Statistical Analysis

Covariate data came from the examination in which HDL‐P and CEC were measured (ie, analytic baseline). Thus, all covariate data and measurements were from the same time point, and samples from analytic baseline were used for the measurement of APOA1, HDL‐P, and CEC. Pearson or Spearman correlation coefficients, as appropriate, were employed to evaluate factors related to these HDL metrics. Descriptive statistics were used to assess differences in participant characteristics by CAD incidence.

Separate complete case Cox proportional hazards analysis models with backward elimination were constructed for HDL‐C and each novel HDL metric as the main independent variable, allowing for factors significantly associated with CAD univariately. Survival time was defined as the time in years from the analytic baseline to the date of a first CAD event or, for noncases, the last available follow‐up. In addition to the univariate model, 3 multivariable models were constructed: the first allowed for lipid risk factors (non‐HDL‐C and HDL‐C) (Model 1), the second allowed for lipid risk factors in addition to all other traditional CVD risk factors (Model 2), and the third model allowed for variables in Model 2 and CEC (for models with HDL‐C or HDL‐P as the main independent variable) or T‐HDL‐P (for models with CEC as the main independent variable) (Model 3). Multivariable analyses were repeated, further allowing for insulin dose per body weight, triglycerides, and APOA1 in the smaller number of study participants who had fasted (n=488).

The proportional hazards assumption was assessed with an interaction term between each covariate of interest and logarithmically transformed time, keeping the fixed covariates in the model. For those covariates that violated the proportionality assumption, a time‐dependent variable, that is, the covariate's interaction with time (covariate×log(time)), in addition to its fixed term, was included in the models. Martingale residuals were used to assess the true functional form of each covariate and logarithmic transformation was used for variables violating the assumption of linearity (ie, insulin dose per body weight, triglyceride concentrations, white blood cell count, and albumin excretion rate). Hazard ratios (HR) are presented per 1 SD increase for CEC, per 1 μmol/L increase for variables reflecting HDL‐P, and per 1 mg/dL increase for HDL‐C. Two‐sided P values <0.05 were considered statistically significant. All analyses were conducted using SAS version 9.4 (SAS Institute, Cary, NC).