Dietary patterns and risk of pancreatic cancer

Abstract

To investigate associations between broad dietary patterns and pancreatic cancer risk, we conducted a case-control study of 585 histologically confirmed pancreatic cancer cases and 4,779 population-based controls in 8 Canadian provinces between 1994 and 1997. Dietary intake was assessed using a FFQ. Major dietary patterns were identified by factor analysis. Unconditional logistic regression was used to describe associations between dietary pattern scores and risk of pancreatic cancer. Three dietary patterns were identified: Western, characterized by high intake of processed meats, sweets and desserts, refined grains and potatoes; fruits and vegetables, characterized by high intake of fresh fruits and cruciferous vegetables; drinker, characterized by high consumption of liquor, wine and beer. After adjustment for age, BMI, smoking, physical activity, province, educational attainment and total energy intake, the fruits and vegetables pattern was associated with a 49% reduction in pancreatic risk among men (OR = 0.51, 95% CI 0.29–0.90, p = 0.004) when comparing the highest and lowest quartiles of dietary pattern scores. No significant relationship was observed with the Western and drinker patterns. Although the response rate for eligible, recruited subjects was relatively low, our results suggest that the fruits and vegetables dietary pattern reduces pancreatic cancer risk among men. © 2004 Wiley-Liss, Inc.

Because of its incidence and poor prognosis, pancreatic cancer is one of the most important human malignancies. Approximately 216,300 new cases occur worldwide each year.1 In Canada in 2004, it is the fourth leading cause of cancer-related death in both men and women.2 Survival is extremely low, with a 5-year rate of <5%3 and a case-fatality ratio of 0.99.2

Few risk factors for pancreatic cancer have been consistently identified. About 30% of cases are attributable to smoking.1, 4 An increased frequency of pancreatic cancer has been suggested among individuals with a long-standing history of diabetes,5, 6 chronic pancreatitis,7, 8 pernicious anemia9 and inheritable syndromes such as familial adenomatous polyposis.10, 11 An elevated risk of pancreatic cancer has also been seen in families with a history of breast cancer and BRCA2 mutations.12, 13 It has been estimated that 30–50% of pancreatic cancers may be due to dietary factors,14 though for most dietary components the overall evidence is equivocal. A number of investigators have recorded a heightened risk in relation to excessive intake of fat,15 grilled and salted/smoked red meats,16, 17 dairy products18 and cholesterol,19, 20 though some of these associations were not confirmed in a large prospective cohort of women.21 In contrast, pancreatic cancer risk has been found to decrease with increased consumption of fresh fruits and vegetables20, 22 as well as foods containing no preservatives or additives19, 20 and with increased carbohydrate and energy intake.23 However, no studies to date have examined broader food consumption, which provides a more comprehensive representation of dietary intake involving a large number of dietary components working together. The present investigation was undertaken to explore the possible role of major dietary patterns in the etiology of pancreatic cancer within the Canadian NECSS.

BMI, body mass index; CI, confidence interval; CNF, Canadian Nutrient File; FFQ, food-frequency questionnaire; I3C, indole-3-carbinol; NECSS, National Enhanced Cancer Surveillance System; OR, odds ratio.

Material and methods

Study population

The NECSS is a multisite, population-based, case-control study involving 21,022 participants with one of 19 types of cancer identified through cancer registries in 8 of the 10 Canadian provinces (Alberta, British Columbia, Manitoba, Newfoundland, Nova Scotia, Ontario, Prince Edward Island and Saskatchewan). The present investigation is based on the pancreatic cancer portion, and the study population has been described elsewhere.24 Briefly, between April 1, 1994, and December 31, 1997, the participating provincial registries identified pancreatic cancer cases as early as possible in the registration process, to minimize loss of subjects because pancreatic cancer is a rapidly fatal disease. All cases included in the NECSS were confirmed histologically and defined according to the WHO's International Classification of Diseases, rubric 157.25 Data were intentionally collected by proxy only in Ontario, where 132 of 271 returned questionnaires (49%) were completed by proxy respondents. Proxy data were obtained for only 8 cases in the other provinces. However, like most other studies of pancreatic cancer that collected information directly from case subjects, the overall proportion of eligible cases that responded was low. Among men diagnosed with pancreatic cancer, 30% had died before interview and consent was not granted by physicians for an additional 15%. Among women, 28% had died before they could be contacted and the attending physician refused consent to approach patients for an additional 16%. Among those sent questionnaires, response rates were 55% for men and 56% for women. The vast majority of cases were ascertained within 1– 3 months of diagnosis; physician consent to send questionnaires to patients was obtained within 1 month, and approximately 70% of questionnaires were returned within 2 months of mailing.

The NECSS used frequency matching in selecting the control population to achieve age and gender distributions similar to those of all cancer cases combined. Based on the projected number of incident cancer cases by province, questionnaires were mailed to 8,117 subjects throughout the 1996 calendar year, using the same protocol as for cases. Questionnaires were returned for 573 controls (7.4%) because of a wrong or old address, with no updated address found. Strategies for control selection varied by province, depending on data accessibility. In Prince Edward Island, Nova Scotia, Manitoba, Saskatchewan and British Columbia, provincial health insurance plans were tapped to obtain a random sample of the population stratified by age and gender. In each of these provinces, >95% of residents are covered by public health-care plans. Active military personnel and their families as well as indigenous peoples were excluded because they were covered by other plans. In Ontario, Ministry of Finance data were used to derive a stratified random sample, while Newfoundland and Alberta adopted a random digit-dialing method to enroll a population sample. A total of 5,039 controls were selected to serve as a common control group for all types of cancer. Response rates of 65% and 71% were obtained from the respective male and female control populations.

Questionnaires, with telephone follow-up for clarification when necessary, were mailed to subjects to obtain information on residential and occupational history as well as other risk factors for cancer. Included were questions on smoking history, height, weight, physical activity, ethnicity and education.

Dietary assessment

Food consumption data were obtained via a semiquantitative FFQ developed after 2 instruments that have been widely validated: the short Block questionnaire26 and the Willett questionnaire.27 Subtle changes were made to the questionnaire items to take into account variations between American and Canadian dietary practices. The FFQ includes questions on 69 different food and beverage items as well as the frequency of consumption and the amounts consumed. Participants were asked how often they had consumed these foods per week in the time period 2 years prior to interview. Daily energy intake was determined by summing the caloric content of each food and beverage item, based on reference values given in the 1997 CNF. Food items were then classified into 38 groups (Appendix Table I). The food grouping design was based on the CNF, similarity of nutrient profiles or processing and culinary usage. Some individual foods were retained either because they did not fit within a broader group (e.g., nuts, eggs and tea) or because they represented distinct nutritional structures (e.g., wine, beer and supplements).

Statistical analysis

In total, data were collected on 630 cases (355 men, 275 women) and 5,039 controls (2,547 men, 2,492 women). We excluded controls with missing age or who were under 30 years of age (n = 226) because pancreatic cancer cases are usually 30 years or older. Furthermore, subjects with daily intake <800 kcal (28 cases, 44 controls) or >4,500 kcal (17 cases, 20 controls) were excluded because such intakes are unrealistic and, hence, of questionable validity. A total of 585 cases (335 men, 250 women) and 4,779 controls (2,422 men, 2,357 women) were eligible for analysis.

Based on the food and beverage items retrieved from the FFQ, major dietary patterns were derived by factor analysis, using an orthogonal rotation (VARIMAX) procedure. This procedure minimizes the correlation between factors and results in a simpler data structure with greater interpretability.28 The correlation matrix, rather than the covariance matrix, was employed since the variables were not all of the same type and magnitude.29 In determining the number of factors to retain, we considered eigenvalues (>1.5), the Scree plot and factor interpretability. This analysis suggested 3 broad dietary patterns in the data. A dietary pattern score was computed for each study subject by summing intakes of each food group weighted by its factor loading,30 which represents the relative contribution of that food group. Positive loading indicates that the dietary variable is positively associated with the factor, whereas negative loading reflects an inverse association with the factor. Labeling of dietary patterns was descriptive, based on our interpretation of the pattern structures.

To assess the relationships of the dietary patterns with selected lifestyle variables, we calculated the χ² for trend and p value with the median value for each variable according to categories of dietary patterns. To determine the associations between dietary pattern scores and pancreatic cancer risk, subjects were divided into 4 categories according to quartiles of dietary pattern scores among the overall study population. ORs and 95% CIs were calculated using unconditional logistic regression. Risk estimates were initially adjusted only for age and total energy intake. Multivariate models were also used to adjust more fully for matching variables (age group and province), lifetime cigarette consumption (0, >0–15 and >15 pack-years), BMI (<23, 23–24.9, 25–26.9, 27–29.9 and ≥30 kg/m²), physical activity (total number of hours/month of moderate and strenuous activity with a weighting factor of 9/5 for strenuous activities),31 educational attainment (years) and total energy intake (as a continuous variable). Furthermore, analysis was run excluding all proxy interviews, and models were fitted for men and women separately. Tests for linear trend in logistic regression analysis were performed with scores derived from the median values of categorized variables and entered into the model as successive integers. All tests of statistical significance were 2-sided. Analyses were conducted with SPSS for Windows (release 10.02, 1999; SPSS, Chicago, IL).

Results

Demographic and other selected characteristics of the study population are presented in Table I. Regardless of gender, the age distributions of cases and controls were fairly similar, and 98% of the study population was Caucasian. Cases were more likely to consume lower amounts of fruit and vegetables, with higher amounts of tobacco and greater total energy intake than controls. Both male and female cases were more likely to be interviewed without the use of a proxy respondent, while there were no appreciable differences between cases and controls with respect to BMI 2 years prior to the diagnosis of cancer.

Table I. Selected Characteristics of the Study Population (N = 5,146), NECSS, Canada, 1994–1997

Characteristics	Male (n = 2,600)				Female (n = 2,546)
	Cases		Controls		Cases		Controls
	n	%	n	%	n	%	n	%
1 Subjects with unrealistically low (<800 kcal) or high (>4,500 kcal) intakes were excluded.
Age (years)
30–34	3	1	122	6	4	2	58	3
35–39	7	2	135	6	4	2	122	5
40–44	7	2	108	5	12	5	217	9
45–49	21	6	111	5	13	5	320	14
50–54	32	10	146	7	23	9	260	11
55–59	46	14	211	9	29	12	254	11
60–64	60	18	336	15	50	20	336	15
65–69	81	24	520	22	60	24	366	16
70–74	78	23	576	25	55	22	363	16
Ethnicity
Caucasian	331	99	2,257	97	242	99	2,237	97
Other	4	1	8	3	8	1	59	3
Cigarette consumption (pack-years)
0	75	22	553	24	107	43	1,148	50
>0–15	75	22	715	32	59	23	680	30
>15	185	56	997	44	84	34	468	20
Use of proxy
Yes	87	26	3	1	54	22	2	1
No	248	74	2,262	99	196	78	2,294	99
BMI 2 years prior to diagnosis(kg/m^2)
<23	62	19	413	18	89	36	813	36
23–24.9	60	18	445	20	44	17	486	21
25–27.9	77	23	574	25	35	14	339	15
28–29.9	71	21	500	22	32	13	352	15
≥30	65	19	333	15	50	20	306	13
				Cases	Controls	Cases		Controls
Mean fruit/vegetable consumption (servings/day)				4.3	4.6	4.7		4.8
Interquartile range (servings/day)				2.6	2.4	2.6		2.7
Mean daily energy intake1 (kcal)				2,062	1,915	1,802		1,752
859	877	654		756

Three major dietary patterns were identified (Table II). The first was labeled “Western” since, as in previous studies,32, 33 it reflected intake of foods related to a Western diet, such as processed meats, sweets and desserts, refined grains and potatoes. The second was labeled “fruits and vegetables” since it was heavily weighted by fresh fruits and cruciferous vegetables, with a modest involvement of green and dark yellow vegetables and less refined grains. The third was labeled “drinker”, reflecting elevated liquor, wine and beer consumption. These 3 patterns explained 16.7% (8.2%, 4.4% and 4.1%, respectively) of the total variation in food intake among individuals.

Table II. Factor-Loading Matrix for the 3 Dietary Patterns in the Study Population, NECSS, Canada, 1994–1997

Food group	Mean intake (g/day)	Dietary pattern
		Western	Fruits and vegetables	Drinker
1 For simplicity and ease of interpretation, factor loadings <0.15 in absolute value are indicated by a dash.
Processed meats	24.30	0.52	—1	—
Sweets and desserts	39.83	0.49	—	—
Refined grains	20.03	0.48	−0.33	—
Potatoes	134.14	0.43	—	—
Processed fish	11.26	0.41	—	—
Organ meats	4.88	0.39	—	—
Soft drinks	263.70	0.36	—	—
Legumes and legume products	28.62	0.36	0.28	—
Snacks	7.13	0.34	—	—
Margarine	5.51	0.33	—	—
Nuts	6.13	0.32	—	—
Fried potatoes	18.68	0.28	—	—
Mayonnaise	6.23	0.26	—	—
Eggs	16.38	0.24	—	—
Oils	0.52	0.24	—	—
High-fat dairy products	331.18	0.23	—	—
Red meats	85.84	0.23	—	—
Fruits	278.95	0.26	0.45	—
Cruciferous vegetables	21.76		0.43	—
Green and other vegetables	32.75		0.38	—
Dark yellow vegetables	37.20	0.26	0.37	—
Low-fat dairy products	444.97	0.18	0.33	—
Rice, pasta	92.44	—	0.32	—
Whole grains	34.71	—	0.32	—
Soups	87.92	—	0.31	—
Poultry products	16.54	—	0.31	—
Cold breakfast cereals	86.84	—	0.29	—
Tomatoes	62.23	0.19	0.28	—
Supplements	0.03	−0.24	0.27	—
Water	681.82	—	0.24	—
Fruit juices	244.61	—	0.24	—
Fish	26.79	—	0.24	—
Liquor	13.07	—	—	0.74
Wine	29.06	—	—	0.73
Beer	146.62	—	—	0.61
Tea	331.18	0.21	—	−0.24
Spices and herbs	0.54	—	−0.17	0.20
Coffee	536.58	—	−0.18	0.20

The relationships between selected lifestyle variables and the 3 dietary patterns are summarized in Table III. Age was positively associated with the Western pattern (p = 0.0002) but not with the fruits and vegetables or drinker pattern. Total energy intake was positively linked with the Western pattern (p = 0.002) but not with the fruits and vegetables or drinker pattern. No relationship between BMI and any of the 3 dietary patterns was apparent. Smoking was positively associated with both the Western (p = 0.0003) and the drinker (p = 0.011) patterns but not with the fruits and vegetables pattern. Physical activity was inversely related to the Western pattern (p = 0.001) but not to the fruits and vegetables or drinker pattern. Educational attainment was inversely associated with the Western pattern only (p = 0.0003).

Table III. Associations Between the 3 Dietary Patterns with Selected Lifestyle Variables, NECSS, Canada, 1994–1997

Quartile1	Number of cases/controls	Age2 (years)	Energy intake2 (kcal/day)	BMI2 (kg/m^2)	Smoking2 (pack-years)	Physical activity2 (hr/month)	Educational attainment2 (years)
1 Subjects were divided into 4 categories according to quartiles of the lifestyle variable in the study population. 2 Median value.
Western pattern
1 (low)1	121/1,163	59	1,484	24.6	2	15.2	13.0
2	137/1,148	60	1,680	25.0	4	12.2	12.0
3	147/1,142	61	1,845	25.5	7	11.0	12.0
4 (high)	180/1,108	63	2,099	25.7	9	10.6	11.0
p		<0.01	<0.01	0.29	<0.01	<0.01	<0.01
Fruits and vegetables pattern
1 (low)	143/1,133	58	1,666	25.6	14	7.8	11.5
2	160/1,117	61	1,638	25.2	5	11.3	12.0
3	148/1,139	61	1,804	25.0	2	13.9	12.0
4 (high)	134/1,172	63	1,900	24.8	1	15.6	12.0
p		0.49	0.22	0.55	0.90	0.60	0.74
Drinker pattern
1 (low)	127/1,173	64	1,817	25.1	1	11.25	11
2	140/1,118	60	1,707	25.2	1	11.62	12
3	132/1,148	58	1,665	25.0	6	13.28	12
4 (high)	176/1,122	60	1,802	25.1	12.9	12.12	12
p		0.86	0.09	0.19	0.01	0.75	0.25

Table IV presents the ORs for pancreatic cancer according to the 3 dietary patterns. After adjustment for age and total energy intake, pancreatic cancer risk for the fruits and vegetables pattern was reduced among men (OR = 0.74, 95% CI 0.52–1.05, p = 0.036) when comparing the highest to the lowest quartile of dietary pattern scores. This reduction in risk was more pronounced in the fully adjusted model (OR = 0.55, 95% CI 0.32–0.93, p = 0.008). Moreover, exclusion of proxy data did not materially affect the results (OR = 0.51, 95% CI 0.29–0.89, p = 0.004). In contrast, pancreatic cancer risk was significantly increased with the drinker pattern (OR = 1.60, 95% CI 1.14–2.34, p = 0.004) after adjusting for age and total energy intake. However, the strength of this association was weakened when additional covariates were included in the model, and the trend was no longer significant (OR = 1.35, 95% CI 0.81–2.25, p = 0.18). No significant relationship was observed among females.

Table IV. OR for Pancreatic Cancer According to Quartile of Dietary Pattern Score, NECSS, Canada, 1994–1997

Adjustment	Quartiles				p1 for trend
	1 (low)	2	3	4 (high)
1 Two-sided Wald tests. 2 Adjusted for age (in 5-year groups), smoking (0, >0–15 and >15 pack-years), BMI (5 categories), physical activity (total number of hr/month of moderate and strenuous activities), province (8 groups), educational attainment (years) and total energy intake (as a continuous variable). 3 Same as above but excluding subjects with proxy interview.
Western pattern
Males
Age- and energy-adjusted OR (95% CI)	1.00	1.10 (0.76–1.58)	0.95 (0.66–1.37)	1.10 (0.76–1.57)	0.72
Fully adjusted model 12: OR (95% CI) (cases/controls)		1.57 (0.94–2.64)	1.14 (0.66–1.96)	1.30 (0.75–2.25)	0.56
	58/461	79/536	80/609	118/659
Fully adjusted model 23: OR (95% CI) (cases/controls)		1.51 (0.88–2.59)	1.23 (0.70–2.14)	1.25 (0.71–2.23)	0.49
	42/496	61/575	61/648	84/700
Females
Age- and energy-adjusted OR (95% CI)	1.00	1.03 (0.70–1.49)	1.34 (0.92–1.94)	1.26 (0.85–1.88)	0.12
Fully adjusted model 1: OR (95% CI) (cases/controls)		1.18 (0.67–2.01)	1.42 (0.82–2.47)	1.40 (0.76–2.59)	0.10
	63/702	58/612	67/533	62/449
Fully adjusted model 2: OR (95% CI) (cases/controls)		1.18 (0.69–2.02)	1.35 (0.78–2.37)	1.36 (0.73–2.54)	0.14
	52/724	46/627	53/545	45/459
Fruits and vegetables dietary pattern
Males
Age- and energy-adjusted OR (95% CI)	1.00	1.21 (0.89–1.64)	0.96 (0.69–1.34)	0.74 (0.52–1.05)	0.04
Fully adjusted model 1: OR (95% CI) (cases/controls)		0.93 (0.61–1.43)	0.78 (0.49–1.24)	0.55 (0.32–0.93)	<0.01
	99/695	97/559	75/504	64/507
Fully adjusted model 2: OR (95% CI) (cases/controls)		0.87 (0.55–1.35)	0.80 (0.50–1.29)	0.51 (0.29–0.90)	<0.01
	81/745	71/605	52/538	44/531
Females
Age- and energy-adjusted OR (95% CI)	1.00	0.99 (0.70–1.40)	1.05 (0.73–1.51)	1.06 (0.72–1.58)	0.59
Fully adjusted model 1: OR (95% CI) (cases/controls)		1.30 (0.73–2.31)	1.30 (0.73–2.30)	1.06 (0.58–1.96)	0.98
	79/681	68/641	59/572	44/402
Fully adjusted model 2: OR (95% CI) (cases/controls)		1.35 (0.75–2.42)	1.30 (0.73–2.34)	1.06 (0.57–1.96)	0.92
	32/452	51/576	62/652	51/675
Drinker dietary pattern
Males
Age- and energy-adjusted OR (95% CI)	1.00	1.42 (0.98–2.06)	1.22 (0.84–1.76)	1.60 (1.14–2.34)	<0.01
Fully adjusted model 1: OR (95% CI) (cases/controls)		1.38 (0.82–2.31)	0.88 (0.52–1.50)	1.51 (0.92–2.47)	0.09
	58/492	72/477	73/576	132/720
Fully adjusted model 2: OR (95% CI) (cases/controls)		1.28 (0.74–2.19)	0.81 (0.47–1.42)	1.35 (0.81–2.25)	0.18
	40/511	53/529	55/621	100/748
Females
Age- and energy-adjusted OR (95% CI)	1.00	0.99 (0.70–1.40)	1.05 (0.73–1.51)	1.06 (0.72–1.58)	0.59
Fully adjusted model 1: OR (95% CI) (cases/controls)		0.86 (0.51–1.43)	0.78 (0.45–1.36)	0.95 (0.53–1.71)	0.80
	79/681	68/641	59/572	44/402
Fully adjusted model 2: OR (95% CI) (cases/controls)		0.81 (0.48–1.36)	0.75 (0.43–1.31)	0.83 (0.46–1.52)	0.53
	63/692	52/671	48/586	33/406

Subgroup analyses of pancreatic cancer risk by age, BMI, educational attainment and smoking status revealed no effect modification (data not shown).

Discussion

The present population-based, case-control study found a significant inverse association between the fruits and vegetables dietary pattern and pancreatic cancer risk in males. Our results are consistent with previous investigations, particularly those that examined associations between consumption of fruits and vegetables and pancreatic cancer risk.

A multicenter, population-based, case-control study in the United States34 discerned a statistically significant inverse trend in pancreatic cancer risk with increasing consumption of cruciferous vegetables in men (OR = 0.5, p = 0.04) and women (OR = 0.4, p = 0.002). In that study, frequent consumption of raw fruits was associated with a nonsignificant 20% decreased risk among men but with little or no risk reduction among women. In a population-based, case-control investigation in Japan,22 raw vegetable consumption was associated with decreased pancreatic cancer risk (OR = 0.71, 95% CI 0.51–0.99). However, when subgroup analysis was performed by gender, the inverse association was no longer significant in both males and females. In a population-based, case-control study in Shanghai,35 a significant inverse association between risk of pancreatic cancer and consumption of both fruits and vegetables was detected only among men. In addition, 3 other investigations,18, 19, 36 which did not take into account effect modification by gender, likewise report that fruits and vegetables may be preventive against pancreatic cancer.

The fruits and vegetables dietary pattern was loaded heavily on fresh fruits (e.g., apples, oranges, bananas and cantaloupes) and cruciferous vegetables (e.g., broccoli, cabbage, cauliflower and Brussels sprouts). These foods are important sources of nutrients, with vitamin C and carotenoids being the main contributors. They also contain high concentrations of unique phytochemicals with potential cancer-preventing properties. Several biologic mechanisms that might explain the protective influence of fruits and vegetables on carcinogenesis have been proposed. Vitamin C has been hypothesized to protect against cancer at many sites by preventing oxidative damage to polyunsaturated fatty acids and DNA through antioxidant and free radical–scavenging actions.37 Vitamin C has also been shown to inhibit N-nitrosation, a process by which carcinogenic substances, such as N-nitroso compounds, are formed in food and in the gastrointestinal tract that may be involved in the etiology of pancreatic cancer.38 Carotenoids are involved in putative cancer-preventing activity via their antioxidant properties, stimulation of gap junction intracellular communication, induction of detoxifying enzymes, inhibition of cellular proliferation39 and enhancement of immune function.40 I3C, a phytochemical abundant in cruciferous vegetables, increases the ratio of 2/16α-(OH)estrone, a metabolite ratio able to prevent or halt carcinogenesis.41 I3C has also been reported to operate through other pathways, including induction of apoptosis,42 inhibition of microtubule assembly43 and decreased effect of 4-(methylnitrosamino)-1-(3-pyridyl)-1-butanone, a known tobacco carcinogen.44 These different mechanisms, which may operate simultaneously, provide support for the hypothesis that high fruits and vegetables dietary pattern scores may be related to reduced pancreatic cancer risk.

It has been argued that consumption of foods commonly considered to be healthy, such as fruits and vegetables, may be part of an overall prudent lifestyle. In this investigation, age, total energy intake, BMI, smoking status, physical activity and educational attainment were not significantly related to the fruits and vegetables dietary pattern. Therefore, it is unlikely that residual confounding from these variables may explain the observed inverse association between this dietary pattern and pancreatic cancer risk.

As expected, the observed relationships between lifestyle and both the Western and drinker dietary patterns were in the expected direction. Higher Western pattern scores were associated with older age, elevated energy intake, heavy cigarette smoking, limited physical activity and lower educational attainment. Higher drinker pattern scores were related to heavy cigarette smoking.

The major strengths of our study include the large number of pancreatic cancer cases, histologic diagnosis, dietary pattern analysis and general population sampling. The large sample size permitted subgroup analyses by gender, age group and smoking status with reasonable statistical power. Histologic confirmation of diagnosis reduced the possibility of disease misclassification. Our dietary pattern analysis extended beyond nutrients and foods and was designed to account for complex interactions among dietary components and their cumulative effects. A population-based approach facilitates extrapolation of results to the general population.

Nevertheless, some limitations may be kept in mind when commenting on our results. Since the assessment of dietary exposure was retrospective, recall bias cannot be completely excluded. A prospective cohort approach has several advantages when studying associations between nutritional factors and cancer risk. However, because of the relatively low incidence of pancreatic cancer, the largest cohort to date examining diet and pancreatic cancer risk confirmed only 178 pancreatic cancer cases in 18 years of follow-up.21 We could not adjust OR estimates for the potentially confounding effect of diabetes mellitus and family history of pancreatic cancer since this information was not collected at baseline. Nevertheless, we expected that diabetes did not confound the association between dietary patterns and pancreatic cancer risk because of the likelihood that diet represents the initial risk factor for both chronic diseases, rather than lying on the causal pathway. In addition, it has been suggested that the significance of diabetes is much weaker if cases of recent onset are excluded.1 As well, it is unlikely that the confounding effect of family aggregation of pancreatic cancer may explain the significantly reduced risk we found because genetic/familial predisposition is relatively rare.6, 45 Although the FFQ was modeled after 2 questionnaires that have been tested in the past on several populations for validity and reproducibility, the NECSS has not validated this instrument with respect to dietary patterns in the Canadian population. As a consequence, it may induce a degree of measurement error and obscure some relationships.

Extensive use was made of proxy respondents in Ontario, with information from 48% of eligible cases collected through proxy interviews. In contrast, only 0.2% of controls were interviewed by proxy. Bias may have been introduced in the present investigation if rapidly fatal cases had a different etiology from cases with longer survival or if the dietary patterns under study influenced survival. However, our findings did not change materially after exclusion of proxy data, indicating that the relative risk associated with dietary patterns was similar for proxy and nonproxy respondents. As in other epidemiologic studies of pancreatic cancer, another limitation is the consequence of early and high case fatality associated with the disease. Cases who died before the questionnaire study were not included. However, since there was no discrimination based on their demographic characteristics and lifestyle factors, such as age, smoking or educational attainment, the survivors were still representative of the study population and bias of this kind, if any, would not be substantial.

Although we were able to generate 3 distinct dietary patterns that are easy to interpret in the study population, the percentage of variance in food consumption among individuals that is explained by these patterns was relatively low. However, this overall variance was somewhat similar to that observed in other studies31, 46 that used 3 dietary patterns to predict cancer and cardiovascular disease risk. In addition, factor analysis requires some arbitrary but important decisions, including the number of factors to retain, their labels and the method of rotation. We conducted sensitivity analysis by performing subgroup analyses by age group (<61 years, ≥61 years) and smoking status (ever/never smokers) and by including subjects with extreme intakes (<800, >4,500 kcal), and we tested oblique instead of orthogonal rotation. These analyses did not significantly alter our findings, indicating that the identified patterns were robust. Labeling of the 3 patterns was somewhat subjective but similar to that documented in other populations.32, 33, 46

In conclusion, our current data suggest that the fruits and vegetables dietary pattern reduces pancreatic cancer risk among male Canadians. A dose–response relationship was apparent. Further epidemiologic investigations assessing the overall role of diet in pancreatic cancer are warranted since the only way to cope with the rapid and inexorable lethality of this cancer is prevention.

Acknowledgements

The present study was supported by the National Cancer Institute of Canada, through a fellowship awarded to A.N., and by the Natural Sciences and Engineering Research Council of Canada, the Social Sciences and Engineering Research Council of Canada and the McLaughlin Foundation, through a research chair awarded to D.K. Health Canada supported the NECSS, in collaboration with the provincial cancer registries, through its program Health and the Environment. The Canadian Cancer Registries Epidemiology Research Group comprises a principal investigator from each of the provincial cancer registries: B. Paulse, Newfoundland Cancer Foundation; R. Dewar, Nova Scotia Cancer Registry; D. Dryer, Prince Edward Island Cancer Registry; N. Kreiger, Cancer Care Ontario; E. Kliewer, Cancer Care Manitoba; D. Robson, Saskatchewan Cancer Foundation; S. Fincham, Division of Epidemiology, Prevention and Screening, Alberta Cancer Board; and N. Le, British Columbia Cancer Agency.

APPENDIX I

Table I. Food Groupings Used for Analysis of Dietary Patterns, NECSS, Canada, 1994–1997

Food group or food	Food items included
Low-fat dairy products	Skim milk, 1% milk
High-fat dairy products	Whole milk, 2% milk, cheese, butter, ice cream
Eggs	Eggs
Spices and herbs	Pepper, salt, garlic and onions added to food
Margarine	Margarine on bread or vegetables
Mayonnaise	Mayonnaise, salad dressing
Oils	Oils used in cooking
Poultry products	Chicken, turkey
Soups	Soups with vegetables
Processed meats	Luncheon meats, sausage, bacon, hot dogs, smoked meat
Cold breakfast cereals	Bran or granola cereals, other cold cereals
Fruits	Apple, orange, banana, cantaloupe and other fruits
Fruit juices	Orange, citrus, grape, apple, other juice drinks or frozen juice
Cruciferous vegetables	Broccoli, cabbage, cauliflower, Brussels sprouts
Dark-yellow vegetables	Carrots, yellow (winter) squash
Tomatoes	Tomatoes, tomato juice, tomato sauce
Potatoes	Potatoes, sweet potatoes
Green and other vegetables	Spinach, other vegetables
Nuts	Nuts
Red meats	Beef, pork or lamb as main dish; hamburger; mixed beef dishes
Organ meats	Beef, pork, turkey and chicken liver
Fish	Fresh, frozen or canned fish
Processed fish	Smoked, salted or dried fish
Legumes and legume products	Tofu, soybeans, beans, lentils, peanut butter
Whole grains	Dark or whole-grain bread, cooked cereals
Refined grains	White bread or rolls
Rice, pasta	Rice, macaroni, spaghetti or noodles
Sweets and desserts	Chocolate, cookies, cake, doughnuts, pastry, pies
Fried potatoes	French fries, fried potatoes
Snacks	Potato chips
Coffee	Coffee
Tea	Tea
Soft drinks	Soda, soft and powdered drinks
Water	Tap water, bottled water
Beer	Beer
Wine	Wine
Liquor	Liquor
Supplements	Multiple vitamin or mineral supplements