How much of GLP-1 blood-sugar improvement comes from weight loss?

Why simultaneous improvements do not establish mediation

Weight and blood sugar can improve during the same period without the weight change explaining all of the glucose change. Mediation analysis asks a narrower question: how much of an estimated treatment effect passes through a specified intermediate variable?

Hyojung Jang and colleagues chose body mass index, or BMI, as that intermediate variable. Their preprint examines the pathway from GLP-1 receptor agonist initiation through BMI change to glycated haemoglobin, or HbA1c. HbA1c is the study's glucose-control outcome. The paper also develops a way to estimate these pathways across hospitals without exchanging individual patient records.

What the 32,146-patient study compared

The application used the Indiana Network for Patient Care and included 32,146 adults with type 2 diabetes across ten hospitals. There were 8,687 initiators of GLP-1 receptor agonists and 23,459 initiators of other glucose-lowering drugs. The comparison therefore concerns treatment initiation in clinical records, with the treatment groups formed through care decisions rather than random assignment.

The authors estimated a total comparative HbA1c effect of about −0.29 percentage points. They divided that estimate into a direct component of −0.282 percentage points and a BMI-mediated component of −0.011 percentage points. The small difference between the sum and the rounded total reflects the displayed precision.

How to interpret the 3.7% BMI-mediated estimate

The model attributed 3.7% of the estimated comparative effect to the pathway represented by BMI. A statistical “direct effect” contains the pathways outside the chosen mediator. It does not identify one particular receptor process or prove that the remainder came entirely from insulin secretion.

BMI is also a limited description of body composition. An analysis with this single mediator cannot disentangle all changes in adiposity, food intake or metabolism. The result supports a question for further study: which other measured pathways account for the estimated effect in this setting?

For the underlying receptor biology, see our GLP-1, GIP and glucagon mechanism review. That biological background and this statistical decomposition answer different questions.

Why the confidence interval needs caution

The paper reports a 1.7–5.6% interval for the mediated proportion, but it also identifies a material uncertainty problem. Patients may appear in more than one hospital's records. Treating those records as independent can make confidence intervals too narrow.

In this application, 34.8% of patients had records across sites. Privacy constraints complicate linking those records. The authors explicitly discuss underestimated uncertainty when this dependence is ignored. The interval should therefore travel with that qualification whenever the finding is quoted.

The causal interpretation also depends on adequate adjustment for confounding and correctly specified models. A large patient count cannot remove bias from an unmeasured factor that affects treatment choice, BMI and HbA1c.

What the privacy-preserving method contributes

Combining hospital data usually raises a practical problem: a pooled analysis may require sharing records that each institution must keep locally. The proposed federated procedure instead exchanges low-dimensional summary statistics. The authors report that its estimates closely reproduced pooled-data estimates in simulations and the clinical application.

That agreement evaluates the distributed calculation. It does not establish that all assumptions behind a causal mediation interpretation are true. Privacy protection, statistical agreement and causal validity are separate properties that each need examination.

What a stronger follow-up would need

A useful follow-up would test the method in other healthcare systems and examine sensitivity to confounding, mediator measurement and cross-hospital record overlap. Measuring additional candidate mediators could clarify which pathways the BMI-only analysis leaves unresolved.

For readers comparing studies, the essential details are the population, alternative treatment, outcome scale and mediator. Keeping those details attached to the 3.7% estimate prevents a study-specific result from becoming a general claim about every GLP-1 drug or patient.