Subjects and study design
NOWAC is a nationwide cohort study set up at the University of Tromsø, Norway in 1991. It currently comprises more than 165,700 women born 1927–65 randomly drawn from the National Central Person Register [
29,
30]. There is a small oversampling from the three most northern counties (
n = 13,674). Due to practical workload, financial restraints and methodological sub-studies [
31] the baseline enrollment was separated into a number of different sub-series and carried out from 1991 to 1997, and from 2003 to 2006. As the follow-up time for the women in the last inclusion period is very short, only women enrolled from 1991 to 1997 are included in the current paper. The women received a mailed letter of invitation requesting informed consent and a self-instructive questionnaire focusing on certain topics (e.g., diet or sun exposure), and also containing a number of core questions. Detailed dietary questions have been included in the questionnaires since 1996. To expand and update the exposure information, a second questionnaire was from 1998 to 2002 sent to the responders of the baseline mailing conducted in 1991–1997. After one or two reminders, the response rate for the first (baseline) questionnaire (1991–1997) was 57% (102,540/179,387), and 81% (80,835/99,541) for the second questionnaire (1998–2002). The present analyses are restricted to women who gave detailed dietary information in the period 1996–1999, either in their first questionnaire (
n = 30,333) or in their second questionnaire (
n = 38,184). The Regional Committee for Medical Research Ethics and the Norwegian Data Inspection Board approved the study.
Dietary assessment and nutrient calculations
Dietary information was collected by means of a semi-quantitative food frequency section in the questionnaire, described in details previously [
28,
32]. In short, the participants were asked to record their average consumption of various food items during the last year ticking off fixed boxes. Consumption of the following dairy items was asked for: glasses of whole milk, low-fat milk and skimmed milk drunk per week or day, packages of yoghurt eaten per week or day, slices of bread with whole fat brown cheese, low-fat/skimmed brown cheese, whole fat white cheese and low-fat/skimmed white cheese eaten per week, and frequency and amount of ice cream eaten per month or week during the summer and during the rest of the year. In addition, the participants were asked to record what kind of fat (butter or margarine) they usually used on bread, and how often and how much melted or solid fat, full fat sour cream, and fat-reduced sour cream they used for fish meals. Number of glasses of milk consumed per day during childhood was recorded in most of the sub-series. Alcohol intake was calculated from three questions on beer, wine, and spirit consumption.
Daily intake of foods, energy and nutrients was computed using a computation program developed at the Institute of Community Medicine, University of Tromsø, for SAS software. Missing frequencies were treated as null intake and missing portion sizes as the smallest portion unit in the questionnaire, giving a conservative intake estimate [
33]. For the women included in the analytic cohort, mode for missing frequencies on the 10 relevant dairy questions was zero; the median was two. The recorded frequency was multiplied with the recorded portion size or a standard portion [
34], and transformed into daily food intake in grams. Total dairy consumption was calculated by summing up the consumption of all the items listed above (Table
1), except brown cheese and fat on bread and for fish meals due to the dissimilar nutrient content of these items. Nutrient intake was calculated by multiplying the daily food intake (in grams) of each item with the nutrient content of the item as given in the official Norwegian food composition table [
35]. Calcium intake was calculated from all food items in the questionnaire (about 80 questions), excluding supplements. Vitamin D intake was calculated from all food items in the questionnaire including cod liver oil (commonly used in Norway), but no other supplements. At the time of the data collection, margarine and butter were the only foods in Norway fortified with vitamin D.
Table 1
Consumption of dairy products (g/day) in the Norwegian Women and Cancer study (n = 64,904)
Total dairy | 174 (220) |
Whole milk | 0 (15) |
Low-fat milk | 0 (81) |
Skimmed milk | 0 (68) |
Yoghurt | 12 (25) |
Full fat white cheese | 6 (13) |
Low-fat/skimmed white cheese | 0 (8) |
Ice cream | 4 (7) |
Full fat sour cream (for fish) | 0 (0.4) |
Fat-reduced sour cream (for fish) | 1 (2) |
The intake of dairy foods and calcium was energy-adjusted by using the residual method [
36]. That is, we added the residuals of the regression of the intake of dairy foods and calcium on total energy intake, to the predicted intake at mean level of energy intake.
Assessment of other exposures
Information on age at invitation was obtained from National Central Person Register. To handle the effect of potential confounding factors and effect modifying factors the following information was derived from the questionnaire: body weight and height, weight change since age 18, level of physical activity, smoking status, years of education, maternal history of breast cancer, mammography practice, age at menarche, number of children and age at first birth, use of oral contraceptives, menopausal status, and use of hormone replacement therapy (HRT) if postmenopausal.
As only one single recording of menopausal status was available for the present analyses, we assigned all women who were premenopausal when filling out the questionnaire to be postmenopausal when they reached the age of 50 during follow-up. This procedure was followed for women reporting hysterectomy, whereas women reporting bilateral oophorectomy were considered postmenopausal throughout. The age of 50 as a dividing line for menopausal status was chosen based on data from an older sub-sample of NOWAC [
38]. Information on self-reported menopausal status and use of HRT has be evaluated against plasma levels of sex hormones and found valid [
39].
When checking for interactions between menopausal status and dairy consumption, we examined both the effect of menopausal status at the time of data recording and the effect of menopausal status at end of follow-up. No significant interaction was seen in either of the analyses.
Analytic cohort
Of the 68,517 women initially in the present cohort, we excluded 2,900 women with a prior cancer diagnosis (any type), one woman with uncertain breast cancer diagnosis, seven women who died and three women who emigrated before the start of the follow-up. We also excluded 104 women who did not answer any of the dairy questions, 585 women for whom the calculated daily energy intake was below 2,500 kJ (n = 516) or above 15,000 kJ (n = 69), and 13 women with implausible age at menopause. Thus, 64,904 women were included in the main analyses. In addition, we performed analyses excluding women who were diagnosed with any kind of cancer (n = 364) or died (n = 41) during the first year of follow-up in order to avoid the possibility that undiagnosed cancer or other severe illness influenced the self-reported data.
Statistical analyses
Person-years of follow-up were calculated as the time elapsed from the date of the returned questionnaire (defined as 3 months after mailing of the invitation letter) to the time of cancer (any type), to time of death or emigration, or to the end of follow-up (31 December 2006), whichever occurred first. Cox proportional hazards regression analyses were carried out to investigate the simultaneous effect of dairy consumption and covariates on breast cancer incidence rate, and hazard rate ratios (HRR) and 95% CI were calculated.
Various categorizations (e.g., quartiles, users–non-users) were applied for the dairy variables. The combined effect of childhood milk consumption and adult dairy consumption was examined by constructing a three-level variable: ‘low consumption,’ defined as no milk consumption as a child or 1st quartile of dairy consumption as adult and not more than next-lowest consumption (1–3 glasses/day)/2nd quartile on the other occasion; ‘high consumption,’ defined as the highest milk consumption as a child (7 or more glasses/day) or 4th quartile of dairy consumption as adult and not less than the next-highest consumption (4–6 glasses/day)/3rd quartile on the other occasions; and ‘moderate consumption,’ defined as all other combinations. The low consumption group was used as the reference category. Calcium intake was examined in quartiles.
Different combinations of the covariates reported at cohort enrollment were employed in multivariable analyses: age at cohort enrollment (5-year categories), height (quartiles: <163, 163–165, 166–169, >169 cm), body mass index (BMI = (wt(kg))/(ht(m)2)) (quartiles: <22.0, 22.0–23.9, 24.0–26.4, >26.4), BMI at age 18 (quartiles: <19.2, 19.2–20.5, 20.6–22.2, >22.2), weight increase since age 18 years (quartiles: <4, 4–9, 10–15, >15 kg), level of physical activity (inactive, moderately active, active), smoking status (current, former, never), education (<11, 11–13, >13 years), maternal history of breast cancer (yes, no), mammography practice (no, every 2nd year or more often, every 2nd year or more seldom), age at menarche (quartiles: <13, 13, 14, >14 years), number of children and age at first birth (nulliparous, 1 child/birth before age 21, 2 or more children/first birth before age 21, 1 child/birth at age 21 or older, 2 or more children/first birth at age 21 or older), use of oral contraceptives (ever, never), and alcohol consumption (0 and tertiles: 0.1–1.52, 1.53–4.13, >4.13 g/day). To assess the influence of HRT (ever, never), we conducted analysis restricted to women who were postmenopausal when completing the questionnaire. Further, we examined any possible confounding effect of intake of fruit, vegetables and potatoes, fat, and vitamin D (all in quartiles). To adjust for energy intake, we categorized the women according to their energy-adjusted intake of each dairy variable (except childhood milk consumption and the combined child/adult consumption variable) and also added total energy intake (in quartiles) to the model.
To test for linear trend in risk, we created a continuous variable by assigning ordinal numbers to each level of exposure and including the continuous variable in the regression models.
Interaction effects were tested using the likelihood ratio test. The assumptions of proportional hazards for the exposures of interest were examined by log–log plots and cumulative hazard plots.
To get an idea about the effect of measurement error in the dietary data, we used the 24-h dietary recalls from the validation study to perform a measurement error correction based on the regression calibration method [
43,
44]. Using the energy-adjusted dairy variables and alcohol on its original continuous scale, we constructed multivariate calibration models by regressing the individual means of the recall values of the dairy variables and alcohol on the questionnaire values. Information about total energy intake, age, height, weight increase since age 18, level of physical activity, years of education, maternal history of breast cancer, mammography practice, and use of oral contraceptives were also included in this model. This was done separately for each of the six dairy exposure variables. The calibration models are slightly smaller than the full models used in the main analyses due to the lower number of women included in the calibration. We excluded variables assumed not to be associated with intake of dairy products or alcohol. Data from 180 women were included in the calibration. This is a somewhat lower number than what was originally included in the validation study (
n = 238). The 180 women included here are those among the 238 who are also included in the main analysis, so that we were able to collect background information to be used in the calibration. Using the publicly available SAS-macro
blinplus (
http://www.hsph.harvard.edu/faculty/spiegelman/blinplus.html), we then performed a measurement error correction based on this calibration model.
All reported p values are two-sided, and a significance criterion of p < 0.05 was used to consider an association as statistically significant. The number of subjects included in the separate analyses varies somewhat due to item non-response. Statistical analyses were done by means of the SAS software package, version 9.1.