Metformin and low levels of thyroid-stimulating hormone in patients with type 2 diabetes mellitus

Data source

We used the Clinical Practice Research Datalink (CPRD) to conduct this study. Established in 1987, the CPRD includes data on more than 13 million patients from over 680 general practices in the United Kingdom. It records information on diagnoses, referrals, lifestyle habits, anthropometric measurements (e.g., body mass index) and prescriptions issued by general practitioners. The CPRD is regularly audited and has been shown to be highly valid.12^,13

The study protocol was approved by the Independent Scientific Advisory Committee of the CPRD (protocol 14040) and by the Research Ethics Committee of the Jewish General Hospital in Montréal.

Study population

We assembled a base cohort of patients at least 40 years of age who began receiving either metformin or sulfonylurea monotherapy between Jan. 1, 1988, and Dec. 31, 2012. Patients were required to have at least 1 year of medical history in the CPRD before the first metformin and sulfonylurea prescriptions, with no evidence of other antidiabetic drugs being prescribed at any time before these first prescriptions. Thus, cohort entry was the date of the first-ever prescriptions for metformin and sulfonylureas during the study period. We excluded patients with a history of cancer (other than nonmelanoma skin cancer) and pituitary disorders at any time before cohort entry. We also excluded women with a diagnosis of polycystic ovary syndrome or gestational diabetes at any time before cohort entry, or pregnancy in the year before cohort entry. Finally, for the purposes of this study, the cohort was restricted to patients whose TSH levels had been measured at least once in the year before cohort entry.

Using the base cohort, we assembled 2 sub-cohorts of patients with treated hypothyroidism and euthyroidism. Patients with treated hypothyroidism were identified by the presence of at least 2 prescriptions for levothyroxine. They were excluded if they were receiving concurrent treatment with antithyroid drugs, or if they had TSH levels outside the reference range (0.4–4.5 mIU/L14) in the year before cohort entry. Euthyroid patients were identified by exclusion of those with records of thyroid-related disorders, drugs or procedures, and of those with TSH levels outside the reference range (0.4–4.5 mIU/L14) in the year before cohort entry.

All patients were followed from cohort entry until a first TSH measurement below 0.4 mIU/L, the initiation (in euthyroid patients) or dosage modification (increase or decrease) of thyroid-related drugs, a TSH measurement above 4.5 mIU/L (such an increase might result in the initiation or intensification of levothyroxine, leading to subsequent normal TSH levels15), death from any cause, end of registration with the general practice or end of the study period (Mar. 31, 2013), whichever occurred first.

Statistical analysis

We used descriptive statistics to summarize the characteristics of patients receiving metformin and sulfonylurea monotherapy, separately for each subcohort. Crude incidence rates of low TSH levels (< 0.4 mIU/L), with 95% confidence intervals (CIs) based on the Poisson distribution, were calculated for each exposure group.

Kaplan–Meier curves were plotted for the cumulative incidence of low TSH levels for each exposure group, stratified by subcohort. For each subcohort, we used Cox proportional hazards models to estimate crude and adjusted hazard ratios (HRs) with 95% CIs for the incidence of low TSH levels associated with the use of metformin monotherapy, compared with sulfonylurea monotherapy. The following potential confounders, which were considered a priori, were included in the model: age, sex, calendar year of cohort entry, body mass index, smoking status, excessive alcohol use (i.e., alcohol-related disorders such as alcoholism, alcoholic cirrhosis, alcoholic hepatitis and hepatic failure), hemoglobin A_1C level (last measure before cohort entry), and use of lithium, amiodarone, glucocorticoids and dopamine agonists17 (measured in the year before cohort entry). Dopamine agonists were not included in the final model because of their low prevalence and instability in the models. Variables with missing information were coded with an “unknown” category.

In secondary analyses, we assessed whether there was a duration–response relation between the use of metformin and the risk of low TSH levels in each subcohort. Using time-dependent Cox proportional hazards models, we estimated HRs for 3 categories of duration of use (< 90 d, 90–180 d, > 180 d). We also estimated HRs for the association between metformin and the risk of suppressed TSH levels (< 0.1 mIU/L14). It was not possible to assess thyroxine and triiodothyronine levels, given their low testing rates (0.037 and 0.013 tests/1000 person-years for patients with treated hypothyroidism and euthyroidism, respectively).

We performed 7 sensitivity analyses. First, we assessed residual confounding by fitting the outcome models with high-dimensional propensity score deciles18 (for additional details, see Appendix 1, available at www.cmaj.ca/lookup/suppl/doi:10.1503/cmaj.140688/-/DC1). Second, we conducted an intention-to-treat analysis, in which the initial treatment was carried forward until a maximum of 2 years. Third, we repeated the primary analysis after excluding first-generation sulfonylureas because they have been associated with an increased risk of hypothyroidism in earlier studies.19^,20 Fourth, to minimize potential residual confounding,21 we further restricted the cohort with treated hypothyroidism to patients who, in the year before cohort entry, had stable doses of levothyroxine (i.e., last 2 prescriptions at the same daily dosage) and whose last 2 TSH measurements were within the reference range (0.4–4.5 mIU/L) and within ± 0.5 mIU/L of each other. Fifth, patients were additionally censored at the time of a new cancer diagnosis, pregnancy or infection (using antibiotics as a proxy), because these conditions may result in transient fluctuation of TSH levels. Sixth, cohort entry was moved to the 30 days after the first prescription of metformin and sulfonylureas to minimize possible misclassifications related to the situation where the actual TSH measurement was recorded in the database after cohort entry, and to account for a minimum 30-day latency period between treatment initiation and outcome. Finally, to assess the effect of missing TSH measurements that may have been measured but not recorded in the database during follow-up, we censored patients with no TSH measurements in the year after cohort entry, as well as patients with no repeat TSH measurements within 1 year of each other. All analyses were conducted using SAS version 9.3 (SAS Institute).