Metformin Start Did Not Lower Severe COVID Risk in Newly Diagnosed Type 2 Diabetes
A large English cohort study has pushed back against the idea that starting metformin offers meaningful protection against COVID-19 outcomes in people with newly diagnosed type 2 diabetes. Using linked NHS electronic health records from OpenSAFELY-TPP, investigators compared adults who began metformin monotherapy within six months of diagnosis with those who started no antidiabetic drug during that period. The analysis was built around a new-user design, a choice meant to avoid the bias that can creep in when long-term medication users are compared with heterogeneous untreated controls.
The central result was not a benefit signal. For the composite outcome of COVID-19-related hospitalisation or death, the maximally adjusted hazard ratio was 1.17 with a 95% confidence interval of 1.02 to 1.35 in the intention-to-treat analysis. Over two years, that translated into a risk difference of 0.36%, with a 95% confidence interval of 0.09 to 0.62. The per-protocol estimate was closer to null, with a hazard ratio of 1.11 and a 95% confidence interval of 0.96 to 1.29.
The pattern is hard to reconcile with a clinically important preventive effect of metformin on severe SARS-CoV-2 outcomes in this setting.
What the OpenSAFELY Cohort Actually Compared
The study included 52,341 people aged 18 to 85 years who were newly diagnosed with diabetes between August 2018 and August 2019. Among them, 19,599, or 37.4%, started metformin within six months. Follow-up began six months after diagnosis and before the UK pandemic, which helped avoid immortal time bias and created a cleaner baseline for comparing later COVID outcomes.
This design matters because metformin has repeatedly looked favorable in some observational studies, but many of those earlier analyses were built on prevalent-user comparisons. In that setting, people who have already tolerated treatment for years can differ in important ways from controls who never received the drug. Survival to treatment initiation, differences in disease severity, prescribing patterns, and healthcare contact can all distort the apparent effect of a medication. A new-user framework reduces some of that noise, though it cannot erase unmeasured confounding.
The investigators also used negative and positive control outcomes to test whether residual bias was still present. Bone fracture served as the negative control outcome and produced a null estimate, with a hazard ratio of 1.00. Diabetes-related death served as the positive control and came back elevated, with a hazard ratio of 1.41 and a 95% confidence interval of 1.09 to 1.81. That mismatch points toward unmeasured confounding still shaping the estimates.
Long COVID Remained Uncertain, but Not Reassuring
Long COVID was also examined, although the estimate was imprecise. The intention-to-treat hazard ratio was 1.39, with a 95% confidence interval of 0.83 to 2.33. Wide intervals like these do not support a clear claim either way, but they do make it difficult to argue that starting metformin meaningfully reduces the burden of post-acute sequelae.
For clinicians, the bedside question is straightforward: should metformin initiation be expected to alter later COVID risk in newly diagnosed diabetes? This dataset does not support that expectation. The direction of effect actually leaned the other way, though the elevated estimates are more plausibly explained by unmeasured differences between treatment groups than by harm from the drug itself. People selected for metformin may differ in glycaemic trajectory, comorbidity burden, clinical follow-up, or early disease severity in ways not fully captured in routine records.
Why Earlier Observational Signals Looked Better
Metformin has a biologically plausible reputation in infectious disease research. It can influence insulin resistance, inflammatory signaling, and metabolic stress, pathways that could conceivably affect host response to SARS-CoV-2. In theory, stabilizing metabolic dysfunction might blunt some of the downstream inflammatory cascade that drives severe COVID-19. Yet biologic plausibility is not the same as a measurable clinical effect.
Randomised evidence has already suggested little or no mortality benefit from metformin in COVID-19. This preprint aligns more closely with that trial-level picture than with the more optimistic observational literature. The difference likely reflects study design rather than a reversal of metformin’s pharmacology. When long-term users are compared with nonusers in nonrandomised data, the healthiest survivors and the most treatment-engaged patients often cluster in ways that can mimic benefit.
The OpenSAFELY analysis tried to address exactly that problem by focusing on initiation rather than historical exposure. Even so, the positive control outcome suggests the residual bias did not disappear. For that reason, the modestly higher hazard seen for severe COVID should not be read as a harmful drug effect. It is better interpreted as a warning that routine data can still leave important confounding on the table.
What This Means for Practice While the Evidence Matures
For adults with newly diagnosed type 2 diabetes, metformin remains a standard glucose-lowering first-line option for metabolic reasons, not as a COVID-prevention strategy. This study offers no support for starting metformin with the expectation of reducing hospitalisation, death, or later Long COVID. The findings also weaken the case for repurposing metformin specifically to prevent severe SARS-CoV-2 outcomes in this population.
The study’s strengths are substantial: population scale, linked primary care and hospital data, a new-user design, and explicit negative and positive control outcomes. The main limitation is also familiar in pharmacoepidemiology: routine data cannot fully capture all the factors that influence who starts treatment and when. Severity of diabetes at baseline, clinical frailty, physician prescribing preferences, and other unmeasured differences may still distort the estimate.
The most useful takeaway is narrow and practical. In this English cohort, starting metformin did not reduce severe COVID-19 risk, did not clearly improve Long COVID outcomes, and produced signals that fit better with confounding than with a preventive drug effect. The authors’ interpretation is conservative for good reason: the result belongs in the growing pile of evidence showing that earlier observational enthusiasm for metformin and COVID-19 likely outpaced the more reliable data.