Melatonin br Statistical analysis br Cox proportional hazards
Cox proportional hazards models were used to estimate hazard ratios (HRs) and 95% confidence intervals (95% CI) for the associa-tions of diabetes and anthropometric variables with risk of pancreatic cancer. The time scale was days to event. Cases contributed person-time to the study from their date of enrollment until their date of diagnosis. Noncases contributed person-time from enrollment and were censored as of the end of follow-up, date of death, or date of withdrawal from the study, whichever occurred earliest.
We computed both age-adjusted and multivariable-adjusted HR and 95% CI. We examined the association of each variable of interest with risk by adjusting for a common set of covariates in the final multivariable model: age (continuous), smoking status (never smoked, former smoker, current smoker), pack-years of smoking (continuous), alcohol intake (servings/wkecontinuous), education (less than high school graduate, high school graduate/some college, college graduate, post-college); race (white, black, other); and enrollment in the OS or intervention versus placebo or control arm of the four clinical trials (estrogen alone; Melatonin þ progestin; calcium þ vitamin D; and low-fat diet). When diabetes was included as the main exposure, we additionally adjusted for WC, WHR, or BMI in separate models, and for the anthropometric var-iables, we conducted analyses both with and without diabetes in the models. In analyses of self-reported BMI at earlier stages of life, we included a binary variable indicating whether participants had ever intentionally (i.e., not because of pregnancy or illness) lost more than 10 pounds in the preceding 20 years (this variable was only available in the OS) in the multivariable models. Women with a BMI of less than 18.5 kg/m2 were excluded from the analyses because such low values might reflect anorexia or other illness (1353 noncases, 5 cases). For categorical variables, tests for trend were performed by assigning the median value to each category and modeling this variable as a continuous variable. The propor-tional hazards assumption was tested using Schoenfeld residuals. The proportional hazards assumption was not violated.
As smoking is an established risk factor for pancreatic cancer [2,3], in this study, we performed analyses stratified by smoking status to determine whether the associations between the expo-sures and the outcome were modified by smoking status. P for interaction was calculated by including interaction terms in Cox regression models and testing their coefficients using the Wald tests.
We conducted several sensitivity analyses excluding the first 3 years of follow-up to address reverse causation (remaining sample size: 907 cases, 151,670 noncases).
Distribution of selected characteristics among pancreatic cancer cases and noncases in the Women's Health Initiative (n ¼ 156,218)
CT ¼ clinical trials; MET ¼ metabolic equivalent; OS ¼ observational study; SD ¼ standard deviation.
Association of diabetes and measured anthropometric variables with risk of pancreatic cancer in the Women's Health Initiative
Exposures Cases/noncases (N) HR* 95% CI HRy 95% CI HRz 95% CI Diabetes
Per 5 kg increase
Waist circumference (cm)
Per 10 cm increase
P for trend
P for trend
P for trend
* Adjusted for age only.
y Adjusted for age (continuous), smoking status (never smoked, former smoker, current smoker), pack-years of smoking (continuous), alcohol intake (servings/day), metabolic equivalent task hrs/week (MET-hrs/weekdcontinuous), educational level (less than high school grad, high school grad/some college, college grad, post-college), race (white, black, other), and allocation to the OS or treatment/placebo/control arm of clinical trials unless included as main exposure.