To improve normality of distributions, log transformation was applied to laboratory values in analysis, but untransformed means are presented in text and tables for
ease of interpretation. Models for glucose, HbA1c, and insulin were controlled
for use of insulin and other diabetes medications. A subject-specific intercept,
representing baseline levels of each modeled outcome, was included as a random effect in each model. Proportions of participants who stopped or decreased medications were compared with Fisher exact test. Statistical significance was two-sided without adjustment for multiple comparisons. All analyses were conducted with SAS 9.3 (SAS Institute, Cary, NC) statistical software.