy.
We estimated propensity scores for each second‐line treatment using multinomial logistic regression, with treatment as the dependent variable and all baseline covariates as independent variables [22]. Continuous covariates (age, BMI, blood pressure, eGFR, GP visits) were modelled using restricted cubic splines with 4 knots, while categorical covariates were modelled based on the categories listed in Table 1. In examining propensity score overlap plots, we found insufficient overlap, suggesting certain treatments may only be given to certain subgroups of patients in our cohort (Figure S1). To ensure that patients had clinical equipoise for receiving all second‐line antidiabetic drugs, we applied the Stürmer trimming approach to restrict the cohort to patients with clinical equipoise across all five treatments [23]. For each treatment, the threshold was defined as the 2nd percentile of the propensity‐score distribution among patients who received that treatment (based on 0.1/5 = 0.02 for five treatment groups). Patients were retained only if their propensity scores for all five treatments exceeded the corresponding thresholds.
TABLE 1.
Characteristics of patients with type 2 diabetes in the final analytic cohort (before weighting), by second‐line antidiabetic medication initiated at study cohort entry, in the United Kingdom between 2013 and 2021.
| Characteristics | Total (n = 64 404) | IS (n = 21 091) | TZD (n = 903) | DPP‐4 inhibitors (n = 27 170) | GLP‐1 receptor agonists (n = 1907) | SGLT‐2 inhibitors (n = 13 333) | Absolute standardized difference | |||
|---|---|---|---|---|---|---|---|---|---|---|
| TZD vs. IS | DPP‐4 inhibitors vs. IS | GLP‐1 receptor agonists vs. IS | SGLT‐2 inhibitors vs. IS | |||||||
| Male, n (%) | 36 421 (56.6) | 11 855 (56.2) | 549 (60.8) | 15 439 (56.8) | 866 (45.4) | 7712 (57.8) | 0.09 | 0.01 | 0.22 | 0.03 |
| Age, (years), mean (SD) | 55.5 (10.3) | 54.7 (10.3) | 55.0 (9.7) | 56.7 (10.3) | 52.8 (11.0) | 54.7 (10.0) | 0.03 | 0.19 | 0.18 | 0.00 |
| Time from first metformin prescription to second‐line treatment initiation (years), mean (SD) | 3.4 (3.1) | 3.1 (3.0) | 3.4 (3.0) | 3.6 (3.1) | 3.2 (3.1) | 3.3 (3.0) | 0.09 | 0.18 | 0.03 | 0.07 |
| Year at study cohort entry, n (%) | ||||||||||
| 2013 | 51 (0.1) | 20 (0.1) | S | 22 (0.1) | S | S | 0.03 | 0.01 | 0.05 | 0.04 |
| 2014 | 6204 (9.6) | 3247 (15.4) | 118 (13.1) | 2215 (8.2) | 216 (11.3) | 408 (3.1) | 0.07 | 0.23 | 0.12 | 0.44 |
| 2015 | 8410 (13.1) | 3819 (18.1) | 156 (17.3) | 3137 (11.5) | 252 (13.2) | 1046 (7.8) | 0.02 | 0.19 | 0.14 | 0.31 |
| 2016 | 9037 (14.0) | 3190 (15.1) | 158 (17.5) | 3992 (14.7) | 251 (13.2) | 1446 (10.8) | 0.06 | 0.01 | 0.06 | 0.13 |
| 2017 | 10 223 (15.9) | 3106 (14.7) | 123 (13.6) | 4718 (17.4) | 276 (14.5) | 2000 (15.0) | 0.03 | 0.07 | 0.01 | 0.01 |
| 2018 | 10 343 (16.1) | 2718 (12.9) | 116 (12.8) | 4802 (17.7) | 271 (14.2) | 2436 (18.3) | 0.00 | 0.13 | 0.04 | 0.15 |
| 2019 | 10 861 (16.9) | 2694 (12.8) | 133 (14.7) | 4651 (17.1) | 319 (16.7) | 3064 (23.0) | 0.06 | 0.12 | 0.11 | 0.27 |
| 2020 | 8220 (12.8) | 2016 (9.6) | 87 (9.6) | 3256 (12.0) | 277 (14.5) | 2584 (19.4) | 0.00 | 0.08 | 0.15 | 0.28 |
| 2021 | 1055 (1.6) | 281 (1.3) | S | S | S | S | 0.02 | 0.01 | 0.06 | 0.09 |
| Medications use, n (%) a | ||||||||||
| Angiotensin converting Enzyme | 25 265 (39.2) | 8132 (38.6) | 364 (40.3) | 10 952 (40.3) | 759 (39.8) | 5058 (37.9) | 0.04 | 0.04 | 0.03 | 0.01 |
| Angiotensin receptor blockers | 8917 (13.8) | 2730 (12.9) | 132 (14.6) | 4034 (14.8) | 269 (14.1) | 1752 (13.1) | 0.05 | 0.06 | 0.03 | 0.01 |
| Statins | 45 169 (70.1) | 14 426 (68.4) | 661 (73.2) | 19 756 (72.7) | 1202 (63.0) | 9124 (68.4) | 0.11 | 0.10 | 0.11 | 0.00 |
| Beta blocker | 11 683 (18.1) | 3798 (18.0) | 117 (13.0) | 5247 (19.3) | 338 (17.7) | 2183 (16.4) | 0.14 | 0.03 | 0.01 | 0.04 |
| Calcium channel blockers | 18 105 (28.1) | 5750 (27.3) | 247 (27.4) | 7994 (29.4) | 534 (28.0) | 3580 (26.9) | 0.00 | 0.05 | 0.02 | 0.01 |
| Smoking status at study cohort entry, n (%) | ||||||||||
| Never smoking | 27 997 (43.5) | 8963 (42.5) | 386 (42.7) | 11 940 (43.9) | 789 (41.4) | 5919 (44.4) | 0.01 | 0.03 | 0.02 | 0.04 |
| Former smoking | 17 880 (27.8) | 5669 (26.9) | S | 7602 (28.0) | 583 (30.6) | 3777 (28.3) | 0.02 | 0.03 | 0.08 | 0.03 |
| Current smoking | 17 924 (27.8) | 6208 (29.4) | 263 (29.1) | 7412 (27.3) | 516 (27.1) | 3525 (26.4) | 0.07 | 0.05 | 0.05 | 0.07 |
| Unknown | 603 (0.9) | 251 (1.2) | S | 216 (0.8) | 19 (1.0) | 112 (0.8) | 0.07 | 0.04 | 0.02 | 0.04 |
| Body mass index (kg/m2) b , n (%) | ||||||||||
| Underweight (< 18.5) | 37 (0.1) | 20 (0.1) | 0 (0.0) | 17 (0.1) | 0 (0.0) | 0 (0.0) | 0.04 | 0.01 | 0.04 | 0.04 |
| Normal weight (18.5–24.9) | 2592 (4.0) | 960 (4.6) | 34 (3.8) | 1176 (4.3) | 11 (0.6) | 411 (3.1) | 0.04 | 0.01 | 0.25 | 0.08 |
| Overweight (25–29.9) | 14 339 (22.3) | 4983 (23.6) | 216 (23.9) | 6360 (23.4) | 86 (4.5) | 2694 (20.2) | 0.01 | 0.01 | 0.57 | 0.08 |
| Obesity (≥ 30) | 47 436 (73.7) | 15 128 (71.7) | 653 (72.3) | 19 617 (72.2) | 1810 (94.9) | 10 228 (76.7) | 0.01 | 0.01 | 0.65 | 0.11 |
| Systolic blood pressure, (mmHg) c , n (%) | ||||||||||
| < 140 | 46 525 (72.2) | 15 273 (72.4) | 656 (72.6) | 19 750 (72.7) | 1373 (72.0) | 9473 (71.0) | 0.01 | 0.01 | 0.01 | 0.00 |
| ≥ 140 | 17 879 (27.8) | 5818 (27.6) | 247 (27.4) | 7420 (27.3) | 534 (28.0) | 3860 (29.0) | 0.01 | 0.01 | 0.01 | 0.00 |
| Diastolic blood pressure, (mmHg) c , n (%) | ||||||||||
| < 90 | 55 751 (86.6) | 18 144 (86.0) | 767 (84.9) | 23 889 (87.9) | 1602 (84.0) | 11 349 (85.1) | 0.03 | 0.06 | 0.06 | 0.03 |
| ≥ 90 | 8653 (13.4) | 2947 (14.0) | 136 (15.1) | 3281 (12.1) | 305 (16.0) | 1984 (14.9) | 0.03 | 0.06 | 0.06 | 0.03 |
| Estimated glomerular filtration rate (eGFR), (mL/min/1.73m2) c , n (%) | ||||||||||
| < 60 | 2148 (3.3) | 601 (2.8) | 24 (2.7) | 1242 (4.6) | 96 (5.0) | 185 (1.4) | 0.01 | 0.09 | 0.11 | 0.10 |
| ≥ 60 | 62 256 (96.7) | 20 490 (97.2) | 879 (97.3) | 25 928 (95.4) | 1811 (95.0) | 13 148 (98.6) | 0.01 | 0.09 | 0.11 | 0.10 |
| Haemoglobin A1c (HbA1c) (%) c , n (%) | ||||||||||
| ≤ 7 | 5961 (9.3) | 1784 (8.5) | 86 (9.5) | 2811 (10.3) | 269 (14.1) | 1011 (7.6) | 0.04 | 0.07 | 0.18 | 0.03 |
| 7–8 | 16 870 (26.2) | 4299 (20.4) | 228 (25.2) | 8616 (31.7) | 364 (19.1) | 3363 (25.2) | 0.12 | 0.26 | 0.03 | 0.12 |
| > 8 | 41 573 (64.6) | 15 008 (71.2) | 589 (65.2) | 15 743 (57.9) | 1274 (66.8) | 8959 (67.2) | 0.13 | 0.28 | 0.09 | 0.09 |
| Comorbidities d , n (%) | ||||||||||
| Alcohol use disorder | 6398 (9.9) | 2193 (10.4) | 101 (11.2) | 2571 (9.5) | 188 (9.9) | 1345 (10.1) | 0.03 | 0.03 | 0.02 | 0.01 |
| Dementia | 232 (0.4) | 71 (0.3) | S | 115 (0.4) | 8 (0.4) | 35 (0.3) | 0.00 | 0.01 | 0.01 | 0.01 |
| Hypertension | 34 981 (54.3) | 11 037 (52.3) | 478 (52.9) | 15 362 (56.5) | 1040 (54.5) | 7064 (53.0) | 0.01 | 0.09 | 0.04 | 0.01 |
| Hyperlipidemia | 11 413 (17.7) | 3598 (17.1) | 139 (15.4) | 5059 (18.6) | 284 (14.9) | 2333 (17.5) | 0.05 | 0.04 | 0.06 | 0.01 |
| Hypoglycaemia | 4098 (6.4) | 1384 (6.6) | 51 (5.6) | 1768 (6.5) | 151 (7.9) | 744 (5.6) | 0.04 | 0.00 | 0.05 | 0.04 |
| Diabetes complications d , n (%) | ||||||||||
| Diabetic neuropathy | 6836 (10.6) | 2303 (10.9) | 83 (9.2) | 3209 (11.8) | 168 (8.8) | 1073 (8.0) | 0.06 | 0.03 | 0.07 | 0.10 |
| Renal disease | 3671 (5.7) | 1110 (5.3) | 45 (5.0) | 1891 (7.0) | 117 (6.1) | 508 (3.8) | 0.01 | 0.07 | 0.04 | 0.07 |
| Retinopathy | 12 067 (18.7) | 3832 (18.2) | 152 (16.8) | 5369 (19.8) | 318 (16.7) | 2396 (18.0) | 0.04 | 0.04 | 0.04 | 0.01 |
| Myocardial infarction | 2546 (4.0) | 857 (4.1) | 20 (2.2) | 1148 (4.2) | 53 (2.8) | 468 (3.5) | 0.11 | 0.01 | 0.07 | 0.03 |
| Ischemic and hemorrhagic stroke | 2572 (4.0) | 852 (4.0) | 34 (3.8) | 1166 (4.3) | 70 (3.7) | 450 (3.4) | 0.01 | 0.01 | 0.02 | 0.04 |
| Peripheral vascular disease | 1696 (2.6) | 578 (2.7) | 16 (1.8) | 743 (2.7) | 46 (2.4) | 313 (2.3) | 0.07 | 0.00 | 0.02 | 0.03 |
| Heart failure | 1717 (2.7) | 608 (2.9) | 17 (1.9) | 769 (2.8) | 52 (2.7) | 271 (2.0) | 0.07 | 0.00 | 0.01 | 0.06 |
| Coronary artery disease | 7132 (11.1) | 2375 (11.3) | 75 (8.3) | 3268 (12.0) | 176 (9.2) | 1238 (9.3) | 0.10 | 0.02 | 0.07 | 0.07 |
| Numbers of GP visit a , mean (SD) | 35.2 (18.6) | 35.7 (19.3) | 33.8 (17.9) | 35.1 (18.5) | 40.6 (21.1) | 34.2 (17.3) | 0.05 | 0.06 | 0.16 | 0.15 |
Note: There were variables with missing values (% of missing values): BMI (1.6%), blood pressure (2.9%), estimated glomerular filtration rate (6.5%) and haemoglobin A1C (1.6%). We used the multiple imputation by chained equations method to impute missing values and generated five complete datasets. The results presented in this table were from one of randomly selected imputed databases. The final analytic cohort was obtained from the study cohort after applying propensity score trimming.
Abbreviations: DPP‐4, dipeptidyl peptidase‐4; GLP‐1, glucagon‐like peptide‐1; GP, general practitioner; IS, insulin secretagogues; S, based on data regulations for CPRD, for cell counts < 5, more than 1 cell needs to be suppressed to avoid being back calculated; SD, standard deviation; SGLT‐2, sodium‐glucose co‐transporter‐2; TZD, thiazolidinediones.
Assessed in the year prior to cohort entry.
Used most recent measure recorded within 5 years prior to cohort entry.
Used most recent measure recorded in the year prior to cohort entry.
Assessed in any time prior to cohort entry.
Unadjusted cumulative incidence curves were used to graphically depict time to insulin initiation by treatment groups. For the primary outcome, we estimated both intention‐to‐treat and per‐protocol effects using Inverse probability weighted (IPTW) Cox models [24] to estimate hazard ratios (HRs) and 95% CIs, with each treatment compared with insulin secretagogues. IPTWs were calculated as the inverse of the probability of the actual treatment received, based on the re‐estimated propensity scores in the trimmed cohort [25]. These weights were stabilized by multiplying by the marginal probability of receiving that treatment in the trimmed cohort. For secondary outcome analysis, we repeated the above analyses using time to treatment modification as the outcome.
Four sensitivity analyses were performed to assess the robustness of the results. First, to account for potentially informative censoring, we repeated our analyses using marginal structural Cox models weighted by the product of stabilized IPTW and inverse probability of censoring weights (IPCW) [26]. The IPCW denominator represented the probability of remaining uncensored conditional on baseline and monthly updated time‐varying covariates (age, BMI, clinical measurements and comorbidities), whereas the numerator was conditional on baseline covariates only and the extreme weights were truncated. Second, we calculated E‐values [27] to assess the minimum strength of association that an unmeasured confounder would need to have with both the treatment and the outcome to explain the observed associations. Third, although baseline HbA1c, BMI and eGFR were modelled using flexible functional forms in the propensity‐score model, imbalance remained after propensity‐score trimming and weighting. We therefore conducted a sensitivity analysis that additionally included these variables, modelled using flexible functional forms, in the weighted Cox models to address residual imbalance in prognostically important covariates [28]. Finally, because the limited numbers of GLP‐1 receptor agonist and TZD initiators may have impacted the composition of the trimmed analytic cohort, we repeated the analyses after excluding these two groups and re‐estimating the propensity‐score model and weights among insulin secretagogue, DPP‐4 inhibitor and SGLT‐2 inhibitor initiators.
Missing data were present for BMI (1.6%), blood pressure (2.9%), eGFR (6.5%) and HbA1c (1.6%). We used multiple imputation by chained equations to generate five complete datasets, assuming data were missing at random. Estimates were pooled using Rubin's rules [29].
All analyses were performed in SAS 9.4 (SAS Institute, Cary, NC) and R software version 3.6.
3. Results
3.1. Patient Characteristics
Our base cohort included 620 081 patients who initiated metformin during the study period; of these, 125 294 initiated a second‐line antidiabetic medication after 2012 and formed the study cohort. Most patients started DPP‐4 inhibitors (41%) or insulin secretagogues (40%), while 15% initiated SGLT‐2 inhibitors, 3% initiated TZDs and 1% initiated GLP‐1 receptor agonists (Table S2, Figure 1).
FIGURE 1.
Flowchart for analytic cohort selection of patients initiated second‐line antidiabetic medications in the United Kingdom, Clinical Practice Research Datalink and Hospital Episode Statistics databases, 2013–2021.
Before propensity‐score trimming, compared with insulin secretagogue initiators, DPP‐4 inhibitor initiators were older but had similar eGFR, whereas TZD initiators were younger, had higher eGFR, used statins more frequently and used beta blockers less frequently. GLP‐1 receptor agonist and SGLT‐2 inhibitor initiators were both younger and had higher diastolic blood pressure and eGFR, as well as a lower prevalence of dementia. GLP‐1 receptor agonist initiators also used statins less frequently and had a lower prevalence of hyperlipidemia, whereas SGLT‐2 inhibitor initiators used beta blockers less frequently. Across all four drug classes, patients entered the cohort in later calendar years, had higher BMI and lower baseline HbA1c compared to insulin secretagogue initiators. For diabetes‐related complications, DPP‐4 inhibitor initiators had similar prevalence, whereas TZD, GLP‐1 receptor agonist and SGLT‐2 inhibitor initiators had lower baseline prevalences (Table S2). After propensity score trimming, the final analytic cohort included 64 404 patients, with a similar distribution of second‐line treatment groups (a slight decrease in insulin secretagogue and increase in SGLT‐2 initiators). The propensity‐score overlap plots for the trimmed cohort showed improved overlap after trimming (Figure S2). Most individual covariates were well balanced after propensity‐score trimming and IPTW, although imbalance remained in BMI, HbA1c and eGFR for some treatment groups (Tables 1 and S3).
3.2. Primary Outcome
The mean follow‐up time for the primary outcome of insulin initiation was 2.9 years (standard deviation, 1.9). The incidence rate per 1000 person‐years was highest among GLP‐1 receptor agonist initiators (2.9, 95% CI, 2.5–3.4) and lowest among TZD (1.3, 95% CI, 1.0–1.8) and SGLT‐2 inhibitor initiators (1.3, 95% CI, 1.2–1.5) (Table 2 and Figure 2). The distributions of the reasons for censoring by treatment group for both analyses are presented in Table S4. In the intention‐to‐treat analyses, compared to insulin secretagogue initiators, initiation of TZDs, DPP‐4 inhibitors or SGLT‐2 inhibitors was associated with a decreased risk of initiating an insulin (HR [95% CI]: 0.59 [0.43, 0.83], 0.75 [0.69, 0.80], 0.62 [0.56, 0.68], respectively). While, initiating GLP‐1 receptor agonist or insulin secretagogue had similar risk of the primary outcome.
TABLE 2.
Risk of insulin initiation and treatment modification among second‐line antidiabetic medications initiated at cohort entry, in the United Kingdom between 2013 and 2021.
| Outcome | No. event | Person‐year of follow‐up | Incidence a (95% CI) | Crude hazard ratio (95% CI) | IPTW‐weighted hazard ratio b (95% CI) | IPTW*IPCW‐weighted hazard ratio c (95% CI) |
|---|---|---|---|---|---|---|
| Primary outcome: Time to insulin initiation (intention to treat analysis) | ||||||
| Insulin secretagogues | 1772 | 69 384 | 2.6 (2.4, 2.7) | Reference | Reference | Reference |
| Thiazolidinediones | 39 | 2995 | 1.3 (1.0, 1.8) | 0.50 (0.36, 0.69) | 0.59 (0.43, 0.83) | 0.60 (0.42, 0.85) |
| DPP‐4 inhibitors | 1197 | 79 986 | 1.5 (1.4, 1.6) | 0.59 (0.55, 0.63) | 0.75 (0.69, 0.80) | 0.77 (0.70, 0.84) |
| GLP‐1 receptor agonists | 156 | 5362 | 2.9 (2.5, 3.4) | 1.14 (0.97, 1.35) | 1.06 (0.88, 1.28) | 1.05 (0.83, 1.32) |
| SGLT‐2 inhibitors | 418 | 31 704 | 1.3 (1.2, 1.5) | 0.51 (0.46, 0.57) | 0.62 (0.56, 0.68) | 0.61 (0.53, 0.70) |
| Primary outcome: Time to insulin initiation (per‐protocol analysis) | ||||||
| Insulin secretagogues | 452 | 20 590 | 2.2 (2.0, 2.4) | Reference | Reference | Reference |
| Thiazolidinediones | S | S | 0.2 (0.1, 0.8) | 0.09 (0.02, 0.37) | 0.13 (0.03, 0.46) | 0.11 (0.03, 0.45) |
| DPP‐4 inhibitors | 210 | 30 703 | 0.7 (0.6, 0.8) | 0.33 (0.28, 0.38) | 0.44 (0.38, 0.52) | 0.43 (0.36, 0.52) |
| GLP‐1 receptor agonists | 39 | 1598 | 2.4 (1.8, 3.3) | 1.04 (0.75, 1.44) | 1.12 (0.79, 1.59) | 1.13 (0.73, 1.75) |
| SGLT‐2 inhibitors | 68 | 12 693 | 0.5 (0.4, 0.7) | 0.24 (0.18, 0.31) | 0.31 (0.24, 0.39) | 0.29 (0.21, 0.40) |
| Secondary outcome: Time to treatment modification (intention to treat analysis) | ||||||
| Insulin secretagogues | 10 267 | 46 499 | 22.1 (21.7, 22.5) | Reference | Reference | Reference |
| Thiazolidinediones | 425 | 1939 | 21.9 (19.9, 24.1) | 0.99 (0.90, 1.09) | 1.06 (0.96, 1.17) | 1.06 (0.95, 1.18) |
| DPP‐4 inhibitors | 12 476 | 52 428 | 23.8 (23.4, 24.2) | 1.06 (1.03, 1.09) | 1.18 (1.15, 1.21) | 1.18 (1.14, 1.21) |
| GLP‐1 receptor agonists | 745 | 3846 | 19.4 (18.0, 20.8) | 0.87 (0.80, 0.94) | 0.88 (0.81, 0.95) | 0.87 (0.78, 0.97) |
| SGLT‐2 inhibitors | 4357 | 23 555 | 18.5 (18.0, 19.1) | 0.82 (0.79, 0.85) | 0.88 (0.85, 0.91) | 0.87 (0.83, 0.90) |
Abbreviations: CI, confidence interval; DPP‐4, dipeptidyl peptidase‐4; GLP‐1, glucagon‐like peptide‐1; IS, insulin secretagogues; S, based on data regulations for CPRD, for cell counts < 5, more than 1 cell needs to be suppressed to avoid being back calculated; SGLT‐2, sodium‐glucose co‐transporter‐2; TZD, thiazolidinediones.
Incidence: per 1000 person‐year.
Covariates included in the propensity score model: sex, angiotensin converting enzyme, angiotensin receptor blockers, statins, beta blocker, calcium channel blockers, smoking, alcohol use disorder, dementia, hypertension, hyperlipidemia, hypoglycaemia, diabetic neuropathy, renal disease, retinopathy, myocardial infarction, ischemic and hemorrhagic stroke, peripheral vascular disease, heart failure, coronary artery disease; age, duration between first metformin and initiation of second line treatment, BMI, systolic blood pressure, diastolic bloop pressure, estimated glomerular filtration rate (eGFR), haemoglobin A1C (HbA1c), GP visits were included as continuous variable with restricted cubic splines.
Covariates included in the censoring model: same set of baseline covariates listed above and time‐varying covariate: age, BMI, clinical measurements and comorbidities with information updated every month.
FIGURE 2.
Cumulative incidence of insulin initiation by second‐line antidiabetic medications initiated at the study cohort entry. Intention to treat approach was used to categorize person times based on the second‐line antidiabetic initiated at the study cohort entry. S, based on data regulations for CPRD, for cell counts < 5, more than 1 cell needs to be suppressed to avoid being back calculated.
In the per‐protocol analyses, the patients were censored when they discontinued their current second‐line antidiabetic agents or when they initiated a new non‐insulin antidiabetic agent. As a result, the follow‐up period was shortened (mean of 1 year), leading to fewer observed events. We observed similar results as in intention‐to‐treat analyses but with stronger associations (Table 2). Compared with insulin secretagogue initiators, the point estimate for GLP‐1 receptor agonist initiators suggested a slightly higher risk of insulin initiation, although the estimate was imprecise and therefore inconclusive (HR, 1.12; 95% CI, 0.79–1.59).
3.3. Secondary Outcome
Among the five second‐line antidiabetic classes, DPP‐4 inhibitor initiators had the highest crude incidence of having treatment modification (23.8, 95% CI: 23.3, 24.2 per 1000 person‐years), while SGLT‐2 inhibitor initiators had the lowest (18.5, 95% CI: 18.0, 19.1 per 1000 person‐years) (Table 2, Figure S3). Figure S4 demonstrates the transition patterns from baseline second‐line antidiabetic medications to newly initiated third‐line treatments in our cohort. For example, patients initiating insulin secretagogues or SGLT‐2 inhibitors were more likely to initiate DPP‐4 inhibitors as third‐line treatment, whereas those starting on DPP‐4 inhibitors were more likely to initiate SGLT‐2 inhibitors.
Compared to insulin secretagogue initiators, those who initiated GLP‐1 receptor agonists or SGLT‐2 inhibitors had a decreased risk of treatment modification (HR [95% CI]: 0.88 [0.81, 0.95] and 0.88 [0.85, 0.91], respectively). In contrast, the point estimate suggested similar risk among TZD initiators and a slightly higher risk among DPP‐4 inhibitor initiators (HR [95% CI]: 1.06 [0.96–1.17] and 1.18 [1.15–1.21], respectively).
3.4. Sensitivity Analyses
When accounting for potential informative censoring using IPCW, we observed no meaningful differences compared to the main analysis results (Table 2). Figure S5 presents bias plots illustrating the strength of associations an unmeasured confounder would need to have with both the exposure and outcome to fully explain the observed associations. For example, an E‐value of 2.12 indicates that an unmeasured confounder would need to be associated with both a 2.12‐fold increased probability of initiating DPP‐4 inhibitors and a 2.12‐fold increased risk of insulin initiation to explain the observed HR of 0.59. E‐values ranged from 1.34 to 19.49; larger values indicate greater robustness to unmeasured confounding, whereas results with lower E‐values may be more susceptible to bias. Results from analyses that additionally included the imbalanced covariates in the outcome models were similar to those from the main IPTW‐weighted analysis. However, after further accounting for IPCW and outcome adjustment, the changes in the estimates were slightly larger, particularly for the GLP‐1 receptor agonist and SGLT‐2 inhibitor groups (Table S5). For treatment modification, the beneficial associations for GLP‐1 receptor agonists and SGLT‐2 inhibitors moved towards the null. Finally, when we restricted the study cohort by excluding the small GLP‐1 receptor agonist and TZD groups, the final trimmed analytic cohort increased by 6%. Propensity‐score overlap improved after trimming (Figures S6 and S7), and the covariates were balanced after trimming and weighting (Tables S6 and S7). The estimates for DPP‐4 inhibitors and SGLT‐2 inhibitors showed no meaningful changes (Table S8).
4. Discussion
In this population‐based cohort study of patients initiating second‐line antidiabetic medications after metformin monotherapy, we found that, compared to insulin secretagogue initiators, those who initiated TZDs, DPP‐4 inhibitors, or SGLT‐2 inhibitors had a 25%–40% lower risk of initiating insulin over a mean follow‐up of 2.9 years. Initiation of GLP‐1 receptor agonists or SGLT‐2 inhibitors was also associated with a reduced risk of discontinuing the current medication or initiating a new antidiabetic class, while TZD and DPP‐4 inhibitor initiators were at increased risk of doing so. These findings suggest that several second‐line antidiabetic agents can extend the duration of glycemic control before insulin is required, offering alternatives to the long‐standing insulin secretagogues.
Our findings on time to insulin initiation align with prior research. A population‐based study in Canada reported that TZDs added to metformin delayed insulin initiation compared to sulfonylureas [14], while a US‐based study found no meaningful difference [15]. Inzucchi et al. [16] reported that sitagliptin compared to sulfonylureas, was associated with a lower risk of insulin initiation over 6 years (HR: 0.76, 95% CI: 0.65–0.90). In contrast, a US electronic medical records‐based study [17] found that DPP‐4 inhibitors were associated with a longer mean time to insulin initiation (7.1 years) compared to GLP‐1 receptor agonists (6.6 years) or sulfonylureas (6.3 years); although, the analysis was descriptive and unadjusted. In a population‐based study in Japan [30], SGLT‐2 inhibitors were associated with a lower risk of insulin initiation than DPP‐4 inhibitors (HR 0.46, 95% CI: 0.28–0.74) over 1.5 years; however, treatment was not restricted to use to the second‐line setting.
Two observational studies have assessed persistence of second‐line antidiabetic agents. Mamza et al. [12], using UK THIN data (2007–2014), compared time to treatment intensification after metformin and found that DPP‐4 inhibitors had the shortest time to intensification, while TZDs showed the most durable glycemic response. Differences in study period may partly explain discrepancies with our findings, as our cohort began in 2013 when SGLT2 inhibitors became available, whereas the previous study largely preceded their uptake. Although our outcome definition did not distinguish these components, a recent US‐based study [31] compared within 1 year, switching, discontinuation and intensification and found compared to sulfonylureas, other second‐line agents were associated with lower risk of intensification but higher risk of switching and discontinuations, with GLP‐1 receptor agonists showing the strongest association.
Of note, SGLT‐2 inhibitors were associated with delayed insulin initiation and greater treatment durability. Beyond glycemic control, they provide cardiovascular and renal protection, including reductions in the risk of major adverse cardiovascular events, heart failure hospitalization and kidney disease progression [3, 32]. Importantly, these effects appear to be independent of baseline glycemic control [33]. GLP‐1 receptor agonists also provide cardiovascular and renal benefits, promote weight loss and achieve substantial reductions in HbA1c [3, 34]. In contrast, DPP‐4 inhibitors have not demonstrated comparable cardiovascular or renal benefits, which may have contributed to switching from DPP‐4 inhibitors to SGLT‐2 inhibitors in our study. The lower insulin initiation but higher risk of treatment modification among TZDs or DPP‐4 inhibitors initiators may also reflect other clinical considerations, including TZD‐related fluid retention and heart failure risk and the limited cardiorenal benefits of DPP‐4 inhibitors. These patterns may also reflect patient and clinician preferences and therapeutic inertia, including reluctance to initiate insulin [3, 35].
Although studies such as the GRADE trial [36] have demonstrated good glycemic efficacy of GLP‐1 receptor agonists, we found no meaningful reduction in insulin initiation but a lower risk of treatment modification compared with insulin secretagogues. These findings suggest that GLP‐1 receptor agonist initiators may have undergone fewer or later noninsulin treatment changes while still having a similar time to insulin during follow‐up. This difference from efficacy trials may partly reflect the outcomes assessed: GRADE defined treatment failure using an HbA1c threshold, whereas our outcomes captured real‐world prescribing decisions. During the study period, GLP‐1 receptor agonists were uncommonly prescribed as second‐line treatment in the UK [37] because contemporaneous NICE guidance [2] generally positioned them later in the treatment pathway and specified continuation criteria. Consequently, the small group of patients who initiated these drugs as second‐line therapy likely represented a highly selected population, and the estimates may have been affected by unmeasured confounding related to disease progression, comorbidities and treatment preferences. For example, patients willing to initiate an injectable GLP‐1 receptor agonist may have been more accepting of injectable treatment than those receiving oral antidiabetic medications. In addition, given the weight‐loss benefits of GLP‐1 receptor agonists, some guidelines [2, 3] recommend their combined use with insulin, although we could not determine whether clinicians had a similar preference for adding insulin to patients already receiving a GLP‐1 receptor agonist. As GLP‐1 receptor agonist use increases, studies using more recent data may include larger, more diverse cohorts and provide better treatment overlap and comparability to address this issue.
Our findings should be interpreted considering several limitations. First, differences in time to insulin initiation may partly reflect clinical inertia [35], which varies across clinical practice and patient characteristics [38]. Because glycemic targets are personalized, we were unable to determine the interval between inadequate glycemic control and insulin initiation, although the secondary outcome provided complementary information on treatment trajectories. Second, our secondary outcome, defined as switching to or adding another antidiabetic drug class, may not exclusively represent treatment failure. Treatment modification may occur because of inadequate glycemic control, adverse effects or intolerance, changes in comorbidities, or patient and clinician preferences. Because the reasons for treatment modification were unavailable, we could not distinguish inadequate effectiveness from tolerability or other clinical considerations. Third, we cannot rule out residual confounding resulting from unmeasured confounders. Restricting the cohort to patients receiving metformin monotherapy, using active comparators [39], and trimming patients outside the region of propensity‐score overlap [40] improved comparability. E‐value analysis suggested that most findings were robust unless affected by a strong unmeasured confounder. Fourth, few participants initiated TZDs. Their limited use may reflect concerns about adverse effects associated with pioglitazone [41], the only TZD available in the UK.
5. Conclusions
In this population‐based cohort of patients with type 2 diabetes receiving metformin monotherapy, initiation of TZDs, DPP‐4 inhibitors or SGLT‐2 inhibitors was associated with a decreased risk of insulin initiation, compared to insulin secretagogues. These findings provide important real‐world evidence on the long‐term comparative effectiveness of the use of these agents as second‐line therapies following metformin monotherapy.
Author Contributions
Y.‐H.Y. and K.B.F. conceived of the study idea in consultation with all other co‐authors. All authors contributed to the study design. Q.Z. conducted the data management and analyses. P.R. provided critical input during the analytic stage. Y.‐H.Y. wrote the initial draft of the manuscript, and all other authors interpreted data and reviewed the manuscript for intellectual content. All authors approved the submitted manuscript. K.B.F. is the guarantor.
Funding
The authors have nothing to report.
Ethics Statement
The study protocol was approved by the CPRD's Independent Scientific Advisory Committee (protocol number: 21_000450) and the Research Ethics Board of the Jewish General Hospital in Montreal, Canada (reference number: 2022–2957).
Conflicts of Interest
R.W.P. has received personal fees from Analysis Group, Biogen, Merck/Organon, Merck KGAA and Pfizer, all outside of the submitted work. K.B.F. has received personal fees from Regeneron and Statlog, all outside of the submitted work.
Supporting information
Figure S1: Distribution of propensity scores for each second‐line antidiabetic medication in the study cohort (before trimming), by treatment actually initiated (colour‐coded).
Figure S2: Distribution of propensity scores for each second‐line antidiabetic medication in the study cohort (after trimming), based on re‐estimated propensity score, by treatment actually initiated (colour‐coded).
Figure S3: Cumulative incidence of treatment modification by second‐line antidiabetic medications initiated at the study cohort entry. Intention to treat approach were used to categorize person times based on the second‐line antidiabetic initiated at the study cohort entry. S, based on data regulations for CPRD, for cell counts < 5, more than 1 cell needs to be suppressed to avoid being back calculated.
Figure S4: Patterns of transition from second‐line to third‐line antidiabetic medications (among those initiating a new class antidiabetic medication during follow‐up).
Figure S5: Bias plots of E‐values for unmeasured confounder.
Figure S6: Distribution of propensity scores for each second‐line antidiabetic medication (excluding GLP‐1 receptor agonists and TZD) in the study cohort (before trimming), by treatment actually initiated (colour‐coded).
Figure S7: Distribution of propensity scores for each second‐line antidiabetic medication (excluding GLP‐1 receptor agonists and TZD) in the study cohort (after trimming), based on re‐estimated propensity score, by treatment actually initiated (colour‐coded).
Table S1: Target trial protocol and observational emulation.
Table S2: Characteristics of patients with type 2 diabetes in the study cohort (before trimming and weighting), by second‐line antidiabetic medication initiated at study cohort entry, in the United Kingdom between 2013 and 2021.
Table S3: Characteristicsa of patients with type 2 diabetes in the final analytic cohort (after weighting), by second‐line antidiabetic medication initiated at study cohort entry, in the United Kingdom between 2013 and 2021.
Table S4: Distribution of reasons for censoring in the primary outcome analysis, by second‐line antidiabetic medication initiated at cohort entry, United Kingdom, 2013–2021.
Table S5: Risk of insulin initiation and treatment modification among second‐line antidiabetic medications initiated at cohort entry, adjusted for imbalanced covariates in the outcome model, in the United Kingdom between 2013 and 2021.
Table S6: Characteristicsa of patients with type 2 diabetes in the final analytic cohort (before weighting), by second‐line antidiabetic medication (excluding GLP‐1 receptor agonists and TZD) initiated at study cohort entry, in the United Kingdom between 2013 and 2021.
Table S7: Characteristicsa of patients with type 2 diabetes in the final analytic cohort (after weighting), by second‐line antidiabetic medication (excluding GLP‐1 receptor agonists and TZD) initiated at study cohort entry, in the United Kingdom between 2013 and 2021.
Table S8: Risk of insulin initiation and treatment modification among second‐line antidiabetic medications (excluding GLP‐1 receptor agonists and TZD) initiated at cohort entry, in the United Kingdom between 2013 and 2021.
Acknowledgements
K.B.F. is supported by a Merit salary support award from the Fonds de recherche du Québec–Santé (FRQS; Quebec Foundation for Health Research) and a William Dawson Scholar award from McGill University. R.W.P. holds the Albert Boehringer I Chair in Pharmacoepidemiology. O.H.Y.Y. is supported by a Junior 1 salary support award from the FRQS. E.Z.‐B. holds a Doctoral Training Scholarship from the FRQS.
Yu Y.‐H., Zhang Q., Zapata‐Bravo E., et al., “Time to Insulin Initiation Among Patients With Type 2 Diabetes Treated With Second‐Line Antidiabetic Drugs,” Diabetes, Obesity and Metabolism 28, no. 10 (2026): 9655–9667, 10.1111/dom.71182.
Handling Editor: Johan Jendle
Data Availability Statement
This study is based in part on data from the Clinical Practice Research Datalink obtained under licence from the UK Medicines and Healthcare products Regulatory Agency. The data are provided by patients and collected by the UK National Health Service as part of their care and support. The interpretation and conclusions contained in this study are those of the author/s alone. Because electronic health records are classified as ‘sensitive data’ by the UK Data Protection Act, information governance restrictions (to protect patient confidentiality) prevent data sharing via public deposition. Data are available with approval through the individual constituent entities controlling access to the data. Specifically, the primary care data can be requested via application to the Clinical Practice Research Datalink (https://www.cprd.com).
References
- 1. Khan M. A. B., Hashim M. J., King J. K., Govender R. D., Mustafa H., and Al Kaabi J., “Epidemiology of Type 2 Diabetes – Global Burden of Disease and Forecasted Trends,” Journal of Epidemiology and Global Health 10, no. 1 (2020): 107–111, 10.2991/jegh.k.191028.001. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 2. McGuire H., Longson D., Adler A., Farmer A., and Lewin I., “Management of Type 2 Diabetes in Adults: Summary of Updated NICE Guidance,” British Medical Journal 353 (2016): i1575, 10.1136/bmj.i1575. [DOI] [PubMed] [Google Scholar]
- 3. American Diabetes Association Professional Practice Committee , ElSayed N. A., McCoy R. G., et al., “9. Pharmacologic Approaches to Glycemic Treatment: Standards of Care in Diabetes—2025,” Diabetes Care 48, no. S1 (2025): S181–S206, 10.2337/dc25-S009. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 4. Diabetes Canada Clinical Practice Guidelines Expert Committee , Lipscombe L., Booth G., et al., “Pharmacologic Glycemic Management of Type 2 Diabetes in Adults,” Canadian Journal of Diabetes 42, no. S1 (2018): S88–S103, 10.1016/j.jcjd.2017.10.034. [DOI] [PubMed] [Google Scholar]
- 5. Fonseca V. A., “Defining and Characterizing the Progression of Type 2 Diabetes,” Diabetes Care 32, no. S2 (2009): S151–S156, 10.2337/dc09-S301. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 6. Home P., Riddle M., Cefalu W. T., et al., “Insulin Therapy in People With Type 2 Diabetes: Opportunities and Challenges?,” Diabetes Care 37, no. 6 (2014): 1499–1508, 10.2337/dc13-2743. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 7. Mearns E. S., Sobieraj D. M., White C. M., et al., “Comparative Efficacy and Safety of Antidiabetic Drug Regimens Added to Metformin Monotherapy in Patients With Type 2 Diabetes: A Network Meta‐Analysis,” PLoS One 10, no. 4 (2015): e0125879, 10.1371/journal.pone.0125879. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 8. Tsapas A., Avgerinos I., Karagiannis T., et al., “Comparative Effectiveness of Glucose‐Lowering Drugs for Type 2 Diabetes: A Systematic Review and Network Meta‐Analysis,” Annals of Internal Medicine 173, no. 4 (2020): 278–286, 10.7326/M20-0864. [DOI] [PubMed] [Google Scholar]
- 9. Bidulka P., Lugo‐Palacios D. G., Carroll O., et al., “Comparative Effectiveness of Second Line Oral Antidiabetic Treatments Among People With Type 2 Diabetes Mellitus: Emulation of a Target Trial Using Routinely Collected Health Data,” British Medical Journal 385 (2024): e077097, 10.1136/bmj-2023-077097. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 10. Ray A., Sreedhara S. K., Paik J. M., et al., “Comparative Effectiveness of Current Glucose‐Lowering Medications in Type 2 Diabetes Mellitus: Emulation of a Modified GRADE Trial,” Journal of General Internal Medicine 41 (2026), 10.1007/s11606-026-10170-7. [DOI] [PubMed] [Google Scholar]
- 11. Schrijnders D., Hartog L. C., Kleefstra N., Groenier K. H., Landman G. W. D., and Bilo H. J. G., “Within‐Sulfonylurea‐Class Evaluation of Time to Intensification With Insulin (ZODIAC‐43),” PLoS One 11, no. 6 (2016): e0157668, 10.1371/journal.pone.0157668. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 12. Mamza J., Mehta R., Donnelly R., and Idris I., “Important Differences in the Durability of Glycaemic Response Among Second‐Line Treatment Options When Added to Metformin in Type 2 Diabetes: A Retrospective Cohort Study,” Annals of Medicine 48, no. 4 (2016): 224–234, 10.3109/07853890.2016.1157263. [DOI] [PubMed] [Google Scholar]
- 13. Jermendy G., Kiss Z., Rokszin G., Abonyi‐Tóth Z., Wittmann I., and Kempler P., “Persistence to Treatment With Novel Antidiabetic Drugs (Dipeptidyl Peptidase‐4 Inhibitors, Sodium‐Glucose co‐Transporter‐2 Inhibitors, and Glucagon‐Like Peptide‐1 Receptor Agonists) in People With Type 2 Diabetes: A Nationwide Cohort Study,” Diabetes Therapy 9, no. 5 (2018): 2133–2141, 10.1007/s13300-018-0483-4. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 14. Carney G. A., Bassett K., Wright J. M., and Dormuth C. R., “Is Thiazolidinediones Use a Factor in Delaying the Need for Insulin Therapy in Type 2 Patients With Diabetes? A Population‐Based Cohort Study,” British Medical Journal Open 2, no. 6 (2012): e001910, 10.1136/bmjopen-2012-001910. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 15. Rascati K. L., Richards K. M., Lopez D., Cheng L. I., and Wilson J. P., “Progression to Insulin for Patients With Diabetes Mellitus Using the Texas Medicaid Database,” Clinical Therapeutics 33, no. 12 (2011): 2016–2020, 10.1016/j.clinthera.2011.10.018. [DOI] [PubMed] [Google Scholar]
- 16. Inzucchi S. E., Tunceli K., Qiu Y., et al., “Progression to Insulin Therapy Among Patients With Type 2 Diabetes Treated With Sitagliptin or Sulphonylurea Plus Metformin Dual Therapy,” Diabetes, Obesity & Metabolism 17, no. 10 (2015): 956–964, 10.1111/dom.12489. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 17. Montvida O., Shaw J., Atherton J. J., Stringer F., and Paul S. K., “Long‐Term Trends in Antidiabetes Drug Usage in the U.S.: Real‐World Evidence in Patients Newly Diagnosed With Type 2 Diabetes,” Diabetes Care 41, no. 1 (2018): 69–78, 10.2337/dc17-1414. [DOI] [PubMed] [Google Scholar]
- 18. Cashin A. G., Hansford H. J., Hernán M. A., et al., “Transparent Reporting of Observational Studies Emulating a Target Trial: The TARGET Statement,” Journal of the American Medical Association 334, no. 12 (2025): 1084–1093, 10.1001/jama.2025.13350. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 19. Wolf A., Dedman D., Campbell J., et al., “Data Resource Profile: Clinical Practice Research Datalink (CPRD) Aurum,” International Journal of Epidemiology 48, no. 6 (2019): 1740, 10.1093/ije/dyz034. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 20. Austin P. C., “Balance Diagnostics for Comparing the Distribution of Baseline Covariates Between Treatment Groups in Propensity‐Score Matched Samples,” Statistics in Medicine 28, no. 25 (2009): 3083–3107, 10.1002/sim.3697. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 21. McCaffrey D. F., Griffin B. A., Almirall D., Slaughter M. E., Ramchand R., and Burgette L. F., “A Tutorial on Propensity Score Estimation for Multiple Treatments Using Generalized Boosted Models,” Statistics in Medicine 32, no. 19 (2013): 3388–3414, 10.1002/sim.5753. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 22. Imbens G. W., “The Role of the Propensity Score in Estimating Dose‐Response Functions,” Biometrika 87, no. 3 (2000): 706–710, 10.1093/biomet/87.3.706. [DOI] [Google Scholar]
- 23. Yoshida K., Solomon D. H., Haneuse S., et al., “Multinomial Extension of Propensity Score Trimming Methods: A Simulation Study,” American Journal of Epidemiology 188, no. 3 (2019): 609–616, 10.1093/aje/kwy263. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 24. Robins J. M., Hernán M. A., and Brumback B., “Marginal Structural Models and Causal Inference in Epidemiology,” Epidemiology 11, no. 5 (2000): 550–560. [DOI] [PubMed] [Google Scholar]
- 25. Stürmer T., Webster‐Clark M., Lund J. L., et al., “Propensity Score Weighting and Trimming Strategies for Reducing Variance and Bias of Treatment Effect Estimates: A Simulation Study,” American Journal of Epidemiology 190, no. 8 (2021): 1659–1670, 10.1093/aje/kwab041. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 26. Joffe M. M., “Administrative and Artificial Censoring in Censored Regression Models,” Statistics in Medicine 20, no. 15 (2001): 2287–2304, 10.1002/sim.850. [DOI] [PubMed] [Google Scholar]
- 27. VanderWeele T. J. and Ding P., “Sensitivity Analysis in Observational Research: Introducing the E‐Value,” Annals of Internal Medicine 167, no. 4 (2017): 268–274, 10.7326/M16-2607. [DOI] [PubMed] [Google Scholar]
- 28. Austin P. C., “An Introduction to Propensity Score Methods for Reducing the Effects of Confounding in Observational Studies,” Multivariate Behavioral Research 46, no. 3 (2011): 399–424, 10.1080/00273171.2011.568786. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 29. Rubin D. B., “Multiple Imputation After 18+ Years,” Journal of the American Statistical Association 91, no. 434 (1996): 473–489, 10.2307/2291635. [DOI] [Google Scholar]
- 30. Suzuki R., Shoji S., Yoshinaga Y., Kosakai Y., and Shintani‐Tachi M., “Risk of Insulin Initiation With Sodium‐Glucose Cotransporter‐2 Inhibitors Versus Dipeptidyl Peptidase‐4 Inhibitors in Patients With Type 2 Diabetes Mellitus: A Real‐World Claims Database Study in Japan,” Diabetes, Obesity & Metabolism 27, no. 4 (2025): 1960–1971, 10.1111/dom.16188. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 31. Liss D. T., Cherupally M., O'Brien M. J., et al., “Treatment Modification After Initiating Second‐Line Medication for Type 2 Diabetes,” American Journal of Managed Care 29, no. 12 (2023): 661–668, 10.37765/ajmc.2023.89466. [DOI] [PubMed] [Google Scholar]
- 32. Zelniker T. A., Wiviott S. D., Raz I., et al., “SGLT2 Inhibitors for Primary and Secondary Prevention of Cardiovascular and Renal Outcomes in Type 2 Diabetes: A Systematic Review and Meta‐Analysis of Cardiovascular Outcome Trials,” Lancet 393, no. 10166 (2019): 31–39, 10.1016/S0140-6736(18)32590-X. [DOI] [PubMed] [Google Scholar]
- 33. Heerspink H. J. L., Desai M., Jardine M., Balis D., Meininger G., and Perkovic V., “Canagliflozin Slows Progression of Renal Function Decline Independently of Glycemic Effects,” Journal of the American Society of Nephrology 28, no. 1 (2017): 368–375, 10.1681/ASN.2016030278. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 34. Li X., Song Y., Guo T., Xiao G., and Li Q., “Effect of Glucagon‐Like Peptide 1 Receptor Agonists on the Renal Protection in Patients With Type 2 Diabetes: A Systematic Review and Meta‐Analysis,” Diabetes & Metabolism 48, no. 5 (2022): 101366, 10.1016/j.diabet.2022.101366. [DOI] [PubMed] [Google Scholar]
- 35. Khunti K. and Millar‐Jones D., “Clinical Inertia to Insulin Initiation and Intensification in the UK: A Focused Literature Review,” Primary Care Diabetes 11, no. 1 (2017): 3–12, 10.1016/j.pcd.2016.09.003. [DOI] [PubMed] [Google Scholar]
- 36. The GRADE Study Research Group , “Glycemia Reduction in Type 2 Diabetes — Glycemic Outcomes,” New England Journal of Medicine 387, no. 12 (2022): 1063–1074, 10.1056/NEJMoa2200433. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 37. Abrahami D., D'Andrea E., Yin H., et al., “Contemporary Trends in the Utilization of Second‐Line Pharmacological Therapies for Type 2 Diabetes in the United States and the United Kingdom,” Diabetes, Obesity and Metabolism 25, no. 10 (2023): 2980–2988, 10.1111/dom.15196. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 38. Khunti K., Nikolajsen A., Thorsted B. L., Andersen M., Davies M. J., and Paul S. K., “Clinical Inertia With Regard to Intensifying Therapy in People With Type 2 Diabetes Treated With Basal Insulin,” Diabetes, Obesity and Metabolism 18, no. 4 (2016): 401–409, 10.1111/dom.12626. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 39. Lund J. L., Richardson D. B., and Stürmer T., “The Active Comparator, New User Study Design in Pharmacoepidemiology: Historical Foundations and Contemporary Application,” Current Epidemiology Reports 2, no. 4 (2015): 221–228, 10.1007/s40471-015-0053-5. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 40. Stürmer T., Rothman K. J., Avorn J., and Glynn R. J., “Treatment Effects in the Presence of Unmeasured Confounding: Dealing With Observations in the Tails of the Propensity Score Distribution – A Simulation Study,” American Journal of Epidemiology 172, no. 7 (2010): 843–854, 10.1093/aje/kwq198. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 41. Alam F., Islam M. A., Mohamed M., et al., “Efficacy and Safety of Pioglitazone Monotherapy in Type 2 Diabetes Mellitus: A Systematic Review and Meta‐Analysis of Randomised Controlled Trials,” Scientific Reports 9, no. 1 (2019): 5389, 10.1038/s41598-019-41854-2. [DOI] [PMC free article] [PubMed] [Google Scholar]
Associated Data
This section collects any data citations, data availability statements, or supplementary materials included in this article.
Supplementary Materials
Figure S1: Distribution of propensity scores for each second‐line antidiabetic medication in the study cohort (before trimming), by treatment actually initiated (colour‐coded).
Figure S2: Distribution of propensity scores for each second‐line antidiabetic medication in the study cohort (after trimming), based on re‐estimated propensity score, by treatment actually initiated (colour‐coded).
Figure S3: Cumulative incidence of treatment modification by second‐line antidiabetic medications initiated at the study cohort entry. Intention to treat approach were used to categorize person times based on the second‐line antidiabetic initiated at the study cohort entry. S, based on data regulations for CPRD, for cell counts < 5, more than 1 cell needs to be suppressed to avoid being back calculated.
Figure S4: Patterns of transition from second‐line to third‐line antidiabetic medications (among those initiating a new class antidiabetic medication during follow‐up).
Figure S5: Bias plots of E‐values for unmeasured confounder.
Figure S6: Distribution of propensity scores for each second‐line antidiabetic medication (excluding GLP‐1 receptor agonists and TZD) in the study cohort (before trimming), by treatment actually initiated (colour‐coded).
Figure S7: Distribution of propensity scores for each second‐line antidiabetic medication (excluding GLP‐1 receptor agonists and TZD) in the study cohort (after trimming), based on re‐estimated propensity score, by treatment actually initiated (colour‐coded).
Table S1: Target trial protocol and observational emulation.
Table S2: Characteristics of patients with type 2 diabetes in the study cohort (before trimming and weighting), by second‐line antidiabetic medication initiated at study cohort entry, in the United Kingdom between 2013 and 2021.
Table S3: Characteristicsa of patients with type 2 diabetes in the final analytic cohort (after weighting), by second‐line antidiabetic medication initiated at study cohort entry, in the United Kingdom between 2013 and 2021.
Table S4: Distribution of reasons for censoring in the primary outcome analysis, by second‐line antidiabetic medication initiated at cohort entry, United Kingdom, 2013–2021.
Table S5: Risk of insulin initiation and treatment modification among second‐line antidiabetic medications initiated at cohort entry, adjusted for imbalanced covariates in the outcome model, in the United Kingdom between 2013 and 2021.
Table S6: Characteristicsa of patients with type 2 diabetes in the final analytic cohort (before weighting), by second‐line antidiabetic medication (excluding GLP‐1 receptor agonists and TZD) initiated at study cohort entry, in the United Kingdom between 2013 and 2021.
Table S7: Characteristicsa of patients with type 2 diabetes in the final analytic cohort (after weighting), by second‐line antidiabetic medication (excluding GLP‐1 receptor agonists and TZD) initiated at study cohort entry, in the United Kingdom between 2013 and 2021.
Table S8: Risk of insulin initiation and treatment modification among second‐line antidiabetic medications (excluding GLP‐1 receptor agonists and TZD) initiated at cohort entry, in the United Kingdom between 2013 and 2021.
Data Availability Statement
This study is based in part on data from the Clinical Practice Research Datalink obtained under licence from the UK Medicines and Healthcare products Regulatory Agency. The data are provided by patients and collected by the UK National Health Service as part of their care and support. The interpretation and conclusions contained in this study are those of the author/s alone. Because electronic health records are classified as ‘sensitive data’ by the UK Data Protection Act, information governance restrictions (to protect patient confidentiality) prevent data sharing via public deposition. Data are available with approval through the individual constituent entities controlling access to the data. Specifically, the primary care data can be requested via application to the Clinical Practice Research Datalink (https://www.cprd.com).