Univariate and bivariate admixture analyses of serum glucose and glycated hemoglobin distributions in a Jerusalem population sample.

Univariate and bivariate analyses of fasting glucose and glycated hemoglobin (HbA1c) levels and of glucose levels 2 hours after an oral glucose load test were performed in a random sample of the Jewish population of Jerusalem, aged 25-64 years. Using ln-transformed data, we found that a mixture of two distributions fits the glucose data significantly better than a single distribution in the age groups 25-44 and 45-64 years. The fasting glucose results indicate that 1.1% of subjects aged 25-44 and 3.7% of subjects aged 45-64 without known diabetes come from an upper distribution with mean values of 154 mg/dl and 224 mg/dl, respectively. Estimates from the analysis of glucose levels after a load test indicate that an additional 2.1% of younger subjects and 3.8% of older subjects belong to a minor distribution with a high mean glucose value. The frequency distribution of HbA1c is also bimodal in all age groups. Yet the bimodality indicates that only 0.1% and 2.3% of subjects in the two age groups, respectively, come from minor distributions with mean levels of 13.0-15.7%, compared with HbA1c values of 5.0% and 5.3% for the main distributions. Using glucose levels, we found that specificity rates are consistently greater than 99%, whereas sensitivity rates vary with age. The use of cutoff points suggested by the National Diabetes Data Group (140 mg/dl for fasting glucose level and 200 mg/dl for glucose level after an oral glucose load test) indicates a lower sensitivity rate in the younger subjects with a minimal improvement among older subjects. A mixture of bivariate log-normal distributions fitted to the fasting and 2-hour glucose levels in subjects aged 45-64 indicates a larger proportion (6.3%) belonging to the minor distribution compared with those obtained when a single variable is used. Yet this combined score shows a low specificity rate. No similar improvement in separating "normal" from "abnormal" subjects is achieved when a mixture of bivariate distributions is fitted to the glucose and HbA1c variables. Admixture of glucose and HbA1c distributions is demonstrated. Bivariate analysis of these distributions does not, however, provide better discrimination of putatively abnormal subjects than univariate analysis.