Dietary patterns and risk of cancer: A factor analysis in Uruguay
A multisite case–control study on factor analysis and several cancer sites (mouth and pharynx, esophagus, stomach, colon, rectum, larynx, lung, breast, prostate, bladder, kidney) was conducted in Uruguay. The study included 3,528 cases and 2,532 controls. Factor analysis (principal components) was modeled among controls. This patterning method retained 4 factors per sex, labeled as prudent, drinker, traditional and Western. Odds ratios for these cancer sites, stratified by sex, were estimated using polytomous regression. Whereas the prudent pattern was mainly negatively associated with cancers of the upper aerodigestive tract, the Western pattern showed a strong increase in breast, lung and colon cancers. The study allowed for the reproducibility of the prudent, drinker and Western patterns, whereas the traditional pattern appears to be country specific. © 2008 Wiley-Liss, Inc.
Uruguay is a developing country with high rates of cancer, showing age-adjusted incidence rates of 386.0 per 100,000 men and 303.2 per 100,000 women for all sites.1 The most frequent sites are female breast (age standardized rate 114.9 per 100,000 women) and lung (Age standardized rate 76.5 per 100,000 men). Smoking and alcohol drinking are the major risk factors for frequent cancers like lung cancer and cancers of the upper aerodigestive tract in the Uruguayan population. Also, a diet low in vegetables and fruits and high in beef consumption are risk factors for digestive tract cancers.2
Since foods are consumed together it could be suggested that the true effect of the diet may only be observed when all components are analyzed simultaneously. In fact, foods could act in synergism or be metabolized jointly.3–5 Patterning methodologies, including factor analysis, may turn the analytical difficulties into an advantage. Factor analysis is used to reduce a large number of foods or nutrients to smaller number of factors for modeling purposes.6 This analytic method was originated in the pioneer paper of Spearman in 1904.7 Since then it has grown explosively mainly in psychology and social sciences. In 1992, Randall et al.8 firstly applied this method in the study of food patterns and colon cancer. Since then, numerous studies on dietary patterns and diverse cancers have been published all over the World.
On the basis of a large dataset of cases and controls, we decided to conduct a multisite study using factor analysis in a high-risk country such as Uruguay. The main objective was to estimate odds ratios (ORs) of several cancer sites by means of the application of the principal components method and multinomial regression.
Material and methods
A large scale study was designed jointly by the International Agency for Research on Cancer, Lyon, France and the Epidemiology Group of the Department of Pathology, Uruguay. This study recruited cases and controls from the 4 major hospitals located in Montevideo, Uruguay in the time period 1996–2004.
Selection of cases
The study included cases with cancers of the oral cavity and pharynx, esophagus (squamous cell type), stomach, colon, rectum, larynx, lung, breast, prostate, bladder and kidney (renal cell carcinoma). All the patients with newly diagnosed and microscopically confirmed cancers of the above-mentioned locations were considered eligible for the study, totaling 3,728 patients. Of these 3,728 patients, 200 patients refused the interview, leaving a final number of 3,528 cases (response rate 94.6%). The relative frequency of these cases is shown in Table I.
Table I. Relative Frequency of Cancer Sites
|Mouth and pharynx||274||10.4||9||1.0||283||8.0|
Selection of controls
In the same time period and in the same hospitals, all patients hospitalized by conditions not related with smoking or alcohol drinking and without recent changes in their diets were considered as eligible for the study (total 2,610). In total, 78 patients refused the interview, leaving a final number of 2,532 controls (response rate 97.0%). No attempts for matching were made, since the number of controls was smaller compared with the number of cases. The controls presented the following diseases: eye disorders (622 patients, 24.6%), abdominal hernia (513, 20.3%), osteoarticular disorders (258, 10.2%), injuries (200, 7.9%), varicose veins (178, 7.0%), acute appendicitis (176, 6.9%), diseases of the skin (160, 6.3%), hydatid cyst (127, 5.0%), urinary stones (119, 4.7%), blood disorders (89, 3.5%), bone diseases (50, 2.0%) and prostate hypertrophy (40, 1.6%).
Interviews and questionnaire
All the interviews were conducted in the hospitals by 2 trained social workers. No proxy interviews were admitted. The patients were administered with a structured questionnaire which included the following sections: sociodemographics (age, sex, residence, urban/rural status, education, monthly income), occupations and their duration in years, self-reported height and weight 5 years before the date of the interview, family history of cancer among first-degree relatives, a complete history of smoking (age at start, age at quit, number of cigarettes smoked per day, type of tobacco, type of cigarette, inhalation practices), a complete alcohol drinking history (age at start, age at quit, number of glasses drunk per day, type of beverage), a complete history of mate drinking (a local herbal tea), a complete history of coffee and tea drinking, a complete history of reproductive events (age at menarche, age at menopause, number of live births, number of abortions, breastfeeding, age at first live birth, age at last live birth) and a food frequency questionnaire (FFQ) on 64 food items. This FFQ allowed the calculation of total energy intake and it represents the usual diet of Uruguayan population. Although this FFQ was not validated it was tested for reproducibility with good results.9
Foods included in the factor analysis
After careful examinations, the following foods and food groups were included in the model for factor analysis: fried red meat, barbecued meat, poultry, fish, processed meat (bacon, sausage, mortadella, salami, saucisson, hot dog, ham, salted meat), dairy foods (cheese, butter, whole milk), total eggs (boiled eggs, fried eggs, mayonnaise), desserts (caramel, rice pudding, custard, marmalade, ice cream, cake), total grains (rice, polenta, pasta, maize, oat, white bread), raw vegetables (carrot, tomato, lettuce, onion), cooked vegetables (garlic, swiss chard, spinach, beetroot, winter squash, cabbage, cauliflower, zucchini, red pepper), all tubers (potato, sweet potato), legumes (chickpea, kidney bean, lentil), total fruits (orange, tangerine, apple, grape, peach, plum, banana, figs, fruit cocktail), beer, wine (white wine, red wine) and hard liquor (grappa, sugarcane hard liquor, whiskey, gin, vodka). All the foods and food groups were log transformed before being entered into the analysis.
Principal component analysis was performed with the control group separately by gender and included 17 foods.6, 10, 11 The analysis retained 4 factors per sex representing 36.3% of the total variance for men and 37.8 for women. These factors were rotated with the orthogonal method (varimax method) and normalized by the Horst procedure.12 Afterward they were scored (means zero and standard error one) using Thomson's regression method.13 High loadings between foods and the factor could be positive or negative. To obtain a simple solution for the model we followed the following rules: (i) each factor should have, at least, the same number of zero values of factors and (ii) each factor should present a small number of high loadings.14
To estimate the relationship between scored patterns and the risk of cancer, we fitted the following equation: age, residence, urban/rural status, education, body mass index, smoking status, years after stopping smoking, number of cigarettes smoked per day among current smokers and total energy intake. The estimation was performed using multiple multinomial (polytomous) logistic regression.15 Since each pattern was conditional on the others,16 the equation included all 3 patterns. Furthermore, the scored patterns were categorized into tertiles following the distribution of the controls. All calculations were performed using the STATA software.17
Relative frequency of cancer sites is shown in Table I. Lung, prostate, laryngeal and oropharyngeal cancers are the most frequent malignancies among men. On the other hand, 52% of female tumors correspond to breast cancer, followed by colon cancer, being very small the number of malignancies originated in the oral cavity (9 cases) and larynx (7 cases). Therefore, we combined cancers of the oral cavity and pharynx, esophagus and larynx in a broader category (cancers of the upper aerodigestive tract).
Factor-matrix loadings among male controls (left column) are shown in Table II. The factor 1 was characterized by high positive loadings for poultry, fish, fresh vegetables, and total fruits. For this reason, this factor was labeled as the prudent pattern explaining 11.1% of the variance. Factor 2 showed high positive loadings for total grains, all tubers, desserts, and dairy foods, explaining 8.6% of the variance. This factor was labeled as the traditional pattern. Factor 3 loaded positively on fried red meat, barbecue and eggs and it was labeled as the Western pattern explaining 8.3% of the total variance. Finally, the fourth factor presented high loadings for alcoholic beverages such as beer, wine and hard liquor. For this reason, it was labeled as the drinker pattern. This pattern explained 8.3% of the total variance of the model.
Table II. Factor-Matrix Loadings Among Controls Stratified by Gender1,2
In the same Table, loadings for females are shown in the right column. The prudent pattern displayed lower loadings for total fruits, whereas the traditional one showed higher loadings for eggs intake. The Western pattern presented higher loadings for processed meat consumption compared with results for men and the drinker pattern loaded only in alcoholic beverages. Therefore, we considered that stratifying the factors by gender did not modify essentially the results. All the factors for men and women presented 4 or more zero loadings. The communalities were reasonably high, ranging from a low value of 0.33 for eggs to a high value of 0.71 for total fruits and beer. The model for men explained 36.3% of the total variance, including the error variance, whereas among females explained 37.8%. We think that we have obtained a simple structure following the law of parsimony.
Demographic and patients characteristics among controls are shown in Table III. The prudent pattern was characterized by an increasing gradient of means of intake of raw vegetables and total fruits. The traditional pattern showed an increased intake of cooked vegetables. The Western pattern displayed a clear gradient of intake of red meat whereas the drinker pattern was characterized by an increase of alcohol drinking. These findings were similar in both sexes.
Table III. Demographic and Participant Characteristics of Controls by Factor Tertile1
|Body mass index (BMI)||1||25.0||26.7||25.6||26.8||25.2||26.6||25.5||27.0|
|Smoking intensity (cigarettes smoked per day)||1||18.4||4.5||18.3||5.0||17.0||3.2||14.4||3.0|
|Alcohol drinking (ml of ethanol consumed per day)||1||124.5||11.6||127.7||5.7||93.2||9.5||15.6||0.0|
|Red meat (servings per week)||1||5.7||5.5||5.7||5.0||2.6||3.0||5.6||6.0|
|Poultry (servings per week)||1||0.4||0.5||0.6||0.8||0.8||1.2||0.9||1.0|
|Fish (servings per week)||1||0.1||0.2||0.6||0.6||0.5||0.7||0.5||0.7|
|Fresh vegetables (servings per week)||1||1.0||1.2||2.2||2.7||1.9||2.2||1.8||2.4|
|Cooked vegetables (servings per week)||1||11.8||10.1||8.2||6.9||12.0||9.9||11.8||10.6|
|Total fruits (servings per week)||1||7.0||8.0||10.1||7.1||11.9||10.4||11.3||12.4|
ORs for the prudent pattern among each cancer site stratified by gender are shown in Table IV. Among men, cancers of the mouth and pharynx, esophagus, stomach, colon, colorectum, larynx, upper aerodigestive tract and all sites were inversely associated with risk for the highest tertile of the prudent pattern, with significant dose–response trends. Among females, only breast cancer and cancer of all sites were significantly protective for the prudent pattern. Finally, cancers of the lung, bladder, prostate and renal cell carcinoma were not associated with the prudent pattern.
Table IV. Odds Ratios of Prudent Pattern for Each Cancer Site Stratified by Gender1,2
The ORs for the traditional pattern are shown in Table V. Cancers of the stomach, and upper aerodigestive tract were directly associated with this pattern among men. All p values for trend were significant. The highest OR was observed among the highest tertile of gastric cancer (OR 2.20, 95% CI 1.27–3.81). On the other hand, breast cancer were negatively associated with the traditional pattern and the dose–response was significant (OR 0.53, 95% CI 0.36–0.71, p = 0.002). Finally, female colorectal cancer displayed an increased risk of 1.87 (95% CI 1.02–3.45, p value for trend = 0.04) for the traditional pattern.
Table V. Odds Ratios of Traditional Pattern for Each Cancer Site Stratified by Gender1,2
ORs for the Western pattern discriminated by cancer sites are shown in Table VI. Among men, this scored pattern was directly associated with cancers of the esophagus, colon, rectum, colorectum, larynx, upper aerodigestive tract, lung, bladder and all sites. The highest OR was observed among colon cancer (OR 2.19, 95% CI 1.20–3.97, p value for trend = 0.005), followed by laryngeal cancer (OR 1.80, 95% CI 1.25–2.59, p value for trend = 0.001). Among females, breast cancer was directly associated with risk (OR 1.81, 95% CI 1.32–2.50), whereas cancers of upper aerodigestive tract were negatively associated with Western pattern.
Table VI. Odds Ratios of Western Pattern for Each Cancer Site Stratified by Gender1,2
ORs of the drinker pattern for cancer sites are shown in Table VII. Among men, cancers of the mouth and pharynx, esophagus, larynx and lung were directly associated with the drinker pattern. The highest tertile for the drinker pattern was observed among cancer of the oral cavity and pharynx (OR 2.27, 95 CI 1.53–3.37, p value < 0.0001). Breast cancer was also directly associated with the drinker pattern (OR 1.40, 95% CI 1.05–1.87, p value for trend = 0.02). On the contrary, females afflicted by rectal and colorectal cancer displayed inverse associations with the drinker pattern (OR for the highest tertile of rectal cancer vs. the lowest tertile 0.40, 95% CI 0.19–0.84, p value for trend = 0.01).
Table VII. Odds Ratios of Drinker Pattern for Each Cancer Site Stratified by Gender1,2
This study identified dietary patterns which were associated either with an increased or a decreased risk of several cancer sites, thus confirming the suitability of this approach to study dietary determinants of cancers. Furthermore, this is the first study which employed patterning methodologies in a multisite case–control study.
Cancer of the oral cavity and pharynx was negatively associated with the prudent pattern, whereas it was directly associated with the drinker and traditional patterns. To our knowledge, the only study so far performed on dietary patterns and risk of oral and pharynx was conducted in Uruguayan men.18 This study identified 2 scored patterns: the stew, and vegetables and fruits diets. In fact, the stew pattern was essentially similar to the traditional pattern of this study. The same similarity between the vegetables and fruits pattern and the prudent pattern was observed, suggesting that this study replicates the previous one. Also the drinker pattern was strongly associated with risk of oral/pharyngeal cancer, an expected finding.19
Squamous cell carcinoma (SCC) of the esophagus in men was negatively associated with the prudent pattern and positively associated with the drinker and Western patterns. The traditional pattern was not associated with risk of SCC of the esophagus. Previous studies on this cancer site through factor analysis were conducted in Sweden20 and Uruguay.21 Both studies showed that the drinker pattern was associated with increased risk. In fact, this finding is expected since alcohol drinking is a strong risk factor in this cell type of esophageal cancer.19 Also the prudent pattern was strongly protective in these studies, since loaded on fish, raw vegetables and fruits.20, 21
In this study, gastric cancer was inversely associated with the prudent pattern and positively associated with the traditional pattern. Both the drinker and Western patterns were not associated with this malignancy. Four previous studies on dietary patterns and gastric cancer were conducted in Italy,22 Sweden,20 Uruguay23 and Japan.24 The Italian study used nutrients instead of foods, making its comparison with the other studies rather difficult. Nevertheless, in the Italian study a traditional pattern was identified,22 which loaded on starch and increased the risk of gastric cancer. In the Uruguayan study, the so-called starchy foods pattern was associated with an increased risk of gastric cancer.23 Similar findings were observed in the Japanese prospective study24 and in other reviews.25 A possible mechanism could be related with the loading on grains and tubers, which are important sources of starch, possible carcinogen in the gastric mucosa.25 It should be remembered that reproducibility of a pattern is a central characteristic of simple structures of factor analysis.14, 26
Colon cancer was characterized by a significant increase in risk for the Western pattern among men, whereas the prudent and drinker patterns displayed lower risks for the tumours of the large bowel. Previous studies have reported conflicting results.8, 27–31 Most of the prospective and case–control studies reported significant increases in risk of colorectal cancer associated with the Western pattern.8, 27, 29–31 Although the precise definition of the Western pattern is difficult to establish, red meat, processed meat and eggs are consistently present in this pattern. According to Kesse et al.31 the drinker pattern was directly associated with colorectal cancer, replicating current evidence which strongly suggests that alcohol drinking is a risk factor for colon cancer, mainly among men.19
To our knowledge, only 1 study on dietary patterns and risk of laryngeal cancer has been published.32 In this study, conducted on men only, the traditional, healthy, drinker and Western patterns were associated with risk for this malignancy. This study reproduced the same patterns with similar risks.
In our study, lung cancer showed that the Western pattern was associated with an increased in risk mainly among men. On the other hand, the prudent and traditional patterns were not associated with risk of lung cancer. Previous studies on dietary patterns and risk of the lung displayed different effects to that observed in our study.16, 33 This applies to lung cancer (all cell types together). The study by Balder et al.16 stratified cell types into Kreyberg I and II types. The salad vegetables pattern showed a protective effect for those tumours included in the first category (Kreyberg I), whereas the pork, processed meat and potatoes pattern increased the risk for the same cancers. Similar results for the healthy and Western patterns were observed in the previous Uruguayan study on dietary patterns.33
In this study, breast cancer was inversely associated with the prudent pattern and showed a strong increase in risk for the Western pattern. Also, breast cancer displayed a significant increase in risk for the drinker pattern and was not associated with the traditional pattern. Previous studies on dietary patterns and risk of breast cancer displayed different results.9, 34–36 The study by Terry et al.34 showed an increased risk of breast cancer for the drinker pattern, whereas the healthy and Western patterns were not associated with this malignancy. In the study conducted in Boston, smokers at baseline presented an increase in risk among postmenopausal breast cancer for the Western pattern.35 On the other hand, oestrogen receptor-negative patients were inversely associated with the prudent pattern. Sieri et al.36 identified a pattern called salad vegetables with high loadings on salads, leafy vegetables, raw tomatoes and raw carrots. This pattern was strongly protective among leaner women. Finally, the Uruguayan study displayed a protective effect for the healthy and traditional patterns and increased risk for the Western pattern.9
In our study, the identified dietary patterns were not associated with prostate cancer. Three previous studies37–39 reported rather conflicting results. Therefore, previous studies on prostate cancer did not appear to be related with dietary patterns. Our findings appear to confirm the paucity of findings of these reports.37–39
In this study, bladder cancer was positively associated with the Western pattern, whereas the remaining patterns were not associated with this malignancy. In a recent study,40 the Western pattern also increased the risk of bladder cancer.
Renal cell carcinoma was inversely associated with the drinker pattern among men, but the dose–response was nonsignificant. The remaining patterns were not associated with this malignancy. Two previous studies on dietary patterns and renal cell carcinoma reported a protective effect of alcohol and an increase in risk with a desserts pattern.41, 42 According to the Swedish study, alcohol consumption improves insulin resistance and may reduce the risk of renal cell carcinoma similarly to the cholesterol-lowering statin drugs.41
Like other case–control studies, this multisite case–control study has limitations and strengths. The major limitations are selection bias, recall bias and the possibility of error due to multiple comparisons. On the other hand, this study has several strengths. Firstly, the power of the study (due to the large number of the participants) allowed the estimation of small ORs as significant. Secondly, all the cases were microscopically confirmed by expert pathologists. Finally, the high response rate for both cases and controls should be considered as strength.
In summary, we have conducted a multisite case–control study on factor analysis (principal components) and several cancer sites. Perhaps the major finding was the reproducibility of different dietary patterns such as the Western and drinker ones. Furthermore, the traditional pattern appears to be a country-specific pattern, confirming the results by Balder et al.43 in a study conducted with 5 cohorts of EPIC.