A re-analysis of case-cohort data from the ADVANCE study with the addition of new Hp phenotype data15 was undertaken. The design, methods, and major findings of the ADVANCE study (ClinicalTrials.gov identifier: NCT00145925) have been reported previously.4 17 Briefly, 11 140 patients with type 2 diabetes were recruited between June 2001 and March 2003 from 215 centers in 20 countries from Asia, Australasia, Europe, and North America. Participants were aged ≥55 years with diagnosed type 2 diabetes at age 30 years or older. Additionally, participants had a history of CVD or one or more CVD risk factors. Participants were randomized to either intensive glycemic treatment or control (usual care) over a median follow-up of 5 years. Information on blood glucose, blood pressure, and lipids was collected at month 6 after randomization and then every 12 months thereafter.4 17

To improve the efficiency of the biomarker studies in the ADVANCE trial, a case-cohort subsample was established as previously described.18 19 Samples for the case-cohort were available from all countries involved in ADVANCE, except China and India. Case-cohorts are an efficient prospective study design with power similar to that of a full cohort.20 Baseline characteristics for the ADVANCE case-cohort were similar to the full ADVANCE cohort.21 Because our outcome for the current analysis is major CAD events, any enriched cases from the previously established case-cohort that were not a major CAD event were excluded from the current case-cohort analysis of major CAD events as previously described15 and only included in the analysis of their relevant secondary outcome. Available samples from the ADVANCE case-cohort (90.7%) were used to determine the Hp phenotype of participants using a validated high-throughput ELISA.22 Less than 4% of data were missing for any baseline variables. Twelve (0.4%) participants had missing HbA1c and thus were excluded from the present analysis, leaving a final sample of 3392 (online supplemental figure 1).

An independent adjudication committee validated all outcome events.4 17 We report our primary outcome of major CAD events according to the original ADVANCE study prespecified diagnosis criteria and definition of ‘major coronary heart disease’ events, which is defined as the first occurrence of non-fatal MI, death from coronary heart disease, or sudden death (otherwise unexplained).4 Due to the biological mechanism linking Hp and CAD in hyperglycemia, the present analysis studied the primary outcome of CAD events rather than the original ADVANCE study primary outcome, which included stroke. Stroke is an end point that has been associated with the Hp1-1 phenotype rather than the Hp2-2 phenotype,23 and stroke is a composite of different subtypes with varying etiologies not always related to atherosclerosis, further suggesting that CAD and composite stroke should be separated from a composite CVD outcome for analyses involving Hp phenotype.

We also investigated the relationship between HbA1c and other ADVANCE study outcomes that were embedded within the previously established ADVANCE case-cohort enriched case definition (non-fatal MI, major macrovascular and microvascular events, CVD mortality, all-cause mortality, and non-fatal stroke) following on previous findings of increased CVD mortality and total mortality in the non-Hp2-2 phenotype.14 For each outcome, appropriate cases were added to the random subcohort from the enriched case pool of the previously established ADVANCE case-cohort, explaining the slightly different sample size for the analyses of secondary outcomes.

Statistical analyses were conducted using Stata/SE software V.18.0 (College Station, Texas, USA) at a two-tailed alpha level of 0.05. Except for when testing for Hardy-Weinberg Equilibrium, participants with the Hp2-1 or Hp1-1 phenotypes (those without the Hp2-2 phenotype, also known as Hp1 allele carriers) were combined to form a group, which is a common approach when studying the Hp phenotype.8 9 14 24

Participants were grouped based on Hp phenotype and baseline HbA1c (<7.0%, 7.0%–7.9%, or ≥8.0%). We chose these cut points because: (1) current guidelines recommend a glycemic target of HbA1c <7.0% in otherwise healthy older adults while a less stringent target of HbA1c <8.0% may be recommended for those with chronic illness2 25 and (2) we have previously found that achieving HbA1c ≥8.0% compared with 7.0%–7.9% was associated with incident CAD among ACCORD study participants with the Hp2-2 phenotype, while no association was observed among participants without the Hp2-2 phenotype and there was no evidence that HbA1c concentration <7.0% prevented more CAD events compared with 7.0%–7.9% in either Hp phenotype group.16 Characteristics were compared using t-tests and Kruskal-Wallis tests for continuous variables and χ2 tests for categorical variables. Variables were presented similarly as in the original ADVANCE study to facilitate assessment of our findings in the context of the original study.

To obtain HRs, Cox proportional hazards regression models with time-varying covariables, Prentice weighting, and robust SE estimates26 were fitted using the STSELPRE procedure for case-cohort analyses to quantify the association between time-varying achieved HbA1c (categorized as <7.0% or ≥8.0% compared with the reference group of 7.0%–7.9%) and CAD in each Hp phenotype group separately.

The multivariable Cox models adjusted for both time-dependent and time-independent variables, and these covariables were selected a priori based on statistical differences between groups at baseline and established confounders commonly included in previous research such as age, sex, and medications.16 27 28 Time-varying variables included HbA1c (%, categorized as <7.0% or ≥8.0% compared with 7.0%–7.9%), total cholesterol (mg/dL), triglycerides (mg/dL), systolic blood pressure (mm Hg), and body mass index (BMI). Time-varying covariables were used to relate the most recent measure for each of those variables to incident outcomes at the time of an event to avoid potential bias from using a single baseline measurement. Cluster variance estimates accounting for within-subject correlation of repeated measures were used. Time-fixed covariables recorded at baseline included: age, sex, race, region (Northern Europe/Continental Europe/Australia and New Zealand/Canada), assignment to the intensive glycemic control intervention (yes/no), assignment to the intensive blood-pressure intervention (yes/no), history of CVD (yes/no, includes a history of major microvascular or macrovascular disease), smoking at baseline (yes/no), alcohol consumption (non-drinker/current drinker/past drinker), diabetes duration (>10 years/≤10 years), any lipid medication use (yes/no), and any diabetes medication use (yes/no) unless stratified by that variable. The proportional hazards assumption was verified using tests of correlation between the Schoenfeld residuals and time.29 The presence of an interaction between HbA1c categories as a continuous variable and Hp phenotype was tested by adding an interaction term to the adjusted model.

Current reporting guidelines recommend disaggregation of results by sex,30 and the frequency of the Hp2-2 phenotype differs among race-based and geographic populations,7 warranting additional a priori stratification. Current diabetes care guidelines also suggest that diabetes duration and established CVD are important factors in glucose management,31 and epidemiological methods stipulate that longitudinal studies of etiology should distinguish people with prevalent disease at baseline (especially if the prevalent disease is the outcome of interest in the study) to prevent bias due to differing risks associated with having prevalent disease.32 Thus, stratified analyses by sex, previous CVD at baseline (history of major microvascular or macrovascular disease), diabetes duration (>10 years), and race were performed by dividing these variables into their respective levels (strata) and running the same model as for the main analysis (but without the respective stratified variable in the model). For the race stratification, we were only able to run the adjusted model in White participants (94% of participants in the study sample) as the sample sizes for the other race-based groups were too small. We did not run the analysis for other race-based groups all together because when diverse populations are collapsed into a single group, racial/social/cultural relevance is lost, the results for this group cannot be interpreted as race-based data, and it is not consistent with current guidelines on reporting race-based data where specific racial categories are recommended over collective terms.33 Interactions were tested between HbA1c as a continuous variable and sex, CVD history at baseline, and diabetes duration, in each phenotype group by adding the respective interaction terms to the adjusted model. Follow-up time was defined as the time from randomization to the date of documented outcome, or until a participant was censored if no event occurred.

To further investigate the linearity of the relationship of continuous HbA1c and CAD in each phenotype group across the range of HbA1c in this study, a restricted cubic spline model was fitted with knots placed at evenly spaced percentiles (5, 27.5, 50, 72.5, and 95)34 of HbA1c and compared with a model with the linear term and no cubic spline terms. The plot from the regression was truncated at the 5th and 95th percentiles. The MEAN command in Stata was used to calculate the mean, SE, and confidence limits for each time point for each phenotype group. T-tests were used to compare the means between phenotype groups at each time point.