The WHS is a randomized, double-blind, placebo-controlled, 2 × 2 factorial trial of low-dose aspirin and vitamin E for the primary prevention of cardiovascular disease and cancer. The original study consisted of 39,876 U.S. female health professionals aged ≥45 years (mean age 53.9 years at baseline) who were free of diabetes, cancer (except nonmelanoma skin cancer), and cardiovascular disease at baseline (12
). Of the original 39,876 participants, baseline blood samples were obtained from 28,345 participants (71%). We restricted our population to 6,574 postmenopausal women who were not using hormone replacement therapy at the time of blood collection. Women who had developed incident diabetes during a median follow-up of 10 years (by February 2005) were matched in a 1:1 ratio to control subjects by age (within 1 year), duration of follow-up (within 1 month), race, and fasting status at time of blood draw (72% provided fasting blood samples, defined as ≥10 h since the last meal). On the basis of these eligibility criteria, a total of 359 case patients and 359 control subjects were included in our analyses from the WHS.
The PHS II is a randomized, double-blind, placebo-controlled 2 × 2 × 2 × 2 factorial trial of vitamin E, vitamin C, β-carotene, and multivitamins in the prevention of cardiovascular disease and cancer in 14,641 U.S. male physicians aged ≥55 years who were free of cancer and cardiovascular disease at baseline (14
). Baseline blood samples were obtained from 11,130 (76%) of the 14,641 PHS II participants. For each incident case occurring during 8 years of follow-up, one appropriate control subject was selected at random from men who provided baseline blood samples at the time of diagnosis in the case patient. In total, 170 case patients were matched with 170 control subjects by age (within 1 year), race, duration of follow-up (within 1 month), and time of blood draw.
Written informed consent was obtained from all participants in both the WHS and PHS II. Both studies were conducted according to the ethics guidelines of Brigham and Women's Hospital, Harvard Medical School, and the UCLA institutional review board.
Ascertainment of diabetes
Details regarding ascertainment of incident type 2 diabetes in our cohorts have been reported previously (15
). After excluding those with diabetes at baseline, all participants were asked annually whether and when they had a diagnosis of diabetes since baseline. With use of the diagnostic criteria of the American Diabetes Association, (16
), all self-reported cases of type 2 diabetes were confirmed by a supplemental questionnaire. In populations of health professionals such as the WHS and PHS II, self-reported diagnosis of diabetes yields high validity in identifying true cases. As confirmation, a small validation study was conducted, in which self-reported diabetes in the WHS was validated against physician-led telephone interviews, supplementary questionnaires, and medical record reviews, all yielding positive predictive values >91% (17
Blood samples were centrifuged and stored in liquid nitrogen freezers. Matched case-control pairs were handled identically and assayed in random order in the same analytical run in each cohort. Laboratory personnel were blinded to case-control status during all assays. A1C was measured using a Food and Drug Administration–and U.S. National Glycohemoglobin Standardization Program–approved immunoassay (Hitachi 911; Roche Diagnostics), as described previously (18
). Plasma levels of resistin were measured by ELISA (ALPCO Diagnostics, Windham, NH). The assay has a sensitivity of 0.2 ng/ml and day-to-day variabilities of the assay at concentrations of 8.95 and 13.08 ng/ml are 8.9 and 7.4%, respectively. Tumor necrosis factor-α receptor II (TNF-RII) was also measured by ELISA (R&D Systems). The day-to-day variabilities of the assay at concentrations of 89.9, 197.0, and 444.0 pg/ml are 5.1, 3.5, and 3.6%, respectively. As described previously (19
), C-reactive protein (CRP) (performed in the WHS only) was assayed using validated methods at a certified laboratory; the average intra-assay coefficient of variation for CRP was 7.8% and the inter-assay coefficients of variations were 2.5 and 5.1%.
Biomarker values were log-transformed in all analyses to enhance compliance with normality assumptions. Baseline characteristics between case patients and matched control subjects were compared using McNemar's χ2 test for categorical variables and a paired t test for continuous variables. Age- and/or BMI-adjusted partial correlation coefficients were estimated to evaluate associations between plasma resistin levels and traditional metabolic risk factors among control subjects. We reported regression estimates after multiple imputation of missing values using the ICE and MIM procedures in STATA 10.0. Variables with the largest number of missing variables were BMI (n = 12) and multivitamin use (n = 11).
To assess the resistin–type 2 diabetes association, we analyzed the WHS and PHS II data sets separately using conditional logistic regression with robust variance estimators. The levels of resistin were categorized into quartiles based on their distributions among control subjects. Conditional logistic regression was applied to estimate the odds ratio (OR) and 95% CI for type 2 diabetes risk in each quartile using the lowest quartile as the referent category. Because risk set sampling was used for our matched case-control pairs, the ORs yielded unbiased estimates of the relative risk (RR), specifically, the rate ratio. Tests of linear trends across increasing quartiles of resistin levels were conducted by assigning the median values within quartiles treated as a continuous variable. We also estimated the RR per 1 SD increase in log-transformed resistin levels, assuming a linear relationship. The following models were prespecified in this analysis. The basic model (model 1) was adjusted for matching factors (age, ethnicity, and fasting status at time of blood draw). The full model (model 2) was further adjusted for established type 2 diabetes risk factors of smoking (current or former/never), alcohol use (<3 drinks/month, 1–6 drinks/week, or ≥1 drink/day), physical activity (<1, 1–3, or ≥4 times/week), and family history of diabetes in a first-degree relative (yes/no). In model 3, we additionally adjusted for BMI (continuous) in the full model to assess the impact of BMI on the resistin–type 2 diabetes association. In model 4, we further adjusted the full model for CRP, but because CRP was not available in the PHS II data set, we adjusted the full model for TNF-RII in the PHS II data set as a surrogate marker for CRP; the Pearson's correlation between TNF-RII and CRP was r = 0.28 in the WHS data set, consistent with findings in other populations. Model 5 was the same as model 4 with the addition of BMI as a covariate.
Because there was insufficient statistical evidence to suggest that the resistin–type 2 diabetes association differed by sex, we subsequently performed subgroup analyses in the pooled data set to examine potential effect modification by levels of prespecified risk factors: sex (male or female), BMI (normal/underweight, overweight, or obese), age (<60 or ≥60 years), family history of diabetes in a first-degree relative (yes/no), physical activity (<1, 1–3, or ≥4 times/week), alcohol intake (<3 drinks/month, 1–6 drinks/week, or ≥1 drink/day), smoking status (nonsmoker, former smoker, or current smoker), A1C (less than median or more than or equal to median), and TNF-RII (less than median or more than or equal to median). The statistical significance of these interactions was tested by using a Wald test (for variables with two levels) or a χ2 test for homogeneity (for variables with more than two levels) of the pooled estimates of the interaction terms.
We performed additional sensitivity analyses to assess the robustness of our estimates. First, we conducted a series of analyses to address the concern that underdiagnosed type 2 diabetes may have biased our findings. In one analysis we excluded diabetes case patients (and their matched control subjects) in whom diabetes was diagnosed within the first 3 years of follow-up (154 women and 148 men excluded). In another analysis we excluded case patients who had elevated risks of developing type 2 diabetes at baseline (defined by A1C >6.5% and BMI >30 kg/m2 at baseline) (462 women and 172 men excluded). Second, we evaluated the influence of adjusting for A1C in our estimates. Third, because CRP was only measured in the WHS, we assessed the influence of adjusting for another inflammatory marker (alone and with BMI), TNF-RII, which was measured in both data sets. Last, we assessed the influence of using a different measure of adiposity by adjusting for waist circumference (instead of BMI); however, waist circumference was only measured in the WHS and was collected 72 months after baseline. Thus, we adjusted for waist circumference in the WHS, adjusted for BMI in the PHS II, and pooled the estimates. All analyses were conducted using Stata 10.0 (StataCorp, College Station, TX).