Metformin Treatment and Cancer Risk

Cox Regression Analysis, With Time-Dependent Covariates, of 320,000 Persons With Incident Diabetes Mellitus

Rachel Dankner; Nirit Agay; Liraz Olmer; Havi Murad; Lital Keinan Boker; Ran D. Balicer; Laurence S. Freedman


Am J Epidemiol. 2019;188(10):1794-1800. 

In This Article

Abstract and Introduction


There is conflicting evidence regarding the association between metformin use and cancer risk in diabetic patients. During 2002–2012, we followed a cohort of 315,890 persons aged 21–87 years with incident diabetes who were insured by the largest health maintenance organization in Israel. We used a discrete form of weighted cumulative metformin exposure to evaluate the association of metformin with cancer incidence. This was implemented in a time-dependent covariate Cox model, adjusting for treatment with other glucose-lowering medications, as well as age, sex, ethnic background, socioeconomic status, smoking (for bladder and lung cancer), and parity (for breast cancer). We excluded from the analysis metformin exposure during the year before cancer diagnosis in order to minimize reverse causation of cancer on changes in medication use. Estimated hazard ratios associated with exposure to 1 defined daily dose of metformin over the previous 2–7 years were 0.98 (95% confidence interval (CI): 0.82, 1.18) for all-sites cancer (excluding prostate and pancreas), 1.05 (95% CI: 0.67, 1.63) for colon cancer, 0.98 (95% CI: 0.49, 1.97) for bladder cancer, 1.02 (95% CI: 0.59, 1.78) for lung cancer, and 0.88 (95% CI: 0.56, 1.39) for female breast cancer. Our results do not support an association between metformin treatment and the incidence of major cancers (excluding prostate and pancreas).


The evaluation of associations between glucose-lowering medications (GLMs) and cancer risk has increased dramatically. Metformin is the predominant drug being investigated in this context. This biguanide has a well-established safety profile and has been used as a treatment for hyperglycemia for more than half a century.[1] Metformin is the most commonly prescribed oral GLM worldwide and is recommended as first-line therapy for type 2 diabetes mellitus.[2] Several studies, starting in 2005, have suggested a possible protective effect of metformin on cancer risk.[3] Wu et al.[4] concluded that there was such an effect in a meta-analysis of 265 studies on GLMs and cancer incidence, 66 of them on metformin; however, heterogeneity between them was high (I2 = 89%), and the association was not supported by a sub–meta-analysis that comprised the 23 randomized controlled trials from the full meta-analysis. Time-related biases and insufficient attention to the natural history of type 2 diabetes have been claimed to be the basis of the negative metformin association demonstrated in observational studies.[5] Accordingly, in a systematic review of observational studies, Farmer et al.[6] showed that among studies with a low possibility of bias, a causal effect of metformin on risk of all-sites or specific types of cancer was not evident.

We investigated the association between metformin treatment as a time-dependent exposure and cancer incidence in a population-based cohort study of patients with incident type 2 diabetes, while accounting for major time-related biases and for various diabetes treatments as they changed over time. An important feature of our investigation was the use of Cox regression that included history of metformin treatment and history of other GLMs as time-dependent covariates. Sylvestre and Abrahamowicz[7] described the use of weighted cumulative exposure (WCE) functions for evaluating the effects of history of use of a medication on the incidence of a disease in continuous time. In this paper, we describe a simplified, discrete-time version of the WCE.