Statistical Analyses

Participants were initially categorized into two groups based on whether one or more serious hypoglycemic episodes occurred in the time between VADT entry and the follow-up computed tomography scan. Between-group differences in normally distributed continuous variables were assessed with mean ± SD and t tests. Median (25th–75th percentile) and Mann-Whitney U tests were used for variables with skewed distributions, and proportions and χ² tests were used for dichotomized variables.

To determine the effect of serious hypoglycemia on progression of CAC, we used a robust regression model that limits the influence of outliers and provides stable results in the presence of outliers in both response and covariate variables. To assess the possibility of effect modification by treatment, pairwise interaction terms between treatment and occurrence of hypoglycemia were evaluated. The interaction between treatment and occurrence of hypoglycemia was significant; thus, we performed stratified analyses by treatment arm.

To identify the best parsimonious predictors of CAC progression within each treatment group, we performed a linear regression model with stepwise variable selection (forward in, backward out). All relevant variables from Table 1 were included in the models. Selection criteria required a P < 0.1 for a variable to enter and be retained in models. Serious hypoglycemia as the primary variable of interest was forced into the models. The stepwise analysis indicated that the best predictors of CAC progression differed in each treatment arm. In addition to serious hypoglycemia, prior hypoglycemia in the standard treatment group and baseline CAC and albumin-to-creatinine ratio in the intensive treatment group were retained in the models. We then ran a series of multivariate linear regression models in each treatment arm, with all the variables selected from the parsimonious models for both treatment arms, and adjusted for additional potential confounding effects of other covariates. We assessed the goodness-of-fit of the models with the multivariate coefficient of determination (R²) and residuals. Model fit did not improve with the inclusion of more than seven variables; thus, to avoid overfitting of the regression models, we limited the maximum number of variables in a model to seven.

Baseline characteristics of participants by serious hypoglycemia

Data are mean ± SD, median (25th–75th percentile), or percentages. Prior hypoglycemia denotes signs and symptoms of hypoglycemia or blood glucose levels <70 mg/dL (∼3.9 mmol/L) during the last 1.5 months before randomization.

We also performed several sensitivity analyses to examine the robustness of these results and to account for key potential confounders. First, we categorized participants on the basis of the number of reported serious hypoglycemic episodes during follow-up. Because the median number of reported episodes in those with serious hypoglycemia was five, we categorized participants into three groups: 1) those with no hypoglycemic episodes, 2) those with less than five episodes, and 3) those with five or more episodes. Next, we performed stratified analyses by the mean HbA_1c value of 7.5% (∼58 mmol/mol) during the study as a level of reasonably good glycemic control and to match HbA_1c values used in a prior study of hypoglycemia and CAC (24). We also analyzed the data by severe hypoglycemia and G<50 separately. Finally, we determined CAC progression in a subset of participants matched on baseline CAC (±10%) to further account for differences in baseline CAC. Statistical analyses were performed with SAS release 9.4 software (SAS Institute, Cary, NC).