This study used data collected in the WHI OS. The details of the scientific rationale, eligibility criteria, and design of the WHI OS have been published (17
). Briefly, 87
724 postmenopausal women aged 50–79 years without a history of breast cancer who self-reported their alcohol use histories were enrolled between October 1, 1993 and December 31, 1998 through 40 clinical centers in the United States. All exposures used in this analysis were collected at the time of entry into the OS. Data were uniformly collected from participants according to a standardized institutional review board approved procedures and protocols by trained study staff. All participants provided written informed consent for this research study at the time of enrollment.
The primary follow-up of WHI OS participants was through self-administered questionnaires that were mailed annually after enrollment through the close-out period, April 2004 to March 2005. In our analysis, we included cohort data ascertained through September 15, 2005, by which time 2.2% of participants were lost to follow-up, 2.5% declined further follow-up, and 6.7% were deceased. Women with breast cancer were initially identified from annual questionnaires. Based on these reports, medical records were obtained and reviewed by trained study adjudicators to verify diagnoses. For the 2944 confirmed invasive breast cancer patients identified through September 15, 2005, information from medical records was forwarded to the WHI coordinating center for central adjudication, and coding of breast cancer stage, size, nodal status, grade, histology, and estrogen receptor (ER) and progesterone receptor (PR) status. Invasive histology was classified as ductal (n = 1805, International Classification of Diseases for Oncology [ICD-O] code 8500) or lobular (n = 720, ICD-O codes 8520 and 8522), and the 419 cancers with other ICD-O histology codes were excluded from our histology specific analyses. Data on ER and PR status were available for 88% of cancers. The 358 cancers with an unknown ER and PR status were excluded from the ER and PR analyses, as were those with ER− and progesterone receptor–positive (PR+) tumors because of insufficient statistical power (n = 37), leaving a total of 2549 invasive cancers included in the analysis focused on ER and PR status.
Cohort members completed baseline self-administered questionnaires covering a wide range of topics including demographic characteristics, medical history, reproductive history, lifestyle characteristics, and family history of various diseases. In addition, baseline height and weight was measured by the study staff. Alcohol consumption at the time of enrollment, our primary exposure of interest, was assessed from two sources: 1) self-administered questionnaires of personal habits at baseline that collected alcohol consumption history and 2) self-administered food-frequency questionnaires (FFQ) completed at enrollment. In the alcohol consumption questionnaires, women were asked whether they ever consumed at least 12 alcoholic drinks of any kind, and those who answered yes were asked whether they still drank alcohol so that never, former, and current drinkers could be distinguished from each other; ever drinkers were also asked how many alcoholic beverages they consumed each day, week, or month over different ages in their lives. In the data collected, one bottle or can of beer, one glass of wine, and one shot of liquor were all considered to be equivalent. Using these two data sources, summary measures of recency and frequency of alcohol consumption were obtained. If there were any discrepancies between these two measures, the FFQ data were given priority. Women were categorized as never drinkers (never consumed 12 or more alcoholic beverages of any kind in their lives), former drinkers (ever drinkers who reported having stopped drinking at the time data were collected), and current drinkers. Among current drinkers, the average number of drinks per week was computed. In our main analysis, frequency of alcohol consumption among current drinkers was then grouped into six consumption categories based on the following number of drinks consumed per week: less than 0.5, 0.5–0.9, 1.0–3.9, 4.0–6.9, 7.0–13.9, and 14.0 or more. Although most studies of alcohol use and breast cancer are limited to categories of less than 7.0 or 7.0 or more drinks per day, because of our sample size finer categories could be used to more clearly evaluate the potential dose–response relationship between alcohol use and breast cancer risk. Two approaches were used to assess the dose–response relationship. First, risk per number of drinks consumed per day among current drinkers was computed by treating the number of drinks per day consumed as a continuous term in the statistical model. Second, P values for trend were calculated by using the number of drinks per day consumed as a continuous variable and restricting the analyses only to women who were categorized as current drinkers. We also conducted a subanalysis of risk by alcohol type (beer, wine, and liquor) with this classification derived from FFQ data collected at baseline.
Cox regression was used to calculate hazard ratios (HRs) and 95% confidence intervals as a measure of the association between history of alcohol use and breast cancer risk. Assumptions of proportionality for the Cox models were confirmed based on scaled Schoenfeld residuals. Time to breast cancer was computed from date of enrollment to date of first breast cancer diagnosis, with times for women without breast cancer censored by date of last study follow-up or September 15, 2005, whichever occurred first. All analyses were adjusted for age, race, and/or ethnicity, and women categorized as never drinkers served as the reference category. Variables considered as potential confounders or effect modifiers included the following categorical baseline characteristics using the categories shown in : education, body mass index, use of menopausal hormone therapy, smoking status, Gail model scores of 5-year breast cancer risk, and number of screening mammograms received in the past 5 years. We present risk estimates from models adjusted simply for age, race, and/or ethnicity, and the ones additionally adjusted for each of these characteristics as categorical variables according to how they are categorized in . Effect modification was assessed using likelihood ratio testing, and none of these variables were observed to be statistically significant effect modifiers (all Pinteraction > .05). P values characterizing the difference in risk estimates between case groups were calculated through comparisons only of case patients using unconditional logistic regression (ie, a logistic regression model was fit restricted to ductal and lobular case patients’ data where those with ductal carcinoma served as the reference group). All analyses were conducted using Stata 9.2 (Stata Corp, College Station, TX), and all P values were from two-sided tests in which values less than .05 were considered statistically significant.
Distribution of demographic and personal characteristics by alcohol use*