Conditional Quantiles Under Parametric Model Misspecification: Predicting Individualized Ranges of A1C Change

• Activity Number: 129 - Quantile and Nonparametric Regression Models
• Type: Contributed
• Date/Time: Monday, July 31, 2017 : 8:30 AM to 10:20 AM
• Sponsor: Section on Nonparametric Statistics
• Author(s): Sergey Tarima* and Bonifride Tuyishimire and Rodney Sparapani and Lisa Rein and John Meurer
• Companies: Medical College of Wisconsin and Medical College of Wisconsin and Medical College of Wisconsin and Medical College of Wisconsin and Medical College of Wisconsin
• Keywords: personalized medicine ; quantile regression ; model misspecification ; minimum variance ; minimum MSE

Abstract

We consider the problem of predicting the expected ranges of A1C following 3, 6, 9 and 12 months ahead of time conditional on person specific characteristics and their history of diabetes. We explore the performance of several conditional quantile modeling techniques: parametric and non-parametric quantile regressions, and the modeling approaches which combine both parametric and non-parametric approaches into a single estimating procedure. The combined modeling targeting the minimum mean squared error (MMSE) showed the best performance in simulation settings and thus justified its use with the A1C modeling. In addition, we split the dataset of predominantly 10,000 type 2 diabetes patients into equally sized two data sets for training and test purposes. We used the multinomial likelihood as a goodness of fit measure to compare the performance of different quantile regressions on the test dataset, and MMSE showed more accurate modeling. Predicting expected ranges of A1C is a valuable tool for realistic goal setting for type 2 diabetes patients.