点击显示 收起
1 From the Department of Cancer Prevention and Control, Roswell Park Cancer Institute, Buffalo, NY (SEM, WEM, C-CH, JRM, and SBE); the Department of Social and Preventive Medicine, University at Buffalo, Buffalo, NY (MT and JLF); and the Department of Epidemiology, Italian National Cancer Institute Regina Elena, Rome, Italy (PM)
2 Supported in part by grants 5K07CA089123 from the National Cancer Institute, DAMD17-96-1-6202 from the United States Army Medical Research and Materiel Command, and R01CA92040 from the NIH. 3 Address reprint requests to SE McCann, Department of Cancer Prevention and Control, Roswell Park Cancer Institute, Elm and Carlton Streets, Buffalo, NY 14263. E-mail: susan.mccann{at}roswellpark.org
ABSTRACT
Background: Reduced rank regression (RRR) has been used to identify dietary patterns that predict variation in a selected risk factor and may be useful in describing dietary exposures associated with glycemic index (GI) and glycemic load (GL).
Objective: To estimate breast cancer risk, we compared the relative utility of RRR-derived dietary patterns predictive of GI and GL with those of simple GI and GL.
Design: RRR was used to identify dietary patterns predicting GI and GL from food-frequency data obtained in the Western New York Exposure and Breast Cancer Study (1166 cases, 2105 controls). Odds ratios (ORs) and 95% CIs were estimated with unconditional logistic regression, adjusted for energy and nondietary breast cancer risk factors.
Results: Sweets, refined grains, and salty snacks explained 34% of the variance in GI and 68% of the variance in GL. In general, breast cancer risks were not associated with GI, GL, or dietary pattern score. However, we observed a significant reduction in postmenopausal breast cancer risk with GI and GL pattern scores combined (OR: 0.68; 95% CI: 0.50, 0.93), especially in women with a body mass index (in kg/m2) 25 (OR: 0.64; 95% CI: 0.44, 0.93). Conversely, in premenopausal women, increased risks were associated with high GL pattern scores only for women with a body mass index 25 (OR: 2.21; 95% CI: 1.04, 4.69).
Conclusions: Although RRR may be useful in studies of diet and disease, our results suggest that RRR dietary patterns based on GI and GL provide similar information regarding the association between breast cancer, GI, and GL.
Key Words: Dietary patterns • glycemic index • glycemic load • breast cancer • body mass index
INTRODUCTION
Many previous studies have investigated the association of glycemic index (GI) and glycemic load (GL) with breast cancer. GI reflects the effect of carbohydrates in individual foods on the postprandial glycemic response, whereas glycemic load (GL) includes both the GI and total carbohydrate intake; thus, it approximates the total glycemic effect of the diet given an adequate assessment of total diet (1). Dietary GI and GL can affect carbohydrate metabolism in vivo: high GI and GL have been associated with hyperinsulinemia, impaired glucose tolerance, and higher circulating insulin-like growth factor (IGF) concentrations (2-6). An increased risk of breast cancer is associated with hyperinsulinemia (7), high IGF concentrations (8-10), and GI and GL in a few (11-13), but not all (14-19), epidemiologic studies.
Because the glycemic effect of an individual food may not be representative of that of a mixed diet, the validity and utility of GI as a measure of the glycemic effect of diet on carbohydrate metabolism and related disease risk has been questioned (20, 21). Construction of the GI relies on the availability of reliable food-composition data for foods contributing to carbohydrate intake in the target population. Although the GI has been widely studied, it is limited in that it does not take into account specific combinations of foods that might affect glycemic response. Thus, the extent to which food combinations might affect GI and GL and subsequent disease risk remains unclear.
Methodology, which takes advantage of the multidimensionality of food consumption, has been developed to describe patterns of food use that can be associated with chronic disease (22, 23). Principal components analysis (PCA), the most widely used dietary patterns method, produces linear combinations of foods with the goal of identifying patterns that explain the largest variation in food use. However, PCA patterns do not necessarily correspond to a physiologic response. Recently, another statistical method, reduced rank regression (RRR), was developed to derive dietary patterns that predict a specific response such as a biomarker or a nutrient that has been previously associated with a chronic disease outcome (24, 25). In this regard, RRR has an advantage over PCA in that the identified dietary patterns not only describe food use, but do so in association with a previously described disease risk factor. Given the inconsistent associations reported for GI and GL and breast cancer, and the methodologic concerns associated with GI as a risk factor, we investigated the use of RRR to identify dietary patterns related to GI and GL and their association with pre- and postmenopausal breast cancer in the Western New York Exposure and Breast Cancer (WEB) Study.
SUBJECTS AND METHODS
Study subjects
The methods for the WEB Study were reported in detail elsewhere (26, 27). Briefly, data were collected between 1996 and 2001 for 1166 women aged 35-79 y with incident, primary, histologically confirmed breast cancer and 2105 controls with no history of cancer other than non-melanoma skin cancer frequency-matched by age, race, and county of residence to cases. Women aged <65 y were randomly selected from the New York State Department of Motor Vehicles driver's license list, and those aged 65 y were randomly selected from the Health Care Finance Administration rolls. The study protocol was approved by the Institutional Review Boards of the University at Buffalo and participating hospitals, and informed consent was obtained from all subjects.
Data collection
Demographic, health, and lifestyle related data were collected during detailed in-person interviews by trained interviewers. Height and weight were obtained by trained personnel using a standardized protocol. Body mass index (BMI) was calculated as weight in kilograms divided by the square of height in meters. We used a slightly modified version of the Health Habits and History food-frequency questionnaire (28) (FFQ) to query about diet 12-24 mo before diagnosis in cases or an interview in controls. Nutrient intake from the FFQ was calculated with DIETSYS (version 3.7) nutrient analysis software developed for use with this instrument and configured to accommodate our modifications (29).
Exposure variables
GI and GL were calculated by using values from published international tables (1). Overall dietary GL was calculated as the product of each food-specific GI, frequency of that food's use, and the carbohydrate content of the food summed across foods. Overall dietary GI was calculated by dividing dietary GL by total carbohydrate intake.
Dietary patterns were derived by using the RRR option of the partial least-squares procedure (PROC PLS) in SAS (version 9.1; SAS Institute Inc, Cary, NC) for WINDOWS. We performed 3 separate analyses with the following dependent variables: GI, GL, and GI and GL. RRR identifies as many factors as there are dependent variables; therefore, one factor each was identified for GI and GL when considered separately. Two factors were identified when GI and GL were considered together.
We considered the frequency of food use as predictor variables in all RRR models. Individual foods were grouped into 49 detailed food groups and foods (Appendix A). Foods were grouped according to nutrient content and use; foods thought to contribute to specific dietary patterns or with unique nutrient contents were left ungrouped (considered as separate items). To identify the patterns, all 49 food groups and individual foods listed in Appendix A were entered into the separate RRR analyses. All models were developed by using random cross validation, and food variable frequencies were centered and scaled before inclusion in the models. Participants were assigned scores on each factor, which were calculated as the product of the frequency of use of each food or food group and its factor loading summed across foods.
View this table:
Appendix A. Foods and food groups included in the derivation of dietary patterns associated with the glycemic index and glycemic load: Western New York Exposure and Breast Cancer Study (1996-2001)
Statistical analyses
All statistical analyses were conducted by using SAS for WINDOWS version 9.1. All tests were 2-sided and were considered statistically significant at P < 0.05. It is generally accepted in the breast cancer research community that breast cancer is most likely a separate disease for pre- and postmenopausal women, with different etiologic pathways. Therefore, all analyses were stratified by menopausal status. To estimate the association of breast cancer with GI, GL, and scores for each dietary pattern, we categorized each continuous index and score into quartiles based on the menopause-specific distribution in the controls. Odds ratios (ORs) and 95% CIs for each quartile referent to the lowest quartile were computed with unconditional logistic regression with adjustment for age, years of education, race, BMI, age at menarche, age at first birth, parity, history of benign breast disease, family history of breast cancer, and total energy intake. Models for postmenopausal women were further adjusted for age at menopause. Because several previous investigations of GI and GL and breast cancer have reported effect modification by BMI, we also conducted analyses as above, stratified by BMI (< 25 and 25).
RESULTS
The characteristics of the WEB Study participants were previously reported (27). Briefly, 40% of the study sample was premenopausal (30% of cases) and most were white (90%). Mean daily GI values were 77 ± 10.5 for premenopausal women and 76 ± 10.7 for postmenopausal women. Mean daily GL values were 253 ± 141 for premenopausal women and 237 ± 125 for postmenopausal women.
As described in the methods, RRR produces as many factors as there are dependent variables. We identified one dietary pattern each for GI and GL; 2 patterns were identified when both GI and GL were included in the model. The foods and factor loadings associated with each pattern are shown in Table 1. In general, dietary patterns predicting 34% of the variance in GI and 68% of the variance in GL, independently as well as simultaneously, consisted primarily of 3 to 4 food groups: sweets, refined grains, salty snacks, and (added) fats. Dark bread, cruciferous vegetables, yellow vegetables, and green beans were negatively associated with the GI factor, which suggests that these foods were less frequently consumed in diets with a higher GI. These foods also loaded negatively on the second factor obtained when we derived dietary patterns simultaneously predicting GI and GL.
View this table:
TABLE 1. Factor loadings for the top 10 foods associated with reduced rank regression-derived dietary patterns associated with glycemic index, glycemic load, or both combined1
The associations between breast cancer and GI, GL, and related dietary pattern scores are shown in Table 2. In general, in premenopausal women, breast cancer was not related to GI, GL, or any of the patterns derived from these indexes. However, although not statistically significant, we observed a suggestive trend toward a decrease in risk of breast cancer for postmenopausal women in the highest versus lowest quartile of GI (OR: 0.80; 95% CI: 0.61, 1.03) and GL (OR: 0.74; 95% CI: 0.53, 1.03). This association was mirrored in the GL-related pattern (OR: 0.78; 95% CI: 0.56, 1.08) and became statistically significant for women in the highest versus lowest quartile of the first dietary pattern predicting GI and GL simultaneously (OR: 0.68; 95% CI: 0.50, 0.93).
View this table:
TABLE 2. Odds ratios (ORs) and 95% CIs for the association of glycemic index, glycemic load, and related dietary patterns with pre- and postmenopausal breast cancer: Western New York Exposure and Breast Cancer Study (1996-2001)
When stratified by BMI, we observed a 2-fold increase in breast cancer risk associated with the highest versus lowest quartile of the GL-related dietary pattern in premenopausal women with a BMI 25 (OR: 2.21; 95% CI: 1.04, 4.69), but no association was observed with GI or GL expressed as an index (Table 3). In postmenopausal women, the decrease in risk in the highest versus lowest quartile of GL appeared to be limited to those women with a BMI 25 (OR: 0.63; 95% CI: 0.42, 0.94); the respective values for a BMI < 25 were as follows: OR: 1.07; 95% CI: 0.60, 1.91. Similarly, decreased risks were somewhat stronger among the overweight postmenopausal women in the highest versus lowest quartile of the first dietary pattern predicting GI and GL simultaneously: BMI < 25 (OR: 0.79; 95% CI: 0.45, 1.39) and BMI 25 (OR: 0.64; 95% CI: 0.44, 0.93).
View this table:
TABLE 3. Odds ratios (ORs) and 95% CIs for the association of glycemic index, glycemic load, and related dietary patterns with pre- and postmenopausal breast cancer by BMI: Western New York Exposure and Breast Cancer Study (1996-2001)
DISCUSSION
We had 2 main aims in the present study. The first was to explore the use of RRR to identify dietary patterns related to GI and GL. Using RRR, we identified refined grains, salty snacks, and added fats as most important in explaining both GI and GL. Although these 3 groups make intuitive sense in predicting GI and GL, none was independently associated with breast cancer risk in these data (data not shown). In fact, although we present factor loadings for the foods most strongly related to each pattern, the advantage of RRR-derived dietary patterns is that all foods and food groups load on each pattern, albeit with varying strengths. Each subject is thus assigned a score representing the sum of the frequency of use of all foods and food groups weighted by the food-specific factor loading and can be subsequently ranked on the strength of adherence to a particular pattern of intake. Thus, the pattern scores represent a total dietary pattern of behavior rather than the consumption of one or more food groups (25). On the other hand, it is entirely possible that some other component of diet associated with GI or GL is responsible for any changes in risk observed with the GI- or GL-related dietary patterns.
Our second goal was to examine associations with breast cancer risk for these dietary patterns and for GI and GL. Similar to several previous reports, we observed little association between breast cancer and GI or GL, whether expressed as an index or dietary pattern score, at least when the data were not stratified by BMI (14-19). In this regard, we might conclude that RRR dietary patterns derived from GI and GL showed no advantage over the simpler GI or GL indexes. However, among overweight premenopausal women, the GL dietary pattern was associated with a 2-fold increase in risk, contrary to the lack of association observed with the GL index in our data. Because excess adiposity may be associated with hyperinsulinemia and hyperglycemia, it is possible that the GL dietary pattern provides a better estimate of the glycemic effect of the diet on breast cancer risk in overweight women because it emphasizes those foods most strongly associated with GL. Alternatively, the GL dietary pattern includes other dietary behaviors related to GL and may be more relevant to total exposure as it relates to breast cancer risk.
RRR has been applied in many investigations of diet and chronic disease. The method represents a logical advance over PCA-derived dietary patterns in that it can identify dietary patterns associated with physiologic effects putatively resulting from these patterns of intake. RRR has been shown to be useful in identifying dietary patterns predictive of all-cause mortality (24), weight change over time (30), coronary heart disease (31), and diabetes (32). Although much of the previous work with RRR has used biomarkers of disease, which presents challenges in the study of cancer, the method does not require a biomarker, but it can be applied with other risk factors, such as nutrient intake. Especially with regard to GI and GL, it might have been advantageous to have used biomarkers of glycemic control—such as serum insulin, C-peptide, and hemoglobin A1c (Hb A1c) or a combination thereof—to identify dietary patterns using this approach, because we used food-frequency data to calculate GI and GL. The use of FFQ data does not capture how foods are combined into meals, which is an important determinant of the true glycemic effect of diet. Whereas the food groups that loaded the most strongly were those foods that tended to have a higher GI (refined grains and salty snacks), it is uncertain how having biomarkers of glycemic control might have improved our pattern identification. On the other hand, we previously showed that although the inclusion of varying degrees of detail in the generation of dietary patterns has a marginal effect on the character of the identified patterns, it may affect the precision of risk estimates associated with the resulting pattern scores (33). Therefore, although we may have identified similar patterns using additional biomarkers of glycemic effect, our estimates of risk associated with the patterns may have differed.
The present study used data from a large case-control study of environmental and dietary exposures in the etiology of breast cancer and, thus, is subject to the biases inherent in case-control designs. We queried about diet in the time period 1-2 y before diagnosis in cases or conducted an interview in controls. It is possible that dietary recall was biased either by a breast cancer diagnosis or by current diet. This source of bias should not have affected the identification of dietary patterns by RRR, although it could have affected the associations with breast cancer case-control status. Nevertheless, several of our observations were consistent with previously published data, which showed an increased risk in premenopausal women, especially at higher BMIs (14). Another potential limitation of the study was the availability of only one measure of diet. It is possible that more recent GI and GL data are not relevant in the etiology of breast cancer, given the relatively long latency of most cancers. If earlier diet is more important, it might explain the lack of positive associations observed in postmenopausal women compared with the increased risks observed in the overweight premenopausal women.
On the other hand, we observed an unexpected inverse association with breast cancer risk in postmenopausal women with higher GI and GL pattern scores, especially in women with a BMI 25. The reasons underlying these observations are unclear, although a recent report by Lin et al (34) examining hemoglobin A1c (HbA1c) concentrations in the Women's Health Study as a marker of glycemic control and insulin levels within the last 2 mo observed an inverse association between HbA1c levels and postmenopausal breast cancer, but no interaction with BMI.
In conclusion, we found that RRR may be useful in identifying dietary patterns associated with GI and GL, and that these patterns worked equally as well as did GI and GL in estimating the association of the dietary glycemic effect with breast cancer. Future studies are warranted to incorporate biomarkers of glycemic control, such as insulin, C-peptide, and Hb A1c, to further improve the identification of related foods and food groups. More targeted identification of specific dietary patterns could help in our development of guidelines to reduce chronic disease risk related to GI and GL.
ACKNOWLEDGMENTS
SEM and WEM designed the analysis plan. JRM provided consultation on the analysis plan. WEM conducted the statistical analyses. SEM, WEM, and C-CH interpreted the data. SEM and WEM drafted the manuscript. JLF, SBE, MT, and PM designed and conducted the original WEB Study and contributed to the analysis plan of the current study. All authors contributed to the revision of the draft to its final form. None of the authors had any conflicts of interest to declare.
REFERENCES