Association of Metformin Use With Risk of Venous Thromboembolism in Adults With Type 2 Diabetes

A General-Population-Based Cohort Study

Tingting Sha; Yuqing Zhang; Changjun Li; Guanghua Lei; Jing Wu; Xiaoxiao Li; Zidan Yang; Chao Zeng; Jie Wei


Am J Epidemiol. 2022;191(5):856-866. 

In This Article


Data Sources

We used data from The Health Improvement Network (THIN), an electronic medical record database derived from the records of general practitioners in the United Kingdom. It draws approximately 17 million participants from 790 general practices, which is representative of the UK population in terms of demographic structure and prevalence of key comorbid conditions.[4] Health-care information, including data on sociodemographic characteristics, anthropometric factors, lifestyle factors, diagnoses, prescriptions, hospital admissions, and referrals, as well as results of laboratory tests, is prospectively gathered and anonymized. The Read classification system[22] is used to code specific diagnoses, whereas a dictionary based on the Multilex classification system (FDB Multilex; First Databank UK Ltd., Exeter, United Kingdom)[23] is used to code medications.[24] The validity of THIN for use in clinical and epidemiologic research studies has been demonstrated previously.[25] The scientific review committee for the THIN database (The Health Improvement Network Ltd., London, United Kingdom) and the institutional review board at Xiangya Hospital (Changsha, China) approved this study, with a waiver of informed consent.

Study Design and Cohort Definition

The risks of DVT and PE increase rapidly with age; both diseases are rare before age 40 years.[26,27] Additionally, since the number of individuals aged 90 years or older in THIN is small, the THIN institutional review board requires that persons aged ≥90 years should not be included in analyses to avoid the potential risk of unintentional (deductive) identification of a specific individual. Thus, we included participants who were 40–90 years of age, had a history of type 2 diabetes based on the Read codes,[28,29] and had at least 1 year of continuous enrollment in a general practice participating in THIN from January 2000 to December 2018. The code list for type 2 diabetes is provided in Web Table 1 (available at We conducted a cohort study to compare the risks of VTE between metformin initiators and sulfonylurea initiators. Initiation of either metformin use or sulfonylurea use was defined as not having a history of use of either medication before entering the study and refilling at least 1 prescription within 60 days after the end of the supply of the first prescription.[30] The date of the first refill was denoted as the index date. We did not exclude the patients with type 2 diabetes who used other types of antidiabetic medication before the first prescription of either metformin or a sulfonylurea. Because metformin is contraindicated for patients with severe chronic kidney disease,[31] we excluded persons with severe chronic kidney disease,[30] defined as an estimated glomerular filtration rate less than 30 mL/minute/1.73 m2 on at least 2 occasions more than 90 days apart or at least 1 Read code for advanced chronic kidney disease.[32] In addition, we also excluded persons who had a history of VTE before the index date or who had no general practitioner visit or specialist referral within 1 year before the index date.

To account for potential secular trends in metformin and sulfonylurea prescriptions in relation to various covariates,[33] we divided calendar time into 19 1-year time blocks ranging from January 2000 to December 2018. Follow-up ended on December 31, 2019. Within each time block, inverse probability of treatment weighting was conducted using a logistic regression model in which the propensity of initiation of metformin versus sulfonylureas was regressed on the following covariates: sociodemographic factors, study entry time, lifestyle factors, type 2 diabetes duration, comorbid conditions, prescriptions, and health-care utilization prior to the index date. Inverse probability of treatment weighting applies weights corresponding to 1/propensity score for patients in the metformin group and [1/(1 − propensity score)] for those in the sulfonylurea group. To exclude patients with extreme weights, we truncated weights at the 95th percentile with this threshold (i.e., replaced any weight greater than the 95th percentile with this threshold).[34,35]

Assessment of Outcomes

The primary outcome of interest was a diagnosis of incident VTE (the combined endpoint of PE and DVT), with secondary outcomes of PE and DVT separately. An individual was considered to have developed the outcome of interest if he or she had a recorded Read code of PE or DVT and received anticoagulant treatment.[36] Since VTE is potentially fatal and some individuals might have died before receiving an anticoagulant, we also considered participants to have developed VTE if they died within 1 month after a recorded code of PE or DVT but without a prescription for an anticoagulant or autopsy results.[37] A code list for PE, DVT, and anticoagulants is available in Web Table 2. This VTE definition has been validated and adopted in previous studies, with a sensitivity rate of 94% in the UK Clinical Practice Research Datalink, which is similar to the THIN database.[4,36,38,39]

Assessment of Covariates

Information on sociodemographic and anthropometric characteristics, lifestyle factors, comorbid conditions, and medication use was selected on the basis of clinical significance before the index date (Table 1). Data on body mass index (weight (kg)/height (m)2) measurements, cigarette smoking status (never, former, or current smoker), alcohol consumption (never, former, or current drinker), Townsend deprivation index, comorbidity (myocardial infarction, stroke, hypertension, liver diseases, chronic kidney diseases, polycystic ovary syndrome, etc.), and prescriptions (statins, nonsteroidal antiinflammatory drugs, opioids, hormone replacement therapy, etc.) were collected before the index date. Data on indicators of health-care utilization (numbers of hospitalizations, general practitioner visits, and specialist referrals) were derived from the past 1 year before the index date.

Statistical Analysis

Person-years of follow-up were calculated as the amount of time from the index date to the date of any of the following events: an incident VTE, death, drug discontinuation (i.e., continuous use of a drug allowing for a 60-day grace period), a switch to or addition of a comparator antidiabetic drug, disenrollment from a general practice participating in THIN, or the end of the study period (i.e., December 31, 2019), whichever came first. We plotted the cumulative incidence curves for VTE, PE, and DVT accounting for the competing risk of death. We performed Cox proportional hazards regression to estimate hazard ratios (HRs) with 95% CIs and conducted the Kolmogorov-Smirnov supremum test to assess the proportional hazards assumption. If the proportional hazards assumption was violated, we then estimated the HR at 2 years, 4 years, 6 years, and 10 years of follow-up.[4] We also estimated the rate differences in the risk of VTE between the metformin group and the sulfonylurea group.

We conducted several sensitivity analyses to assess the robustness of our results. First, we performed an intention-to-treat analysis to assess the relationship between initiation of metformin use and the risk of VTE (or PE/DVT). Specifically, person-years of follow-up were calculated from the index date to the date of an outcome of interest, death, disenrollment from a general practice participating in a THIN general practice, or the end of the study period (i.e., December 31, 2019), irrespective of any treatment switching, augmentation, or discontinuation. Second, since hemoglobin A1c (HbA1c) is a sensitive proxy for blood glucose control in type 2 diabetes and is a potential risk factor for VTE, we further adjusted for HbA1c level prior to the index date to control for potential confounding. Third, to minimize potential selection bias, we repeated the primary analyses among the incident cases of type 2 diabetes instead of the prevalent cases. Fourth, since cancer is a strong risk factor for VTE, we excluded persons with a history of cancer before the index date from the analyses. Fifth, to evaluate the robustness of the assumptions about the prescription refill period, we varied the prescription refill period from 60 days (a 60-day supply) to 90 days (a 90-day supply).[40] Sixth, because hematocrit might be associated with metformin use and might affect the results, we conducted a sensitivity analysis by adjusting the level of hematocrit prior to the index date to control its potentially confounding effect. Finally, we also added potential covariates to the Cox proportional hazards model as a sensitivity analysis to evaluate the robustness of our results.

We also evaluated the associations of antidiabetic therapy with 2 positive control outcomes—namely, all-cause mortality and incident cardiovascular disease (CVD)—to further assess the robustness of our study findings.[41,42] Death was confirmed by the patient's death date as recorded in THIN, linked to the UK National Health Service.[24] Incident CVD was defined as newly developed myocardial infarction, stroke, or heart failure if the participant had a recorded corresponding Read code during the following year.[43]

All P values were 2-sided, and P < 0.05 was considered statistically significant for all tests. Statistical analyses were conducted using SAS, version 9.4 (SAS Institute, Inc., Cary, North Carolina) and R, version 3.5.1 (R Foundation for Statistical Computing, Vienna, Austria).