• Paul J Christine, Rebekah Young, Sara D Adar, Alain G Bertoni, Michele Heisler, Mercedes R Carnethon, Rodney A Hayward, and Ana V Diez Roux.
• Department of Epidemiology, University of Michigan School of Public Health, Ann Arbor, Michigan. Electronic address: pjchris@umich.edu.
• Am J Prev Med. 2017 Aug 1; 53 (2): 201209201-209.

IntroductionThe purpose of this study was to evaluate if adding SES to risk prediction models based upon traditional risk factors improves the prediction of diabetes.MethodsRisk prediction models without and with individual- and area-level SES predictors were compared using the prospective Multi-Ethnic Study of Atherosclerosis. Cox proportional hazards models were utilized to estimate hazard ratios for SES predictors and to generate 10-year predicted risks for 5,021 individuals without diabetes at baseline followed from 2000 to 2012. C-statistics were used to compare model discrimination, and the proportion of individuals reclassified into higher or lower risk categories with the addition of SES predictors was calculated. The accuracy of risk prediction by SES was assessed by comparing observed and predicted risks across tertiles of the SES variables. Statistical analyses were performed in 2015-2016.ResultsOver a median of 9.2 years of follow-up, 615 individuals developed diabetes. Individual- and area-level SES variables did not significantly improve model discrimination or reclassify substantial numbers of individuals across risk categories. Models without SES predictors generally underestimated risk for low-SES individuals or individuals residing in low-SES areas (underestimates ranging from 0.31% to 1.07%) and overestimated risk for high-SES individuals or individuals residing in high-SES areas (overestimates ranging from 0.70% to 1.30%), and the addition of SES variables largely mitigated these differences.ConclusionsStandard diabetes risk models may underestimate risk for low-SES individuals and overestimate risk for those of high SES. Adding SES predictors helps correct this systematic misestimation, but may not improve model discrimination.Copyright © 2017. Published by Elsevier Inc.

20 keywords...

hide…