Annals of Oncology Advance Access published online on October 7, 2008
Annals of Oncology, doi:10.1093/annonc/mdn594
| ||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
Clustering dietary habits and the risk of breast and ovarian cancers
1 Department of Medicine and Surgery, "G. A. Maccacaro" Institute of Medical Statistics and Biometry of the University of Milan
2 Department of Epidemiology, The Mario Negri Institute for Pharmacological Research
3 The National Institute of Tumors IRCSS Foundation, Milan, Italy
4 The International Agency for Research on Cancer (IARC), Lyon, France
5 Department of Epidemiology and Biostatistics, Aviano Cancer Center, Aviano
6 Department of Medicine, Surgery and Dentistry, University of Milan, Milan, Italy
* Correspondence to: Dr V. Edefonti, Istituto di Statistica Medica e Biometria ‘Giulio A. Maccacaro’, Facoltà di Medicina e Chirurgia, Università degli Studi di Milano, via Venezian 1, 20133 Milan, Italy. Tel: +390250320873; Fax: +390250320866; E-mail: valeria.edefonti{at}unimi.it
 |
Abstract |
|---|
 Top Abstract introductionpatients and methodsresultsdiscussionfundingReferences |
|---|
Background: Limited information is available on the relationship between dietary patterns and breast and ovarian cancers.
Patients and methods: Cases were 2569 breast cancers and 1031 ovarian cancers hospitalized in four Italian areas from 1991 to 1999. Controls were 3413 women in hospital for acute non-neoplastic diseases. Dietary habits were investigated through a validated food-frequency questionnaire. Dietary patterns were obtained from a K-means clustering on factor scores from factor analysis. Odds ratios (ORs) for both cancers were estimated using unconditional multiple logistic regression models on clusters of patients. Floating absolute risk method was used for reporting 95% floating confidence intervals (FCIs).
Results: We identified five groups of subjects. The G3 cluster, including subjects with the lowest intakes of any food group, was used as reference. The G5 cluster, including subjects mainly consuming bread and pasta, was unfavorable for both cancers (OR = 1.23, 95% FCI = 1.08–1.38 for breast cancer, OR = 1.21, 95% FCI = 1.03–1.42 for ovarian cancer). The G1 group, including subjects mainly consuming fruits and vegetables, was protective against ovarian cancer (OR = 0.81, 95% FCI = 0.67–0.98).
Conclusions: A diet mainly based on bread and pasta is unfavorable for breast and ovarian cancers; a diet rich in fruits and vegetables may be associated with a reduced risk of ovarian cancer.
breast cancer, clustering, cluster analysis, dietary patterns, factor analysis, ovarian cancer
|
introduction |
|---|
|
TopAbstractintroductionpatients and methodsresultsdiscussionfundingReferences |
|---|
Because of the complexity of diet and the potential for interaction between food components, approaches that focus on individual nutrients may miss information on the role of diet in cancer etiology. Due to their ability to capture the variations in overall food intake in a given population, dietary patterns have been used to describe associations between diet and disease [1, 2]. Moreover, by characterizing a healthy diet in a population, they allow for dissemination of dietary recommendations in a more practical way. Different multivariate statistical methods are available to define a posteriori dietary patterns. The aim of principal component and factor analyses is to reduce the dimensionality of the data, by transforming an original larger set of correlated foods or nutrients into a smaller and easily interpretable set of uncorrelated variables. In contrast, the aim of cluster analysis is to identify potential mutually exclusive subgroups of patients with specific dietary behaviors, on the basis of a similarity in food or nutrient intake. The two approaches can be combined in an overall statistical strategy for data reduction and clustering, though only a few studies have tried to do it [1, 3, 4]. This allows the joint identification of both a partition of different subjects and of some key nutritional features of the identified clusters. It also avoids the arbitrary practice of referring to an a priori knowledge, coming from the Food Guide Pyramid [5], the American Dietetic Association's Exchange List for Meal Planning [6] or personal judgments on similarities/differences in nutrient content, to define the restricted set of food groups used as input for the clustering procedure [1].
We have applied the described strategy to two multicentric case–control studies of breast and ovarian cancers conducted in Italy. In the current work, we carry out a cluster analysis starting from factor scores obtained from a principal component factor analysis (PCFA) on a selected set of nutrients.
Previous analyses on single nutrients for those data showed a direct relationship with starch intake [7, 8] and measures of glycemic index and load [9, 10] and an inverse relationship with monounsaturated and polyunsaturated fatty acids, calcium and selected micronutrients [11, 12]. Moreover, the PCFA carried out on those data [13] identified four major dietary patterns named ‘Animal products’, ‘Vitamins and fiber’, ‘Unsaturated fats’ and ‘Starch-rich’. The Starch-rich pattern emerged as an unfavorable indicator of risk of both breast and ovarian cancers, whereas the ‘Vitamins and fiber’ one (rich in vegetables and fruit) was associated with a reduced risk of ovarian cancer [13].
|
patients and methods |
|---|
|
TopAbstractintroductionpatients and methodsresultsdiscussionfundingReferences |
|---|
design and participants
The data for the present analyses were derived from two case–control studies conducted from 1991 to 1999 in Italy. Details on the study population and design have been already presented [7, 12, 13]. Briefly, the case–control study on breast cancer was conducted from 1991 to 1994 in the urban areas of Milan and Genoa in north-west Italy; the provinces of Pordenone, Gorizia and Forlì in north-east Italy; the province of Latina in central Italy and the urban area of Naples in southern Italy [11]. The case–control study on ovarian cancer was conducted from 1992 to 1999 in the urban areas of Milan in north-west Italy, the provinces of Pordenone and Padua in north-east Italy, the province of Latina in central Italy and the urban area of Naples in southern Italy [8].
Cases were women with incident, histologically confirmed breast or ovarian cancer diagnosed no longer than 1 year before the interview, and with no previous diagnosis of cancer in any site. They were identified in major teaching and general hospitals of the study areas. A total of 2569 women with breast cancer, aged 23–74 years (median age 55), and 1031 women with ovarian cancer, aged 18–79 years (median age 56), were included. Controls were patients with no history of cancer admitted to hospitals with the same catchment areas as those where cases were identified. A total of 3413 nonoverlapping controls, aged 17–79 years (median age 57), were included. On average, 
4% of cases and 4% of controls invited to take part in the interview during their hospital stay refused to participate.
food-frequency questionnaire
Centrally trained interviewers administered a structured questionnaire to cases and controls during their hospital stay. The questionnaire included information on sociodemographic characteristics, anthropometric variables, lifestyle factors, a problem-oriented personal medical history, family history of cancer and menstrual and reproductive history.
Dietary habits were investigated through a food-frequency questionnaire (FFQ). Subjects were asked to indicate quantity and average weekly frequency of consumption of 78 foods, food groups and beverages, consumed 1 year before diagnosis for cases, or hospital admission for the controls. The FFQ was satisfactorily reproducible and valid [14]. The case–control studies were approved by the local ethics committees. Italian food composition sources [15] were used to calculate intakes of total energy and various nutrients.
statistical analysis
Exploratory PCFA was carried out on the correlation matrix of a selected set of 30 macro- and micronutrients. We retained four factors, which explained 75.7% of the total variance in the original data set [13]. Factor scores were calculated from factor analysis for each subject and for each of the 4 retained factors. They indicate the degree to which a subject's diet conforms to each of the identified dietary patterns. Those four-dimensional points were then used as input for the clustering procedure in the current work.
We carried out our cluster analysis with the K-means clustering method [16]. This method assumes a certain number of clusters, K, fixed a priori and produces a separation of the objects into nonoverlapping groups coming from Euclidean distances minimized at each step of an iterative procedure.
To find the most reasonable number of clusters, we ran a series of cluster analyses with predefined cluster numbers from 2 to 17. We compared a set of statistics built on 12 different indices between all runs. The indices were mainly based on the sum of squares within or between the clusters or on the scatter matrix of the data points and the sum of the scatter matrices in each cluster [17]. Cluster numbers equal to either 3 or 5 resulted to be the most suitable in all the comparisons. We chose the 5 solutions after having scrutinized the food-intake patterns of those solutions to see which set provided the clearest separation that was nutritionally meaningful. We refer to the 5 clusters as G1, G2, G3, G4 and G5 from now on. Each cluster corresponds to a specific dietary pattern.
Once the clustering partition was defined, we characterized our groups by examining their average consumption of selected standardized nutrients. We also referred to suitable food groups and sociodemographic and lifestyle factors to characterize the identified groups. Separate unconditional multiple logistic regression models for breast and ovarian cancers were used to assess the association between the main dietary habits of each group and those cancers. We took the G3 group as our reference category, since it had the highest number of subjects and the highest percentage of controls. The confounding variables included in the models were age, education, geographic area, parity, menopausal status, family history of hormone-related cancers, body mass index (BMI) and total energy intake.
The odds ratios (ORs) were estimated for each of the four groups. Corresponding 95% confidence intervals (CIs) were estimated referring to floating absolute risks [18]. This method ascribes a floating standard error (SE) to each group category that is independent of the choice of the reference group. A CI for the OR between any two groups can then be calculated from the floating SEs and is indicated as ‘floating confidence interval’ (FCI). Floating SE estimates have been derived from a covariance structure model applied to the covariance matrix of the log relative risk estimates [19]. We checked for the adequacy of this model by looking at how accurate the CIs were for our risk contrasts of interest, which were only the treatment contrasts. Acceptable limits for the accuracy of the relative SEs may be judged by considering a nominal 95% CI, the length of which is proportional to the SE. If the length is accurate to within 5%, then the coverage probabilities are close to 95%.
We checked for robustness and solution stability both by searching for potential outliers and by applying different clustering methods. Because clustering is sensible to outliers, we checked for them before starting the analysis. We ran the analyses with a predefined number of 300 clusters and tried to remove clusters with less than or equal to either 5 or 10 subjects each. Since our results were so similar to the ones obtained starting from the entire set of data, we decided to work on the full data set. We carried out the K-means clustering using different algorithms, proposed by Hartigan and Wong, MacQueen, Lloyd and Forgy [20]. We also tried several combinations of random starts. Moreover, we applied other nonhierarchical clustering methods, such as PAM (Partitioning Around Medoids) with the Euclidean distance specification and CLARA (Clustering Large Applications) with both Euclidean and Manhattan distances [20]. We obtained comparable results.
Calculations were carried out using the open-source statistical computing environment R [21], its library cclust [22] and cluster [23] and its functions kmeans() and glm(), pam() and clara(), which were entered into a specialized code reflecting the procedure previously described. The code is available upon request.
|
results |
|---|
|
TopAbstractintroductionpatients and methodsresultsdiscussionfundingReferences |
|---|
Table 1 gives the distribution of breast and ovarian cancer cases and controls according to age, years of education, parity and other selected variables. Cases of both cancers reported higher levels of education and energy intake and had more frequently a history of hormone-related cancers in their family. Breast cancer cases had fewer full-term pregnancies and were more likely to be pre- or perimenopausal. They also reported higher levels of alcohol intake and more frequent use of oral contraceptives.
|
Table 2 gives a description of the identified clusters in terms of distribution of cases and controls and cluster center. The G3 cluster showed the highest number of subjects and the highest percentage of controls. The second group as to the number of subjects is G5, which showed a lower percentage of controls compared with cases. All the other groups had a similar number of subjects and composition, in terms of both cases and controls. Each cluster showed an extreme behavior in one of the center coordinates, except for G3. G1 center was extreme on the ‘Vitamins and fiber’ factor, G2 center was extreme on the ‘Animal products’ factor, whereas G4 center and G5 center were extremes on the ‘Unsaturated fats’ and on the ‘Starch-rich’ factors, respectively.
|
Figure 1 represents a kernel density estimation of the factor scores univariate distributions, plotted along with cluster centers and quartiles of factor scores (calculated among the controls). The bandwidths were
0.13. In each plot, there was always a center coordinate either smaller than the first quartile or bigger than the third one and that was our extreme coordinate: G2 coordinate is extreme for ‘Animal products’, G1 coordinate for ‘Vitamins and fiber’, G4 coordinate for ‘Unsaturated fats’ and G5 coordinate for ‘Starch-rich’ factor. Moreover, all the factor scores had a slightly asymmetric distribution.
|
Figure 2 shows the scatter plots of the factor scores, plotted with a different color according to the corresponding group. Each factor was able to separate one group from the others. The ‘Animal products’ factor was able to distinguish the G2 group from the others, the ‘Vitamins and fiber’ factor the G1 group, the ‘Unsaturated fats’ factor distinguished the G4 group and the ‘Starch-rich factor’ the G5 group.
|
Table 3 shows the mean daily intake of selected standardized nutrients in each cluster. Cluster G1 had the highest mean intakes of total fiber, vitamin C, soluble carbohydrates, potassium, total folate, β-carotene equivalents and vitamin B6. Cluster G2 had the highest mean intakes of animal protein, animal fat, calcium, riboflavin, phosphorus, cholesterol, saturated fatty acids and zinc. No dominant nutrients were identified for cluster G3 which had the lowest mean intakes for each nutrient. Cluster G4 had the highest mean intakes of vegetable fat, vitamin E, monounsaturated and polyunsaturated fatty acids. Cluster G5 had the highest mean intakes of starch, vegetable protein and sodium.
|
Table 4 shows the median weekly intake of selected food groups by cluster. We chose the group medians instead of the corresponding means because the distributions of the food groups within each cluster were generally asymmetric and even multimodal in some cases. The G1 cluster was characterized by the highest median intakes of citrus fruits and other fruits, pulses, raw and cooked vegetables. The G2 cluster was characterized by the highest median intakes of red meat, pork and processed meats, cheese, eggs, cakes and desserts, milk, sugar and candies and butter. The G3 cluster was characterized by the lowest median intakes of any food group, except milk. The G4 cluster was characterized by the highest median intakes of olive oil, mixed or unspecified seed oils, red meat, raw and cooked vegetables and potatoes. The G5 cluster was characterized by the highest median intakes of bread and pasta. Both citrus fruits and pasta median intakes were high across all the groups, being more than 6 and 4 times per week, respectively.
|
Table 5 considered the association of the identified dietary patterns with selected risk factors for breast and ovarian cancers. Patients in the G1 group were more likely to have
12 years of education, to come from north-eastern Italy, to spend 5 h/week of physical activity in their free time. They were also less likely to drink alcohol usually. Patients in the G2 group were more likely to come from northern Italy and to be in the highest quintile of energy intake. Patients in the reference cluster, G3, were more likely to be >65 years, to come from north-western Italy, to be postmenopausal, to have a low level of both occupational and leisure-time physical activity and to be in the lowest quintiles of energy intake. They were also less likely to have used oral contraceptives and to drink alcohol usually. Patients in the G4 group were more likely to be younger, to come from central Italy, to have three or more children, to have a BMI >30, to have a high level of occupational physical activity, to drink >25.9 g/day of alcohol (1 drink 
13 g/day) and to be in the highest quintile of energy intake. They were less likely to have 12 years of education. Patients in the G5 group were more likely to be younger, to come from the south of Italy and to have a BMI <25. They were also less likely to have 12 years of education.
|
Table 6 gives the ORs and corresponding 95% FCIs for breast and ovarian cancers for the 5 dietary patterns derived from cluster analysis. Belonging to the G5 group was unfavorable compared with belonging to G3 for both breast and ovarian cancers: the OR was 1.23 (95% FCI = 1.08–1.38) for breast cancer and 1.21 (95% FCI = 1.03–1.42) for ovarian cancer. Belonging to the G1 group was protective against ovarian cancer: the OR was 0.81 (95% FCI = 0.67–0.98). The minimum and maximum values for the relative SEs on all the treatment contrasts were 0.97–1.05 for breast cancer and 0.97–1.04 for ovarian cancer. The lengths were accurate to within 8.2% and 7.2%, respectively, for breast and ovarian cancers.
|
|
discussion |
|---|
|
TopAbstractintroductionpatients and methodsresultsdiscussionfundingReferences |
|---|
The present analysis identified 5 clusters in a sample of 7013 patients from two Italian multicentric case–control studies on breast and ovarian cancers. We obtained coherent evidence from the interpretation of center coordinates, mean intakes of standardized nutrients and median intakes of food groups. The G1 cluster included subjects who were more likely to consume fruits and vegetables, which gave them higher intakes of various antioxidants, soluble carbohydrates and fiber. It had its extreme coordinate on the ‘Vitamins and fiber’ factor dimension. The G2 cluster included subjects who were more likely to consume meat, eggs, dairy products, sugar and desserts, which gave them higher intakes of saturated fats, calcium, riboflavin, protein and fat from animal source. It had its extreme coordinate on the ‘Animal products’ factor dimension. The G3 cluster included subjects who were more likely to have the lowest intakes of food groups and nutrients. The G4 cluster included subjects who were more likely to consume vegetables seasoned with either olive oil or seed oils, which gave them higher intakes of monounsaturated and polyunsaturated fatty acids and vitamin E. In this Italian population, unsaturated fats are indeed largely derived from olive oil, whereas vitamin E is derived from seed oils [24]. The G4 cluster had its extreme coordinate on the ‘Unsaturated fats’ factor dimension. The G5 cluster included subjects who were more likely to consume bread and pasta dishes, which gave them higher intakes of starch, sodium and vegetable protein. It had its extreme coordinate on the ‘Starch-rich’ factor dimension. The association of the identified clusters with selected risk factors is also in agreement with general existing evidence on sociodemographic and lifestyle variable distributions by clusters.
Most studies provide names for the observed dietary patterns. These names, however, could easily be misleading, as they rarely correspond exactly to similarly named patterns in other studies. For this reason, we avoided to refer to the clusters with any name.
To our knowledge, no study deriving dietary patterns from cluster analysis has been published on either breast or ovarian cancers. Dietary patterns from cluster analysis have been identified for adenocarcinoma of the esophagus and distal stomach [2], for lung cancer [25] and for the adenoma–carcinoma sequence of colorectal cancer [4], whereas another study [26] focused on patients with different types of advanced cancers. Only one of them [4] showed a principal component analysis carried out before clustering.
Moreover, in the wide literature focused on dietary patterns derived from clustering, we found other two papers [1, 3], applying an initial principal component analysis before clustering. The first paper [1] proposed an empirical approach based on the intuitive link between number and type of retained principal components and number of clusters. It chose which components to retain changing the set of retained components until it obtained the desired number of three clusters identified by preliminary exploratory analyses. In contrast, we carried out factor and cluster analysis independent of one another. Methods that carry out data reduction and clustering simultaneously should outperform both this and our solution and deserve extra attention [27]. Principal component and factor analyses result indeed in linear combinations and do not allow evaluation of the original variables anymore. In addition, it has been shown that the leading components do not necessarily contain the most part of the information about cluster structure [28]. The second paper [3] identified five factors that accounted for only 23.7% of the variance in the original data. The data reduction carried out by factor analysis is potentially misleading, thus biasing the subsequent clustering algorithm. In the present study, the factors used as input for the clustering algorithm explained 76% of the total variance in the original data.
The main advantage of cluster analysis is its ability to pinpoint and to focus dietary concerns within a group. Insight into dietary behaviors of distinct clusters within a population can help target health promotion messages and design intervention strategies to address specific behavioral concerns. Cluster analysis has some well-known limitations. The selection of clusters is largely subjective. However, we carried out the analysis with different nonhierarchical methods and with varying numbers of clusters, finding similar results. Moreover, the nutritional implications of the dietary patterns were generally understandable, making them useful for dietary guidance. Although it always terminates, the K-means algorithm does not necessarily find the optimal configuration corresponding to the global objective function minimum. The algorithm is also significantly sensitive to the initial randomly selected cluster centers. We ran it multiple times to reduce this effect and obtained similar results.
The data for the present analysis have the usual strengths and weaknesses of hospital-based case–control investigations. In general, case–control studies tended to report stronger associations between dietary factors and breast and ovarian cancers than cohort ones [29] and randomized clinical trials [30]. In our studies, however, selection bias should be limited, on account of the high participation rate and of the comparable catchment areas of cases and controls. The comparability of the recall between cases and controls was improved by interviewing all the subjects in a hospital setting [31].
Moreover, we specifically examined the evidence on breast/ovarian cancers and dietary patterns coming from cohort studies. Referring to breast cancer, five studies [32–36] out of 10 found significant associations in the population under study. In all those studies, only one dietary pattern was identified to be significantly associated with breast cancer. The same conclusion also held in the ovarian cancer literature, where we identified only one cohort study assessing the association between dietary patterns and ovarian cancer [37]. Although negative findings seem to dominate the accumulation of evidence on dietary patterns and breast and ovarian cancers, both unsolved methodological issues and deep differences in the way studies were conceived and carried out may alter the corresponding epidemiological results and prevent a fair evaluation.
No other study except the one by Edefonti et al. [13] has examined whether a common dietary pattern may be related to risk of breast and ovarian cancers occurrence. The present investigation used data from the same case–control studies to explore the role that a different definition of dietary patterns might have in the etiology of both those cancers. Cluster analysis still highlighted the detrimental effect of a pattern mainly based on bread and pasta for both breast and ovarian cancers, and the beneficial effect of a pattern rich in vegetables and fruit for ovarian cancer. These conclusions were in agreement with previous evidence from several analyses on single foods and nutrients on the same data set [7, 12]. Factor and cluster analyses have, however, emerged as methods able to estimate cancer risk more comprehensibly than others based on single foods or nutrients. Intakes of different foods or nutrients may indeed interact to increase or decrease cancer risk.
The direct association between the G5 cluster and both breast and ovarian cancers is of specific interest given that the Italian population has one of the highest starch intakes among wealthy countries [38]. The increased risk could be related to reduced intakes of beneficial substances, including micronutrients and food components, which were poorly consumed by subjects in the G5 cluster. Moreover, high intakes of refined starch may be responsible for glycemic overload, which has been associated with ovarian cancer [39, 40] and may increase the risk of breast cancer among Italian women [41, 42].
|
funding |
|---|
|
TopAbstractintroductionpatients and methodsresultsdiscussionfundingReferences |
|---|
Italian Association for Cancer Research; Italian League against Cancer; Italian Ministry of Education (PRIN 2004, PRIN 2005).
Received for publication February 12, 2008. Revision received July 21, 2008. Accepted for publication July 25, 2008.
|
References |
|---|
|
TopAbstractintroductionpatients and methodsresultsdiscussionfundingReferences |
|---|
1. Villegas R, Salim A, Collins MM, et al. Dietary patterns in middle-aged Irish men and women defined by cluster analysis. Public Health Nutr (2004) 7:1017–1024.[Web of Science][Medline]
2. Chen H, Ward MH, Graubard BI, et al. Dietary patterns and adenocarcinoma of the esophagus and distal stomach. Am J Clin Nutr (2002) 75:137–144.
3. Engeset D, Alsaker E, Ciampi A, et al. Dietary patterns and lifestyle factors in the Norwegian EPIC cohort: the Norwegian Women and Cancer (NOWAC) study. Eur J Clin Nutr (2005) 59:675–684.[CrossRef][Web of Science][Medline]
4. Rouillier P, Senesse P, Cottet V, et al. Dietary patterns and the adenoma carcinoma sequence of colorectal cancer. Eur J Nutr (2005) 44:311–318.[CrossRef][Web of Science][Medline]
5. US Department of Agriculture, Center for Nutrition Policy and Promotion. The Food Guide Pyramid. In: Home and Garden Bulletin No. 252 (1996) Washington, DC: US Department of Agriculture.
6. American Diabetes Association and American Dietetic Association. Exchange Lists for Meal Planning (1989) Chicago, IL: American Dietetic Association.
7. Franceschi S, Favero A, Decarli A, et al. Intake of macronutrients and risk of breast cancer. Lancet (1996) 347:1351–1356.[CrossRef][Web of Science][Medline]
8. Bidoli E, La Vecchia C, Montella M, et al. Nutrient intake and ovarian cancer: an Italian case-control study. Cancer Causes Control (2002) 13:255–261.[CrossRef][Web of Science][Medline]
9. Augustin LS, Dal Maso L, La Vecchia C, et al. Dietary glycemic index and glycemic load, and breast cancer risk: a case-control study. Ann Oncol (2001) 12:1533–1538.
10. Augustin LS, Polesel J, Bosetti C, et al. Dietary glycemic index, glycemic load and ovarian cancer risk: a case-control study in Italy. Ann Oncol (2003) 14:78–84.
11. Negri E, La Vecchia C, Franceschi S, et al. Intake of selected micronutrients and the risk of breast cancer. Int J Cancer (1996) 65:140–144.[CrossRef][Web of Science][Medline]
12. Bidoli E, La Vecchia C, Talamini R, et al. Micronutrients and ovarian cancer: a case-control study in Italy. Ann Oncol (2001) 12:1589–1593.
13. Edefonti V, Decarli A, La Vecchia C, et al. Nutrient dietary patterns and the risk of breast and ovarian cancers. Int J Cancer (2008) 122:609–613.[CrossRef][Web of Science][Medline]
14. Decarli A, Franceschi S, Ferraroni M, et al. Validation of a food-frequency questionnaire to assess dietary intakes in cancer studies in Italy. Results for specific nutrients. Ann Epidemiol (1996) 6:110–118.[CrossRef][Web of Science][Medline]
15. Gnagnarella P, Parpinel M, Salvini S, et al. The update of the Italian Food Composition Database. J Food Compost Anal (2004) 17:509–522.[CrossRef]
16. Johnson RA, Wichern DW. Applied Multivariate Statistical Analysis (2002) Upper Saddle River, NJ: Prentice-Hall.
17. Dimitriadou E, Dolnicar S, Weingessel A. An examination of indexes for determining the number of clusters in binary data sets. Psychometrika (2002) 67:137–160.[CrossRef][Web of Science]
18. Easton DF, Peto J, Babiker A. Floating absolute risk: an alternative to relative risk in survival and case-control analysis avoiding an arbitrary reference group. Stat Med (1991) 10:1025–1035.[Web of Science][Medline]
19. Plummer M. Improved estimates of floating absolute risk. Stat Med (2004) 23:93–104.[CrossRef][Web of Science][Medline]
20. Kaufman L, Rousseeuw PJ. Finding Groups in Data. In: An Introduction to Cluster Analysis (2005) New York: John Wiley & Sons.
21. R Development Core Team. R: A Language and Environment for Statistical Computing (2008) Vienna, Austria: R Foundation for Statistical Computing. http://www.R-project.org.
22. Dimitriadou E. cclust: Convex Clustering Methods and Clustering Indexes. R package version 0.6-14 2008 http://www.R-project.org.
23. Maechler M, Rousseeuw P, Struyf A, et al. Cluster analysis basics and extensions (2005) http://www.R-project.org.
24. Favero A, Salvini S, Russo A, et al. Sources of macro- and micronutrients in Italian women: results from a food frequency questionnaire for cancer studies. Eur J Cancer Prev (1997) 6:277–277.[Web of Science][Medline]
25. Tsai YY, McGlynn KA, Hu Y, et al. Genetic susceptibility and dietary patterns in lung cancer. Lung Cancer (2003) 41:269–271.[CrossRef][Web of Science][Medline]
26. Hutton JL, Martin L, Field CJ, et al. Dietary patterns in patients with advanced cancer: implications for anorexia-cachexia therapy. Am J Clin Nutr (2006) 84:1163–1170.
27. Tadesse MG, Sha N, Vannucci M. Bayesian variable selection in clustering high-dimensional data. JASA (2005) 100:602–617.
28. Chang WC. On using principal components before separating a mixture of two multivariate normal distributions. Appl Stat (1983) 32:267–275.[CrossRef]
29. World Cancer Research Fund/American Institute for Cancer Research. Food, nutrition, physical activity and the prevention of cancer: a global perspective (2007) Washington, DC: AICR.
30. Van Horn L, Manson JE. The women's health initiative: implications for clinicians. Cleve Clin J Med (2008) 75:385–390.
31. D'Avanzo B, La Vecchia C, Katsouyanni K, et al. An assessment, and reproducibility of food frequency data provided by hospital controls. Eur J Cancer Prev (1997) 6:288–293.[CrossRef][Web of Science][Medline]
32. Terry P, Suzuki R, Hu FB, et al. A prospective study of major dietary patterns and the risk of breast cancer. Cancer Epidemiol Biomarkers Prev (2001) 10:1281–1285.
33. Sieri S, Krogh V, Pala V, et al. Dietary patterns and risk of breast cancer in the ORDET cohort. Cancer Epidemiol Biomarkers Prev (2004) 13:567–572.
34. Sant M, Allemani C, Sieri S, et al. Salad vegetables dietary pattern protects against HER-2-positive breast cancer: a prospective Italian study. Int J Cancer (2007) 121:911–914.[CrossRef][Web of Science][Medline]
35. Mannisto S, Dixon LB, Balder HF, et al. Dietary patterns and breast cancer risk: results from three cohort studies in the DIETSCAN project. Cancer Causes Control (2005) 16:725–733.[CrossRef][Web of Science][Medline]
36. Schulz M, Hoffmann K, Weikert C, et al. Identification of a dietary pattern characterized by high-fat food choices associated with increased risk of breast cancer: the European Prospective Investigation into Cancer and Nutrition (EPIC)-Potsdam Study. Br J Nutr (2008) doi:10.1017/S0007114508966150, published online by Cambridge University Press 17 April 2008.
37. Chang ET, Lee VS, Canchola AJ, et al. Dietary patterns and risk of ovarian cancer in the California Teachers Study cohort. Nutr Cancer (2008) 60:285–291.[CrossRef][Web of Science][Medline]
38. Serra-Majem L, La Vecchia C, Ribas-Barba L, et al. Changes in diet and mortality from selected cancers in southern Mediterranean countries, 1960—1989. Eur J Clin Nutr (1993) 47:S25–S34.
39. Silvera SA, Jain M, Howe GR, et al. Glycaemic index, glycaemic load and ovarian cancer risk: a prospective cohort study. Public Health Nutr (2007) 10:1076–1081.[Web of Science][Medline]
40. Augustin LS, Franceschi S, Jenkins DJ, et al. Glycemic index in chronic disease: a review. Eur J Clin Nutr (2002) 56:1049–1051.[CrossRef][Web of Science][Medline]
41. Sieri S, Pala V, Brighenti F, et al. Dietary glycemic index, glycemic load, and the risk of breast cancer in an Italian prospective cohort study. Am J Clin Nutr (2007) 86:1160–1166.
42. Levi F, Pasche C, Lucchini F, et al. Glycaemic index, breast and colorectal cancer. Ann Oncol (2002) 13:1688–1689.
CiteULike 
Connotea 
Del.icio.us What's this?
This article has been cited by other articles:
|
|
 |
|
  S. A. Quandt, H. Chen, R. A. Bell, M. R. Savoca, A. M. Anderson, X. Leng, T. Kohrman, G. H. Gilbert, and T. A. Arcury Food Avoidance and Food Modification Practices of Older Rural Adults: Association With Oral Health Status and Implications for Service Provision Gerontologist, July 2, 2009; (2009) gnp096v1. [Abstract] [Full Text] [PDF]
|
|
| ||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||