Time-Related Bias Changes the Metformin Signal in Breast Cancer Research
Metformin has long attracted attention as more than a glucose-lowering drug. Laboratory studies, observational cohorts, and mechanistic hypotheses have all fed the idea that it might also blunt cancer risk, especially in hormone-sensitive tumors such as breast cancer. This new systematic review and meta-analysis forces a harder look at that narrative by separating studies that dealt properly with time-related bias from those that did not.
The difference is not academic. Immortal time bias and time-window bias can make a treatment appear protective simply because the exposed group had to survive long enough to be classified as exposed. In a cancer-prevention question, where the outcome may develop years after drug initiation, those design flaws can easily create a false sense of benefit.
Investigators searched PubMed and earlier meta-analysis reference lists for cohort and case-control studies reporting breast cancer-specific estimates for metformin use and incident breast cancer. They identified 368 PubMed records and 28 additional reports, then narrowed the evidence to 22 reports from 15 source populations. One report per source population was pooled, and 12 independent studies entered the main random-effects analysis.
The pooled relative risk was 0.89, with a 95% confidence interval of 0.74 to 1.07 and substantial heterogeneity across studies. On its face, that point estimate still hints at a modest reduction. Once the studies were stratified by design quality, however, the pattern changed sharply.
In nine studies that addressed time-related biases through a new-user design, time-varying exposure, or target trial emulation, the relative risk was 1.00, with a 95% confidence interval of 0.94 to 1.08 and no heterogeneity. In three studies that did not address those biases, or did so unclearly, the relative risk was 0.69, with a 95% confidence interval of 0.56 to 0.84. The between-group difference was statistically significant.
That abrupt split is the central clinical finding.
Why Better Epidemiology Erases the Apparent Protective Effect
Breast cancer prevention studies are especially vulnerable to misclassification of drug exposure over time. Patients must first remain event-free long enough to be counted as exposed, and that interval can create an artificial survival advantage if analysis does not properly assign person-time. New-user designs reduce this problem by anchoring follow-up at treatment initiation, while time-varying exposure models prevent earlier unexposed time from being incorrectly credited to the drug.
The authors then tested how stable the signal remained when one ambiguous study was reclassified. Once that study was moved into the bias-adjusted group, the statistical difference between strata disappeared. The result weakens any claim that metformin itself lowers incident breast cancer risk, at least on the basis of currently available observational data.
Mechanistically, that is not surprising. Metformin affects hepatic gluconeogenesis, insulin signaling, and downstream metabolic pathways that could, in theory, alter tumor promotion. Yet a plausible mechanism does not rescue a biased epidemiologic signal. If the observed protection comes from study design artifacts rather than a biological effect, then the laboratory rationale may be true in principle and still fail in population-level practice.
The practical consequence reaches beyond metformin alone. Clinicians and researchers looking for chemopreventive effects in common metabolic drugs must demand methods that align exposure timing with risk time. Without that discipline, a favorable relative risk may reflect analytic convenience more than actual cancer prevention.
Imeglimin Has No Clinical Breast Cancer Dataset Yet
The review also examined imeglimin, a structurally related glucose-lowering agent approved in Japan in 2021. Interest in imeglimin as a cancer-preventive candidate rests largely on molecular resemblance to metformin, but structural similarity does not guarantee shared cancer biology.
Here the evidence base was far thinner. A PubMed search returned 218 records and 10 registered studies, yet no clinical study reported breast cancer or any other cancer outcome. No meta-analysis could be performed because no human outcome data exist.
What remains is preclinical inference, and even that is weak. The available evidence consisted of one in vitro study in breast cancer cells, one docking study, and models of other cancers using concentrations at least about 40-fold above clinical plasma levels. Those findings do not establish a preventive role in humans. They only show that the hypothesis has been explored in the laboratory.
A drug can look promising in silico or in a petri dish and still do nothing in patients. For cancer prevention, exposure duration, biologic plausibility, and real-world event capture all have to line up before a signal becomes credible.
What a Credible Study Would Need to Show
The investigators estimated that a comparative cohort would need about 631 breast cancer events to detect a hazard ratio of 0.80. That number gives a sense of how much clinical experience is required before a modest preventive effect can be judged with any confidence. Short follow-up after imeglimin’s approval makes that threshold unreachable for now.
A future study would need to start with new users, use an active comparator, and handle time at risk explicitly. It would also need enough breast cancer events to distinguish a true preventive effect from background noise and treatment-selection bias. Without those features, the same methodological pitfalls that clouded the metformin literature could easily recur.
For metformin, the bottom line is more restrained than the popular narrative once suggested: in studies that addressed time-related biases, use of the drug was not associated with incident breast cancer. For imeglimin, the evidence base is too early and too indirect to support any breast cancer conclusion at all.
The remaining question is not whether these drugs are interesting. It is whether future cohorts can finally answer the prevention question with methods strong enough to survive scrutiny.