Study population and design
Women included in this study were selected from participants of the NHS and NHSII cohorts. These prospective cohorts followed registered nurses in the USA who were 30–55 years (NHS) or 25–42 years old (NHSII) at enrollment. After administration of the initial questionnaire, the information on breast health risk factors and any cancer diagnoses was updated biennially. More detailed description of the cohort has been published elsewhere [18].
We used a nested case–control study design to examine the association between aspirin and other NSAIDs with breast tumor subtypes defined based on the COX-2 expression status. Details of this nested case–control study have been previously described [19]. Briefly, using incidence density sampling, women who did not have any type of cancer at the time of the cases’ breast cancer diagnosis (controls) were matched with women diagnosed with in situ or invasive breast cancer (cases) during the follow-up period from June 1, 1989, through June 30, 2004, for NHS and from June 1, 1996, to June 30, 2007, for NHSII [18]. Breast cancer cases were confirmed through medical record review by trained personnel. Because the original study was designed to evaluate associations between circulating biomarkers and breast cancer risk, the cases were matched with controls on age, menopausal status, postmenopausal hormone use (current vs. not current) at blood collection, and day and time of blood collection. We restricted our analysis to 421 cases and 3,166 controls with available data on COX-2 status and important covariates, including mammographic breast density, a well-established and strong breast cancer risk factor.
The study protocol was approved by the institutional review boards of the Brigham and Women’s Hospital and Harvard T.H. Chan School of Public Health, and those of participating registries as required. Consent was obtained or implied by return of questionnaires.
Assessment of aspirin intake and NSAIDs
The methods of assessing exposure to aspirin and other NSAIDs have been described in detail elsewhere [20]. Briefly, information on aspirin use in NHS was first obtained in 1980 and biennially thereafter except in 1986. In 1980, participants were asked whether they currently took aspirin in most weeks and, if yes, what was the weekly amount and years of aspirin use. Information on aspirin dose and frequency of use was also collected beginning in 1982 and 1984, respectively. In NHSII, on the baseline questionnaire in 1989, participants were asked if they regularly (≥ 2 times per week) used aspirin, or other anti-inflammatory drugs in three separate questions and this was updated biennially from 1993. Beginning in 1993 (for aspirin) or 1995 (for other anti-inflammatory drugs), women were asked to report frequency of use (categorized as either days per week or days per month). Beginning in 1999, participants were additionally asked about quantity used (tablets per week) in each category.
Women were classified as current users at each questionnaire in which current use was reported and were considered current users for the subsequent two-year follow-up period (or the 4-year follow-up period from 1989 to 1993). For participants who missed a questionnaire, drug use information was carried forward from the previous cycle. The women who ceased reporting use were classified as past users, but they were eligible to become current users in subsequent follow-up years. Women were classified as nonusers if they did not report analgesic use at baseline or on any of their follow-up questionnaires. Duration of use of each drug was calculated from baseline (1980 for aspirin, 1990 for other NSAIDs for NHS, and 1989 for NHSII) to the reference date (date of diagnosis for cases and their matched controls) [20]. To better represent long-term use, we calculated the cumulative average dose (standard 325-mg tablet) and frequency (days per week) for each woman who was classified as a past or current user as the average of current use and all previous follow-up cycles. Status, quantity, and frequency of use were carried forward one cycle to replace missing data, and cumulative average quantity, cumulative average frequency, and duration of use were calculated from these variables with the carried-forward data.
Tumor tissue analyses
We requested formalin-fixed paraffin-embedded tissue samples from hospitals throughout the US where women underwent primary breast tumor resection. Tumor microarrays (TMAs) were constructed at the Dana Farber Harvard Cancer Center Tissue Microarray Core Facility, Boston, MA. As described previously [21], TMAs were assembled by taking three 0.6-mm-diameter cores from each breast cancer sample and inserting cores into a recipient TMA block.
Immunohistochemistry for COX-2 was conducted on 5 μm paraffin sections of TMA blocks. Cores with fewer than 100 cells were excluded from all analyses. For NHS, the staining was first performed with monoclonal antibodies from Thermo Fisher Scientific (SP21 clone, Waltham, Massachusetts, USA). These stained NHS TMAs were subsequently stained with a second monoclonal antibody from Cayman Chemical (CX229 clone, Ann Arbor, Michigan, USA) at the Pepper Schedin’s laboratory (Oregon Health and Science University, Portland, Oregon, USA), as previously described [22].
Expression of COX-2 for NHSII TMAs was assessed in Dr. Pepper Schedin’s laboratory (Oregon Health and Science University, Portland, OR) [22]. Based on previous reports of differing COX-2 staining patterns for the monoclonal antibodies produced by Cayman Chemical (CX229 clone, Ann Arbor, Michigan, USA) and Thermo Fisher Scientific (SP21 clone, Waltham, Massachusetts, USA) [23], TMA slides were dual-stained at a single time point using both antibodies and labeled with distinct chromogens so that the two COX-2 signals could be distinguished. COX-2 staining results from both NHS and NHSII TMAs were analyzed using the Aperio co-localization image analysis algorithm and expressed as percentages of positively stained area for each antibody. In primary analyses, we examined the mean % area across all three cores that stained positive for at least one of the two antibodies. In supplemental analyses, we examined the mean % area for each antibody separately. Tumor expression status was defined as either negative or positive using the median % positivity as a cutoff.
Covariate information
Information on breast cancer risk factors was obtained from the biennial questionnaires closest to the date of the mammogram. Women were considered to be postmenopausal if they reported: (1) no menstrual periods within the 12 months before blood collection with natural menopause, (2) bilateral oophorectomy, or (3) hysterectomy with one or both ovaries retained and were 54 years or older for ever-smokers or 56 years or older for never-smokers [24, 25].
To quantify mammographic density, the craniocaudal views of both breasts for all screening mammograms in the NHS and for the first two batches of mammograms in the NHSII were digitized at 261 μm per pixel with a Lumisys 85 laser film scanner (Lumisys, Sunnyvale, California). The third batch of NHSII mammograms was digitized using a VIDAR CAD PRO Advantage scanner (VIDAR Systems Corporation; Herndon, VA) and comparable resolution of 150 dots per inch and 12 bit depth. The Cumulus software (University of Toronto, Toronto, Canada) was used for computer-assisted determination of the absolute dense area, non-dense area, and percent mammographic density on all mammograms [26, 27]. Percent breast density was measured as percentage of the total area occupied by epithelial/stromal tissue (absolute dense area) divided by the total breast area. Because breast densities of the right and left breast for an individual woman are strongly correlated [26], the average density of both breasts was used in this analysis.
Statistical analysis
We used unconditional logistic regression to analyze the association between anti-inflammatory drug use and breast cancer risk, while adjusting for the following potential confounders in the fully adjusted logistic regression models: age at diagnosis (continuous, years), body mass index (BMI, continuous, kg/m2), percent breast density (< 10%, 10 to < 25%, 25 to < 50%, ≥ 50%), age at menarche (< 12, 12, 13, > 13 years), parity and age at first birth (nulliparous, parous/ < 25 years, and parous/ ≥ 25 years), PMH use (never, ever, unknown), family history of breast cancer (yes, no), alcohol consumption (0, < 5, 5 to < 15, ≥ 15 g/day), and study cohort (NHS, NHSII).
Differences in the association of breast density with COX2-defined tumor subtypes were investigated using polychotomous logistic regression [27]. In this analysis, the outcome has three levels which include controls and two breast cancer subtypes defined based on the COX2 status (positive and negative). We used a likelihood ratio test to compare a model with separate anti-inflammatory drug use slopes in each case group with a model with a common slope. This method has been described in detail elsewhere [19]. In this analysis, the drug intake variables were modeled using respective medians within each of the categories. For all analyses, the level of statistical significance was assessed at α equal to 0.05. All tests were two-sided. All analyses except the test of heterogeneity were performed using SAS software (version 9.2, SAS Institute, Cary, NC, USA). The test of heterogeneity from polychotomous logistic regression models was done using STATA version 11.0 (Stata Corp, College Station, TX, USA).