Non-dietary factors as risk factors for breast cancer, and as effect modifiers of the association of fat intake and risk of

Cancer Causes and Control, 1997, 8, pp. 49-56 Non-dietary factors as risk factors for breast cancer, and as effect modifiers of the association of fat intake and risk of breast cancer Cancer Causes and Control. Vol 8. 1997 David J. Hunter, Donna Spiegelman, Hans-Olov Adami, Piet A. van den Brandt, Aaron R. Folsom, R. Alexandra Goldbohm, Saxon Graham, Goeffrey R. Howe, Lawrence H. Kushi, James R. Marshall, Anthony B. Miller, Frank E. Speizer, Walter Willett, Alicja Wolk, and Shiaw-Shyuan Yaun (Received 25 March 1996; accepted in revised form 17 October 1996) To assess more precisely the relative risks associated with established risk factors for breast cancer, and whether the association between dietary fat and breast cancer risk varies according to levels of these risk factors, we pooled primary data from six prospective studies in North America and Western Europe in which individual estimates of dietary fat intake had been obtained by validated food-frequency questionnaires. Based on information from 322,647 women among whom 4,827 cases occurred during follow-up: the multivariate-adjusted risk of late menarche (age 15 years or more compared with under 12) was 0.72 (95 percent confidence interval [CI] = 0.62-0.82); of being postmenopausal was 0.82 (CI = 0.69-0.97); of high parity (three or more births compared with none) was 0.72 (CI = 0.61-0.86); of late age at first birth (over 30 years of age compared with 20 or under) was 1.46 (CI = 1.22-1.75); of benign breast disease was 1.53 (CI = 1.41-1.65); of maternal history of breast cancer was 1.38 (CI = 1.14-1.67); and history of a sister with breast cancer was 1.47 (CI = 1.27-1.70). Greater duration of schooling (more than high-school graduation compared with less than high-school graduation) was associated significantly with higher risk in ageadjusted analyses, but was attenuated after controlling for other risk factors. Total fat intake (adjusted for energy consumption) was not associated significantly with breast cancer risk in any strata of these non-dietary risk factors. We observed a marginally significant interaction between total fat intake and risk of breast cancer according to history of benign breast disease, with fat intake being associated nonsignificantly positively with risk among women with a previous history of benign breast disease; no other significant interactions were observed. Risks for reproductive factors were similar to those observed in case-control studies; relative risks for family history of breast cancer were lower. We found no clear evidence in any subgroups of a major relation between total energy-adjusted fat intake and breast cancer risk. Cancer Causes and Control 1997, 8, 49-56 Key words: Breast cancer, diet, reproductive factors, women. Drs Hunter, Speizer, Willett, and Ms Yaun are with the Channing Laboratory, Department of Medicine, Brigham and Women s Hospital and Harvard Medical School, Boston, MA, USA. Authors are also affiliated with the Harvard School of Public Health Department of Epidemiology, Boston, MA (Drs Hunter, Spiegelman, Willett), Department of Nutrition (Dr Willett), and Department of Biostatistics (Dr Spiegelman); NCIC Epidemiology Unit, Department of Preventive Medicine and Biostatistics, Faculty of Medicine, University of Toronto, Toronto, ON, Canada (Drs Howe, Miller); Division of Epidemiology, School of Public Health, University of Minnesota, Minneapolis, MN (Drs Folsom, Kushi); Department of Epidemiology, Maastricht University, The Netherlands (Dr van den Brandt); Department of Epidemiology, TNO Nutrition and Food Research Institute, Zeist, The Netherlands (Dr Goldbohm); Department of Social and Preventive Medicine, State University of New York at Buffalo, NY (Dr Graham); Arizona Cancer Center, University of Arizona, College of Medicine, Tucson, AZ (Dr Marshall); Department of Cancer Epidemiology, University Hospital, Uppsala, Sweden (Drs Adami, Wolk). Address correspondence to Dr Hunter, Channing Laboratory, 181 Longwood Ave, Boston, MA 02115, USA. This project is funded by research grants NIH CA55075 and CA50597 and by a Faculty Research Award (FRA-455) to Dr Hunter from the American Cancer Society. 1997 Rapid Science Publishers Cancer Causes and Control. Vol 8. 1997 49

D.J. Hunter et al Introduction A number of reproductive and other risk factors for breast cancer are considered established ; however, uncertainty still exists about the magnitude of the relative risks associated with different levels of these factors. 1 The likelihood of recall and selection bias in case-control studies suggests that pooled information from prospective studies with high follow-up rates should offer the most accurate assessment of the magnitude of these associations. Among the risk factors for which uncertainty exists, the relation of dietary fat intake with breast cancer is among the most controversial. 2,3 Although a significant positive association between total fat intake and breast cancer was observed in a pooled analysis of 12 casecontrol studies 4 (including 4,312 cases and 5,978 controls), in the largest case-control study to date, 5 of 2,024 cases and 1,463 controls (which was not included in the pooled analysis), an association with total fat intake was not observed. Prospective studies generally have yielded null or only weakly positive results. 6 To provide the best available summary of the prospective evidence, we pooled the primary data from six prospective studies, representing follow-up of 322,647 women among whom 4,827 incident cases of breast cancer (including 443 cases of carcinoma in situ) were diagnosed. No association was observed between total fat intake and breast cancer risk in either premenopausal or postmenopausal women; 7 however, the possibility remains that a positive association might be present among certain subgroups defined by established breast-cancer risk factors. Here, we present the associations with breast cancer of established and other non-dietary breast-cancer risk factors from this pooled analysis of prospective studies, and assess whether the association with fat intake or types of fat varies according to levels of these risk factors. Because of the complexity of the associations of body mass index (wt/ht 2 ) (BMI) and height with breast cancer risk, 1 these variables will be the subject of another report. Materials and methods We identified six cohort studies 8-13 that met the following pre-defined criteria: (i) at least 200 incident cases of breast cancer available for analysis; (ii) diet assessed at baseline with a comprehensive instrument that assessed food and nutrient intake over a medium- to long-term period and permitted an estimation of total energy intake; and (iii) data were available from a validation study of the diet assessment instrument in either the cohort itself or a closely related population. A relatively small cohort study, the Seventh-day Adventist Health Study, 14 was included in categorical analyses in our assessment of the main effect of dietary fat. These investigators had calculated a nutrient ranking index rather than estimated absolute nutrient intake however, and this index was not suitable for our analyses of interaction requiring a continuous estimate of nutrient intake (see below). Table 1. Characteristics of cohort studies included in the pooled prospective analysis of non-dietary risk factors and diet interactions in breast cancer risk Study Location Duration No. Name of follow-up Baseline cohort size a Age at No. of cases a baseline (yrs) (No. carcinoma in situ) 1. Canadian Breast Canada 1982-87 56,837 40-59 514 (85) Screening Study 8 2. Iowa Women s Health Iowa (USA) 1986-91 34,406 55-69 723 (70) Study 11 3. Netherlands Cohort Netherlands 1986-89 62,412 55-69 434 (0) c Study 12 4. New York State Cohort 10 New York (USA) 1980-87 18,475 50-93 376 (9) 5. Nurses Health Study(a) 9 USA 1980-86 89,046 34-59 1,094 (71) 6. Nurses Health Study(b) 9 USA 1986-91 68,817 40-65 911 (105) 7. Sweden Mammography Sweden 1987-93 61,471 40-76 775 (103) Cohort 13 Total 322,647 b 4,827 443 a b c Excluding subjects with previous cancer (other than non-melanoma skin cancer), incomplete dietary data, or outlying values for energy intake (in Canadian Breast Screening Study these exclusions were made for cases and controls in the nested case-control study, and in the Netherlands Cohort Study exclusions were made for cases and subcohort members). The 68,817 women in the Nurses Health Study(b) were also members of the Nurses Health Study(a) cohort. Carcinoma in situ cases were not ascertained in the Netherlands Cohort Study. 50 Cancer Causes and Control. Vol 8. 1997

Dietary fat and breast cancer Because the duration of follow-up available from the Nurses Health Study (10 years) was substantially longer than for the other studies, and to take advantage of the fact that this study has more than one round of diet assessment available, we divided it into two cohorts, Nurses Health Study(a) based on a food frequency questionnaire administered in 1980 with follow-up through 1986, and Nurses Health Study(b) based on a 1986 questionnaire with follow-up through 1991. Basic information about each study is listed in Table 1 and summarized in the study-specific publications. 8-13 Non-dietary Exposure variables. Non-dietary exposure information was self-reported by women in each study on selfadministered questionnaires. Although information on the validity of self-reports of variables such as age at menarche, parity, age at first birth, and education is not available, the face validity of these variables is high. In one of the component studies, the Nurses Health Study, 15 self-reported information on menopausal status was found to be highly accurate. Information on self-reported family history of breast cancer has been found to be acceptably valid when prospectively obtained. 16 Questions on benign breast disease did vary substantially among cohorts, ranging from enquiry about any history of breast lumps, to questions limited to biopsy-proven benign breast disease. Differences in the prevalence of benign breast disease in the cohorts (Table 2) probably are due partially to differences in the specificity of these questions. Statistical analyses Exclusion criteria. In addition to exclusion criteria originally applied to individual studies, we excluded subjects whose estimate of total energy intake was more than three standard deviations from the studyspecific log e-transformed mean of the baseline population. We also excluded the small percentage of subjects who had been diagnosed with cancer prior to baseline (other than non-melanoma skin cancer), as their recent dietary patterns may have been influenced by cancer or its treatment. Because of these exclusions, and extended follow-up in the Iowa Women s Health Study 11 and Nurses Health Study, for most studies the baseline cohort size and number of cases is slightly different in our analysis (Table 1) than in the original published analyses. Selection of cases and sampling of risk sets. To reduce computational burdens, we analyzed five studies (Iowa Women s Health Study, 11 New York State Cohort, 10 Nurses Health Study(a), 9 Nurses Health Study(b), 9 and Sweden Mammography Cohort 13 ) as nested casecontrol studies with a matching ratio of 10 controls for each case. Cases were assigned to the calendar year in which they were diagnosed and their follow-up ceased in that year. For each case, 10 controls were selected from the risk set of women with the same year of birth, who were alive, not known to have out-migrated from the study, and who had not been diagnosed with breast cancer by the beginning of the year in which the case was diagnosed. Controls were sampled without replacement within each year, but were eligible to be chosen again or to become cases in subsequent years. A similar design was used for the Canadian Breast Screening Study, 8 but data were only available for two controls per case. In the Netherlands Cohort Study, 12 a case-cohort design was used; 17 cases were ascertained from the entire cohort (the numerator information for incidence rates), while the accumulated person-years of the entire cohort were estimated using a subcohort (providing the denominator information) of 1,812 women randomly sampled at baseline. Nested case-control and case-cohort studies, shown to be an efficient and unbiased alternative to full cohort analysis, are not susceptible to the recall and selection biases which may arise in conventional case-control studies. 18 Models and analyses. The basic model for these analyses is the proportional hazards model. 19 For the five studies for which nested case-control sampling was used, conditional logistic regression analysis was used to fit this model, with SAS PROC PHREG. 20 For the Netherlands Cohort Study, Cox regression analysis was used, with variance modified as required for the case-cohort design using EPICURE software. 21 To estimate the rate ratio, or relative risk (RR), we exponentiated the appropriate conditional logistic regression coefficient multiplied by a 25 g increment for total fat intake, 10 g increments for saturated, polyunsaturated and monounsaturated fat, and 100 mg increment for dietary cholesterol intake. We used indicator variables for categorical analyses of the nondietary risk factors. In the analyses of non-dietary risk factors, indicator variables were included for the relatively low numbers of subjects with missing data for these variables; studies for which individual variables were missing entirely are indicated in Table 2. Two-sided 95 percent confidence intervals (CI) are presented throughout. To estimate the effect of total and type of fat intake at different levels of non-dietary risk factors, we included an interaction term between nutrient intake (grams/day) and each level (excluding the referent level) of the nondietary risk factor, controlling for confounding by all other risk factors; subjects with missing values of the covariate were deleted. The RR for a 25 g difference in Cancer Causes and Control. Vol 8. 1997 51

D.J. Hunter et al 52 Cancer Causes and Control. Vol 8. 1997

Dietary fat and breast cancer total fat intake, for a 10 g difference in saturated, monounsaturated, or polyunsaturated fat intake, or 100 mg of dietary cholesterol, then was calculated at each level of the risk factor by combining the coefficients for the main effect of nutrient and the interaction term at that level. The study-specific test for interaction was calculated from the likelihood ratio test comparing models with and without a single interaction term. The pooled P-value for interaction was obtained by squared Wald statistics, constructed by dividing the pooled interaction term by its pooled standard error, referred to a chi-square distribution with one degree of freedom. Energy-adjustment. To provide information on the effect of dietary composition comparable to that obtained in an isocaloric metabolic study, we adjusted nutrient intakes for total energy intake using the residual method 22 (in which the log e-transformed nutrient is regressed against log e-transformed energy intake and the residual standardized to a median energy intake of 1,600 kcal); the residual represents nutrient intake independent of energy intake. Pooling of relative risks. We used the random effects model methods developed by DerSimonian and Laird 23 to combine log RRs from multiple studies. Fixed effects models assume that the only source of between-study variability is random within-study sampling variation, while random effects models allow for additional between-study variation as well. Because the test for between-studies heterogeneity is believed to have limited power, we wish to allow for the possibility of some between-studies variation and use the random effects model. Random effects models result in wider CIs and a more conservative interpretation of the data. Results Non-dietary risk factors Study-specific and pooled estimates for non-dietary risk factors, and their study-specific prevalence, are presented in Table 2. For all but one variable (maternal history of breast cancer), the test for heterogeneity was not statistically significant, suggesting that the pooled estimates are appropriate summaries of the study-specific data. In general, study-specific estimates are in the expected direction, although in certain instances (e.g., maternal history in the Sweden Mammography Cohort) the magnitude is less than expected. The pooled estimates, however, are associated with much smaller confidence limits; indeed, in many cases, the study-specific estimates are not statistically significant although their magnitudes are similar to the pooled estimate (e.g., postmenopausal status in the Canadian Breast Screening Study). Both later age at menarche (age 15 years or more compared with under 12) and higher parity (three or more births compared with none) are associated with reductions in risk of breast cancer of about 25 percent, while among parous women, later age at first birth (over 30 years compared with 20 or under) increases risk by about 50 percent. Higher levels of education were associated with increased risk in age-adjusted analyses, although this association was attenuated substantially in multivariate analyses. A history of benign breast disease was associated with an approximately 50 percent increase in risk, as was history of breast cancer in mother or sister(s). Interaction of fat intake with non-dietary risk factors For each nutrient, and energy intake, we assessed whether the association with breast cancer was modified across the categories of the covariates listed in Table 2 (see footnote to Table 3 for categories). There were no significant interactions between energy intake and these covariates. For total fat (adjusted for energy consumption), only one marginally significant interaction was observed. Among women with a history of benign breast disease, the RR for each 25 g increase in total fat was 1.29 (CI = 0.96-1.72), while among women with no history of benign breast disease, this risk was 0.95 (CI = 0.87-1.05); P value for interaction = 0.05. A similar interaction with benign breast disease was observed for saturated fat (P value for interaction = 0.06), monounsaturated fat (P value for interaction = 0.09) and polyunsaturated fat (P value for interaction = 0.14), but not for cholesterol intake (P value for interaction = 0.63). In several instances, study-specific interactions were significant, but the overall pooled estimates were not. For example, in the Netherlands Cohort Study, a highly significant interaction (P < 0.001) was observed between total fat intake and age at first birth: the RR for 25 g of energy-adjusted total fat for women with age at first birth 20 years or under was 2.34 (CI = 0.37-14.74); whereas for women with age at first birth over age 30, this RR was 0.40 (CI = 0.19-0.84). A weaker interaction (P = 0.07) was observed in Nurses Health Study(b), but the effect modification was in the opposite direction; the RR for 25 g of energy-adjusted total fat for women with age at first birth 20 or under was 0.61 (CI = 0.18-2.05), whereas for women with age at first birth over 30, this RR was 1.32 (CI = 0.73-2.39). Across all six studies, the stratum-specific pooled RRs offered no evidence of interaction (P value for interaction = 0.73). Overall, of 60 pooled interactions assessed (10 covariates tested for each of total, saturated, monounsaturated, polyunsaturated fat, cholesterol, and energy), none were statistically significant at the P < 0.05 level, and one was marginally significant (P = 0.05 for total fat and benign breast disease as above). Cancer Causes and Control. Vol 8. 1997 53

D.J. Hunter et al Table 3. Pooled multivariate relative risks (RR) and 95% confidence intervals (CI) for a 25 g increase in energy-adjusted total fat consumption, according to levels of non-dietary risk factors in multivariate analyses a of data from the pooled analysis of prospective studies Variable Overall b (CI) Category of variable Interaction P-value Age at menarche < 12 yrs 12 yrs 13 yrs 14 yrs 15 yrs 1.02 1.00 1.03 0.98 1.16 0.95 0.29 (0.91-1.14) (0.75-1.34) (0.88-1.19) (0.79-1.22) (0.94-1.41) (0.75-1.20) Menopausal status Pre- Post- 0.97 0.98 0.96 0.95 (0.88-1.07) (0.85-1.13) (0.86-1.08) Parity Null 1-2 3 1.02 0.90 1.10 0.99 0.82 (0.94-1.10) (0.76-1.06) (0.99-1.22) (0.88-1.11) Age at first birth (yrs) 20 > 20-25 > 25-30 > 30 1.04 1.09 1.09 1.05 0.96 0.73 (0.94-1.16) (0.87-1.36) (0.89-1.34) (0.89-1.24) (0.75-1.24) Education < HS HS > HS 1.08 1.03 1.12 1.09 0.75 (0.95-1.21) (0.88-1.21) (0.93-1.34) (0.93-1.27) Benign breast disease No Yes 1.01 0.95 1.29 0.05 (0.90-1.13) (0.87-1.05) (0.96-1.72) Mother with breast cancer 1.02 1.02 1.00 0.94 (0.91-1.14) (0.90-1.15) (0.77-1.29) Sister with breast cancer 1.05 1.05 1.02 0.86 (0.94-1.18) (0.94-1.18) (0.74-1.42) Alcohol intake (g/day) 0 0- < 1.5 1.5- < 5 5- < 15 15- < 30 30 0.95 1.02 1.09 1.03 0.97 1.00 1.20 0.77 (0.94-1.12) (0.93-1.28) (0.84-1.27) (0.81-1.17) (0.83-1.19) (0.93-1.55) (0.47-1.28) a Multivariate model includes terms for: age (year of birth); calendar time (single years); age at menarche ( 11, 12, 13, 14, 15); menopausal status (pre-, post-); parity (0, 1-2, 3); age at first birth ( 20, 21-25, 26-30, 30); body mass index (wt/ht 2 ) ( 21, 21-22, 23-24, 25-29, > 29 kg/m 2 ); height (< 1.60, 1.60-< 1.64, 1.64-< 1.68, 1.68 m); education (< high school graduate, high school graduate, > high school graduate); history of benign breast disease (no, yes); maternal history of breast cancer (no, yes); sister(s) history of breast cancer (no, yes); oral contraceptive use, ever (no, yes); fiber intake (quintiles); alcohol intake (0, 0-< 1.5, 1.5-< 5, 5-< 15, 15-< 30, 30 gms/day); energy intake (continuous). b Estimate for total fat was calculated after excluding missing values for each variable, and studies which are uniform for the variable (e.g., Iowa Women s Health Survey, Netherlands Cohort Study, New York State Cohort for menopause). Thus, the estimates vary due to differences in the number of subjects included. Discussion Using the largest set of prospective data available, we confirmed and quantified the associations of established risk factors with breast cancer; in almost all strata of these risk factors, no association between dietary fat intake and breast cancer was observed. Although differential recall by cases and controls of major events such as age at first birth and parity seem unlikely, inappropriate selection or incomplete participation of controls may lead to biased estimates for these risk factors in case-control studies. Recall bias for other risk factors such as family history and dietary fat intake has been demonstrated to occur and results in over-estimates of the magnitude of these associations. 7,24,25 Thus, prospective information is preferable in assessing both the direction and the magnitude of these associations. The associations we observed with late age at menarche, high parity, late age at first birth, and menopause underscore the importance of these reproductive variables in 54 Cancer Causes and Control. Vol 8. 1997

Dietary fat and breast cancer determining breast cancer risk, and provide more precise estimates of their magnitude than previously available. The pooled RR estimates for these variables are in good agreement with those observed in large case-control studies such as the Cancer and Steroid Hormone Study 26 and a multicenter international case-control study. 27,28 Interestingly, even for these established RRs, the modest associations observed in some studies are not statistically significant (e.g., age at menarche in the Nurses Health Study [a] and [b]), while the pooled estimates are highly significant. This demonstrates one of the advantages of pooling these data from multiple studies when heterogeneity is minimal. Our results suggest that later age at menarche (15 or over cf under 12 years) is associated with a 28 percent reduction in breast cancer risk, that higher parity (three or more births cf none) also predicts a 28 percent reduction in risk, and early age at first birth among parous women (age 20 or under cf over 30) is associated with a 32 percent reduction in risk. These data reinforce the fact that reproductive events are important predictors of breast cancer incidence, and that individual breast cancer risk is substantially determined relatively early in life. The 50 percent increase in risk for history of benign breast disease emphasizes the importance of this risk factor in identifying a group of women at substantially higher risk of subsequent breast cancer. Questionnairebased information on benign breast disease, however, masks substantial variation in the risks associated with certain histologic subtypes of benign breast disease; the presence of atypical hyperplasia in benign breast disease lesions confers an RR of 2.6-5.3, and for proliferative disease without atypia, 1.6-2.0, compared with biopsies without evidence of proliferative disease; 29 unfortunately, we did not have this detailed histologic information available in the cohorts. The 40 to 50 percent elevations in risk associated with history of breast cancer in mother or sister are lower than the RRs of 2.0 or greater previously reported, mostly from case-control studies. 16 While some of this difference may be accounted for by chance, and variation in age of enrollment (risk associated with family history declines with age), 16 it is likely that some of the lower risk may be due to the fact that recall bias in reporting of family history of breast cancer could occur in studies in which family history is assessed retrospectively. The heterogeneity in the study-specific estimates of the risk associated with maternal history of breast cancer was surprising, and was not obviously attributable to older age of the women enrolled in the studies in which the association was weak. This heterogeneity suggests that results from any one study should be interpreted with caution. All of these exposure variables were self-reported and not independently confirmed. However, validity of variables such as age at menarche, age at first birth, and parity is likely to be high. Information on benign breast disease is certainly much less reliable, although the questions asked in most studies were typical of these which could be asked in a clinical or screening setting. We have shown elsewhere 7 that the pooled main effects of total fat intake on breast cancer risk, and the fat subtypes considered individually, are close to unity. A further major goal of this study was to assess whether certain subgroups of women may be at higher risk of breast cancer if they consume diets relatively high in fat. Assessment of an interaction typically requires a sample size four or more times the size needed to detect a main effect of the same size; 30 thus, most individual cohorts are limited in their ability to study interactions. On the other hand, significant interactions may arise by chance due to multiple comparisons, particularly when many nutrients and many covariates are involved. An advantage of pooling the primary data from these studies is that interactions with classical breast cancer risk factors can be examined using uniform categories, and interactions observed in one data set can be immediately compared with the results from other data sets. In the example of age at first birth for instance, a highly significant interaction with total fat intake in Netherlands Cohort Study was not observed in other data sets. In fact, in Nurses Health Study(b), the interaction was almost significant in the opposite direction, and the overall pooled estimates offered little evidence of interaction. Overall, out of the 60 studyspecific pooled interactions examined, only one was marginally statistically significant, less than the five percent expected under the global null hypothesis of no effect modification whatsoever. Results for the nondietary risk factors and interactions with fat intake were similar when we excluded the small proportion (nine percent) of cases diagnosed with carcinoma in situ rather than invasive disease. We did observe a marginally significant increased risk with higher intake of dietary fat among women with a history of benign breast disease. Although increased dietary fat has been suggested to be a risk factor for benign breast disease, we are not aware of published data suggesting that an adverse effect of dietary fat might be limited to women with benign breast disease. Rose et al 31 in a metabolic study of women with cystic breast disease, observed that a low-fat diet reduced serum estradiol levels; however, parallel information from women without cystic breast disease was not reported. Although these data are intriguing, in the absence of corroborating information, this finding should be treated as preliminary in view of the relatively large number of hypotheses we tested. A limitation of these data derived from studies in North America and Europe is that relatively few women consumed a very low fat diet (e.g., less than 20 percent Cancer Causes and Control. Vol 8. 1997 55

D.J. Hunter et al of calories from fat). In the overall analysis we had sufficient power to show that there was no reduction in risk associated with this very low fat diet; 7 however, in these interaction analyses, we had less power to examine this association within smaller subgroups of levels of the individual risk factors. In general however, the RR estimates we calculated for a 25 g increase in total fat intake were close to unity in most of these subgroups, implying that any substantial reduction in risk at very low fat intakes would represent a departure from linearity of the association of fat intake and breast cancer risk. The international correlation studies of fat intake and breast cancer risk do not show evidence of such threshold effects. In summary, we confirmed several expected relations of lifestyle with breast cancer, and provided more precise estimates of these associations from prospective studies than previously available. Although we observed a suggestion that women with a history of benign breast disease may be at higher risk of breast cancer if they consume a diet higher in fat, we did not observe any clear evidence for a major relation between total fat intake (adjusted for energy consumption) and breast cancer risk within strata of other breast cancer risk factors. Acknowledgements The authors thank Tracey Corrigan for manuscript preparation, and Laura Newcomer and Walkyria Pas de Almeida for computer programming. References 1. Kelsey JL. Breast cancer epidemiology: Summary and future directions. Epidemiol Rev 1993; 15: 256-63. 2. Prentice RL, Sheppard L. Dietary fat and cancer: consistency of the epidemiologic data, and disease prevention that may follow from a practical reduction in fat consumption. Cancer Causes Control 1990; 1: 81-97. 3. Willett WC, Stampfer MJ. Dietary fat and cancer: another view. Cancer Causes Control 1990; 1: 103-9. 4. Howe GR, Hirohata T, Hislop TG, et al. Dietary factors and risk of breast cancer: combined analysis of 12 casecontrol studies. JNCI 1990; 82: 561-9. 5. Graham S, Marshall J, Mettlin C, et al. Diet in the epidemiology of breast cancer. Am J Epidemiol 1982; 116: 68-75. 6. Howe GR. Dietary fat and breast cancer risk; an epidemiologic perspective. Cancer (Suppl) 1994; 74: 1078-94. 7. Hunter DJ, Spiegelman D, Adami H-O, et al. Cohort studies of fat intake and the risk of breast cancer a pooled analysis. N Engl J Med 1996; 334: 356-61. 8. Howe GR, Friedenreich CM, Jain M, Miller AB. A cohort study of fat intake and risk of breast cancer. JNCI 1991; 83: 336-40. 9. Willett WC, Hunter DJ, Stampfer MJ, et al. Dietary fat and fiber in relation to risk of breast cancer: an 8-year follow-up. JAMA 1992; 268: 2037-44. 10. Graham S, Zielezny M, Marshall J, et al. Diet in the epidemiology of postmenopausal breast cancer in the New York State cohort. Am J Epidemiol 1992; 136: 1327-37. 11. Kushi LH, Sellers TA, Potter JD, et al. Dietary fat and postmenopausal breast cancer. JNCI 1992; 84: 1092-9. 12. van den Brandt PA, van t Veer P, Goldbohm RA, et al. A prospective cohort study on dietary fat and the risk of postmenopausal breast cancer risk. Cancer Res 1993; 53: 75-82. 13. Holmberg L, Ohlander EM, Byers T, et al. Diet and breast cancer risk: results from a population-based, case-control study in Sweden. Arch Intern Med 1994; 154: 1805-11. 14. Mills PK, Beeson WL, Phillips RL, Fraser GE. Dietary habits and breast cancer incidence among Seventh Day Adventists. Cancer 1989; 64: 582-90. 15. Colditz GA, Stampfer MJ, Willett WC, et al. Reproducibility and validity of self-reported menopausal status in a prospective cohort of women. Am J Epidemiol 1987; 126: 319-25. 16. Colditz GA, Willett WC, Hunter DJ, et al. Family history, age, and risk of breast cancer: prospective data from the Nurses Health Study. JAMA 1993; 270: 338-43. 17. Prentice RL. A case-cohort design for epidemiologic cohort studies and disease prevention trials. Biometrika 1986; 73: 1-12. 18. Langholz B, Thomas DC. Nested case-control and casecohort methods of sampling from a cohort: a critical comparison. Am J Epidemiol 1990; 131: 169-76. 19. Cox DR. Regression models and life-tables. J R Stat Soc Series B 1972; 34: 187-220. 20. SAS Institute, Inc. SAS/STAT Software. The PHREG Procedure. Preliminary Documentation. Cary, NC (USA): SAS Institute, 1991. 21. Hirosoft International Corporation. EPICURE/PEANUTS Software. The PEANUTS Program. EPICURE User s Guide. Seattle, WA (USA): Hirosoft, 1993. 22. Willett WC, Stampfer MJ. Total energy intake: implications for epidemiologic analyses. Am J Epidemiol 1986; 124: 17-27. 23. DerSimonian R, Laird N. Meta-analysis in clinical trials. Controlled Clin Trials 1986; 7: 177-88. 24. Floderus B, Barlow L, Mack TM. Recall bias in subjective reports of familial cancer. Epidemiology 1990; 1: 318-21. 25. Love RR, Evans AM, Josten DM. The accuracy of patient reports of a family history of cancer. J Chronic Dis 1985; 38: 289-93. 26. Layde PM, Webster LA, Baughman AL, et al. The independent associations of parity, age at first full term pregnancy, and duration of breastfeeding with the risk of breast cancer. J Clin Epidemiol 1989; 42: 963-73. 27. Trichopoulos D, MacMahon B, Cole P. Menopause and breast cancer risk. JNCI 1972; 48: 605-13. 28. MacMahon B, Cole P, Lin TM, et al. Age at first birth and breast cancer risk. Bull World Health Organ 1970; 43: 209-21. 29. Bodian, CA. Benign breast diseases, carcinoma in situ, and breast cancer risk. Epidemiol Rev 1993; 15: 177-87. 30. Smith PG, Day NE. The design of case-control studies: the influence of confounding and interaction effects. Int J Epidemiol 1984; 13: 356-65. 31. Rose DP, Boyer AP, Cohen L, et al. Effect of a low-fat diet on hormone levels in women with cystic breast disease. I. Serum steroids and gonadotropins. JNCI 1987; 78: 623-6. 56 Cancer Causes and Control. Vol 8. 1997