Learn more about Peter Attia’s digital program
In 2014, a retrospective study was published by Bannister et al. which laid the foundation for metformin – traditionally a first-line treatment for type II diabetes – to be regarded as a potential “anti-aging” drug for the general population. The authors of the U.K.-based study reported that diabetic patients on metformin monotherapy were found to have lower mortality than non-diabetic controls, despite the obvious baseline health bias favoring the control group. This remarkable finding suggests that this drug has benefits for lifespan beyond management of diabetes – benefits which could improve longevity for anyone, diabetic or not. The results certainly grabbed popular attention and have come up a number of times on The Drive (for instance, in my first and second interviews with Dr. Nir Barzilai).
But the study has also had its share of critics, and concerns over biased methodology led a research group in Denmark to reassess the 2014 findings with an independent cohort study of similar design. The new results by Keys et al., which were published in the International Journal of Epidemiology this past fall, failed to replicate the broad protective effects of metformin on lifespan. So how were these studies different? And which, if either, should we believe?
Criticism of Bannister et al.
Diabetes is a progressive disease which generally requires increasingly aggressive treatment over time. This means that most patients who start on metformin monotherapy will eventually need to add other, complementary medications to maintain adequate glycemic control. But diabetic patients in the Bannister et al. study who followed this typical progression to multitherapy would, as a result, have no longer met the criteria for the metformin monotherapy group and thus would have been excluded from analysis. The metformin group therefore would have selected only for patients whose diabetes did not progress to a level requiring additional medications during the follow-up period (2.8 years on average).
This is not simply saying that the metformin group was biased toward diabetic patients who were healthiest at baseline. Those who became progressively sicker would also have been much more likely to die from diabetes complications over the course of the study. Thus, the final metformin group was defined by a variable that related directly to the survival outcome of interest (a source of bias known as informative censoring), largely negating the longevity disadvantage conferred by the presence of diabetes. (An analogy: if we designed a survival study on smokers vs. non-smokers and decided that anyone in the smoking group who died of cancer or lung disease was disqualified, smoking would probably appear to confer a survival advantage simply because we’ve eliminated the most relevant causes of death for smokers.)
An additional – though likely less significant – criticism of Bannister et al. is that the non-diabetic control group may have been biased toward a less health-conscious population. According to the study, controls were matched to treatment groups in age, gender, prior cancer status, and smoking status and were eligible if they had never been exposed to metformin and had no diabetes diagnosis prior to or throughout the study. But an absence of diagnosis can be caused by two things: an absence of disease or an absence of medical attention to diagnose a disease. Those who are less likely to have regular health check-ups or seek attention for symptoms may be more likely to have undiagnosed diabetes or other undiagnosed morbidities. This potential source of bias tends to be most impactful when the disease in question is very prevalent in a society, as is the case with type 2 diabetes in the U.K.
What did Keys et al. do differently?
Keys et al. sought both to replicate Bannister et al.’s methods and to adopt strategies to estimate and reduce the potential impact of bias. These strategies included performing sensitivity analyses to determine the robustness of results to the presence or absence of informative censoring based on treatment regimen: in one version of the analysis, patients were excluded from the metformin monotherapy group if they changed treatments or advanced to additional medications (as in the Bannister et al. study), while in another analysis, these patients remained included in analyses. The authors also conducted an analysis of a same-sex twins discordant for metformin use – a more robust and tightly matched variation on the standard matched cohort study design.
In their replication of the methodology employed for the 2014 study, Keys et al. found significantly higher mortality among diabetic patients on metformin monotherapy than among non-diabetic controls (HR = 1.39, 95% CI: 1.22 – 1.57). This effect – the inverse of the effect reported by Bannister et al. – was also observed when the authors omitted informative censoring of the metformin group based on changes to treatment regimen (HR = 1.48, 95% CI: 1.32 – 1.64), though, as expected, the effect was stronger in this case. Metformin was likewise associated with higher mortality in the analysis of twins discordant for metformin use (HR = 1.80, 95% CI: 1.11 – 2.91).
How reliable are these data?
It might be tempting to view the additional analyses conducted by Keys et al. as proof that their results are reliable and represent the “right answer” to the question of metformin’s effects on longevity, but it’s important to remember that both this and the Bannister et al. study were strictly observational, meaning that both were subject to potential confounder effects. Exacerbating this problem, their retrospective designs prohibited collection of certain critical covariate data, such as participants’ BMIs. While the Keys et al. discordant twin analysis may be assumed to correct for some confounds (particularly those related to prenatal and childhood environments), it cannot account for all possible variables relevant to survival, as evidenced by the very fact that twin pairs could be discordant for diabetes status.
A related limitation of both studies is the fact that they compared inherently unequal groups: those with diabetes on metformin monotherapy vs. those without diabetes. In other words, the groups were defined by a variable other than metformin use (i.e., diabetes status) that was likely to impact the readout variable (i.e., survival). So even if metformin did improve longevity, the effect could conceivably be masked by the opposing effect of diabetes or other health-related confounds.
Additionally, though Keys et al.’s conducted a sensitivity analysis to determine the impact of (and correct for) informative censoring, we must take results of this analysis with an enormous grain of salt due to one critical flaw. In omitting censoring of participants based on changes to treatment regimen, Keys et al. not only retained patients who added medications to a metformin regimen but also patients who stopped metformin altogether. This leaves us uncertain about whether the metformin group was even truly receiving metformin for any significant period of time. Further, it’s unclear how this problem might affect results, as regimen changes might indicate either an improvement or worsening in disease management.
Looking forward
So how can we gain more reliable insight into metformin’s effects on longevity?
A randomized, controlled trial is the only way to approach this question without the plague of endless confounds, particularly those related to diabetes. Dr. Barzilai is currently leading efforts to conduct such a trial in individuals without type II diabetes, thereby assessing the drug’s aging effects independently of its effects on glycemic control. The Targeting Aging with Metformin (TAME) trial, which he and I have discussed in detail on the podcast, aims to elucidate how metformin impacts lifespan and a host of chronic diseases, including cardiovascular disease, cancer, and dementia.
The TAME trial was in part inspired by the remarkable results by Bannister et al.. While the recent Keys et al. study certainly casts doubt over those 2014 findings, the new data aren’t enough to put the question to bed and don’t nullify the rationale for TAME. In the “Ask Me Anything” episode coming out next week, I discuss in more detail the strengths and weaknesses of existing evidence for metformin as a geroprotective molecule and why, until we have results from a randomized clinical trial, its efficacy in extending lifespan is likely to remain a topic of speculation more than fact.
For a list of all previous weekly emails, click here.
