Fig. 1. Flowchart of data cleaning and use.

^aExcluded because glucose metabolism changes during pregnancy. ^bData from the first available measurement were used for these participants. ^cSome surveys only measured glycemic biomarker on a subset of participants for logistic or budget reasons. ^dExcluded because glycemic measurements in these participants were systematically different from the rest from the same study, possibly because the specific area had high prevalence of thalassemia⁹⁴. ^eExcluded because such values are more likely to be due to data recording error than values within the range. ^fWe removed participants for implausible pairs of FPG and HbA1c using the method of local outlier factor (LOF)⁹⁵. This approach detects data combinations that are extremes in the joint density of the variable pairs (for example, a participant with FPG of 5 mmol l⁻¹ and HbA1c of 17%, or with FPG of 28 mmol l⁻¹ and HbA1c of 5%). We identified extremes as those measurements whose measure of local density by LOF method is less than half of the average of their 100 nearest neighbors. ^gIncluding all 2,436 participants from four studies that did not measure BMI. ^hIncluding all 3,455 participants from four studies in which all individuals without previously diagnosed diabetes had FPG < 7.0 mmol l⁻¹ and HbA1c < 6.5%.