Cancer Epidemiology 39 (2015) 66–72
Contents lists available at ScienceDirect
Cancer Epidemiology The International Journal of Cancer Epidemiology, Detection, and Prevention journal homepage: www.cancerepidemiology.net
Nutrient-based dietary patterns and endometrial cancer risk: an Italian case–control study Francesca Bravi a, Paola Bertuccio a, Federica Turati b, Diego Serraino c, Valeria Edefonti a, Luigino Dal Maso c, Adriano Decarli a,b, Maurizio Montella d, Antonella Zucchetto c, Carlo La Vecchia a, Cristina Bosetti e,*, Monica Ferraroni a a
Department of Clinical Sciences and Community Health, Universita` degli Studi di Milano, via G. Venezian, 1, 20122 Milan, Italy Unit of Medical Statistics, Biometry and Bioinformatics, Fondazione IRCCS Istituto Nazionale Tumori, via G. Venezian, 1, 20122 Milan, Italy c Unit of Epidemiology and Biostatistics, Centro di Riferimento Oncologico, IRCCS, via F. Gallini, 2, 33080 Aviano, PN, Italy d Department of Epidemiology, Istituto Nazionale Tumori IRCCS – Fondazione Pascale, via M. Semmola, 80131 Naples, Italy e Department of Epidemiology, IRCCS – Istituto di Ricerche Farmacologiche ‘‘Mario Negri’’, via Giuseppe La Masa 19, 20156 Milan, Italy b
A R T I C L E I N F O
A B S T R A C T
Article history: Received 17 September 2014 Received in revised form 2 December 2014 Accepted 7 December 2014 Available online 30 December 2014
Diet has been suggested to have a role on endometrial cancer risk, but few data are available on the role of dietary patterns on this neoplasm. A case–control study was carried out in Italy, including 454 women with endometrial cancer and 908 hospital controls admitted to the same hospitals for acute, nonneoplastic diseases. Dietary information was based on a reproducible and valid food frequency questionnaire. A posteriori dietary patterns were obtained using principal component factor analysis on 28 nutrients. Odds ratios (ORs) and the corresponding 95% confidence intervals (CIs) were obtained from multiple logistic regression models conditioned on age and study center, and adjusted for major known confounding factors. Positive associations were found for the ‘‘Western-type diet’’ (OR = 1.63, 95% CI: 1.12–2.38, for the highest versus the lowest quartile category) and the ‘‘Animal-derived nutrients and polyunsaturated fatty acids’’ patterns (OR = 1.76, 95% CI: 1.23–2.52). The corresponding risk estimates among women with a body mass index 30 were 2.08 (95% CI: 0.92–4.69) and 2.30 (95% CI: 1.03–5.16) for the two patterns, respectively. No association was found for the other three patterns (i.e., ‘‘Vitamins and fiber’’, OR = 0.96, 95% CI: 0.67–1.37, ‘‘Starch-rich’’, OR = 0.99, 95% CI: 0.69–1.42, and ‘‘Other fats’’, OR = 1.03, 95% CI: 0.70–1.52). This study indicates that dietary habits characterized by high intakes of animal products increase endometrial cancer risk, the association being appreciably stronger for obese women. ß 2014 Published by Elsevier Ltd.
Keywords: Diet Dietary patterns Factor analysis Endometrial cancer Case–control study
1. Introduction Endometrial cancer is the sixth most common female cancer in terms of incidence [1] and the third most prevalent cancer type in women worldwide [2]. Major recognized risk factors are related to prolonged exposure to high estrogen levels, and include early age at menarche, nulliparity, late age at menopause, unopposed estrogen hormone replacement therapy (HRT) use, as well as obesity and history of diabetes, whereas oral contraceptive (OC) use is inversely related to this neoplasm [3]. Diet has been suggested to have a role on endometrial cancer independently from obesity, although the evidence for specific
* Corresponding author. Tel.: +39 0239014526; fax: +39 0233200231. E-mail address: [email protected] (C. Bosetti). http://dx.doi.org/10.1016/j.canep.2014.12.003 1877-7821/ß 2014 Published by Elsevier Ltd.
dietary components is still limited and inconclusive [3,4]. A few prospective cohort studies found a positive association with glycemic load [5–8], and an inverse one with coffee consumption [9–13], whereas two other cohort studies did not find any significant role of diet on endometrial cancer risk [14,15]. In 2013, the World Cancer Research Fund judged as probable the evidence of a beneficial role of coffee consumption and a detrimental role of glycemic load on endometrial cancer; other foods and nutrients were considered in a few studies, but no definite evidence was found [4]. A few case–control studies from North America have considered the role of diet in terms of combinations of foods – identified through a posteriori derived dietary patterns or a priori scores [16– 19]. In a case–control study from the USA including 232 cases and 936 controls an inverse association was observed between endometrial cancer risk and the a posteriori ‘‘Healthy’’ pattern,
F. Bravi et al. / Cancer Epidemiology 39 (2015) 66–72
rich in vegetables and fruits, whereas no relationship was observed with the ‘‘High-Fat’’ pattern [19]. In a Canadian study including 506 cases and 981 controls, a reduced risk was observed for the a posteriori ‘‘Plant’’ pattern, characterized by vegetables, fruits, legumes and whole grain, while no association was found for the ‘‘Sweet’’ and ‘‘Meat’’ patterns [16]. Another study from the USA, including 488 cases and 936 controls, did not find a meaningful relation with any of the a posteriori patterns (i.e., ‘‘Plant-based’’, ‘‘Western’’, ‘‘Ethnic’’, and ‘‘Phytoestrogen-rich’’) [18]. In the same study, different a priori ‘‘Mediterranean Diet’’ scores were not associated to endometrial cancer, whereas an increase in risk was observed for an ‘‘High fat diet’’ (defined a priori as more than 33% of calories from fat per day), independently from fruit and vegetable consumption. In a case–control study from USA based on 424 cases and 398 controls, the adherence to the specific recommendations of the Dietary Guidelines for Americans and the Food Guide Pyramid was evaluated through the Healthy Eating Index-2005, and no significant association was observed [17]. Previous analyses on our multicentric Italian case–control study showed positive associations with red meat, cholesterol, and possibly saturated fatty acids, and a beneficial role of vegetables, some carotenoids, and cereals [20–22]. To further contribute to the issue, we derived a posteriori dietary patterns using an exploratory principal component factor analysis (PCFA) on data from this case– control study, and analyzed their association with endometrial cancer. 2. Materials and methods 2.1. Design and participants We carried out a case–control study on endometrial cancer between 1992 and 2006 in three Italian areas, including the greater Milan area, the provinces of Pordenone and Udine in northern Italy, and the urban area of Naples in southern Italy [21]. Cases were 454 women (median age: 60 years, range: 18–79 years) with incident, histologically confirmed endometrial cancer, admitted to major teaching and general hospitals of the study areas. Controls were 908 women (median age: 61 years, range: 19–80) admitted to the same hospitals for a wide spectrum of acute, non-neoplastic conditions. Women admitted for gynecological or hormonerelated diseases, with a previous history of hysterectomy, or any medical condition related to long-term dietary modifications were excluded. Cases and controls were matched by 5-years groups of age and study center, with a case to control ratio of 1:2. Thirty-six percent of the controls were admitted for traumas, 32% for other orthopedic disorders, 9% for acute surgical conditions, and 23% for miscellaneous other diseases, including nose, eye, ear or skin disorders. Less than 5% of both cases and controls approached refused to participate to the study. The study was approved by the local ethical committees, and cases and controls who agreed to participate in the study signed an informed consent. Study participants were interviewed during their hospital stay by trained interviewers using a structured questionnaire, including information on sociodemographic characteristics, anthropometric measures, tobacco smoking, alcohol drinking and other lifestyle habits, personal medical history, menstrual and reproductive characteristics, use of OC and HRT, and family history of cancer in first-degree relatives. The subjects’ usual diet during the two years before cancer diagnosis (or hospital admission for controls) was assessed through a food frequency questionnaire (FFQ) including 78 foods and beverages, grouped according to 6 categories: milk and hot beverages, bread and cereal dishes (first courses), meat and other main dishes (second courses), vegetables (side dishes), fruits, and sweet, dessert and soft drinks. An additional section concerned
67
alcoholic beverages. The participants were asked to indicate the average weekly frequency of consumption for each dietary item; intakes at least once a month but less than once a week were coded as 0.5 per week. A commonly used unit or serving size was prespecified for 40 items; for the remaining, the portion was defined as small, medium or large. For selected vegetables and fruits, seasonal consumption and the corresponding duration were asked. To estimate intakes of total energy and various nutrients we used an Italian food composition database, integrated with other sources when needed [23,24]. The FFQ had been previously validated using two 7-day dietary records [25] and tested for reproducibility, with the majority of nutrients showing a correlation coefficient 0.60 [26]. 2.2. Statistical analysis We carried out the analyses on an exhaustive set of 28 nutrients, including minerals, macro-, and micro-nutrients representative of the Italian diet. Exploratory PCFA was performed on the correlation matrix of the original nutrients of both cases and controls, in order to identify a few underlying unobservable factors, known as dietary patterns [27]. We standardized the original absolute nutrients before carrying out the factor analysis, in order to avoid problems related to the use of different measurement units and amounts of intake. We obtained the standardized nutrients by subtracting from each observed nutrient the mean of the nutrient and dividing it by its standard deviation, both calculated on all the subjects. We evaluated if the correlation matrix was factorable, through both visual inspection and statistical procedures, including the Bartlett’s test of sphericity, the Keiser–Meyer–Olkin measure, and individual measures of sampling adequacy [28]. We chose the number of factors to retain using the following criteria: interpretability, factor eigenvalue >1, and visual inspection of the scree plot. We applied an orthogonal (i.e., varimax) rotation to the factor loading matrix to achieve a simpler structure with easier interpretability. We used nutrients with an absolute rotated factor loading value 0.63 on a given factor to interpret and label that factors [29]. Since the contribution that a factor gives to a nutrient’s sample variance is equal to the square of its loading, we chose a priori this cut-off, which corresponds to a contribution of approximately 0.40. We calculated the factor scores – which represent the subject’s degree of adherence to the identified patterns – using the weighted least squares method. To evaluate the robustness of the identified dietary patterns, we carried out additional analyses including principal axis factor analysis, and PCFA with varimax rotation in two equally sized subgroups, obtained through random allocation of the subjects. Given the robustness of the patterns identified in these complementary analyses, all the subsequent analyses were based on the factor scores obtained from the overall PCFA with varimax rotation. To assess the reliability and refine the identified factors, we evaluated the internal consistency of the nutrients that loaded 0.40 on any factor using the standardized Cronbach’s coefficient alpha [28]. This is a measure of reliability which represents the proportion of total variance in a given factor that can be attributed to a common source. Finally, we calculated the coefficient alpha for each factor and the coefficient alpha when item deleted (i.e., calculated excluding each nutrient, one by one) for each factor and for each nutrient loading j0.40j. In order to facilitate the interpretation of the identified patterns, we calculated the Spearman rank correlation coefficients between the continuous factor scores derived from PCFA and the weekly consumption of 29 food groups and seasonings, obtained summing up the 78 original foods according to homogeneous categories. For each factor, we grouped participants into four categories according to quartiles of factor scores among the controls. We estimated the odds ratios (ORs), and the corresponding 95%
F. Bravi et al. / Cancer Epidemiology 39 (2015) 66–72
68
Table 1 shows the factor loading matrix for the five retained dietary patterns, which together accounted for 80% of the variance of the original nutrients. The first pattern, named ‘‘Western-type diet’’, had the greatest loadings on (i.e. the higher contribution) from calcium, riboflavin, phosphorus, animal protein, saturated fatty acids, cholesterol, and zinc. The second pattern, labeled ‘‘Vitamins and fiber’’, was characterized by high loadings on vitamin C, total fiber, potassium, total folate, beta-carotene equivalents, and soluble carbohydrates. The third pattern, named ‘‘Starch-rich’’, had the highest loadings on starch, vegetable protein, and sodium. The fourth pattern, named ‘‘Animal-derived nutrients and polyunsaturated fatty acids’’, was characterized by the greatest loadings on vitamin D, other polyunsaturated fatty acids, and niacin. The fifth pattern, labeled ‘‘Other fats’’, had the highest loadings on linoleic acid, linolenic acid, and vitamin E. All the nutrients had at least one loading greater than 0.30, thus confirming a relevant role of each nutrient in the original list. Standardized Cronbach’s coefficient alphas were high (at least 0.90) for all the factors and most of the standardized coefficient alphas when nutrient deleted were lower than the corresponding coefficient alpha for the factor, indicating a good internal consistency of the nutrients on the identified factors (data not shown). Baseline characteristics of participants according to quartiles of distribution of the identified dietary patterns are shown in Table 2. OC use was positively related to the ‘‘Animal-derived nutrients and polyunsaturated fatty acids’’ pattern. HRT use was positively related to the ‘‘Vitamins and fiber’’ and inversely related to the ‘‘Other fats’’ pattern. Occupational physical activity was inversely related to the ‘‘Vitamins and fiber’’ and the ‘‘Other fats’’ patterns, and positively related to the ‘‘Starch-rich’’ pattern. BMI
confidence intervals (CIs), for each quartile category compared to the first one using conditional multiple logistic regression models [30], conditioned on age and study center, and adjusted for period of interview, education, body mass index (BMI, based on selfreported anthropometric data), history of diabetes, age at menarche, menopausal status, parity, OC and HRT use. We fitted both separate models for each factor and a composite model that included all the factors simultaneously. We also examined the possible confounding effect of tobacco smoking and alcohol drinking; however, since the inclusion of these covariates in the models did not materially modify the risk estimates, we presented results based on the more parsimonious model described above. We computed tests for linear trend for all these models using the within category medians calculated among the controls. Further, we carried out stratified analysis of risk by age and BMI. We performed the analyses using SAS software, version 9.2 (SAS Institute, Inc., Cary, NC). 3. Results The correlation matrix of the original nutrients was suitable for factor analysis. Each nutrient had at least 10 correlation coefficients greater than 0.30 (data not shown), thus allowing to carry out the analyses on the 28 nutrients. Bartlett’s test of sphericity was highly significant (p < 0.0001), allowing to reject the null hypothesis that the correlation matrix is an identity matrix (Table A1). The individual measures of sampling adequacy were generally very high, with 24 nutrients having measures of at least 0.70. The overall measure of Keiser–Meyer–Olkin was 0.82, confirming that the sample size was adequate for performing factor analysis [28].
Table 1 Factor loading matrixa and explained variances for the five major dietary patterns identified by factor analysis. Italy, 1992–2006. Nutrients
Dietary patterns Western-type diet
Animal protein Vegetable protein Cholesterol Saturated fatty acids Monounsaturated fatty acids Linoleic acid Linolenic acid Other polyunsaturated fatty acids Soluble carbohydrates Starch Sodium Calcium Potassium Phosphorus Iron Zinc Thiamin (vitamin B1) Riboflavin (vitamin B2) Vitamin B6 Total folate Niacin Vitamin C Retinol Beta-carotene equivalents Lycopene Vitamin D Vitamin E Total fiber Proportion of explained variance (%) Cumulative explained variance (%)
Vitamins and fiber
Starch-rich
Animal-derived nutrients and PUFAs
Other fats
0.76 0.17 0.65 0.73 0.31 0.15 0.33 0.10 0.46 0.19 0.52 0.88 0.45 0.79 0.38 0.63 0.53 0.80 0.49 0.39 0.31 0.13 0.36 – – 0.13 0.15 0.13
0.13 0.37 – 0.16 0.30 – 0.12 – 0.65 0.10 – 0.27 0.74 0.34 0.52 0.29 0.52 0.43 0.59 0.72 0.39 0.87 0.11 0.72 0.33 – 0.46 0.82
0.18 0.86 0.28 0.27 0.29 0.14 0.15 0.14 0.16 0.92 0.74 0.10 0.26 0.33 0.38 0.41 0.42 0.12 0.31 0.30 0.34 – – – 0.37 – 0.21 0.38
0.45 0.14 0.46 0.15 0.27 0.15 0.16 0.83 – – 0.10 – 0.26 0.25 0.47 0.44 0.27 0.22 0.39 0.26 0.65 – 0.39 – 0.13 0.85 0.20 –
0.26 0.18 0.31 0.43 0.48 0.87 0.81 0.34 – 0.18 0.16 0.13 0.19 0.21 0.21 0.28 0.22 – 0.27 0.15 0.24 – – 0.23 0.30 0.10 0.78 –
21.74 21.74
19.88 41.62
13.63 55.25
12.61 67.86
12.18 80.04
PUFAs: polyunsaturated fatty acids. a Estimated from a principal component factor analysis performed on 28 nutrients. The magnitude of each loading indicates the importance of the corresponding nutrient to the factor. Loadings 0.63 were used to name the factor and are shown in bold; loadings
Contents lists available at ScienceDirect
Cancer Epidemiology The International Journal of Cancer Epidemiology, Detection, and Prevention journal homepage: www.cancerepidemiology.net
Nutrient-based dietary patterns and endometrial cancer risk: an Italian case–control study Francesca Bravi a, Paola Bertuccio a, Federica Turati b, Diego Serraino c, Valeria Edefonti a, Luigino Dal Maso c, Adriano Decarli a,b, Maurizio Montella d, Antonella Zucchetto c, Carlo La Vecchia a, Cristina Bosetti e,*, Monica Ferraroni a a
Department of Clinical Sciences and Community Health, Universita` degli Studi di Milano, via G. Venezian, 1, 20122 Milan, Italy Unit of Medical Statistics, Biometry and Bioinformatics, Fondazione IRCCS Istituto Nazionale Tumori, via G. Venezian, 1, 20122 Milan, Italy c Unit of Epidemiology and Biostatistics, Centro di Riferimento Oncologico, IRCCS, via F. Gallini, 2, 33080 Aviano, PN, Italy d Department of Epidemiology, Istituto Nazionale Tumori IRCCS – Fondazione Pascale, via M. Semmola, 80131 Naples, Italy e Department of Epidemiology, IRCCS – Istituto di Ricerche Farmacologiche ‘‘Mario Negri’’, via Giuseppe La Masa 19, 20156 Milan, Italy b
A R T I C L E I N F O
A B S T R A C T
Article history: Received 17 September 2014 Received in revised form 2 December 2014 Accepted 7 December 2014 Available online 30 December 2014
Diet has been suggested to have a role on endometrial cancer risk, but few data are available on the role of dietary patterns on this neoplasm. A case–control study was carried out in Italy, including 454 women with endometrial cancer and 908 hospital controls admitted to the same hospitals for acute, nonneoplastic diseases. Dietary information was based on a reproducible and valid food frequency questionnaire. A posteriori dietary patterns were obtained using principal component factor analysis on 28 nutrients. Odds ratios (ORs) and the corresponding 95% confidence intervals (CIs) were obtained from multiple logistic regression models conditioned on age and study center, and adjusted for major known confounding factors. Positive associations were found for the ‘‘Western-type diet’’ (OR = 1.63, 95% CI: 1.12–2.38, for the highest versus the lowest quartile category) and the ‘‘Animal-derived nutrients and polyunsaturated fatty acids’’ patterns (OR = 1.76, 95% CI: 1.23–2.52). The corresponding risk estimates among women with a body mass index 30 were 2.08 (95% CI: 0.92–4.69) and 2.30 (95% CI: 1.03–5.16) for the two patterns, respectively. No association was found for the other three patterns (i.e., ‘‘Vitamins and fiber’’, OR = 0.96, 95% CI: 0.67–1.37, ‘‘Starch-rich’’, OR = 0.99, 95% CI: 0.69–1.42, and ‘‘Other fats’’, OR = 1.03, 95% CI: 0.70–1.52). This study indicates that dietary habits characterized by high intakes of animal products increase endometrial cancer risk, the association being appreciably stronger for obese women. ß 2014 Published by Elsevier Ltd.
Keywords: Diet Dietary patterns Factor analysis Endometrial cancer Case–control study
1. Introduction Endometrial cancer is the sixth most common female cancer in terms of incidence [1] and the third most prevalent cancer type in women worldwide [2]. Major recognized risk factors are related to prolonged exposure to high estrogen levels, and include early age at menarche, nulliparity, late age at menopause, unopposed estrogen hormone replacement therapy (HRT) use, as well as obesity and history of diabetes, whereas oral contraceptive (OC) use is inversely related to this neoplasm [3]. Diet has been suggested to have a role on endometrial cancer independently from obesity, although the evidence for specific
* Corresponding author. Tel.: +39 0239014526; fax: +39 0233200231. E-mail address: [email protected] (C. Bosetti). http://dx.doi.org/10.1016/j.canep.2014.12.003 1877-7821/ß 2014 Published by Elsevier Ltd.
dietary components is still limited and inconclusive [3,4]. A few prospective cohort studies found a positive association with glycemic load [5–8], and an inverse one with coffee consumption [9–13], whereas two other cohort studies did not find any significant role of diet on endometrial cancer risk [14,15]. In 2013, the World Cancer Research Fund judged as probable the evidence of a beneficial role of coffee consumption and a detrimental role of glycemic load on endometrial cancer; other foods and nutrients were considered in a few studies, but no definite evidence was found [4]. A few case–control studies from North America have considered the role of diet in terms of combinations of foods – identified through a posteriori derived dietary patterns or a priori scores [16– 19]. In a case–control study from the USA including 232 cases and 936 controls an inverse association was observed between endometrial cancer risk and the a posteriori ‘‘Healthy’’ pattern,
F. Bravi et al. / Cancer Epidemiology 39 (2015) 66–72
rich in vegetables and fruits, whereas no relationship was observed with the ‘‘High-Fat’’ pattern [19]. In a Canadian study including 506 cases and 981 controls, a reduced risk was observed for the a posteriori ‘‘Plant’’ pattern, characterized by vegetables, fruits, legumes and whole grain, while no association was found for the ‘‘Sweet’’ and ‘‘Meat’’ patterns [16]. Another study from the USA, including 488 cases and 936 controls, did not find a meaningful relation with any of the a posteriori patterns (i.e., ‘‘Plant-based’’, ‘‘Western’’, ‘‘Ethnic’’, and ‘‘Phytoestrogen-rich’’) [18]. In the same study, different a priori ‘‘Mediterranean Diet’’ scores were not associated to endometrial cancer, whereas an increase in risk was observed for an ‘‘High fat diet’’ (defined a priori as more than 33% of calories from fat per day), independently from fruit and vegetable consumption. In a case–control study from USA based on 424 cases and 398 controls, the adherence to the specific recommendations of the Dietary Guidelines for Americans and the Food Guide Pyramid was evaluated through the Healthy Eating Index-2005, and no significant association was observed [17]. Previous analyses on our multicentric Italian case–control study showed positive associations with red meat, cholesterol, and possibly saturated fatty acids, and a beneficial role of vegetables, some carotenoids, and cereals [20–22]. To further contribute to the issue, we derived a posteriori dietary patterns using an exploratory principal component factor analysis (PCFA) on data from this case– control study, and analyzed their association with endometrial cancer. 2. Materials and methods 2.1. Design and participants We carried out a case–control study on endometrial cancer between 1992 and 2006 in three Italian areas, including the greater Milan area, the provinces of Pordenone and Udine in northern Italy, and the urban area of Naples in southern Italy [21]. Cases were 454 women (median age: 60 years, range: 18–79 years) with incident, histologically confirmed endometrial cancer, admitted to major teaching and general hospitals of the study areas. Controls were 908 women (median age: 61 years, range: 19–80) admitted to the same hospitals for a wide spectrum of acute, non-neoplastic conditions. Women admitted for gynecological or hormonerelated diseases, with a previous history of hysterectomy, or any medical condition related to long-term dietary modifications were excluded. Cases and controls were matched by 5-years groups of age and study center, with a case to control ratio of 1:2. Thirty-six percent of the controls were admitted for traumas, 32% for other orthopedic disorders, 9% for acute surgical conditions, and 23% for miscellaneous other diseases, including nose, eye, ear or skin disorders. Less than 5% of both cases and controls approached refused to participate to the study. The study was approved by the local ethical committees, and cases and controls who agreed to participate in the study signed an informed consent. Study participants were interviewed during their hospital stay by trained interviewers using a structured questionnaire, including information on sociodemographic characteristics, anthropometric measures, tobacco smoking, alcohol drinking and other lifestyle habits, personal medical history, menstrual and reproductive characteristics, use of OC and HRT, and family history of cancer in first-degree relatives. The subjects’ usual diet during the two years before cancer diagnosis (or hospital admission for controls) was assessed through a food frequency questionnaire (FFQ) including 78 foods and beverages, grouped according to 6 categories: milk and hot beverages, bread and cereal dishes (first courses), meat and other main dishes (second courses), vegetables (side dishes), fruits, and sweet, dessert and soft drinks. An additional section concerned
67
alcoholic beverages. The participants were asked to indicate the average weekly frequency of consumption for each dietary item; intakes at least once a month but less than once a week were coded as 0.5 per week. A commonly used unit or serving size was prespecified for 40 items; for the remaining, the portion was defined as small, medium or large. For selected vegetables and fruits, seasonal consumption and the corresponding duration were asked. To estimate intakes of total energy and various nutrients we used an Italian food composition database, integrated with other sources when needed [23,24]. The FFQ had been previously validated using two 7-day dietary records [25] and tested for reproducibility, with the majority of nutrients showing a correlation coefficient 0.60 [26]. 2.2. Statistical analysis We carried out the analyses on an exhaustive set of 28 nutrients, including minerals, macro-, and micro-nutrients representative of the Italian diet. Exploratory PCFA was performed on the correlation matrix of the original nutrients of both cases and controls, in order to identify a few underlying unobservable factors, known as dietary patterns [27]. We standardized the original absolute nutrients before carrying out the factor analysis, in order to avoid problems related to the use of different measurement units and amounts of intake. We obtained the standardized nutrients by subtracting from each observed nutrient the mean of the nutrient and dividing it by its standard deviation, both calculated on all the subjects. We evaluated if the correlation matrix was factorable, through both visual inspection and statistical procedures, including the Bartlett’s test of sphericity, the Keiser–Meyer–Olkin measure, and individual measures of sampling adequacy [28]. We chose the number of factors to retain using the following criteria: interpretability, factor eigenvalue >1, and visual inspection of the scree plot. We applied an orthogonal (i.e., varimax) rotation to the factor loading matrix to achieve a simpler structure with easier interpretability. We used nutrients with an absolute rotated factor loading value 0.63 on a given factor to interpret and label that factors [29]. Since the contribution that a factor gives to a nutrient’s sample variance is equal to the square of its loading, we chose a priori this cut-off, which corresponds to a contribution of approximately 0.40. We calculated the factor scores – which represent the subject’s degree of adherence to the identified patterns – using the weighted least squares method. To evaluate the robustness of the identified dietary patterns, we carried out additional analyses including principal axis factor analysis, and PCFA with varimax rotation in two equally sized subgroups, obtained through random allocation of the subjects. Given the robustness of the patterns identified in these complementary analyses, all the subsequent analyses were based on the factor scores obtained from the overall PCFA with varimax rotation. To assess the reliability and refine the identified factors, we evaluated the internal consistency of the nutrients that loaded 0.40 on any factor using the standardized Cronbach’s coefficient alpha [28]. This is a measure of reliability which represents the proportion of total variance in a given factor that can be attributed to a common source. Finally, we calculated the coefficient alpha for each factor and the coefficient alpha when item deleted (i.e., calculated excluding each nutrient, one by one) for each factor and for each nutrient loading j0.40j. In order to facilitate the interpretation of the identified patterns, we calculated the Spearman rank correlation coefficients between the continuous factor scores derived from PCFA and the weekly consumption of 29 food groups and seasonings, obtained summing up the 78 original foods according to homogeneous categories. For each factor, we grouped participants into four categories according to quartiles of factor scores among the controls. We estimated the odds ratios (ORs), and the corresponding 95%
F. Bravi et al. / Cancer Epidemiology 39 (2015) 66–72
68
Table 1 shows the factor loading matrix for the five retained dietary patterns, which together accounted for 80% of the variance of the original nutrients. The first pattern, named ‘‘Western-type diet’’, had the greatest loadings on (i.e. the higher contribution) from calcium, riboflavin, phosphorus, animal protein, saturated fatty acids, cholesterol, and zinc. The second pattern, labeled ‘‘Vitamins and fiber’’, was characterized by high loadings on vitamin C, total fiber, potassium, total folate, beta-carotene equivalents, and soluble carbohydrates. The third pattern, named ‘‘Starch-rich’’, had the highest loadings on starch, vegetable protein, and sodium. The fourth pattern, named ‘‘Animal-derived nutrients and polyunsaturated fatty acids’’, was characterized by the greatest loadings on vitamin D, other polyunsaturated fatty acids, and niacin. The fifth pattern, labeled ‘‘Other fats’’, had the highest loadings on linoleic acid, linolenic acid, and vitamin E. All the nutrients had at least one loading greater than 0.30, thus confirming a relevant role of each nutrient in the original list. Standardized Cronbach’s coefficient alphas were high (at least 0.90) for all the factors and most of the standardized coefficient alphas when nutrient deleted were lower than the corresponding coefficient alpha for the factor, indicating a good internal consistency of the nutrients on the identified factors (data not shown). Baseline characteristics of participants according to quartiles of distribution of the identified dietary patterns are shown in Table 2. OC use was positively related to the ‘‘Animal-derived nutrients and polyunsaturated fatty acids’’ pattern. HRT use was positively related to the ‘‘Vitamins and fiber’’ and inversely related to the ‘‘Other fats’’ pattern. Occupational physical activity was inversely related to the ‘‘Vitamins and fiber’’ and the ‘‘Other fats’’ patterns, and positively related to the ‘‘Starch-rich’’ pattern. BMI
confidence intervals (CIs), for each quartile category compared to the first one using conditional multiple logistic regression models [30], conditioned on age and study center, and adjusted for period of interview, education, body mass index (BMI, based on selfreported anthropometric data), history of diabetes, age at menarche, menopausal status, parity, OC and HRT use. We fitted both separate models for each factor and a composite model that included all the factors simultaneously. We also examined the possible confounding effect of tobacco smoking and alcohol drinking; however, since the inclusion of these covariates in the models did not materially modify the risk estimates, we presented results based on the more parsimonious model described above. We computed tests for linear trend for all these models using the within category medians calculated among the controls. Further, we carried out stratified analysis of risk by age and BMI. We performed the analyses using SAS software, version 9.2 (SAS Institute, Inc., Cary, NC). 3. Results The correlation matrix of the original nutrients was suitable for factor analysis. Each nutrient had at least 10 correlation coefficients greater than 0.30 (data not shown), thus allowing to carry out the analyses on the 28 nutrients. Bartlett’s test of sphericity was highly significant (p < 0.0001), allowing to reject the null hypothesis that the correlation matrix is an identity matrix (Table A1). The individual measures of sampling adequacy were generally very high, with 24 nutrients having measures of at least 0.70. The overall measure of Keiser–Meyer–Olkin was 0.82, confirming that the sample size was adequate for performing factor analysis [28].
Table 1 Factor loading matrixa and explained variances for the five major dietary patterns identified by factor analysis. Italy, 1992–2006. Nutrients
Dietary patterns Western-type diet
Animal protein Vegetable protein Cholesterol Saturated fatty acids Monounsaturated fatty acids Linoleic acid Linolenic acid Other polyunsaturated fatty acids Soluble carbohydrates Starch Sodium Calcium Potassium Phosphorus Iron Zinc Thiamin (vitamin B1) Riboflavin (vitamin B2) Vitamin B6 Total folate Niacin Vitamin C Retinol Beta-carotene equivalents Lycopene Vitamin D Vitamin E Total fiber Proportion of explained variance (%) Cumulative explained variance (%)
Vitamins and fiber
Starch-rich
Animal-derived nutrients and PUFAs
Other fats
0.76 0.17 0.65 0.73 0.31 0.15 0.33 0.10 0.46 0.19 0.52 0.88 0.45 0.79 0.38 0.63 0.53 0.80 0.49 0.39 0.31 0.13 0.36 – – 0.13 0.15 0.13
0.13 0.37 – 0.16 0.30 – 0.12 – 0.65 0.10 – 0.27 0.74 0.34 0.52 0.29 0.52 0.43 0.59 0.72 0.39 0.87 0.11 0.72 0.33 – 0.46 0.82
0.18 0.86 0.28 0.27 0.29 0.14 0.15 0.14 0.16 0.92 0.74 0.10 0.26 0.33 0.38 0.41 0.42 0.12 0.31 0.30 0.34 – – – 0.37 – 0.21 0.38
0.45 0.14 0.46 0.15 0.27 0.15 0.16 0.83 – – 0.10 – 0.26 0.25 0.47 0.44 0.27 0.22 0.39 0.26 0.65 – 0.39 – 0.13 0.85 0.20 –
0.26 0.18 0.31 0.43 0.48 0.87 0.81 0.34 – 0.18 0.16 0.13 0.19 0.21 0.21 0.28 0.22 – 0.27 0.15 0.24 – – 0.23 0.30 0.10 0.78 –
21.74 21.74
19.88 41.62
13.63 55.25
12.61 67.86
12.18 80.04
PUFAs: polyunsaturated fatty acids. a Estimated from a principal component factor analysis performed on 28 nutrients. The magnitude of each loading indicates the importance of the corresponding nutrient to the factor. Loadings 0.63 were used to name the factor and are shown in bold; loadings
