Sex and BMI Alter the Benefits and Risks of Sulfonylureas and Thiazolidinediones in Type 2 Diabetes: A Framework for Evaluating Stratification Using Routine Clinical and Individual Trial Data

Limited guidance is available in type 2 diabetes to help clinicians and patients choose between the different glucose-lowering therapy options recommended after metformin (1–3). Guidelines suggest a discussion of the benefits, adverse effects, and costs of therapy to select the most appropriate medication for a particular patient (1). Estimates of important clinical outcomes, such as HbA_1c, weight change, and risk of side effects, are at present derived from whole trial populations, and a key question is whether they vary across patient subgroups defined by simple characteristics (1). If estimates do vary by simple characteristics, this may provide a starting point for a stratified approach in type 2 diabetes of “targeting treatment according to the biological or risk characteristics shared by patients” (4).

Sulfonylureas and thiazolidinediones are recommended second- and third-line therapy options in all major type 2 diabetes guidelines (1,2). They represented 50% of new second-line prescriptions in 2016 in the U.S. (sulfonylureas, 46%; thiazolidinediones, 4%) (5). As the only generic oral agents, they are more than 10-fold cheaper than the common alternatives dipeptidyl peptidase 4 (DPP-4) inhibitors and sodium–glucose cotransporter 2 (SGLT-2) inhibitors (1,6). Glycemic response, weight change, and common side effects have been well described in whole trial populations for both therapies (7–11). Differences in glycemic response by sex and BMI with thiazolidinediones and sulfonylureas have been previously suggested in observational studies (12,13), but no study has systematically compared whether the benefits and risks of these therapies vary across subgroups defined by simple clinical patient characteristics.

Sulfonylureas and thiazolidinediones have, in contrast to newer therapies, been evaluated head-to-head in two long-term, randomized trials, ADOPT (A Diabetes Outcome Progression Trial) and RECORD (Rosiglitazone Evaluated for Cardiovascular Outcomes and Regulation of Glycaemia in Diabetes) (7,14). ADOPT showed there was a greater durability of response up to 5 years with the thiazolidinedione rosiglitazone compared with the sulfonylurea glyburide or metformin (7). The full individual participant data of both trials are now available through ClinicalStudyDataRequest.com (15), and how to improve the output of secondary research projects using such shared trial data sets is currently being debated (16). In this study, we present a practical and cost-effective framework for stratification research using shared trial data sets alongside routine clinical data. We applied this framework to systematically evaluate whether simple clinical patient characteristics can be used to stratify therapy with sulfonylureas and thiazolidinediones.

In discovery analysis, we explored routine clinical data to identify simple characteristics associated with glycemic response to sulfonylureas and thiazolidinediones and used the results to define patient subgroups likely to show differential response. In validation analysis, we evaluated differences in response within subgroups as a prespecified hypothesis in ADOPT and RECORD, the two largest head-to-head randomized trials of sulfonylureas and thiazolidinediones available via ClinicalStudyDataRequest.com (7,9,14,17,18). We also evaluated the secondary outcomes of weight change and risk of the common side effects of hypoglycemia, edema, and fracture within each subgroup (see Supplementary Fig. 1 for our framework for stratification research using routine clinical and shared trial data).

We analyzed four data sets. Due to the large sample size, discovery analysis was conducted in routine clinical data from U.K. Clinical Practice Research Datalink (CPRD), with validation in trial data sets (ADOPT and RECORD) and a further routine clinical data set from GoDARTS (Genetics of Diabetes Audit and Research in Tayside Scotland). Scientific approval for the use of CPRD data was granted by the CPRD Independent Scientific Advisory Committee (ISAC 13_177R), and permission to use the GoDARTS data was granted by the East of Scotland Regional Ethics Committee (09/21402/44). Data for the ADOPT and RECORD trials were accessed through the Clinical Trial Data Transparency Portal under approval from GlaxoSmithKline (Proposal 930).

CPRD is the world’s largest database of anonymized primary care electronic health records (19). Our study protocol for CPRD data ascertainment has been previously reported (20). We studied 22,379 noninsulin-treated patients with type 2 diabetes and prescription records for a sulfonylurea or thiazolidinedione from the February 2014 build of CPRD (see Supplementary Data for CPRD product codes). We included patients with a duration of diabetes of more than 1 year (to minimize effect of lifestyle change after diagnosis) and at least 1 year on therapy without change in coprescribed glucose-lowering therapy (see Supplementary Fig. 2 for CPRD patient flow diagram) (20).

ADOPT and RECORD were prospective type 2 diabetes trials over at least 5 years of, respectively, glycemic durability and cardiovascular outcomes, in participants randomized to thiazolidinedione, sulfonylurea, or metformin therapy (7,9,14,17,18). In ADOPT, we included participants in the intention-to-treat population with a valid baseline BMI randomized to sulfonylurea (glibenclamide) or thiazolidinedione (rosiglitazone) therapy (n = 2,725). In RECORD, we included participants in the intention-to-treat population on background metformin randomized to sulfonylurea (glibenclamide [18%], gliclazide [30%], or glimepiride [52%], according to local practice) or thiazolidinedione (rosiglitazone) add-on therapy (n = 2,222).

GoDARTS contains information from the medical records of 18,276 people resident in eastern Scotland. We examined 1,977 patients with type 2 diabetes and valid prescription records for a sulfonylurea or thiazolidinedione.

The primary outcome was 1-year glycemic response in patients starting therapy with a sulfonylurea (any) or thiazolidinedione (pioglitazone or rosiglitazone) for the first time.

We extracted HbA_1c at therapy start and at 1 year to calculate initial HbA_1c response (1-year HbA_1c – baseline HbA_1c) (see Supplementary Data for CRPD HbA_1c codes) and baseline clinical characteristics of sex, BMI, age at diagnosis, duration of diabetes, and estimated glomerular filtration rate (eGFR) (20). Baseline HbA_1c was defined as the closest HbA_1c to the drug start date in the 91 days before the drug start date, and 1-year HbA_1c was defined as the closest HbA_1c to 1 year after drug start date (±3 months). HbA_1c response was only valid if there were no changes to diabetes medications between 60 days before the baseline HbA_1c and the date of the 1-year HbA_1c (20). No adjustment was made for dose. To evaluate the secondary outcomes of long-term response and side effects, we extracted measures of body weight, HbA_1c, and records of fracture and edema (see Supplementary Data for CPRD fracture and edema codes) over 5 years from the start of therapy. Patients with a fracture or edema record in the 2 years before the drug start date were excluded from fracture and edema analyses. We defined adherence as a medication possession ratio (the number of days of available medication divided by the number of days between the first and last prescription dates, multiplied by 100). Due to the association between adherence and response (21), only patients issued sufficient prescriptions (medical possession ratio of between 80% and 120%) were included in analysis.

We used individual participant data from the trials to validate initial findings in CPRD. Based on the CPRD results, we prespecified four subgroups defined by sex and obesity (BMI ≥30 kg/m²). For each subgroup we compared average glycemic response by therapy over 5 years as the difference in area under the HbA_1c response curve. This is equivalent to the time-updated HbA_1c measure used in the UK Prospective Diabetes Study outcomes model (22). At years 1, 3, and 5, we also estimated the difference between therapies in average glycemic response. We assessed annual weight change (percentage change from baseline) using the same approach. We also compared durability of response by therapy as measured by time to therapy failure. Failure was defined as in the original trials (ADOPT: confirmed fasting plasma glucose ≥180 mg/dL; RECORD: confirmed HbA_1c ≥8.5%). To evaluate side effects over 5 years, we estimated the on-therapy risk of fracture (any), clinically determined peripheral edema (all events, moderate/severe events [as defined in the original trials as sufficient to, respectively, interfere with or prevent normal everyday activities]), and clinically determined hypoglycemia (all, moderate/severe as defined in the original trials) (9,17). In ADOPT we excluded patients with a history of edema from edema analysis; in RECORD history of edema was not available.

We evaluated average glycemic response by therapy over 5 years using the same approach used for CPRD.

We assessed associations between baseline clinical characteristics (BMI, sex, age at diagnosis, duration of diabetes, eGFR) and 1-year glycemic response in linear regression models. A series of baseline HbA_1c-adjusted models examined each clinical characteristic in turn, separately for each therapy (23). We conducted a complete case analysis for each variable of interest, including all patients with valid data even if they had missing data for other clinical characteristics. Diagnostic plots of residuals were examined to check that model assumptions were met. Based on the initial analysis, we defined four subgroups by sex and obesity (BMI >30 vs. BMI ≤30 kg/m²) and for each therapy calculated baseline HbA_1c-adjusted least square mean estimates of 1-year response for each subgroup. To test for an overall effect of heterogeneity by sex and obesity subgroup, we used a likelihood ratio test to compare a model with a drug-subgroup interaction with a nested model without an interaction term.

We compared how each outcome was altered by therapy in each subgroup separately. We conducted response and weight change analysis in each trial separately but pooled the data for side effects to increase study power. To estimate glycemic response over time, we fitted baseline-adjusted repeated measures mixed-effect models using on-therapy HbA_1c values at each study visit (n = 22 ADOPT, n = 19 RECORD) up to 5 years, including fixed effects for study visit, baseline HbA_1c, therapy, visit by therapy interaction, and visit by baseline HbA_1c interaction, and patient-level random effects with an unstructured covariance matrix. Missing on-therapy HbA_1c records were assumed to be missing at random. We calculated point estimates and 95% CIs for the difference in average glycemic response by therapy at years 1, 3, and 5 through contrasts of least squares mean HbA_1c change. We tested for an overall effect of heterogeneity by subgroup using the same interaction test as in CPRD. Weight change was modeled using the same approach.

To measure the net difference in HbA_1c response between therapies, we calculated the cumulative area under the HbA_1c response curve (AUC) for each participant at every study visit using the trapezoidal rule. Participant AUC was then used as the outcome in repeated measures mixed-effects models of the same structure as for glycemic response. A least squares mean point estimate (95% CI) was calculated at year 5 to contrast overall response by therapy.

Time to therapy failure and side effects were estimated using the Kaplan-Meier method and Cox proportional hazards regression. Proportional hazards assumptions were evaluated using Schoenfeld residuals and were satisfied for all analyses. For each side effect, the hazard ratio contrasting thiazolidinedione therapy with sulfonylurea therapy was estimated for each subgroup using an individual participant meta-analysis of data from both trials.

We replicated analyses in CPRD using the same models as described above for all outcomes except hypoglycemia, which is poorly captured in primary care records. For analysis of long-term HbA_1c response, we extracted all HbA_1c records between 60 days before the drug start date up to 5 years after the drug start date while on unchanged therapy. HbA_1c records were categorized to 3-month intervals (nearest HbA_1c record ±1.5 months) to enable comparison with the trials. Where data points were missing, results were interpolated to ensure each time point reflected the same population of patients. The same approach was used for weight change, but with weights extracted at 6-month intervals (±3 months). For time-to-failure analysis, therapy failure was defined as two consecutive HbA_1c measures ≥8.5% or one HbA_1c ≥8.5%, followed by the addition of another therapy (the same definition of glycemic failure used in RECORD). Data were censored if prescription records ended before a change in therapy. We excluded patients with changes to diabetes therapy without a prior HbA_1c ≥8.5% because these changes were unlikely to relate to glycemic failure.

CPRD data extraction was conducted using Stata 13.0 software. All other analyses were conducted using R software.

In CPRD, we examined clinical factors associated with 1-year glycemic response among 22,379 eligible patients (10,960 with thiazolidinedione, 11,419 with sulfonylurea) (see Supplementary Table 1 for baseline characteristics). Sex and BMI showed the greatest differential response to therapy (Supplementary Fig. 3). Compared with males, females had a greater response with thiazolidinediones but a lesser response with sulfonylureas (both P < 0.001). Higher BMI was associated with greater response with thiazolidinediones but a lesser response with sulfonylureas (both P < 0.001). Older age at diagnosis and lower eGFR were associated with a greater response to both therapies, and there was greater response to thiazolidinediones with shorter diabetes duration and greater response to sulfonylureas with longer diabetes duration and higher HDL (Supplementary Fig. 3).

As sex and BMI showed the greatest differential response, we specified four subgroups defined by sex and obesity (BMI >30 vs. BMI ≤30 kg/m²) for use in subsequent analysis. We found evidence of heterogeneity of response by subgroup (P < 0.001). Figure 1 shows 1-year glycemic response by therapy for the four subgroups. Nonobese males had a greater 1-year response with sulfonylureas than with thiazolidinediones (baseline adjusted change in HbA_1c −13.2 vs. −9.7 mmol/mol, P < 0.001), whereas obese females had a greater 1-year response with thiazolidinediones than with sulfonylureas (−13.8 vs. −9.4 mmol/mol, P < 0.001). Obese males and nonobese females showed similar responses with both therapies (both P = 0.6). Results were consistent for pioglitazone and rosiglitazone when analyzed separately and for gliclazide and non-gliclazide sulfonylureas (Supplementary Fig. 4).

We went on to assess whether the sex- and obesity-defined subgroups also showed differential response when randomly allocated to therapy in the ADOPT (n = 2,725) and RECORD (n = 2,222) trials. Randomization resulted in well-matched patients for each therapy within each subgroup (see Supplementary Tables 2 and 3 for baseline characteristics). There were marked differences in response with both therapies in the four subgroups, with a clear similarity between the two trials (test for heterogeneity in ADOPT and RECORD both P < 0.001) (Fig. 2A and B). Over 5 years there was a greater overall glycemic response for nonobese males with sulfonylureas (both trials P < 0.001), relating to the greater earlier benefit with sulfonylureas over thiazolidinediones that persisted beyond 2 years in both trials. In contrast, there was a greater overall glycemic response for obese females with thiazolidinediones over sulfonylureas (both trials P < 0.001), and there was little early benefit with sulfonylureas.

We assessed the risk of monotherapy failure in ADOPT and dual-therapy failure in RECORD. In both trials, there was no difference in the 5-year risk of failure on the two therapies for nonobese males, but all other subgroups were less likely to fail with thiazolidinediones than with sulfonylureas (hazard ratios 0.23–0.72, test for heterogeneity ADOPT P < 0.001 and RECORD P = 0.01) (Table 1 and Supplementary Figs. 5 and 6). In ADOPT, risk of monotherapy failure at 5 years with thiazolidinediones was lower for obese females (11%) than for nonobese males (22%), whereas with sulfonylureas failure risk was lower for nonobese males (22%) than for obese females (42%) (Table 1).

Weight was increased for all subgroups with thiazolidinediones compared with sulfonylureas, but this was much more marked in obese females (Fig. 2C and Supplementary Fig. 7). Edema was more common with thiazolidinediones than with sulfonylureas for all subgroups. This resulted in the largest difference in absolute risk for obese females, who are most likely to develop edema regardless of therapy (Table 2 and Supplementary Fig. 8).

Fracture was more common with thiazolidinediones compared with sulfonylureas but only for females. Absolute risk was similar for obese and nonobese females (Table 2 and Supplementary Fig. 9).

Sulfonylureas, compared with thiazolidinediones, increased the risk of moderate/severe hypoglycemia for all subgroups (Table 2 and Supplementary Fig. 10). Hazard ratios for hypoglycemia of any severity were consistent with those for moderate/severe events (Supplementary Tables 4 and 5). The differences between therapies for all side effects when the trials were analyzed separately were similar (Supplementary Tables 4 and 5).

In CPRD and GoDARTS (see Supplementary Table 6 for GoDARTS baseline characteristics), 5-year glycemic response results were consistent with the trials(Supplementary Figs. 11 and 16). In CPRD, differences by therapy in time to failure were similar to the trials, although absolute failure rates were higher (Supplementary Fig. 12). Weight gain, edema, and fracture results in CPRD were comparable to trial data (Supplementary Figs. 13–15).

Subgroup data summaries of glycemic response, weight change, and risk of side effects estimates specific to each sex- and obesity-defined subgroup are provided in the Supplementary Data.

We have robustly demonstrated across four data sets that sex and BMI alter the benefits and risks of type 2 diabetes therapy with sulfonylureas and thiazolidinediones. We show in nonobese males the glycemic response with sulfonylureas is better on average in the first 5 years than on thiazolidinediones, without excess weight gain but with an increased risk of hypoglycemia. For obese females, there is a clear glycemic benefit over the first 5 years with thiazolidinediones compared with sulfonylureas, but there is increased weight gain and susceptibility to edema and fracture. Our findings will allow for much more informed discussion of the benefits and risks of these therapies than the present “one size fits all” approach (see the Subgroup Data Summary in the Supplementary Data for estimates specific to each sex- and obesity-defined subgroup).

Our results provide the first example of stratification of therapy in type 2 diabetes based on simple patient characteristics (24). A recent data-driven cluster analysis proposed five subgroups of diabetes with differing disease progression and risk of complications but did not evaluate whether subgroups differed in their response to therapy (25). To date, successful stratification in other conditions has involved expensive genetic testing, as applied in cancer and single-gene diseases such as monogenic diabetes (26,27). Expensive testing is unlikely to become practical or justified in type 2 diabetes, a highly prevalent condition treated with relatively inexpensive therapy. Type 2 diabetes genetic studies have identified polymorphisms associated with drug response, but the impact of these, at present, is too small to guide clinical management, in contrast to our results (28–32).

This study is an early and important demonstration of how shared trial data can be harnessed to meaningfully benefit patients (16). We propose a novel and cost-effective framework to use shared trial data in stratification research. Our framework can be applied to study other type 2 diabetes therapies and to study stratification in other chronic conditions. It has great potential to improve the output of future studies using shared trial data.

Although no existing studies have systematically assessed whether both the benefits and risks of these two therapies are altered by clinical characteristics, previous analyses have suggested sex and BMI are associated with glycemic response to both therapies. In ADOPT, the risk of therapy failure was lower for obese and female subgroups with thiazolidinediones compared with sulfonylureas, but an interaction was not tested for and the difference in glycemic trajectory was not examined (7). Observational studies have found increased glycemic response for obese female patients with thiazolidinediones and for male patients with sulfonylureas (12,13), but the effect of this in terms of stratification has not been assessed. We previously showed that markers of insulin resistance, including BMI, are associated with reduced glycemic response to DPP-4 inhibitors but not glucagon-like peptide 1 (GLP-1) receptor agonists (33,34), but evidence for other agents is limited (35,36).

Previous studies have also found sex and BMI alter the risk of side effects. The increase in fracture risk with thiazolidinediones applies mainly to postmenopausal females and is consistent within trials (10,11,14,37). We found hypoglycemia risk with sulfonylureas was similar across subgroups even though glycemic response differed, and this needs further investigation. Although our study shows the absolute risk of edema with thiazolidinediones was highest in the obese female subgroup that had the greatest response, further study is required to fully evaluate the association between glycemic response and the risk of common side effects for these therapies.

Our study has limitations. The results do not allow prediction at an individual level; however, we present subgroup estimates that will better reflect the likely outcome for an individual patient within that subgroup than outcome estimates derived from whole trial populations. Rosiglitazone, the thiazolidinedione used in both trials analyzed in our study, has been withdrawn in many countries because of concerns over cardiovascular safety (38). Routine clinical data support a thiazolidinedione class effect of differential response by sex and obesity, but trial data were not made available to repeat our analysis for pioglitazone. Previous meta-analyses suggest that the risks of edema and fracture are similar with both drugs, further supporting the generalizability of our findings to pioglitazone (37,39). For sulfonylureas, a similar pattern of results was observed in ADOPT (glibenclamide), RECORD (52% glimepiride, 30% gliclazide, and 18% glibenclamide), and routine clinical data (including a gliclazide-only analysis), supporting a sulfonylurea class effect. For the 1-year glycemic response analysis in CPRD, we excluded nonadherent patients and those whose antihyperglycemic therapy was altered (potentially due to poor response, a very good response, or poor tolerance) within the first year, and this could have accounted for the differences we observed when comparing sulfonylurea and thiazolidinedione therapy. However, we saw a similar pattern of glycemic response differences using time-to-failure and mixed-effect models, both of which included all patients with at least one on-therapy HbA_1c measure for up to 5 years. The CPRD time-to-failure analysis was also limited because patients whose treatment was intensified below the HbA_1c failure threshold of 8.5% were censored rather than defined as experiencing therapy failure.

A strength of the CPRD analysis is the demonstration of consistent results with all three analytical approaches, each with its own strengths and weaknesses. Measured or unmeasured baseline differences between patients could have explained findings in the routine data but are very unlikely to explain the differences we observed in the randomized clinical trials, further highlighting the strength of our study design. More than 90% of patients in the data sets studied were white Caucasian, limiting the applicability of our findings to other racial groups, a common problem with trials in type 2 diabetes. Additional data would be required to answer whether there are differences in patients of South Asian, Hispanic, or black origin, where fat distribution can be markedly different and a different obesity cutoff may be appropriate (40).

The ideal stratified approach would be based on cardiovascular end points rather than the intermediary measure of glycemic response. In this analysis, we were underpowered to detect differences for cardiovascular outcomes in RECORD (the primary trial analysis showed no difference between rosiglitazone and sulfonylureas or metformin) or rarer adverse effects such as heart failure (14). Given that the two recent trials demonstrating cardiovascular benefits with SGLT2 inhibitors and GLP-1 receptors agonists each required more than 7,000 high-risk participants (41,42), it may be that impractically large trials are required for stratification of cardiovascular end points.

Evaluation of the risks and benefits of newer therapies, such as DPP-4 inhibitors, SGLT2 inhibitors, and GLP-1 receptor agonists, will require routine clinical data from large numbers of patients alongside shared head-to-head drug trial data and will be possible in the near future. Given the greater expense of newer therapies, cost-effectiveness evaluation will be necessary in this work. The ongoing Glycemia Reduction Approaches in Diabetes: A Comparative Effectiveness Study (GRADE) will give important long-term head-to-head comparative effectiveness data on second-line treatment with insulin, DPP-4 inhibitor, GLP-1 receptor agonist, and sulfonylurea therapy (43).

Further mechanistic studies are required to interrogate the mechanisms underlying differential response to sulfonylureas and thiazolidinediones. Thiazolidinediones act through the adipocyte, and so any increase in the number of adipocytes is likely to improve glycemic response. This provides a potential explanation for our findings, because females, compared with males, have a higher whole-body percentage fat mass and thus have more adipocytes, and obese subjects have more adipocytes than nonobese subjects (44). The reduced insulin sensitivity seen in obesity is likely to explain the reduced response to sulfonylureas that predominantly stimulate insulin secretion by the β-cell. The consistently better response seen in males to sulfonylureas was unexpected, and further studies are required to define the mechanism of this observation.

The sex and obesity subgroup–specific estimates presented in this study will allow a much more informed discussion between clinicians and patients of the benefits and risks of sulfonylureas and thiazolidinediones, at no cost. We recommend this discussion with an individual patient be based around the appropriate sex and obesity subgroup–specific estimates presented for the two therapies in the Subgroup Data Summary in the Supplementary Data. Whether this alters a decision on therapy will depend on the individual circumstances of the patient, because the trade-off between early response, long-term durability, and risk of side effects will be different.

Simple patient characteristics alter the benefits and risks of type 2 diabetes therapy with sulfonylureas and thiazolidinediones. Our novel and practical framework for stratification research can be applied in type 2 diabetes and other chronic conditions and has great potential to improve output from future studies using shared trial data.

Funding. The MASTERMIND (MRC APBI STratification and Extreme Response Mechanism IN Diabetes) consortium is supported by the Medical Research Council (U.K.) (MR/N00633X/1). A.G.J. is supported by a National Institute for Health Research (NIHR) Clinician Scientist award. N.S. acknowledges support by Innovative Medicines Initiative Joint Undertaking under grant agreement number 115372, the resources of which comprise financial contribution from the European Union’s Seventh Framework Programme (FP7/2007–2013) and European Federation of Pharmaceutical Industries and Associations companies’ in-kind contribution. R.R.H. and A.T.H. are NIHR Senior Investigators. E.R.P. is a Wellcome Trust New Investigator (102820/Z/13/Z). B.M.S. and A.T.H. are supported by the NIHR Exeter Clinical Research Facility. A.T.H. is a Wellcome Trust Senior Investigator.

The views expressed are those of the authors and not necessarily those of the NHS, the NIHR, or the Department of Health. The funders had no role in any part of the study or in any decision about publication.

Duality of Interest. W.E.H. declares a grant from Quintiles. N.S. declares personal fees from Boehringer Ingelheim, Eli Lilly, Novo Nordisk, and Janssen and a grant from AstraZeneca. S.J. is an employee and stockholder of GlaxoSmithKline. R.R.H. declares research funding from Bayer, AstraZeneca, and Merck Sharp & Dohme and honoraria from Amgen, Bayer, Elcelyx, Jannsen, Intarcia, Merck Sharp & Dohme, Novartis, Novo Nordisk, and Servier. E.R.P. declares personal fees from Eli Lilly, Novo Nordisk, and AstraZeneca. Representatives from GlaxoSmithKline, Takeda, Janssen, Quintiles, AstraZeneca, and Sanofi attend meetings as part of the industry group involved with the MASTERMIND consortium. No industry representatives were involved in the writing of the manuscript or analysis of data. For all authors, these are outside the submitted work. No other potential conflicts of interest relevant to this article were reported.

Author Contributions. J.M.D., W.E.H., M.N.W., M.L., L.R.R., A.G.J., W.T.H., N.S., S.J., R.R.H., E.R.P., B.M.S., and A.T.H. approved the final article. J.M.D., W.E.H., A.G.J., W.T.H., S.J., E.R.P., B.M.S., and A.T.H. designed the study. J.M.D., M.L., and B.M.S. analyzed the data. J.M.D. and B.M.S. drafted the article. W.E.H., A.G.J., N.S., R.R.H., E.R.P., and A.T.H. provided support for the analysis and interpretation of results and critically revised the article. M.N.W., L.R.R., and B.M.S. extracted the data from the CPRD. B.M.S. and A.T.H. are the guarantors of this work and, as such, had full access to all the data in the study and take responsibility for the integrity of the data and the accuracy of the data analysis.

Prior Presentation. Parts of this study were presented in abstract form at the Diabetes UK Professional Conference, Glasgow, U.K., 2–4 March 2016; at the 76th Scientific Sessions of the American Diabetes Association, New Orleans, LA, 10–14 June 2016; and at the European Association for the Study of Diabetes 2016 Annual Meeting, Munich, Germany, 12–16 September 2016.