The NIH–AARP Diet and Health Study design and methodology have been described in detail 34. Briefly, the NIH–AARP Study is a prospective cohort study of diet, health-related behaviors and cancer.

The study was initiated in 1995 and 1996 when a questionnaire was mailed to 3.5 million members of the AARP (formerly known as the American Association of Retired Persons) aged 50 to 71 years, who resided in one of six states (California, Florida, Pennsylvania, New Jersey, North Carolina, and Louisiana) or two metropolitan areas (Atlanta, Georgia and Detroit, Michigan) that have large AARP membership, large minority populations, and high-quality cancer registries. The questionnaire captured diet history, demographic characteristics, reproductive history, menopausal status, and family history of cancer. Of the 617,119 men and women (17.6%) who returned the questionnaire, 567,169 satisfactorily completed it.

In 1996 and 1997 a second questionnaire was sent to collect additional information on family history of cancer, anthropometry, and use of menopausal hormone therapy (HT) and other medications, including NSAIDs. A total of 337,074 men and women completed this questionnaire.

After excluding individuals who died or moved out of the cancer registry ascertainment area before their second questionnaire was received and scanned (n = 2,166), excluding proxy respondents to the baseline (n = 6,959) or second (n = 3,424) questionnaires, and excluding 188,117 men, the study population included 136,408 potentially eligible women

The study was approved by the Special Studies Institutional Review Board of the US National Cancer Institute.

The second questionnaire asked whether aspirin products (generic aspirin, Bayer, Bufferin, Anacin, Ecotrin, or Excedrin) and nonaspirin NSAIDs (generic ibuprofen, Advil, Nuprin, Motrin, Naprosyn, and so on) had been used in the past 12 months. As acetaminophen is an analgesic with weak anti-inflammatory activity 35, participants were instructed not to include 'Tylenol, acetaminophen, or any other pain relievers' in nonaspirin NSAIDs. Aspirin or nonaspirin NSAID users were asked to indicate their frequency of usual use: fewer than 2 times per month, 2 to 3 times per month, 1 to 2 times per week, 3 to 4 times per week, 5 to 6 times per week, 1 time per day, or 2 or more times per day – which we collapsed to categories of nonuse, <1/week, 1 to 6/week, and ≥1/day. The dose, duration, and indication for use were not collected. In the current report, 'any NSAID' combines the use of aspirin and nonaspirin NSAIDs.

While the demographic characteristics, lifestyle, and reproductive history were largely obtained through the first questionnaire, the second questionnaire ascertained a self-reported history of hypertension, mammogram screening, and vigorous physical activity. Menopausal hormone therapy use and history of breast cancer in a first-degree relative were derived from information collected on both the first and second questionnaires.

Cohort members were followed annually for mailing address changes and vital status. Address changes were identified through linkage to the US Postal Service's National Change of Address database, through US Postal Service updates received with undeliverable mail, through use of other address change update services, and through participants' notifications. Vital status was updated through linkage to the Social Security Administration Death Master File and was verified by the National Death Index.

Incident in situ breast cancers and invasive breast cancers were initially identified through probabilistic linkage to eight state cancer registries using first and last name, address, sex, date of birth, and Social Security Number from the baseline questionnaire. The cancer registry ascertainment area was recently expanded to include three additional states (Texas, Arizona, and Nevada) to capture cancer occurring among participants who moved to those states during follow-up. Each registry has been certified by the North American Association of Central Cancer Registries for meeting the highest standards of data quality.

Breast cancer estrogen receptor (ER) status was coded as described in the American Joint Committee on Cancer's Collaborative Staging Site-Specific Factors Manual, with a threshold of >10 fmol cytosol protein per milligram for a positive tumor; however, the ER status was not reported by the Florida, Michigan, and Pennsylvania cancer registries. Histology was defined using the International Classification of Diseases for Oncology codes, third edition 36. A previous validation study in this cohort estimated that registry linkage validly identified approximately 90% of all incident cancers 37. The date of death for fatal cancers was identified through linkage to the National Death Index.

Because information regarding NSAID use was not collected on the baseline questionnaire, we limited analyses to the 136,408 women who completed the second questionnaire. We excluded 9,022 women who reported a personal cancer history other than nonmelanoma skin cancer on either questionnaire (including 942 breast cancers), 1,259 women who were missing information on NSAID use (including 52 breast cancers), and three women with no follow-up. Thus, 126,124 women were included in the present analyses.

Through 31 December 2003, 4,501 women developed breast cancer (781 in situ, 3,703 invasive, and 17 missing stage). Among the 2,282 invasive breast cancer cases from states that collected the ER status, 1,439 were coded as ER-positive and 280 as ER-negative (the ER status was unavailable for the remaining 563 invasive cases). There were no substantial risk factor differences between women from states where ER information was available and states where it was unavailable. The majority of invasive breast cancers were ductal carcinomas (n = 2,409), followed by lobular (n = 400), mixed (n = 310), and other (n = 584) histologic types.

Cox proportional hazards models were used to estimate hazard ratios and 95% confidence intervals (CIs) for breast cancer associated with NSAID use; age was the time scale 38, and ties were handled by enumeration 39. Follow-up began at the age at which the second questionnaire was received and scanned, and continued through the earliest of the following dates: participant diagnosed with breast cancer, moved out of her registry catchment area, died from any cause, or 31 December 2003. To test the proportional hazards assumption, we generated time-dependent covariates by including interactions of each predictor with the natural log of age (the time metric); probability values for all time-dependent covariates were >0.05, consistent with hazards that are proportional.

Multivariate models were used to control for age at entry (years), race/ethnicity (white versus other/unknown), age at first birth (nulliparous, <20 years, 20 to 24 years, 25 to 29 years, ≥30 years, or unknown), HT use (never used, estrogen only, estrogen–progestin, or other/unknown), number of breast biopsies (0, 1, 2, ≥3, or unknown), alcoholic drinks per day (0, <1, 1 to 3, ≥3, or unknown), history of hypertension (no, yes, or unknown), and family history of breast cancer in a first-degree relative (no, yes, or unknown). Models examining frequency of aspirin use and breast cancer risk also included terms for frequency of nonaspirin NSAID use and vice versa. Tests for linear trends across the NSAID exposure categories were calculated by treating these categorical variables as ordinal variables.

In subsequent models, we adjusted for calendar time and several additional factors, including ages at menarche and menopause, mammogram in the past 3 years, self-reported heart disease, self-rated health quality, physical activity, and body mass index (BMI). The results were essentially the same and are not shown here.

We used a likelihood ratio test, comparing models with and without the interaction terms, to separately examine effect modification by age at entry (<57 years, 57 to 60 years, 61 to 64 years, 65 to 68 years, or ≥69 years), HT use (never, estrogen only, estrogen–progestin, or other/unknown), BMI (<25 kg/m², 25 to 29 kg/m², ≥30 kg/m², or unknown), and smoking status (never, former, current, or unknown). In addition, we examined whether the relationship between NSAIDs and breast cancer incidence differed by ER status (positive or negative), by stage at diagnosis (in situ or invasive disease), and by histologic type (ductal, lobular, or mixed).

Probability values <0.05 were considered statistically significant. All tests of statistical significance were two-tailed. Analyses were performed using SAS software release 9.1.3 (SAS Institute Inc., Cary, NC, USA).