A food pattern that is predictive of flavonol intake and risk of pancreatic cancer

Abstract

Background

In the Multiethnic Cohort (MEC) study, we showed inverse associations between flavonols and pancreatic cancer risk.

Objective

We aimed to define a food pattern associated with intakes of quercetin, kaempferol, and myricetin; to examine the association of that pattern with pancreatic cancer risk; and to investigate the associations in an independent study.

Design

Reduced rank regression was applied to dietary data for 183 513 participants in the MEC. A food group pattern was extracted and simplified and applied to dietary data of 424 978 participants in the European Prospective Investigation into Cancer and Nutrition (EPIC) study. Dietary intake in both studies was assessed by using specially developed questionnaires. Multivariate Cox proportional hazards models were used to estimate relative risks for pancreatic cancer in the MEC (610 cases) and the EPIC (517 cases) studies.

Results

The food group pattern consisted mainly of tea, fruit, cabbage, and wine. In the MEC, inverse associations with pancreatic cancer in smokers were observed for the food group pattern [relative risk: 0.59 (95% CI: 0.31, 1.12) when extreme quintiles were compared; P for trend = 0.03]. In the EPIC study, the simplified pattern was not associated with pancreatic cancer risk (P for trend = 0.78).

Conclusions

A food pattern associated with the intake of quercetin, kaempferol, and myricetin was associated with lower pancreatic cancer risk in smokers in a US-based population. However, failure to replicate the associations in an independent study weakens the conclusions and raises questions about the utility of food patterns for flavonols across populations.

INTRODUCTION

A recent analysis of the Multiethnic Cohort (MEC) study showed an inverse association between the intake of flavonols and the risk of exocrine pancreatic cancer (1). Flavonols are polyphenols that are ubiquitous in plant foods and that have been said to exert cancer preventive effects (2–4). Evidence of a reduced risk of cancer from epidemiologic studies with increasing consumption of these plant constituents is scarce (5), with our recent study (1) being 1 of 2 prospective investigations examining associations between dietary intake of 3 specific flavonols—quercetin, kaempferol, and myricetin—and pancreatic cancer (6).

Complementary studies of nutrients, foods, and food patterns have been recommended to enhance understanding of the relation of diet and health (7, 8). However, probably because of the widespread occurrence of individual flavonols in many different kinds of plant foods, previous analyses by our group of the food groups vegetables (9) and fruit (10) and of single food items (1) did not provide a good basis for following up the relation of flavonols to pancreatic cancer risk.

Dietary pattern analysis tries to address the complexity of dietary intake by deriving informative combinations of foods. Many dietary patterns have been developed and analyzed with respect to disease risks (11, 12). A new method introduced into nutritional epidemiology for dietary pattern analysis is reduced rank regression (RRR) (13). RRR differs from other typically used methods to derive dietary patterns in that it makes use of prior knowledge about the hypothesized diet-disease pathway. It is therefore neither strictly exploratory, as factor or cluster analysis is, nor strictly hypothesis-oriented. With the use of biomarkers and nutrients as response variables, RRR has been successfully applied in several different contexts so far (14–21).

Among the concerns about dietary pattern analysis are that exploratory patterns are population-specific and that results of etiologic studies may be data-driven and not reproducible in other populations (8, 12, 22). This viewpoint may be counter-acted by the construction of so-called simplified dietary patterns (8, 23). Simplified food patterns are unweighted sums of a limited number of foods, standardized as z scores, that represent the most informative foods of the pattern analysis. Such food pattern indexes have a clear meaning in terms of real food consumption and could be investigated in the initial study sample as well as in other study populations. On the basis of our group’s prior finding of inverse associations between flavonols and pancreatic cancer, we designed the present study to define a food pattern associated with the intake of quercetin, kaempferol, and myricetin; to examine associations of the food pattern with pancreatic cancer risk; and to confirm associations of a simplified food pattern with pancreatic cancer in an independent study [the European Prospective Investigation into Cancer and Nutrition (EPIC) study] population.

SUBJECTS AND METHODS

Study design

The Multiethnic Cohort study

The MEC was established to investigate lifestyle exposures, especially diet, in relation to cancer incidence (24). In brief, between 1993 and 1996, >215 000 men and women 45–75 y old were recruited to the cohort. Most of the study participants belonged to 1 of the 5 targeted racial-ethnic groups: African Americans (16%), Japanese Americans (26%), Latinos (22%), Native Hawaiians (6%), and whites (23%). Participants initially completed a self-administered, comprehensive questionnaire that included a detailed dietary assessment and sections on demographic factors; body weight and height; lifestyle factors other than diet, including smoking history; history of medical conditions; and family history of cancer. Follow-up of the cohort entails computerized linkages to cancer registries and death certificate files.

As explained in the invitation to participate, return of a completed questionnaire indicated a subject’s written informed consent. The institutional review boards of the University of Hawaii and the University of Southern California approved the study proposal.

The European Prospective Investigation into Cancer and Nutrition

The EPIC study is a multicenter prospective cohort study (25, 26). In brief, >500 000 men and women 35–70 y old were recruited to the study between 1994 and 2000 in 23 study centers in Norway, Sweden, Denmark, the United Kingdom, the Netherlands, Germany, France, Italy, Spain, and Greece.

Eligible subjects gave written informed consent; they also completed a lifestyle questionnaire, including questions on education and socioeconomic status, occupation, history of previous diseases, lifetime history of smoking, and physical activity, and completed a comprehensive dietary assessment. Approval for the EPIC study was obtained from the ethics review boards of the International Agency for Research on Cancer and of all local institutions at which subjects had been recruited.

Study population

The Multiethnic Cohort study

For the analysis of the MEC, study participants who did not belong to the targeted racial-ethnic groups were excluded, because the quantitative food-frequency questionnaire (QFFQ) was not specifically designed for these groups. In addition, subjects with implausible energy intakes (27), with a pancreatic cancer diagnosis before baseline, with missing information on smoking history or history of diabetes mellitus, and with missing or implausible [ie, body mass index (BMI; in kg/m²) < 15 or > 50] information on body height and weight were also excluded. Data on 183 513 MEC participants were available for this analysis.

The European Prospective Investigation into Cancer and Nutrition

In the confirmation sample of the EPIC study, subjects with prevalent cancer or no follow-up information, subjects without diet or non-diet questionnaires or with an extreme ranking for energy intake, subjects with missing history of diabetes or missing information on smoking history, and subjects with BMIs of <15 or >50 were excluded from the analysis. Furthermore, we excluded the data from Norway from this analysis, because tea consumption was not assessed in the questionnaire administered in that country. For the final analysis, data on 424 978 EPIC study participants were available.

Dietary assessment

The Multiethnic Cohort study

In the MEC, dietary intake during the previous year was assessed by using a comprehensive QFFQ that was specially designed and validated for this multiethnic population (28). To estimate the intakes of 3 flavonols—quercetin, kaempferol, and myricetin—a food-composition table that was developed and that has been maintained at the Cancer Research Center of Hawaii was used. The flavonol content of 51 foods commonly consumed by cohort members were analyzed in Hawaii and added to the food-composition table (29). The flavonol content of some foods was based on values in the literature. Flavonols, as well as flavanols and isoflavones, are a class of polyphenols (flavonoids). Characteristic chemical substances for flavanols and isoflavones are catechin or genistein, respectively.

The European Prospective Investigation into Cancer and Nutrition

In the EPIC study, dietary intake during the previous year was assessed by means of country-specific instruments that had been developed and validated in a series of studies (25, 26). In addition, a highly standardized reference dietary measurement, which served as a calibration reference, was taken from an age-stratified random sample, 8% of the entire cohort, by using a computerized 24-h dietary recall (30, 31). Food intakes were analyzed as predicted by regression calibration (32, 33). Notably, the food-composition tables used for the EPIC study did not include values for flavonols; thus, flavonols as risk factors for pancreatic cancer per se could not be analyzed for the EPIC study.

Total food intakes in the MEC and EPIC studies were allocated to 39 food groups that were identical in the 2 studies. In both studies, composite food items were deconstructed to the commodity level; for example, tomatoes on pizza were allocated to the vegetable group, and salami on pizza was allocated to the meat group.

Identification of pancreatic cancer cases

The Multiethnic Cohort study

Incident exocrine pancreatic cancer cases were identified by record linkages to the Hawaii Tumor Registry, the Cancer Surveillance Program for Los Angeles County, and the California State Cancer Registry. All 3 registries are members of the National Cancer Institute’s Surveillance, Epidemiology, and End Results (SEER) Program. Case ascertainment was complete through December 31, 2003. Diagnoses coded C25.0–C25.3 and C25.7–C25.9 according to the International Classification of Diseases for Oncology, Second Edition, were defined as exocrine pancreatic cancer. Although not included as cases, persons with endocrine pancreatic cancers were censored from follow-up at the date of diagnosis. A total of 610 pancreatic cancer cases were available for the present analysis. Mortality data were obtained from linkage to death certificate files for the states of Hawaii and California and to the National Death Index, which covers the entire United States.

The European Prospective Investigation into Cancer and Nutrition

Case ascertainment in the EPIC study is based on linkages to cancer registries (Denmark, Italy, the Netherlands, Norway, Spain, Sweden, and the United Kingdom) or active follow-up in combination with health insurance records and cancer and pathology registries (France, Germany, and Greece), depending on the options available in various countries. Mortality data also were obtained from either cancer registries or regional or national mortality registries (25). Incident pancreatic cancer cases reported to the central EPIC study database by April 2007 were eligible for inclusion in the present study. A total of 517 primary exocrine pancreatic cancer cases in the EPIC study were available for the present analysis.

Statistical analysis

To derive the food pattern score based on the MEC study data, the statistical method RRR was employed by using the PLS procedure in SAS. A detailed description of this method was provided by Hoffmann et al (13). RRR derives a score—ie, a linear combination of predictor variables (food groups or food items)—that maximizes the variation in response variables (eg, quercetin, kaempferol, and myricetin). An RRR model was performed on the multivariate outcomes intake of quercetin, kaempferol, and myricetin, which were log transformed. The number of extracted pattern scores equals the number of response variables; ie, 3 pattern scores were extracted. For predictor variables, we compared 2 approaches: first, we used food groups that reflected the total diet, and, second, we used QFFQ line items (food items). The food group approach enabled comparison of simplified patterns between the 2 independent studies. However, flavonols generally are not found in foods allocated to a single commonly used food group, but they belong to different food groups: eg, apples belong to the fruit group, and tea belongs to the beverages group. Therefore, we explored the use of QFFQ food items to construct the dietary pattern. We selected the 43 food items that contributed ≥1% mean intake of ≥1 flavonol in the MEC study population. Notably, a difference between the 2 approaches is that the food group approach contains all foods—even those not contributing any flavonols to the diet (eg, animal products)—whereas the food items approach contains only those food items actually contributing flavonols. All intakes—food groups, food items, and flavonols—were analyzed energy adjusted as densities, ie, per 1000 kcal/d.

Two simplified dietary patterns were created, one based on the food group analysis and one based on the food item analysis. To determine the simplified dietary patterns, unweighted standardized food variables (groups or items) that loaded high on the corresponding food pattern were summed (23). The food variables to be summed for each simplified pattern were determined by linear regression of the standardized food group variables on the factor score resulting from the appropriate RRR statistical analysis. All food variables with a standardized parameter estimate of >0.1 in the linear regression were selected.

To evaluate associations between the dietary patterns and pancreatic cancer in the MEC study, we applied Cox proportional hazard models by using age as the time metric to calculate relative risks (RR). Person-times were calculated from the date of cohort entry, defined as questionnaire completion, or the date of the 45th birthday of the few subjects who were <45 y old at baseline. Person-time ended at the earliest of the following dates: date of pancreatic cancer diagnosis, date of death, or December 31, 2003, the closure date for the study. We adjusted for sex and ethnicity as stratum variables to allow for different baseline hazard rates. All Cox models were additionally stratified by follow-up time, which was categorized as ≥2, >2–5, and >5 y. Dietary pattern variables were investigated in disease models in terms of quintiles based on the distribution across the entire cohort. A trend variable to test dose-response relations was assigned pattern score quintiles and analyzed as a continuous variable in the Cox models. Age at cohort entry, history of diabetes mellitus, history of familial pancreatic cancer, smoking status (never, former, or current), pack-years of smoking (assigned zero for nonsmokers), energy intake (logarithmically transformed), and BMI were used as adjustment factors in all multivariate models.

The statistical significance of interactions was determined by analyzing interaction terms in Cox regression models. In a previous analysis by our group (1), the interaction between smoking status and the intake of kaempferol was statistically significant (P < 0.05).

To confirm associations of the simplified pattern score in the EPIC study, RRs for quintiles of the food pattern score were estimated as hazard rate ratios by using Cox regression models with center and age at enrollment in 1-y categories as stratum variables both to control for differences in questionnaire design, follow-up procedures, and other nonmeasured center effects and to be more robust against any violation of the proportionality assumption. Age was used as the primary time variable; entry time was defined as the subject’s age (in d) at recruitment, and exit time was defined as the subject’s age (in y) at pancreatic cancer diagnosis, death, or censoring (lost to follow-up or end of follow-up period). Multivariate regression models were adjusted for sex, smoking status [defined as never, former (quit >15 y ago; quit ≤15 y ago, or unknown), or current smoker], number of cigarettes (0 for never and former smokers), history of diabetes mellitus, BMI, and energy intake (continuous). As a test for trend, the simplified pattern score based on food groups was analyzed as a continuous variable in the Cox regression models.

All analyses were performed by using SAS statistical software (version 9; SAS Institute Inc, Cary, NC). All statistical tests were 2-sided.

RESULTS

The pattern scores

For both approaches to dietary patterns—using food groups or food items—the first extracted factor accounted for a much higher proportion of explained variation in quercetin, kaempferol, and myricetin intake than did the other 2 factors. The explained variation was 68.3%, 5.8%, and 3.6% for factors 1–3 extracted by using food groups and 55.6%, 7.9%, and 6.9% for factors 1–3 extracted by using food items. Therefore, we used only the first factors in subsequent analyses. Variance explained by the first food group factor was largest for myricetin and smallest for quercetin, whereas variance explained by the first food item factor was largest for quercetin and smallest for myricetin.

Participant characteristics

The characteristics of MEC study participant by quintile of the food group pattern score are shown in Table 1. Intakes of quercetin, kaempferol, and myricetin were positively associated with the food group pattern score, as was the intake of tea, the most important food group for the food group pattern, according to the magnitude of the parameter in the linear regression on the first factor. Energy intake was slightly negatively associated with the pattern score. Age at cohort entry was positively associated with the food group pattern score, and the percentage of men and of current smokers decreased with increasing quintiles of the score. BMI was inversely associated with the pattern score. Cross-sectional associations of participant characteristics with the food item pattern score were similar (data not shown).

TABLE 1.

Participant characteristics in the Multiethnic Cohort study by quintile (Q) of the food pattern score¹

	Food pattern score
Characteristic	Q1	Q2	Q3	Q4	Q5
Intake
Flavonols (mg/d)2	7.4 ± 4.33	10.8 ± 5.9	13.6 ± 7.5	16.9 ± 9.1	25.3 ± 13.3
Quercetin (mg/d)	5.4 ± 3.4	7.9 ± 4.6	9.8 ± 5.8	11.7 ± 6.7	15.6 ± 8.1
Kaempferol (mg/d)	1.7 ± 1.3	2.5 ± 1.7	3.2 ± 2.3	4.4 ± 2.9	7.6 ± 4.7
Myricetin (mg/d)	0.3 ± 0.2	0.4 ± 0.4	0.6 ± 0.5	0.9 ± 0.7	2.1 ± 1.4
Energy (kcal/d)	2283 ± 1095	2233 ± 1083	2203 ± 1083	2197 ± 1002	1855 ± 826
Tea (g/d)	8.1 ± 20.4	19.0 ± 37.5	39.8 ± 64.0	94.5 ± 108.6	347.4 ± 285.7
Age at cohort entry (y)	58.0 ± 8.8	59.1 ± 8.9	60.0 ± 8.7	60.8 ± 8.7	61.3 ± 8.7
Men (%)	63	53	45	37	29
Follow-up (y)	9.1 ± 2.0	9.2 ± 1.9	9.2 ± 1.9	9.2 ± 1.8	9.2 ± 1.7
BMI (kg/m2)	26.4 ± 4.9	26.4 ± 4.8	26.3 ± 4.8	25.8 ± 4.8	24.9 ± 4.7
Smoking status (%)
Never	33	42	46	50	53
Former	41	40	40	38	36
Current	26	18	14	12	11
Race-ethnicity (%)
African American	17	19	18	17	12
Japanese American	27	25	25	29	39
Latino	21	25	27	22	15
Native Hawaiian	10	8	7	7	5
White	25	24	24	25	28
History of diabetes mellitus (%)	10	12	13	12	11
Family history of pancreatic cancer (%)	1.3	1.3	1.7	2.0	2.0

¹Trends across quintiles were significant (P < 0.0001) for all variables according to Cochran-Armitage test for categorical variables and the t test for slope in linear regression models of mean values on pattern score for continuous variables.

²Sum of quercetin, kaempferol, and myricetin.

³x̄ ± SD (all such values).

Tea, cabbage, fresh fruit, and wine were those food groups selected for the simplified food group pattern. Black tea, dark leafy greens, apples and applesauce, red wine, and chili were selected for the simplified food item pattern. These foods explained 79% and 58% of the food group and food item pattern scores, respectively. Both food patterns correlated highest with tea or black tea (Table 2). The correlation coefficients of the pattern scores with flavonols were slightly higher for the food group pattern than for the food item pattern. Correlation coefficients of the food groups that were most important in the food group pattern score with the simplified patterns ranged from 0.44 to 0.62, and respective correlation coefficients for the food items that were most important in the food items pattern score ranged from 0.38 to 0.48. The simplified food patterns tended to have more uniform correlations with the food items score than with the original food pattern score. Therefore, the contribution of tea to the simplified patterns was down-weighted.

TABLE 2.

Correlation coefficients of the food pattern score and the simplified food pattern score with food intake or flavonols in the Multiethnic Cohort study

	Food pattern score	Simplified food pattern score1
Food group pattern
Tea	0.90	0.54
Cabbages	0.37	0.62
Fresh fruit	0.30	0.58
Wine	0.13	0.44
Quercetin	0.79	0.66
Kaempferol	0.82	0.64
Myricetin	0.87	0.64
Food items pattern
Chili	0.05	0.38
Dark leafy greens	0.40	0.47
Apples and  applesauce	0.30	0.48
Red wine	0.14	0.44
Black tea	0.77	0.45
Quercetin	0.79	0.62
Kaempferol	0.78	0.53
Myricetin	0.67	0.52

¹A linear combination of standardized intakes of the food groups tea, cabbage, fresh fruit, and red wine for the food group pattern or a linear combination of the standardized intakes of the food items black tea, dark leafy greens, apples and applesauce, red wine, and chili for the food items pattern.

Relative risks for food pattern scores and pancreatic cancer risk

The Multiethnic Cohort study

The RRs for food group and food item pattern quintiles 2–5 compared with the lowest quintile for pancreatic cancer in MEC are shown in Table 3. Although the RR for the highest pattern scores was <1, neither the food group nor the food item pattern was significantly associated with pancreatic cancer risk in the overall cohort. There was some evidence of statistical interaction between the food patterns and smoking status (food item pattern, P for interaction = 0.06; food group pattern, P for interaction = 0.25). Stratification of the models by smoking status showed significantly inverse associations of both food patterns with pancreatic cancer risk among current smokers. Tests for trend in multivariate-adjusted models were significant with the use of either the food group or the food item pattern in this subgroup. The associations were stronger for the pattern using food items; there was a statistically significant risk reduction of 55% in the highest quintile.

TABLE 3.

Multivariate relative risks (and 95% CIs) for food pattern scores based on food groups or food items that were derived to explain variations in intake of quercetin, kaempferol, and myricetin and the incidence of pancreatic cancer in the Multiethnic Cohort study¹

	Food pattern score
	Q1	Q2	Q3	Q4	Q5	P for trend2
Total cohort
Food group pattern score
Cases/noncases (n)	119/36 583	127/36 576	133/36 570	114/36 589	117/36 585
Adjusted for sex and race-ethnicity	1	1.02 (0.79, 1.31)	1.02 (0.79, 1.31)	0.83 (0.64, 1.08)	0.83 (0.63, 1.08)	0.05
Multivariate-adjusted	1	1.06 (0.82, 1.36)	1.08 (0.84, 1.39)	0.88 (0.68, 1.15)	0.88 (0.67, 1.15)	0.14
Food item pattern score
Cases/noncases (n)	115/36 587	121/36 582	150/36 552	117/36 586	107/36 596
Adjusted for sex and race-ethnicity	1	1.00 (0.77, 1.29)	1.21 (0.95, 1.55)	0.93 (0.72, 1.21)	0.84 (0.64, 1.10)	0.18
Multivariate-adjusted	1	1.04 (0.80, 1.34)	1.28 (1.00, 1.64)	0.99 (0.76, 1.29)	0.90 (0.68, 1.18)	0.39
Never smokers3
Food group pattern score
Cases/noncases (n)	36/12113	41/15 378	56/16 794	49/18 291	58/19 510
Adjusted for sex and race-ethnicity	1	0.85 (0.54, 1.33)	1.01 (0.66, 1.54)	0.76 (0.49, 1.18)	0.83 (0.54, 1.28)	0.33
Multivariate-adjusted	1	0.84 (0.54, 1.32)	1.00 (0.66, 1.53)	0.76 (0.49, 1.18)	0.81 (0.53, 1.26)	0.30
Food item pattern score
Cases/noncases (n)	34/13 130	38/15 564	67/16 905	53/17 920	48/18 567
Adjusted for sex and race-ethnicity	1	0.90 (0.56, 1.43)	1.42 (0.94, 2.15)	1.04 (0.67, 1.61)	0.92 (0.59, 1.44)	0.84
Multivariate-adjusted	1	0.90 (0.57, 1.44)	1.44 (0.95, 2.18)	1.04 (0.67, 1.61)	0.92 (0.59, 1.44)	0.82
Former smokers
Food group pattern score
Cases/noncases (n)	41/15 047	51/14 711	58/14 540	50/14 020	46/13 178
Adjusted for sex and race-ethnicity	1	1.23 (0.81, 1.86)	1.36 (0.91, 2.03)	1.16 (0.76, 1.77)	1.09 (0.71, 1.68)	0.85
Multivariate-adjusted	1	1.23 (0.81, 1.85)	1.34 (0.90, 2.01)	1.15 (0.75, 1.74)	1.09 (0.71, 1.69)	0.86
Food item pattern score
Cases/noncases (n)	39/14 669	54/14 524	59/14 413	46/14 115	48/13 775
Adjusted for sex and race-ethnicity	1	1.32 (0.88, 2.00)	1.46 (0.97, 2.19)	1.16 (0.75, 1.78)	1.21 (0.79, 1.87)	0.66
Multivariate-adjusted	1	1.33 (0.88, 2.00)	1.46 (0.97, 2.19)	1.15 (0.74, 1.77)	1.21 (0.79, 1.87)	0.67
Current smokers
Food group pattern score
Cases/noncases (n)	42/9423	35/6487	19/5236	15/4278	13/3897
Adjusted for sex and race-ethnicity	1	1.11 (0.71, 1.75)	0.74 (0.43, 1.28)	0.66 (0.36, 1.20)	0.61 (0.32, 1.15)	0.04
Multivariate-adjusted	1	1.11 (0.70, 1.74)	0.74 (0.43, 1.28)	0.65 (0.36, 1.18)	0.59 (0.31, 1.12)	0.03
Food item pattern score
Cases/noncases (n)	42/8788	29/6494	24/5235	18/4551	11/4253
Adjusted for sex and race-ethnicity	1	0.89 (0.55, 1.44)	0.87 (0.52, 1.44)	0.76 (0.43, 1.33)	0.46 (0.23, 0.90)	0.03
Multivariate-adjusted	1	0.90 (0.56, 1.45)	0.86 (0.52, 1.44)	0.75 (0.43, 1.32)	0.45 (0.23, 0.89)	0.03

¹Cox regression models are stratified for sex, race-ethnicity, and follow-up time and adjusted for BMI, prevalent diabetes mellitus, family history of pancreatic cancer, age at cohort entry, energy intake (logarithmically transformed), pack-years of smoking, and smoking status (when appropriate). Q, quintile.

²Tests for trend were based on analysis of quintile number as continuous variable in Cox regression models.

³Tests for interaction were based on interaction terms between variables in regression models. Food group pattern score × current smoking interaction, P = 0.25; food item pattern score × current smoking interaction, P = 0.06.

Because of the predominance of tea in the food patterns, we analyzed associations between the food group tea and pancreatic cancer along with the simplified patterns for food groups and food items in the MEC study (Table 4). For comparison of the highest category of tea consumption with the category of no tea consumption, the association was inverse, but it was not statistically significant. As expected, none of the simplified patterns was associated with pancreatic cancer in the total cohort. However, inverse associations for tea consumption and the simplified food item pattern were observed in the subgroup of current smokers. The RR (95% CI) for the highest consumption level compared with the lowest was 0.62 (0.33, 1.18; P for trend = 0.03) for tea and 0.48 (0.25, 0.92; P for trend = 0.02) for the simplified food item pattern. The association for the simplified food group pattern was inverse but not statistically significant (RR = 0.60; 0.32, 1.11; P for trend = 0.07).

TABLE 4.

Multivariate relative risk (and 95% CI) for simplified patterns using food groups and pancreatic cancer in the Multiethnic Cohort study¹

	Category2
	1	2	3	4	5	P for trend3
Total cohort
Tea
Cases/noncases (n)	241/69 568	85/28 341	107/28 319	87/28 339	90/28 336
RR (95% CI)	1	0.94 (0.73, 1.21)	1.12 (0.89, 1.42)	0.85 (0.66, 1.10)	0.83 (0.64, 1.07)	0.151
Simplified food group pattern4
Cases/noncases (n)	108/36 594	110/36 593	131/36 571	147/36 556	114/36 589
RR (95% CI)	1	0.92 (0.70, 1.20)	1.02 (0.79, 1.33)	1.09 (0.84, 1.42)	0.85 (0.64, 1.13)	0.669
Simplified food item pattern5
Cases/noncases (n)	110/36 592	145/36 558	115/36 587	134/36 569	106/36 597
RR (95% CI)	1	1.34 (1.04, 1.71)	1.05 (0.81, 1.36)	1.21 (0.94, 1.56)	0.97 (0.74, 1.27)	0.567
Never smokers6
Tea
Cases/noncases (n)	75/27 149	35/12 480	53/13 665	33/13 988	44/14 804
RR (95% CI)	1	1.09 (0.72, 1.64)	1.44 (1.01, 2.05)	0.82 (0.54, 1.25)	0.98 (0.66, 1.45)	0.721
Simplified food group pattern
Cases/noncases (n)	21/12 295	44/15 110	54/17 059	63/18 465	58/19 157
RR (95% CI)	1	1.47 (0.87, 2.48)	1.43 (0.85, 2.38)	1.44 (0.87, 2.40)	1.26 (0.75, 2.13)	0.759
Simplified food item pattern
Cases/noncases (n)	35/14 391	57/15 672	46/16 562	56/17 735	46/17 726
RR (95% CI)	1	1.51 (0.99, 2.31)	1.14 (0.73, 1.77)	1.26 (0.82, 1.92)	1.03 (0.66, 1.60)	0.647
Former smokers
Tea
Cases/noncases (n)	99/28 525	30/11 085	37/10 706	46/10 941	34/10 239
RR (95% CI)	1	0.81 (0.53, 1.22)	0.97 (0.66, 1.42)	1.07 (0.74, 1.54)	0.81 (0.54, 1.22)	0.657
Simplified food group pattern
Cases/noncases (n)	39/14 000	41/14 898	57/14 671	67/14 117	42/13 810
RR (95% CI)	1	0.84 (0.54, 1.30)	1.09 (0.72, 1.65)	1.26 (0.84, 1.90)	0.83 (0.53, 1.32)	0.836
Simplified food item pattern
Cases/noncases (n)	34/14 254	58/14 164	46/14 385	60/14 197	48/14 496
RR (95% CI)	1	1.69 (1.10, 2.58)	1.31 (0.84, 2.05)	1.70 (1.12, 2.60)	1.40 (0.90, 2.18)	0.211
Current smokers
Tea
Cases/noncases (n)	67/13 894	20/4776	17/3948	8/3410	12/3293
RR (95% CI)	1	0.92 (0.55, 1.53)	0.85 (0.49, 1.46)	0.43 (0.20, 0.91)	0.62 (0.33, 1.18)	0.026
Simplified food group pattern
Cases/noncases (n)	48/10 299	25/6585	20/4841	17/3974	14/3622
RR (95% CI)	1	0.70 (0.43, 1.14)	0.68 (0.40, 1.17)	0.67 (0.38, 1.19)	0.60 (0.32, 1.11)	0.074
Simplified food item pattern
Cases/noncases (n)	41/7947	30/6722	23/5640	18/4637	12/4375
RR (95% CI)	1	0.83 (0.52, 1.34)	0.73 (0.44, 1.22)	0.67 (0.38, 1.17)	0.48 (0.25, 0.92)	0.016

¹RR, relative risk; PY, person-years. Cox regression models were stratified for sex, race-ethnicity, and follow-up time and were adjusted for BMI, prevalent diabetes mellitus, family history of pancreatic cancer, age at cohort entry, energy intake (logarithmically transformed), pack-years of smoking, and smoking status (when appropriate).

²Categories for tea: 1, 0; 2, >0–11.2; 3, >11.2–36.8; 4, >36.8–111.4; and 5, >111.4 g/1000 kcal/d. Categories for simplified pattern score are quintiles.

³Tests for trend were based on analysis of quintile number as a continuous variable in Cox regression models.

⁴Linear combination of standardized intake of the food groups tea, cabbage, fresh fruit, and red wine.

⁵Linear combination of standardized intake of the food items black tea, dark leafy greens, apples and applesauce, red wine, and chili.

⁶Tests for interaction were based on interaction terms between variables in regression models. Current smoking × tea interaction, P = 0.22; simplified food group pattern, P = 0.46; and simplified food item pattern, P = 0.61.

The European Prospective Investigation into Cancer and Nutrition

The simplified food group pattern was calculated for the EPIC study data. Cross-sectional associations of participant characteristics showed that age at recruitment increased across quintiles of the simplified pattern score, and the proportions of men and current smokers declined. BMI was not associated with the simplified pattern (data not shown). The correlation coefficients with the simplified food group pattern were 0.63, 0.67, 0.36, and 0.44 for the food groups tea, cabbages, fresh fruit, and wine. We focused on current smokers, because the inverse association between the pattern and pancreatic cancer risk in the MEC study was present for that group only. The inverse association observed in the MEC study could not be confirmed in the EPIC study (Table 5). The estimated RR for pancreatic cancer with the highest simplified food group pattern score category was <1; it was not statistically significant, and neither were the tests for trend. Notably, associations for white smokers in the MEC study were inverse between the simplified food group pattern and pancreatic cancer, but the numbers were too small for meaningful analysis of this stratum only (data not shown). No associations between the patterns and pancreatic cancer in the entire EPIC study cohort or among never or former smokers were observed (data not shown). No statistical interaction was present.

TABLE 5.

Multivariate relative risks (and 95% CIs) for the simplified pattern score and pancreatic cancer in the European Prospective Investigation into Cancer and Nutrition¹

	Categories2
	1	2	3	4	5	P for trend3
Current smokers4
Tea						0.111
Cases/noncases (n)	52/31 488	27/23 059	32/15 990	23/12 260	17/10 971
RR (95% CI)	1	0.91 (0.53, 1.55)	1.19 (0.73, 1.94)	1.36 (0.78, 2.36)	1.53 (0.80, 2.92)
Simplified food group pattern5						0.784
Cases/noncases (n)	54/27 654	30/20 370	23/17 095	29/15 412	15/13 237
RR (95% CI)	1	1.00 (0.62, 1.61)	0.94 (0.54, 1.62)	1.37 (0.80, 2.36)	0.88 (0.45, 1.73)

¹RR, relative risk. Cox regression models were adjusted for age, sex, and center as strata variables and for diabetes mellitus at baseline, BMI, energy intake, smoking status (4 categories), and the number of cigarettes as covariates.

²Categories for tea: 1, 0; 2, >0–49.0; 3, >49.0–124.0; 4, >124.0–273.8; and 5, >273.8 g/1000 kcal/d. Categories for the simplified pattern score are quintiles.

³Tests for trend were based on analysis of quintile number as continuous variable in Cox regression models.

⁴Tests for interaction were based on interaction terms between variables in regression models. Tea × current smoking interaction, P = 0.95; simplified food group pattern and current smoking, P = 0.37.

⁵Simplified pattern score was calculated as the sum of the standardized consumption of food groups tea, cabbage, fresh fruit, and red wine and for food group intake after calibration and adjustment for energy intake as densities, ie, per 1000 kcal/d.

DISCUSSION

Food patterns associated with the intakes of quercetin, kaempferol, and myricetin were inversely associated with pancreatic cancer in smokers in the MEC study. A stronger inverse association was observed by using a simplified food item pattern defined by black tea, dark leafy greens, apples and applesauce, red wine, and chili than was observed by using a simplified food group pattern consisting of tea, cabbage, fresh fruit, and wine. However, the inverse association for current smokers was not reproduced in an independent study population, the EPIC study.

To our knowledge, only 2 studies—one prospective (34) and one case-control (35) study—analyzed dietary patterns as risk factors for pancreatic cancer. Neither the prudent nor the Western dietary pattern determined by factor analysis (principal component) was associated with risk in a US-based prospective study (34). The prudent pattern was characterized by high fruit and vegetable intakes, whereas the Western pattern was characterized by high meat and fat intakes. In the case-control study (35), 3 patterns were derived by applying factor analysis: a Western pattern, a fruit and vegetables pattern, and a drinker pattern. The fruit and vegetables pattern was significantly and inversely associated with pancreatic cancer in men. In contrast to factor analysis, RRR uses a priori information about the hypothesized disease pathway. Because of this, the patterns derived in the present study are characterized by different foods than were reported in the above-referenced studies. A key element in our analysis was that the pattern analysis selected foods that explained variation in the intakes of quercetin, kaempferol, and myricetin. Therefore, the fact that black tea was a major determinant of the food patterns derived in our study is plausible (36). Other studies reporting on major food sources for flavonols in the US diet reported tea to be the most important source (37–40); additional sources identified were onions, apples, broccoli, green salad, and dark leafy greens (38, 40). In a study in the Netherlands in the early 1990s, black tea, apples, onions, and kale were large contributors to the intake of flavonols (36). A Finnish study reported apples and onions to be the largest contributors to the intake of quercetin and white cabbage to be the largest contributor to the intake of kaempferol (41). Intakes of tea and wine were not assessed in that study. Generally, the food groups and food items selected by the RRR analysis have also been identified as contributors to flavonol intake in other US-based and in European studies.

Epidemiologic studies focusing on dietary sources for flavonols and cancer risk found inverse associations of apples and onions with lung cancer in a Finnish cohort study (41) and a US case-control study (42). However, in the same US study, no association with lung cancer risk was found for black tea or red wine, although these foods were among the main sources of quercetin. One reason, as argued by Le Marchand et al (42), could have been the much larger bioavailability of flavonols from onions than from tea. Another Finnish study found an inverse association between tea consumption and lung cancer risk (43). Black tea was not inversely associated with lung cancer risk in a case-control study in smokers in the United States (44). In studies of cancer at other sites, no associations were found for tea, onions, broccoli, apples, or tomatoes and colorectal cancer (40) or for tea, onions, apples, string beans, broccoli, green peppers, or blueberries and breast cancer (45).

Epidemiologic studies of flavonols and the risk of cancer have produced inconsistent findings (1). One recent study on flavonols and pancreatic cancer found results similar to the MEC study results: that is, inverse associations among smokers in a Finnish intervention trial of α-tocopherol and β-carotene, who belonged to the placebo group, but not for those in the overall or the intervention groups (6).

In the present study, inverse associations were stronger in smokers, and some evidence of statistical interaction was present. Stronger associations with intake of flavonols among smokers have been found in some (6, 42, 46) but not all (41) other studies. The modulation of enzyme activities has been proposed as one mechanism by which flavonols might act as a cancer preventive (47), such as by the inhibition of certain cytochrome P450 enzymes involved in the bioactivation of chemical carcinogens. This approach may have more effect in smokers, who are exposed to carcinogens from smoke, than in never smokers who lack this exposure. Le Marchand et al (42) reported an effect modification by the CYP1A1 genotype, which encoded for P450 enzymes, and by the intake of onions on lung cancer risk.

It was suggested in the past that the results of etiologic studies using exploratory dietary pattern analyses may not be reproducible in other populations (8, 12, 22), because of the data-driven nature of the pattern. We attempted to overcome this limitation and to confirm food pattern and pancreatic cancer associations in an independent study population. Because of differences in the FFQs used in the MEC and EPIC studies, the harmonization of food classifications into identical food groups was crucial to this analysis, and a confirmatory analysis on the level of food items was impossible. The idea underlying our approach was that the food pattern that we derived in the MEC study is the pattern for flavonols—ie, the combination of foods giving the most information about flavonol intake. Accordingly, the food pattern can be investigated in other studies and for different endpoints. However, we were not able to confirm in the EPIC study the association for the food group pattern observed in the MEC study. This outcome merits some discussion.

One question is whether food sources for flavonols in European countries are different from those in the United States. As discussed above, according to published studies, major food sources of flavonols in the United States and European countries seem to be very similar. Indeed, an RRR analysis of data of the EPIC-Potsdam study of 27 542 participants, which included estimates of flavonol intake, also selected tea, fresh fruit, and wine as the food groups most associated with the intakes of quercetin, kaempferol, and myricetin (data not shown). This finding supports the assumption that the food group pattern derived in the MEC study also qualifies as a pattern that is explanatory of flavonol intake in the EPIC study. As for absolute levels of intake, mean tea consumption was higher in the EPIC than in the MEC study. Unfortunately, the lack of food-composition data including flavonols for the whole EPIC study rendered the separate derivation of a food pattern for flavonols in Europe impossible at this time and also precluded an analysis of flavonols per se and pancreatic cancer risk in the EPIC study. Relations between flavonols and total flavonoids, which may differ between populations, also could not be examined.

The fact that the food item pattern showed stronger associations in the MEC study than did the food group pattern gives support to the notion that commonly used food groups are of limited use for reflecting flavonol intakes. However, although associations for the food group pattern generally were weaker than those for the food item pattern, we would have expected, at the least, to observe the same direction for the associations in both studies. The failure to confirm the flavonol food pattern finding in an independent study raises doubt about the external validity and utility of the pattern across populations. As our group reported earlier (1), an association with flavonols seems to be present in the MEC study, and the fact that the strongest association seen in current smokers could be reproduced on the food level by using RRR food pattern analysis further strengthens this observation. Notably, the associations for the food pattern in general were weaker than those for the actual flavonols, which indicates, on the one hand, that the pattern did not detect synergy of constituents and, on the other hand, that values for flavonols themselves apparently provide the best information for analysis. Considering the loss of information when the pattern approach, rather than the single nutrient approach, is applied in this particular case and considering the weakness of the association between the simplified food group pattern and pancreatic cancer in the MEC study, the lack of confirmation of the pattern may not be suprising. This possibility suggests that analyses based on flavonol intake itself, rather than on flavonol food patterns, are preferable for elucidating associations with disease outcomes and especially for comparisons across different study populations. However, it must be kept in mind that associations on the level of foods in general are easier to communicate from a public health perspective.

Clearly, the present study was somewhat limited by dietary data that relied on questionnaire assessments of intake, and some degree of measurement error in dietary intake was certainly present. However, validation studies for the MEC study showed reasonably good performance (28), and the EPIC study made use of its integrated calibration study to obtain food group data (31). Furthermore, the use of energy adjustment in assessing intakes should have reduced the measurement error to some extent (48). Our results also should be interpreted with caution because of the low number of cases in some categories in the stratified analyses.

The large sample sizes of both studies and the prospective designs are major strengths of this analysis. The prospective design ruled out recall bias, which can be a major drawback of case-control studies, especially for pancreatic cancer, in which proxy interviews are usually necessary because of the rapid fatality of the disease.

In conclusion, we observed an association between a food pattern consisting of tea, cabbage, fresh fruit, and wine and pancreatic cancer in smokers in the US-based MEC study, presumably due to the intake of flavonols in these foods. However, the associations of the food pattern could not be confirmed in an independent study in a European population (the EPIC study), which brings into question the utility of the food pattern approach for analyses into flavonols in the light of the small effect on pancreatic cancer risk across different populations.