OBJECTIVE: To determine whether results from a diabetes progression model were consistent with electronic medical records for UK patients with suspected diabetes. METHODS: A data driven simulation was conducted using an existing stochastic model of diabetes progression. The model uses UKPDS equations to calculate annual transition probabilities to death and in five health dimensions (neuropathy, nephropathy, retinopathy, CHD, and stroke). Although equations in multiple dimensions may include the same factors (e.g. blood pressure, A1c), transitions in the various dimensions are calculated independently. The database contained over 100 million encounter, patient or therapy records for 183,119 patients with suspected diabetes between 1982 and 2005. Initial validation was attempted by creating a cohort of patients from the database for whom gender, birth year, diagnosis date, A1c, and blood pressure were available. Any diagnosis in the five health dimensions, prior to the diabetes diagnosis, was noted to assign non-zero levels to the simulated patients' starting state. After initial poor fit, more rigorous cleaning was done, the time frame was limited, and A1c was imputed from blood glucose values when possible. RESULTS: Although the fit was adequate for most events in the health dimensions, the model predicted far more deaths than occurred in the cohort from the database. Compared to patients without A1c measurements in the database, those with A1c had 0.4 relative risk of death. The median birth year was eight years later for those with an A1c test. Moreover, the proportion of patients with an A1c was < 6% through 1992, from 11% to 26% from 1993-8, rising to 65% in 2001, and exceeding 90% the last two years. CONCLUSION: There are strong temporal interactions between year of birth or diagnosis and A1c testing rate. Modelers should consult ISPOR task force reports on retrospective databases before assembling cohorts from longitudinal databases.