Metformin and Cancer: Solutions to a Real-World Evidence Failure

The recent randomized placebo-controlled trial of 3,649 women with breast cancer who were followed for 5 years to assess the effectiveness of metformin as adjuvant therapy for this cancer found no benefit on disease-free survival or overall survival with this drug (1). This large trial likely terminates an extensive pursuit of metformin as a potential treatment for cancer, which began in 2005 with an observational study from Scotland that reported a reduction in the incidence of cancer with metformin use (2). This observational study generated considerable excitement and interest in investigating metformin for cancer prevention and treatment in patients with or without diabetes. Several in vitro and in vivo studies found that metformin reduces cancer cell proliferation in several cancer cell types. In parallel, numerous observational studies and meta-analyses of these studies generally reported quite spectacular reductions in cancer incidence and cancer outcomes with metformin use. On the other hand, the ensuing randomized controlled trials (RCTs) did not find metformin to be efficacious as a treatment for different cancers.

This quest of repurposing metformin, an antidiabetes drug, as an anticancer agent echoes the 21st Century Cures Act enacted by the U.S. Congress in 2016 (3,4). The Act included a mandate for the U.S. Food and Drug Administration to develop guidance to evaluate the use of real-world evidence (RWE) in, among other things, supporting the approval of a new indication for a drug already approved for another indication. This Act, which defines RWE as data derived from sources other than randomized trials, popularized observational studies based on existing health care databases to assess the real-world effects of medications, including the identification of new indications for older drugs. While the RCT is the standard of proof of a drug’s effectiveness, the much larger observational studies have provided greater power to detect less frequent and latent adverse effects while representing the broader spectrum of patients who use this drug in clinical practice. Beyond safety, this Act now also supports the use of observational studies to evaluate the effectiveness of existing drugs in other indications. The use of metformin in the prevention and treatment of various cancers is an example.

In this article, we review the evidence behind the metformin–cancer saga, focusing on the observational and randomized trials. We discuss reasons for the discrepancies in findings and describe some methods to better conduct future observational RWE studies that can help reduce such discrepancies.

Several studies suggest that metformin has anticancer effects by affecting various pathways. Metformin has been shown to act on protein synthesis via the AMP-activated protein kinase (AMPK) and mTOR signaling pathways. Metformin inhibits the mitochondrial respiratory chain complex on tumor cells, increasing the activation of the AMPK pathway, which plays a role in cell proliferation and protein synthesis (5). The activation of the AMPK pathway leads to glycolysis, fatty acid oxidation, and inhibition of fatty acid and protein synthesis (6). Increased mTOR-dependent protein synthesis is one of the main pathways of breast tumorigenesis in humans (7). The anticancer effects of metformin could also occur via lowering of glycemia and reduction in insulin resistance, resulting in lower insulin and insulin-like growth factor 1 (IGF-1) levels, which may inhibit cancer cell growth (5,8). Indeed, growth factors and hormones such as insulin have been shown to play a role in tumorigenesis by activating phosphatidylinositol 3-kinase signaling (7). By affecting these pathways, metformin has been shown to affect cancer cell proliferation in in vitro and in vivo studies.

Metformin was shown to reduce cell proliferation and colony formation in a variety of breast cancer cell lines independent of the estrogen receptor, progesterone receptor, human epidermal growth factor receptor 2 (HER-2), or p53 status (9–11). It was also shown to inhibit proliferation of other cancer cell line types, such as colon (12), prostate (13), endometrial (14), ovarian (15), lung (16), and brain (17) cancers. The findings from in vitro studies are consistent with those in animal studies using mouse models of various cancers, such as breast (7,18) and lung (19), whereby metformin treatment given in a dose comparable to that used in humans decreased tumor growth. Metformin was also shown to decrease tumor development by 25% in mice heterozygous for the tumor suppressor gene PTEN, which results in tumor development in various organs (20). However, the dose of metformin used in this study was >10 times higher than the standard dose used in the clinical setting. Indeed, many of these experimental studies were conducted using suprapharmacological doses of metformin, with 10 to 100 times higher plasma levels than that achievable in humans (21). As such, findings from these studies may not translate to the same effects in humans (21,22).

While metformin has been shown to act in pathways that affect tumorigenesis, reviews of the data from older type 2 diabetes randomized trials of metformin did not observe such effects on cancer incidence. A meta-analysis of seven randomized trials involving metformin as a treatment for type 2 diabetes that reported data on cancer adverse events found no association between metformin treatment and the risk of developing any malignancy (odds ratio [OR] 0.98; 95% CI 0.81–1.19), although cancer was not a prespecified end point in these trials (23). Similarly, in the A Diabetes Outcome Progression Trial (ADOPT), treatment of type 2 diabetes with metformin did not reduce the risk of any malignancies (excluding nonmelanoma skin cancers) compared with rosiglitazone (hazard ratio [HR] 0.92; 95% CI 0.63–1.35) or with glibenclamide (HR 0.78; 95% CI 0.53–1.14) (24). Consistent findings were observed in the Rosiglitazone Evaluated for Cardiac Outcomes and Regulation of Glycaemia in Diabetes (RECORD) trial, where patients treated with metformin did not have a lower risk of malignancies than those treated with rosiglitazone (HR 1.22; 95% CI 0.86–1.74) (24).

Soon after the signals of a potential benefit of metformin as a cancer treatment, calls were made for randomized therapeutic trials in several cancers (25–27). However, these trials that assessed metformin as adjuvant therapy for the treatment of various cancers did not find clinical benefits on cancer outcomes. A recent meta-analysis of nine phase 2 randomized trials of metformin as an adjuvant to standard of care for different advanced or metastatic cancers found no increase in tumor response (OR 1.23; 95% CI 0.89–1.71), progression-free survival (HR 0.95; 95% CI 0.75–1.21), or overall survival (HR 0.97; 95% CI 0.80–1.16) compared with standard of care alone (28).

A meta-analysis of five phase 2 randomized trials of women with breast cancer found that metformin additive therapy was not associated with improved progression-free survival (HR 1.00; 95% CI 0.70–1.43) and overall survival (HR 1.00; 95% CI 0.71–1.39) (29). Recently, a large 5-year phase 3 placebo-controlled trial assessing the effect of adjuvant metformin compared with standard therapy on invasive disease-free survival among 3,649 women with high-risk nonmetastatic breast cancer was completed (1). The incidence rates for invasive disease-free survival events were not different between women treated with adjuvant metformin and women given standard therapy (2.78 vs. 2.74 per 100 patient-years; HR 1.01; 95% CI 0.84–1.21), as were the mortality rates (1.46 vs. 1.32 per 100 patient-years; HR 1.10; 95% CI 0.86–1.41).

Randomized trials were also conducted to assess metformin in the treatment of other cancers. A phase 2 trial of 54 patients with unresected locally advanced non–small-cell lung cancer found that metformin in combination with chemoradiotherapy was associated with a higher incidence of locoregional disease progression (HR 2.42; 95% CI 1.14–5.10) and a higher mortality (HR 3.80; 95% CI 1.49–9.73) than no metformin (30). A meta-analysis of four randomized trials in patients with non–small-cell lung cancer also did not find that metformin improved overall survival (risk ratio [RR] 0.86; 95% CI 0.68–1.08) or progression-free survival (RR 0.92; 95% CI 0.77–1.10) (31). Metformin as adjuvant therapy was also studied for prostate cancer in a randomized trial of 124 patients with high-risk locally advanced or metastatic hormone-sensitive prostate cancer (HSPC) who were randomized to metformin in combination with standard of care or to standard of care alone (32). Prostate cancer-free survival was significantly higher among participants treated with metformin (29 months [95% CI 25–33 months] vs. 20 months [95% CI 16–24 months]) but not overall survival.

Studies of metformin as added therapy for colorectal cancer that suggest some clinical benefits are based on single-arm studies, also called “single-arm trials” (33). However, such studies that use historical comparators are fraught with potential confounding and time-related biases (34,35). For example, a phase 2 single-arm trial was conducted in 41 patients with refractory colorectal cancer to determine whether a combination treatment of metformin and irinotecan (a chemotherapy used for treatment of colorectal cancer) was associated with an improved disease control rate (36). Within 12 weeks of follow-up, 41% of participants, all treated with metformin and irinotecan, attained disease control, which was higher than the corresponding 13% reported in the placebo arm of the Regorafenib Monotherapy for Previously Treated Metastatic Colorectal Cancer (CORRECT) phase 3 trial (37). Such comparisons with external control subjects are subject to several potential biases and must be interpreted with great caution (34,35). Similarly, a phase 2 single-arm clinical trial of 50 participants with metastatic colorectal cancer treated with metformin in combination with leucovorin observed that 22% attained disease control at 8 weeks, with no direct comparator group (38). Nonetheless, the study conveyed that the median overall survival of 7.9 months was longer than the median 5 months reported in the placebo arms of two phase 3 trials. In these studies, participants treated with metformin had a higher risk of adverse events than participants not given metformin (1,30,36), which led to metformin discontinuation in some cases (1,30).

The very first human study of metformin and cancer was a 2005 observational case-control analysis conducted within a clinical database of over 11,000 patients with type 2 diabetes from Tayside, Scotland (2). A comparison of 923 patients who developed cancer with 1,846 matched control subjects who did not resulted in a significant 23% reduction in the incidence of any cancer associated with metformin use, thus advancing the metformin–cancer hypothesis. This “research pointers” report was followed by many observational studies conducted using various clinical and claims databases worldwide.

The first series of observational studies that assessed the preventive potential of metformin on cancer incidence or mortality reported remarkable reductions in the risk of various cancers, including breast (39,40), colon (41), liver (41,42), lung (43), and prostate (44) cancer. In the first such study, the use of sulfonylurea compared with metformin was associated with an increased risk of cancer-related mortality (HR 1.3; 95% CI 1.1–1.6), concluding that these results may reflect a “protective effect of metformin” on cancer-related mortality (45). A meta-analysis of this and other observational studies found that metformin was associated with an overall ∼30% reduction in cancer incidence (46). However, many of the studies included in this meta-analysis have been shown to be affected by various time-related biases that tend to exaggerate the benefit of a drug (47,48).

The most common time-related bias affecting these studies is immortal time bias, which is introduced in these cohort studies by a misclassified definition of metformin exposure, either at the design or the analysis stage (47,49–51). Many of these studies classified subjects who entered the cohort on another treatment and initiated metformin only later as “metformin users” throughout. However, the time between cohort entry and metformin initiation, called “immortal” since the patient must be alive to initiate metformin, is misclassified as metformin use instead of nonuse (Fig. 1). This exposure misclassification leads to immortal time bias by artificially extending the time to cancer incidence or mortality of the metformin users.

The concern for immortal time bias in observational studies that address the benefits associated with metformin use on cancer prevention was illustrated using a cohort of 41,533 individuals newly diagnosed with type 2 diabetes to assess the association between metformin use and colorectal cancer incidence (52). When the analysis was conducted using a time-fixed analysis, which misclassifies exposure and is subject to immortal time bias, it found a remarkable protective effect of metformin on incident colorectal cancer (HR 0.37; 95% CI 0.29–0.46). However, this effect disappeared when data analysis techniques were used that avoid immortal time bias by properly classifying exposure. As awareness for time-related biases increased, more recent observational studies that accounted for these biases found no association between metformin use and the risk of any incident cancer (53,54) as well as lung (55), colorectal (56) and bladder (57) cancer.

Observational studies of metformin as adjuvant therapy for cancer treatment have also reported that metformin use is associated with positive outcomes, including overall survival and decreased cancer progression among patients with various cancers. In a meta-analysis of mainly observational studies (26 retrospective studies and 1 prospective study), metformin use was associated with longer recurrence-free survival (HR 0.63; 95% CI 0.47–0.85), overall survival (HR 0.69; 95% CI 0.58–0.83), and cancer-specific survival (HR 0.58; 95% CI 0.39–0.86) among studies assessing patients with early-stage colorectal cancer (58). Metformin use was also shown to increase recurrence-free survival among individuals with prostate cancer treated with radical radiotherapy (HR 0.45; 95% CI 0.29–0.70) (58). However, metformin use was not found to have a survival benefit among individuals with breast and urothelial cancers (58).

The large magnitude of clinical benefit in cancer mortality associated with metformin use in observational studies also results from time-related biases, especially immortal time bias, which led to an overestimation of the effects of metformin (47,48). This concern was illustrated in a meta-analysis where the pooled RR from 6 studies that were conducted without immortal time bias found no association between metformin and mortality in patients with pancreatic cancer (RR 0.93; 95% CI 0.82–1.05), whereas pooled RR from 9 studies with immortal time bias showed a biased protective association between use and pancreatic cancer mortality (HR 0.76; 95% CI 0.69–0.84) (59). Similarly, prior observational studies with time-related biases demonstrated significantly decreased prostate cancer mortality associated with metformin use (60,61). When these biases were addressed, metformin use was not found to be associated with prostate cancer mortality (RR 1.09; 95% CI 0.51–2.33) (62). Immortal time bias was also avoided in a reanalysis of data from a phase 3 randomized trial of individuals with resected stage 3 colon cancer treated with adjuvant FOLFOX (folinic acid, 5-flurouracil, and oxaliplatin)–based chemotherapy randomized to be treated with or without cetuximab by measuring metformin use at baseline, not after (63). This analysis found no association between metformin use and recurrence (HR 0.87; 95% CI 0.56–1.35) or overall survival (HR 0.99; 95% CI 0.65–1.49).

Unfortunately, immortal time bias in metformin–cancer observational studies is still prevalent today. A recent observational study of the effect of metformin use on colorectal cancer patients with type 2 diabetes, using data on 290 patients from a clinical service, reported that metformin use was associated with a significant 55% reduction in mortality and 69% reduction in recurrence over 2 years of follow-up (64). By defining exposure to metformin as “taking oral metformin for at least 90 days during the follow-up period after surgery,” the study will generate at least 90 days of immortal time, as the patients must remain alive for this duration to be classified as metformin users, in addition to the time from surgery to metformin initiation, resulting in significant immortal time bias. Several other observational studies using designs affected by immortal time bias were published in 2022. These studies also reported significant reductions in mortality with metformin among patients treated for a variety of cancers, such as lung, breast, colorectal, prostate, or pancreatic cancer (65), gastric cancer (66), and prostate cancer (67,68).

The time-related biases identified in many observational studies that reported remarkable benefits of metformin for cancer, especially immortal time bias, are easily correctable without additional data. The studies that used a proper approach to study design, which classifies metformin exposure correctly over time, did not find such benefits. Besides these biases, a major challenge in conducting observational studies is the problem of confounding bias, which is intrinsically addressed by the random allocation in randomized trials.

New-user cohort designs provide an approach to observational studies that inherently avoids immortal time bias and other time-related biases. The incident new-user cohort design, also called active-comparator new-user design, identifies new users of metformin and new users of a comparator drug, including only patients who were treatment naive to both drugs (69,70). For example, sulfonylurea could be used as an active comparator for metformin studies on cancer, as they have not been associated with cancer incidence or prognosis (71). To reduce the effects of confounding, the two groups can be matched on the propensity score of initiating metformin versus initiating sulfonylurea as a function of covariates measured prior to initiation of the two drugs, thus mimicking a randomized trial (Fig. 2) (72–74). Indeed, the use of a propensity score has been shown to balance covariates between treatment groups (75).

Cohort entry is taken as the date of the first metformin prescription and of the matched comparator, with follow-up until cancer occurrence or cancer outcomes of interest. This incident new-user cohort design has been used extensively to study the effectiveness and safety of several antidiabetes drugs. Some examples include studies that evaluated the risk of cardiovascular events and of severe hypoglycemia associated with initiating treatment of type 2 diabetes with sulfonylurea compared with metformin (76,77) as well as the risk of acute pancreatitis associated with incretin-based drugs (78). This approach was also used to assess the risk of several outcomes associated with the use of sodium–glucose cotransporter 2 inhibitors (SGLT2i) (79,80). It was used to assess cancer incidence associated with the initiation of glargine versus human NPH insulin, using inverse probability of treatment weights estimated from propensity scores to adjust for confounding (81). Finally, the incident new-user cohort design was also used to successfully predict the findings of randomized trials before they were completed, such as the Cardiovascular Outcome Study of Linagliptin Versus Glimepiride in Patients With Type 2 Diabetes (CAROLINA), which compared linagliptin with glimepiride in patients with type 2 diabetes (82,83).

In observational studies where the comparison is with nonusers rather than with active comparators, or when it is challenging to identify a valid active comparator, the prevalent new-user design is an appropriate approach (81). It is notable that most observational studies of metformin and cancer incidence and outcomes affected by immortal time bias compared metformin users with nonusers. Indeed, comparisons of new users of metformin with nonusers presents a particular challenge regarding time zero: while new users of metformin start follow-up at the time of initiation, it is unclear where follow-up should start for nonusers. The prevalent new-user design addresses this issue by first forming a base cohort of all subjects with an indication for metformin such as type 2 diabetes (84).

In lieu of matching, various methods based on propensity score weighting have also been used to further adjust for confounding when conducting pharmacotherapy studies (85). Computer programs to conduct propensity score analyses have been developed in a number of statistical software packages such as R (86), SAS (87), and Python (88). Although propensity scores have been shown to reduce confounding in observational studies, there are limitations to consider, including the possibility of unmeasured confounders that were not taken into account in the propensity score (89) and the limited ability of propensity scores to estimate the effects of time-varying exposures (90).

Each new user of metformin is then matched to subjects who belong to the same exposure set as the metformin user, namely those subjects who were not initiated on metformin at the time the user started but had a physician visit, perhaps receiving another treatment. The matching is done on time-conditional propensity scores computed as a function of time-dependent covariates measured at the time of the exposure set and with conditioning on the exposure set. This design can be thought of as mimicking a placebo-controlled randomized trial (Fig. 3). Cohort entry is then taken as the date of the first metformin prescription and the corresponding date for the matched nonuser comparator, with follow-up until cancer occurrence or other outcomes. The prevalent new-user design can also be effective in comparisons of new users of a study drug with users of an older drug, especially if incident new users of these drugs are scarce (73,84). For example, in comparisons of the newer glucagon-like peptide 1 receptor agonists (GLP-1 RA) with the older sulfonylureas on the risk of heart failure, a cohort study found that 75% of the 6,196 new users of GLP-1 RA had previously used a sulfonylurea (84). An incident new-user design would therefore include only 1,633 new users of GLP-1 RA. On the other hand, the prevalent new-user design used all subjects, matching each of the 6,196 GLP-1 RA new users to a sulfonylurea user on the time-conditional propensity scores, finding an HR of heart failure with GLP-1 RA use of 0.73 (95% CI 0.57–0.93). This finding was supported by a subsequent meta-analysis of several randomized trials (HR 0.89; 95% CI 0.82–0.98) (91). This design was particularly useful in evaluating the risk of cardiovascular and hypoglycemic events associated with adding or switching to sulfonylureas compared with remaining on metformin monotherapy in patients with type 2 diabetes treated initially with metformin (92).

The prevalent new-user approach was used in a recent study that assessed the association between SGLT2i use and the risk of major adverse cardiovascular event (MACE) outcomes in individuals with type 2 diabetes (93). When assessing this risk, a suitable clinically relevant comparator drug was dipeptidyl peptidase 4 inhibitors (DPP-4i), given that DPP-4i are prescribed at the same level in the treatment paradigm as SGLT2i for type 2 diabetes management. In this case, the prevalent new-user design first involved matching new SGLT2i users with new DPP-4i users (incident users). Second, the subjects who initiated SGLT2i after switching from DPP-4i treatment were matched to individuals who were treated with DPP-4i for the same duration and were still DPP-4i users in the exposure set (prevalent users). By using this approach, the study population included 103,797 pairs of incident new users and 106,070 pairs of prevalent new users, thus increasing the power substantially. This study found SGLT2i use was associated with a decreased risk of MACE outcomes (HR 0.76, 95% CI 0.69–0.84) (93), with an estimate similar to the 20% reduction in MACE reported in randomized trials (94).

Alternative approaches to study design that avoid time-related biases can be used to address these questions (74,95). For example, observational study designs that emulate a target trial (95) and use of an appropriate comparator group (74) are two approaches that reduce time-related biases. Other approaches include marginal structural models, which adjust for time-dependent confounding and allow estimation of the effect of time-varying exposures (96). Finally, other approaches, such as targeted maximum likelihood estimator, use nonparametric machine learning to estimate the treatment and outcome model. This approach has been shown to decrease bias and SE in observational studies (75). Overall, these approaches have been shown to reduce bias and confounding compared with adjusting covariates in multiple regression analyses (75).

The quest to identify metformin as an agent for cancer prevention or treatment, which began in 2005 with an observational study that reported a reduction in the incidence of cancer with metformin use, generated a considerable amount of basic, clinical, observational, and experimental research (2). Many of these observational studies were shown, as early as 2012, to be affected by time-related biases, which result in an exaggeration of the benefit of metformin on cancer incidence and outcomes (47). Subsequent observational studies that avoided these time-related biases found no such cancer benefits with metformin use. These null findings were confirmed by all randomized trials conducted so far, which found no beneficial effects of metformin as adjuvant therapy on cancer outcomes, including the largest and longest in early breast cancer (1).

Nonetheless, the effects of metformin studied in the setting of cancer treatment in observational studies and RCTs may have been affected by the variations in studied populations. Indeed, observational studies assessed individuals with type 2 diabetes and cancer, whereas most RCTs studied individuals without underlying diabetes. As such, the clinical benefits for cancer outcomes associated with metformin use may be different among individuals with and without underlying diabetes. Furthermore, the anticancer effect of metformin may differ for certain cancers at the tissue and molecular levels. For example, in the recent large RCT for metformin use and breast cancer, women with human epidermal growth factor receptor 2 (ERBB2)-positive breast cancer with any C allele of the rs11212617 single nucleotide variant had an increased pathological response when metformin was given in combination with chemotherapy and ERBB2-targeted therapy compared with women not given metformin treatment (1). Similarly, the effects of metformin may vary by metabolic phenotype. BMI may modulate the clinical benefits associated with metformin use and cancer outcomes (97). However, this has not been well studied. Metformin may not have effects on all types of cancers, as certain cancers may not be responsive to insulin levels (28). The effects that metformin may have on carcinogenesis may also differ depending on the cancer stage when metformin is initiated. Thus, further studies of metformin treatment for cancer should focus on such targeted questions.

The acceptance of real-world evidence by regulatory agencies and the direct access to large health care databases has led to a proliferation of observational studies that assess the real-world effects of medications not only as indicated but also to identify new indications. As a result, several initiatives were launched to assess the nearness of results on the effectiveness of medications between randomized trials and observational studies (98,99). As such, the field of diabetes treatment has been one of the most prolific recipients of observational studies that employ the new-user cohort designs we described, including the use of propensity score matching or weighing and exploiting real-world data to assess the effectiveness and safety of older and newer treatments as well as to accurately predict the findings of ongoing trials.

As for the repurposing of metformin to prevent and treat various cancers, the many erroneous observational studies affected by time-related biases led to a 17-year quest whereby large, randomized trials found no benefit in cancer outcomes associated with metformin treatment. Regrettably, while time-related biases were known in 2012 to have affected the early observational studies, it is surprising that, in 2022, several studies still use approaches susceptible to immortal time bias, continuing to suggest remarkable benefits of metformin as a treatment for various cancers. A November 2022 search of ClinicalTrials.gov on the terms “cancer” and “metformin” revealed around 109 registered ongoing clinical trials that were studying the use of metformin in the treatment of various cancers. While the evidence to date suggests that metformin does not provide significant benefits in reducing cancer incidence and outcomes, further research, if any, should target specific promising phenotypic or genotypic subgroups.

Funding. S.S. is the recipient of the Distinguished James McGill Chair award.

Duality of Interest. S.S. attended scientific advisory committee meetings or received speaking fees from AstraZeneca, Atara Biotherapeutics, Boehringer-Ingelheim, Bristol Myers Squibb, Merck, Novartis, Panalgo, Pfizer, and CSL Seqirus. O.H.Y.Y. attended a scientific advisory board meeting for Novo Nordisk. No other potential conflicts of interest relevant to this article were reported.

Author Contributions. Both authors were involved in the conception, design, and conduct of the study and the analysis and interpretation of the results. O.H.Y.Y. wrote the first draft of the manuscript, and both authors edited, reviewed, and approved the final version of the manuscript. S.S. is the guarantor of this work and, as such, had full access to all the data in the study and takes responsibility for the integrity of the data and the accuracy of the data analysis.