An A1c estimates average glucose. For some people the estimate is systematically off.

An A1c result is the percentage of haemoglobin in the blood that has become glycated — chemically bound to glucose — and it is used as a measure of chronic glycaemia, the average blood sugar over the preceding months [s1]. It is not a direct glucose measurement, and the conversion between the two is a regression line fitted to a specific set of people, which means it can be systematically wrong for people whose red blood cells behave differently.

Where the conversion comes from

The equation that turns an A1c into an "estimated average glucose" was derived from the A1C-Derived Average Glucose study, published in Diabetes Care in 2008 [s1]. It enrolled 507 subjects across 10 international centres — 268 with type 1 diabetes, 159 with type 2, and 80 without diabetes [s1].

Each participant's A1c, measured in a central laboratory at the end of three months, was compared against their average glucose over those same three months [s1]. That average was not estimated loosely: it combined weighted results from at least two days of continuous glucose monitoring performed four times, plus seven-point daily fingerstick self-monitoring on at least three days per week, producing roughly 2,700 glucose values per subject [s1].

Linear regression gave the tightest fit at AG (mg/dL) = 28.7 × A1C − 46.7, with R² = 0.84 and P < 0.0001 [s1]. The study reported that the regression equations did not differ significantly across subgroups defined by age, sex, diabetes type, race/ethnicity or smoking status, and concluded that A1c can be expressed as an estimated average glucose for most patients with type 1 and type 2 diabetes [s1].

Two things in that sentence do the work. "Most patients" is a real qualifier. And an R² of 0.84 means about 16% of the variation in average glucose is not explained by A1c — a good fit for a population instrument, and a looser one for an individual reading.

Sickle cell trait shifts the reading

Sickle cell trait is carried by a substantial minority of people of African ancestry and, unlike sickle cell disease, causes no symptoms. It does affect this measurement.

A retrospective cohort study published in JAMA used data from two community-based cohorts, the Coronary Artery Risk Development in Young Adults study and the Jackson Heart Study, covering 7,938 participants; the analytic sample was 4,620 people with a mean age of 52.3 years, 61.3% women, of whom 367 (7.9%) had sickle cell trait, contributing 9,062 concurrent measures of fasting glucose and A1c [s2].

For a given fasting glucose, A1c values were significantly lower in those with sickle cell trait than without — 5.72% against 6.01%, a mean difference of −0.29 percentage points (95% CI −0.35 to −0.23) [s2]. Using two-hour glucose measurements instead gave 5.35% against 5.65%, a difference of −0.30 percentage points (95% CI −0.39 to −0.21) [s2]. The gap was larger at higher glucose concentrations [s2].

The consequence is visible in classification. Defined by A1c, prediabetes was present in 29.2% of observations from participants with sickle cell trait against 48.6% without, and diabetes in 3.8% against 7.3% [s2]. The authors conclude that A1c may systematically underestimate past glycaemia in Black patients with sickle cell trait [s2].

And a second, separate shift in the other direction

A prospective 12-week observational study across 10 US diabetes centres compared 104 Black and 104 White participants aged 8 or older who had had type 1 diabetes for at least two years, with A1c between 6.0% and 12.0%, measuring mean glucose by continuous glucose monitoring [s3].

Mean A1c was 9.1% in Black participants and 8.3% in White participants [s3]. For a given A1c, mean glucose was significantly lower in Black participants (P = 0.013), which the authors express the other way round: for a given mean glucose, A1c ran 0.4 percentage points higher in Black participants (95% CI 0.2 to 0.6) [s3].

Two alternative measures of glycaemia, glycated albumin and fructosamine, showed no significant racial differences in their relationship with mean glucose (P > 0.20 for both) [s3] — which points toward haemoglobin glycation itself rather than glucose control as the source of the gap.

The authors are careful about what this does and does not license. They note that race only partially explains the observed A1c differences between Black and White people with diabetes, and call for research into the barriers to improved glycaemic control rather than treating the measurement artefact as the whole story [s3]. The study also had too few participants with A1c below 6.5% to generalise to that range [s3].

Note that these two findings point in opposite directions and were measured in different populations. Sickle cell trait lowers A1c relative to glucose [s2]; the trial in type 1 diabetes found A1c running higher for a given glucose in Black participants overall [s3]. They are not contradictory — the first is a specific genetic effect, the second a population-average difference that was not resolved into its causes — but they illustrate that "A1c is inaccurate in group X" is too crude a summary in either direction.

What the number is good for

A1c remains the measure used to judge the adequacy of diabetes treatment and adjust therapy, and the ADAG study exists precisely because clinicians wanted a way to express it in the same units patients see on a glucose meter [s1]. Its strengths are that it does not require fasting, is not thrown off by a single bad day, and integrates months of exposure into one figure.

Its limits follow from the same properties. It is an average, so it conceals variability — the same A1c can come from stable glucose or from wide swings. It depends on the lifespan and behaviour of red blood cells, which is why haemoglobin variants shift it [s2]. And the conversion to an average glucose carries the residual scatter of the original regression [s1].

Interpreting a particular A1c — including whether an alternative measure such as glycated albumin or fructosamine is more appropriate for a given person — is a clinical judgement that depends on information a number alone does not carry.