Introduction
Globally, colorectal cancer (CRC) accounts for 9.7% of all cancer cases and 8% of all cancer-related deaths (GLOBOCAN, 2008). The established risk factors for CRC include personal or family history of CRC or adenomatous polyps, chronic bowel inflammatory disease and presence of any of the hereditary syndromes (Cruz-Bustillo, 2004; Butterworth et al., 2006; Xie and Itzkowitz, 2008). There is extensive published literature reporting statistically significant associations between sporadic CRC and other risk factors. Several components of the diet have been linked with increased (red and processed meat intake) or decreased (dietary fibre, fruit and vegetables) CRC risk (Food, Nutrition, Physical Activity and the Prevention of Cancer, 2007). In addition, lifestyle habits like high energy intake, high body weight, low physical activity, smoking and high alcohol intake have been found to be associated with increased CRC risk (Satia-Abouta et al., 2003; Cho et al., 2004; Food, Nutrition, Physical Activity and the Prevention of Cancer, 2007; Moghaddam et al., 2007; Moskal et al., 2007; Botteri et al., 2008; Renehan et al., 2008). Intake of NSAIDs, including aspirin, has been shown to be associated with a reduced CRC and colorectal adenoma risk in case–control and cohort studies (Dube et al., 2007; Flossmann and Rothwell, 2007; Din et al., 2010). Finally, hormone replacement therapy in postmenopausal women has been linked to decreased CRC risk (Chan et al., 2006; Theodoratou et al., 2008).
The aim of this analysis was to investigate the relationship between CRC and the lifestyle and dietary risk factors that were measured in the large prospective Scottish CRC Study (SOCCS study). Univariate analysis was carried out initially, followed by analysis using forward and backward stepwise regression models (FSR and BSR); finally 1000 bootstrap samples were used to check model stability.
Materials and methods
Ethics statement
Ethical approval was obtained from the Multicentre Research Ethics Committee for Scotland and relevant local research ethics committees. All participants provided written informed consent.
Study population
The study population included 2062 patients and 2776 controls from Scotland. We aimed to recruit all patients aged between 16 and 79 years with adenocarcinoma of the colorectum, presenting to surgical units in Scotland between 1999 and 2006. Patients who died before ascertainment, those who were too ill to participate, those with recurrence or those unable to give informed consent because of learning difficulties or other medical conditions were excluded. We recruited about 40% of all incident cases in Scotland over the study period. During the same period, controls were selected from a population-based register and were invited to participate. All controls were free from colorectal cancer at the time of recruitment and were frequency matched for area of residence, sex and age. Participation rates among those approached were ∼52% for patients and an estimated 39% for controls. Questionnaire completion was sufficient for analysis in 68% of patients and 88% of controls recruited (less than 10 blank lines in the food frequency questionnaire and no blanks in the lifestyle and cancer questionnaire). Over 99% of the study participants were White (see Theodoratou et al., 2007 for further recruitment details).
Lifestyle and dietary data
Participants self-completed one questionnaire about their general lifestyle and one semiquantitative food frequency questionnaire (Scottish Collaborative Group FFQ, version 6.41; http://www.foodfrequency.org). The main characteristics of these questionnaires and data on FFQ validity and repeatability have been described previously (Masson et al., 2003; Theodoratou et al., 2007; Jia et al., 2008). Intake of dietary energy, macronutrients and micronutrients (including sugar, protein and total fat intake) was calculated using the UK National Nutrient Databank, based on ‘The Composition of Foods (5th edition) and related supplements’ by McCance and Widdowson. The variables that were included in the overall and stepwise regression analysis were 83 demographic risk factors, lifestyle variables, food variables and nutrients (Supplementary Tables 1a and 1b).
Statistical analysis
All food and nutrient variables were residually energy adjusted (except for the food groups tea and coffee and the nutrient flavones). The distribution of each variable between patients and controls was examined, and the distributions were tested for significance using the t-test, Wilcoxon’s rank-sum test (continuous variables) and Pearson’s χ2-test (categorical variables). In addition, a correlation analysis for all continuous variables was carried out using Spearman’s rank correlation test. Dietary and nondietary variables that were measured on a continuous scale were grouped into four categories using quartiles as the cutoff points (on the basis of the combined distribution of patients and controls). Univariate logistic regression models were fitted for each demographic, lifestyle, dietary, food and nutrient variable. Odds ratios (ORs) and 95% confidence intervals (CIs) were calculated for each quartile of continuous variables and each category of categorical variables. P-values for trend were calculated for the quartile and continuous forms of the quantitative explanatory variables.
The simplest data-driven model-building approach is the FSR. In this approach, variables are added to the model one at a time, and at each step each variable that is not already included in the model is tested for inclusion. The most significant of these variables is added to the model, as long as its P-value is below some preset significance level. Thus, the first variable to be included in the model is the one that is the most significant in the initial analysis. The procedure of adding variables continues until all the variables are added in the model or none of the remaining variables have a P-value below the preset level when added to the model (Katz, 2006). An alternative approach is BSR. Under this approach, all the variables of interest are fitted in the model and the least significant variable is dropped, as long as it is not significant at our chosen preset significance level. Reduced models are successively refitted and the same rule is applied until all remaining variables are statistically significant (Katz, 2006).
In both FSR and BSR the quartile form of the continuous variables was included. The P-value threshold for a variable to automatically enter the model (FSR) or to remain in the model (BSR) was 0.10. FSR and BSR were applied after tumour location (colon vs. rectal cancer) and sex stratification. To examine the stability of the resultant models the bootstrap method was applied. A bootstrap sample is a sample of the same size as the original sample (i.e. 2062 patients and 2776 controls in our case) chosen with replacement. Thus, a given participant of the original sample may occur in a specific bootstrap sample many times, only once, or not at all. A total of 1000 bootstrap samples were selected. Once a bootstrap sample was selected, FSR and BSR models were applied. The P-value threshold for a variable to enter the model or remain in the model was again 0.10. For each resultant model, the selected variables were noted together with their effect estimates and SEs of the estimates. Results across the variable selection models were compared and this procedure was repeated for all 1000 bootstrap samples. For each variable, the number of times that it was included in a regression model was calculated. In addition, the agreement between the type and number of variables included in the models after applying FSR and BSR was determined. Finally a summary of their ORs and SEs was calculated. STATA version 10.1 (StataCorp, College Station, Texas, USA) was used for statistical analyses. P-values are two-tailed. In this study, we performed 83 independent tests, and the adjusted P-value threshold using the Bonferroni correction method is equal to 0.0006.
Results
Descriptive analysis
Results from the descriptive analysis of all the explanatory variables are presented in Supplementary Tables 2 and 3. Spearman rank correlation coefficients (r) for all the continuous explanatory variables are presented in Supplementary Table 4. Finally, results of the univariate logistic regression models of each explanatory variable are presented in Supplementary Table 5.
Stepwise regression analysis
Application of FSR to the original sample resulted in a model including 20 risk factors (Table 1), whereas application of BSR resulted in a model including 22 risk factors (Table 2). The risk factors with the strongest associations that were selected to be included after application of both FSR and BSR were family history risk [OR (95% CI), P-value: 20.24 (13.59, 30.14), 1.6×10−49], intake of NSAIDs [0.72 (0.63, 0.83), 3.9×10−6] and dietary intake of fruit/vegetable juice [1.19 (1.11, 1.28), 3.7×10−7], high-energy snack foods [1.18 (1.11, 1.26), 4.3×10−7] and eggs [1.15 (1.08, 1.22), 1.1×10−5] (results presented here are from FSR; Table 1).
;)
Table 1: Forward stepwise regression built modela using the quartile form of the continuous variables
Table 2: Backward stepwise regression built modela using the quartile form of the continuous variables
FSR then applied to 1000 bootstrap samples resulted in 1000 unique regression models, and the same was observed after applying BSR. Over the 1000 bootstrap samples the mean percentage of agreement (SD) between the variables selected by forward and backward selection was 84.7% (7.5%). In addition, FSR and BSR resulted in models with 16–40 variables [mean (SD): 27.0 (3.7)] and 20–40 variables [mean (SD): 30.5 (3.5)], respectively (Figs 1 and 2). The mean of the included variables in BSR models was found to be significantly larger than that in FSR models (t-test P-value: <5×10−5). Finally, the mean, SD, minimum and maximum values of the effect estimates for the selected variables after FSR and BSR are presented in Table 3. Results when FSR and BSR were applied after tumour location (colon vs. rectal cancer) and sex stratification are presented in Supplementary Tables 6 and 7.
Fig. 1: Number of times each variable was selected by forward or backward stepwise regression. BSR, backward stepwise regression model; FH, family history; FSR, forward stepwise regression model; HESF, high-energy snack foods; LowCal, low-caloric soft drinks; MUFAS, monounsaturated fatty acids; n3PUFAS, n-3 polyunsaturated fatty acids; n6PUFAs, n-6 polyunsaturated fatty acids; PA, physical activity; ProcMeat, processed meat; SFAs, saturated fatty acids; SSBs, sugar-sweetened beverages; tFAs, trans-fatty acids; tMUFAs, trans-monounsaturated fatty acids.
Fig. 2: Number of variables in the final model for the 1000 bootstrap samples after applying forward or backward stepwise regression.
Table 3: Mean, SD, minimum and maximum values of the effect estimates for the selected variables after forward and backward stepwise regression
FSR and BSR were also applied for two additional sets of variables: (i) one that included 31 demographic risk factors, lifestyle variables and food variables (subset 1; Supplementary Table 8) and (ii) one that included 52 demographic risk factors, lifestyle variables and food variables (subset 2; Supplementary Table 8). Results are presented in Supplementary Tables 9–11 and Supplementary Figures 1–4. For subset 1, over the 1000 bootstrap samples, the mean and SD of the percentage of agreement between the variables selected by forward and backward selection was 97.8% (3.9%). In addition, FSR resulted in models of 9–23 variables [mean (SD): 16.8 (2.0)] and BSR resulted in models of 9–23 variables [mean (SD): 17.2 (2.0)], respectively (Supplementary Figures 1 and 2). The mean number of included variables in BSR models was statistically significantly larger than that in FSR models (t-test P-value: <5×10−5). For subset 2, over the 1000 bootstrap samples, the mean (SD) percentage of agreement between the variables selected by forward and backward selection was 83.7% (8.9%). In addition, FSR and BSR resulted in models with 10–33 variables [mean (SD): 19.1 (3.2)] and 13–31 variables [mean (SD): 21.9 (3.1)], respectively (Supplementary Figures 3 and 4). The mean of the number of included variables in BSR models was statistically significantly larger than that in FSR models (t-test P-value: <5×10–5).
Discussion
Strengths and limitations
The strengths of our study include the very large sample size and the use of a validated FFQ (Masson et al., 2003). General limitations of our study have been described previously (Theodoratou et al., 2007). Briefly, despite the careful design of this case–control study, many patients and controls refused to take part in the study and participation rates were 52 and 39%, respectively. One possible explanation might be that collection of biological specimens was required. Under-representation of patients who were very ill at the time of presentation might limit external validity of the results as well. In addition, lower participation rates among controls might be because of the recruitment procedure, as eligible controls were contacted only through mail by their general practitioners. In addition, the difference in participation rates between patients and controls might be because of the fact that patients are more eager than population controls to participate in a study that investigates their disease. Therefore in our case, the control participants who agreed to participate might have had a healthier diet and lifestyle and therefore were more eager to participate in a case–control study asking about their lifestyle choices and dietary habits. Given these low participation rates, participation bias cannot be ruled out, and therefore results should be interpreted with caution. Limitations of observational studies using FFQs include misclassification bias due to imprecise measures of dietary intake and residual confounding after attempts to control for confounders and of case–control studies include recall and selection bias. However, we attempted to limit these problems by careful adjustment, adoption of identical study procedures in patients and controls, use of an FFQ, which had been validated in both younger (Bolton-Smith et al., 1991; Masson et al., 2003) and older adults (Jia et al., 2008), use of images of portion sizes and careful instructions to improve accuracy of reporting diet and by adoption of a recall period 1 year before diagnosis or recruitment to reduce recall bias. Finally, questionnaire completion rates among patients were lower than those among controls, which is likely to be because patients were readmitted to the hospital or were otherwise too ill to fully cooperate in the study.
The drawbacks of FSR are that each addition of a new variable may render one or more of the already included variables nonsignificant or that one variable might be significantly associated with the outcome only when a group of other variables is also in the model. BSR also has drawbacks. For instance, variables that are dropped could have been significant if added to the final reduced model. In addition, BSR should not be used when the sample size is small, considering the number of independent variables that are included, or when there might be issues of multicollinearity. As in BSR all variables are included in the initial model, an unstable initial model (either because of the small sample size or multicollinearity) might produce spurious results. Finally, the main drawback of the bootstrap sampling methods is that at least 1000–10 000 bootstrap samples need to be produced to get reliable results and this can be both computationally and time restrictive.
Stepwise regression analysis (original sample)
After applying FSR and BSR to the original sample, the variables that were automatically selected to be included were family history, dietary energy, ‘high-energy snack foods’, eggs, fruit/vegetable juice, sugar-sweetened beverages (SSBs) and white fish (associated with an increased CRC risk) and NSAIDs, coffee and magnesium (associated with a decreased CRC risk); their selection was verified by bootstrap sampling. Most of the variables that were selected for inclusion in all models of the whole sample were also selected after tumour location or sex stratification (Supplementary Tables 6 and 7). However, coffee intake was selected only in the colon cancer models, and alcohol and trans-monounsaturated fatty acids (tMUFAs) were selected only in the rectal cancer models. Similarly, energy intake was only selected in male models and tMUFAs only in female models (see Supplementary Tables 6 and 7 for further details).
Stepwise regression analysis (bootstrap samples)
In the bootstrap analysis, the variables that were selected for the final models were highly dependent on the participants included in each bootstrap sample, as 1000 models were chosen once. Our findings also show that the agreement between FSR and BSR models in the same bootstrap sample decreased as the number of variables increased. The mean percentage of agreement for subset 1 (31 variables), subset 2 (52 variables) and the original set (83 variables) was 97.7, 83.7, 84.7%, respectively. In addition, the number of variables that was selected to be included in the models of the 1000 bootstrap samples was smaller for subset 1 analysis (9–23 variables) than for the subset 2 and the original set analyses (10–33 and 16–44 variables, respectively). Resultant models derived from using the set of variables that included only nutrients and not food groups (subset 2) were not as stable as those derived from using the set of variables that included only food groups (subset 1). One possible explanation might be that nutrients are usually highly correlated with each other. Therefore, multicollinearity issues, when attempting to fit highly correlated variables in the same model, might lead to unstable resultant models. Finally, more variables were selected by BSR than by FSR.
Comment on main findings of stepwise regression analysis
Demographic and lifestyle factors
Family history has been considered as one of the main risk factors for CRC and its effect on CRC is well established (Butterworth et al., 2006). Although the effect of dietary energy intake on CRC risk has been investigated in several observational studies, findings are generally inconsistent (Satia-Abouta et al., 2003). In particular, findings from case–control studies suggest that there is a positive and dose-dependent association between dietary energy intake and CRC risk, whereas findings from prospective studies do not support strong inverse associations, suggesting that the case–control findings might be biased (Stemmermann et al., 1985; West et al., 1989; Gaard et al., 1996; Giovannucci and Goldin, 1997; Kato et al., 1997; Slattery et al., 1997; Franceschi et al., 1998; Levi et al., 2002).
With respect to the observed association with intake of NSAIDs, two large randomized controlled trials (Gann et al., 1993; Cook et al., 2005) failed to show a protective benefit of low-dose aspirin on risk for CRC in men and women, perhaps because of either low doses or insufficient duration of the treatment, and results from a recent secondary analysis of data pooled from two other randomized trials (Peto et al., 1988; Farrell et al., 1991) support this argument (Flossmann and Rothwell, 2007). In addition, results from an 18-year follow-up study reported that regular use of aspirin over a long term reduces the risk for CRC among men; however, the beneficial effects of aspirin require at least 6 years of continuous and consistent use to establish (Chan et al., 2008). Finally, a systematic review of randomized controlled trials, case–control studies and cohort studies (Dube et al., 2007), as well as a meta-analysis of observational studies, including data from 19 case–control studies and 11 cohort studies (Flossmann and Rothwell, 2007), reported that regular use of aspirin or NSAIDs was consistently associated with a reduced risk for CRC, especially at high doses and after use for more than 10 years.
Food groups
Average household purchase of the food groups ‘high-energy snack foods’, soft drinks, fruit/vegetable juice, eggs, fish and coffee in the UK per person per week for the years 1977–2007 is presented in Supplementary Figure 5.
To our knowledge this is the first study to report an association between ‘high-energy snack foods’ and CRC. ‘High-energy snack foods’ was a summary variable for high-fat and high-sugar foods, including: pudding and deserts; chocolates, nuts and crisps; and biscuits and cakes. This summary variable represents an unhealthy dietary pattern and is correlated with dietary energy intake (r=0.61, P-value<10−5). Several observational studies that involve joint analyses of foods that are consumed together in clusters of individuals with similar dietary habits (cluster analysis; Wirfalt et al., 2008) have reported associations between CRC and food patterns. The two patterns that appear in the majority of the studies are: (i) a pattern of high intake of fruits, vegetables and other healthy food (‘healthy’ pattern) and (ii) a pattern of high intake of meat, high-fat food and high-sugar food (‘western’ pattern; Wirfalt et al., 2008). In most studies, the ‘healthy’ dietary pattern was found to be associated with a decreased CRC risk (Slattery et al., 1998; Terry et al., 2001; Harnack et al., 2002; Mizoue et al., 2005; Wirfalt et al., 2008), whereas the ‘western’ dietary pattern has been found to be associated with an increased risk for CRC (Slattery et al., 1998; Fung et al., 2003; Dixon et al., 2004). Results from the EPIC study in 10 European countries showed that ‘high-energy snack foods’ accounted for 15% (range 10–20%) of daily energy intake in women and 12% (range 7–19%) in men (Ocke et al., 2009). In our population, ‘high-energy snack foods’ accounted for 20% of the daily energy intake. The finding of a positive association between the intake of ‘high-energy snack foods’ and CRC is novel and remains significant after physical activity or BMI stratification (data not shown).
Similarly, the positive association between the intake of SSBs and CRC remained significant after stratification by levels of physical activity or BMI (data not shown). A recent study has reported that SSB consumption by adults in the USA has increased from the years 1988–1994 to the years 1999–2004 (Bleich et al., 2009). Epidemiologic evidence supports the hypothesis that SSB consumption is linked to obesity and type 2 diabetes (Bleich et al., 2009). In our population, SSBs accounted for 0.9% of the daily energy intake and their caloric index is ∼42 kcal/100 ml. The finding of the positive association between SSBs and CRC might be linked to the high caloric index of these drinks, similar to the association between ‘high-energy snack foods’ and CRC.
The finding of the positive association between fruit/vegetable juice (which included both fresh and ready-to-drink juice) and CRC would appear to be counter-intuitive at first as this would appear contrary to general health promotional messages on diet and cancer. This food group has rarely been investigated separately in studies on diet and CRC risk (Smith-Warner et al., 2002). The positive association was similar for low and high physical activity groups (data not shown). However, after BMI stratification, high intake of fruit/vegetable juice was associated with an increased CRC risk only in the high BMI group (P-value, 0.0001; data not shown). Generally, fruit and vegetable juices have different properties compared with the whole fruit or vegetable they come from, as the majority of them contain sugars, preservatives and other additives (Food, Nutrition, Physical Activity and the Prevention of Cancer, 2007). In addition, it has been shown that pure fruit juice raises blood sugar, and hence insulin, by a greater magnitude than a whole fruit eaten over the same time period, which could account for different metabolic effects (Haber et al., 1977). In our population fruit or vegetable juice intake accounted for 0.7% of the daily energy intake. The caloric index of these juices is ∼38 kcal/100 ml, which is very similar to that of SSBs (42 kcal/100 ml). Therefore, fruit juices can be considered as a class of high-energy drinks, similar to SSBs, and thus might increase CRC risk because of their high sugar or energy content. These reported increased risks associated with intake of ‘high-energy snack foods’ and high-energy drinks (both SSBs and fruit/vegetable juices) should be investigated further as their consumption has been reported to be increasing in industrialized countries (Bleich et al., 2009; Ocke et al., 2009; Supplementary Figure 5).
High consumption of eggs has been hypothesized to be associated with an increased risk for CRC, mainly because of their high fat and cholesterol content. However, the results from case–control and cohort studies have been inconsistent (Food, Nutrition and the Prevention of Cancer: a global prospective, 1997; Järvinen et al., 2001; Chiu et al., 2003; Marques-Vidal et al., 2006; Food, Nutrition, Physical Activity and the Prevention of Cancer, 2007).
The finding of the positive association between high intake of white fish and CRC was unexpected. The majority of observational studies have investigated the associations between total fish intake (white fish, oily fish and shellfish) and the findings are inconsistent (Engeset et al., 2007; Food, Nutrition, Physical Activity and the Prevention of Cancer, 2007; Geelen et al., 2007). A possible explanation for our findings is that 64.3% of white fish consumed was either fried, cooked in butter or smoked, whereas only 24.3% was grilled or poached. These preparations of white fish generally have a high content of fat (both saturated and trans fat), heterocyclic amines (formed during frying) or N-nitroso compounds (from smoking), all of which have been hypothesized to be positively associated with CRC (Knekt et al., 1999; Engeset et al., 2007).
Finally, coffee may be associated with a decreased CRC risk either because it contains particular anticarcinogenic substances, such as phenolic compounds, or because it increases the motility in the large bowel (Larsson et al., 2006). A recent review and meta-analysis of case–control studies has suggested that coffee may be inversely associated with CRC risk, with those who consume four or more cups per day having a 24% lower risk for CRC (Giovannucci, 1998; Tavani and La Vecchia, 2004). However, a recent meta-analysis of 12 prospective studies showed no significant effects of coffee consumption on CRC risk [relative risk (95% CI): 0.91 (0.81, 1.02)] (Je et al., 2009).
Nutrients
The main food sources of magnesium are green leafy vegetables, nuts, bread, fish, meat and dairy foods. Reported inverse associations between magnesium and CRC risk have been inconsistent (Larsson et al., 2005; Folsom and Hong, 2006; Van den Brandt et al., 2007; Ma et al., 2010), possibly because of different levels of magnesium intake among the different populations.
Conclusion
The variables that were included in the majority of regression models (from both original and bootstrap samples) were family history and intake of NSAIDs, high-energy snack foods, fruit/vegetable juice, eggs, dietary energy, white fish, coffee and magnesium. The positive association of ‘high-energy snack foods’ and high-energy drinks (SSBs and fruit juices) with CRC is novel and merits further investigation as such snacks and drinks are increasingly important contributors to diets in industrialized country settings. Bootstrap sampling indicated that the stability of FSR and BSR models is generally low, results should be treated with caution and interpretation of results should consider model stability using methods such as bootstrap sampling.
Acknowledgements
The authors thank Ruth Wilson, Rosa Bisset, Gisela Johnstone and all those who contributed to recruitment, data collection and data curation for the COGS and SOCCS studies.
Conflicts of interest
There are no conflicts of interest.
References
Bleich SN, Wang YC, Wang Y, Gortmaker SL.Increasing consumption of sugar-sweetened beverages among US adults: 1988–1994 to 1999–2004.Am J Clin Nutr2009;89:372–381.
Bolton-Smith C, Smith WC, Woodward M, Tunstall-Pedoe H.Nutrient intakes of different social-class groups: results from the Scottish Heart Health Study (SHHS).Br J Nutr1991;65:321–335.
Botteri E, Iodice S, Raimondi S, Maisonneuve P, Lowenfels AB.Cigarette smoking and adenomatous polyps: a meta-analysis.Gastroenterology2008;134:388–395.
Butterworth AS, Higgins JP, Pharoah P.Relative and absolute risk of colorectal cancer for individuals with a family history: a meta-analysis.Eur J Cancer2006;42:216–227.
Chan AT, Giovannucci EL, Meyerhardt JA, Schernhammer ES, Wu K, Fuchs CS.Aspirin dose and duration of use and risk of colorectal cancer in men.Gastroenterology2008;134:21–28.
Chan JA, Meyerhardt JA, Chan AT, Giovannucci EL, Colditz GA, Fuchs CS.Hormone replacement therapy and survival after colorectal cancer diagnosis.J Clin Oncol2006;24:5680–5686.
Chiu BC, Ji BT, Dai Q, Gridley G, McLaughlin JK, Gao YT, et al..Dietary factors and risk of colon cancer in Shanghai, China.Cancer Epidemiol Biomarkers Prev2003;12:201–208.
Cho E, Smith-Warner SA, Ritz J, van den Brandt PA, Colditz GA, Folsom AR, et al..Alcohol intake and colorectal cancer: a pooled analysis of 8 cohort studies.Ann Intern Med2004;140:603–613.
Cook NR, Lee IM, Gaziano JM, Gordon D, Ridker PM, Manson JE, et al..Low-dose aspirin in the primary prevention of cancer: the Women’s Health Study: a randomized controlled trial.JAMA2005;294:47–55.
Cruz-Bustillo CD.Molecular genetics of colorectal cancer.Rev Esp Enferm Dig2004;96:48–59.
Din FV, Theodoratou E, Farrington SM, Tenesa A, Barnetson RA, Cetnarskyj R, et al..Effect of aspirin and NSAIDs on risk and survival from colorectal cancer.Gut2010;59:1670–1679.
Dixon LB, Balder HF, Virtanen MJ, Rashidkhani B, Männistö S, Krogh V, et al..Dietary patterns associated with colon and rectal cancer: results from the Dietary Patterns and Cancer (DIETSCAN) Project.Am J Clin Nutr2004;80:1003–1011.
Dube C, Rostom A, Lewin G, Tsertsvadze A, Barrowman N, Code C, et al..The use of aspirin for primary prevention of colorectal cancer: a systematic review prepared for the U.S. Preventive Services Task Force.Ann Intern Med2007;146:365–375.
Engeset D, Andersen V, Hjartaker A, Lund E.Consumption of fish and risk of colon cancer in the Norwegian Women and Cancer (NOWAC) study.Br J Nutr2007;98:576–582.
Farrell B, Godwin J, Richards S, Warlow C.The United Kingdom transient ischaemic attack (UK-TIA) aspirin trial: final results.J Neurol Neurosurg Psychiatry1991;54:1044–1054.
Flossmann E, Rothwell PM.Effect of aspirin on long-term risk of colorectal cancer: consistent evidence from randomised and observational studies.Lancet2007;369:1603–1613.
Folsom AR, Hong CP.Magnesium intake and reduced risk of colon cancer in a prospective study of women.Am J Epidemiol2006;163:232–235.
1997American Institute for Cancer Research, World Cancer Research Fund.
2007American Institute for Cancer Research, World Cancer Research Fund.
Franceschi S, La Vecchia C, Russo A, Favero A, Negri E, Conti E, et al..Macronutrient intake and risk of colorectal cancer in Italy.Int J Cancer1998;76:321–324.
Fung T, Hu FB, Fuchs C, Giovannucci E, Hunter DJ, Stampfer MJ, et al..Major dietary patterns and the risk of colorectal cancer in women.Arch Intern Med2003;163:309–314.
Gaard M, Tretli S, Loken EB.Dietary factors and risk of colon cancer: a prospective study of 50,535 young Norwegian men and women.Eur J Cancer Prev1996;5:445–454.
Gann PH, Manson JE, Glynn RJ, Buring JE, Hennekens CH.Low-dose aspirin and incidence of colorectal tumors in a randomized trial.J Natl Cancer Inst1993;85:1220–1224.
Geelen A, Schouten JM, Kamphuis C, Stam BE, Burema J, Renkema JM, et al..Fish consumption, n-3 fatty acids, and colorectal cancer: a meta-analysis of prospective cohort studies.Am J Epidemiol2007;166:1116–1125.
Giovannucci E, Goldin B.The role of fat, fatty acids, and total energy intake in the etiology of human colon cancer.Am J Clin Nutr1997;666 Suppl1564S–1571S.
Giovannucci E.Meta-analysis of coffee consumption and risk of colorectal cancer.Am J Epidemiol1998;147:1043–1052.
2008Cancer Incidence and Mortality Worldwide in 2008. Available at:
http://globocan.iarc.fr/factsheets/cancers/colorectal.asp [Accessed 6 April 2010].
Haber GB, Heaton KW, Murphy D, Burroughs LF.Depletion and disruption of dietary fibre. Effects on satiety, plasma-glucose, and serum-insulin.Lancet1977;2:679–682.
Harnack L, Nicodemus K, Jacobs DR Jr, Folsom AR.An evaluation of the Dietary Guidelines for Americans in relation to cancer occurrence.Am J Clin Nutr2002;76:889–896.
Järvinen R, Knekt P, Hakulinen T, Rissanen H, Heliövaara M.Dietary fat, cholesterol and colorectal cancer in a prospective study.Br J Cancer2001;85:357–361.
Je Y, Liu W, Giovannucci E.Coffee consumption and risk of colorectal cancer: a systematic review and meta-analysis of prospective cohort studies.Int J Cancer2009;124:1662–1668.
Jia X, Craig LC, Aucott LS, Milne AC, McNeill G.Repeatability and validity of a food frequency questionnaire in free-living older people in relation to cognitive function.J Nutr Health Aging2008;12:735–741.
Kato I, Akhmedkhanov A, Koenig K, Toniolo PG, Shore RE, Riboli E, et al..Prospective study of diet and female colorectal cancer: the New York University Women’s Health Study.Nutr Cancer1997;28:276–281.
Katz MH.Multivariable analysis: a practical guide for clinicians2006:2nd edn..USA:Cambridge University Press.
Knekt P, Jarvinen R, Dich J, Hakulinen T.Risk of colorectal and other gastro-intestinal cancers after exposure to nitrate, nitrite and
N-nitroso compounds: a follow-up study.Int J Cancer1999;80:852–856.
Larsson SC, Bergkvist L, Wolk A.Magnesium intake in relation to risk of colorectal cancer in women.JAMA2005;293:86–89.
Larsson SC, Bergkvist L, Giovannucci E, Wolk A.Coffee consumption and incidence of colorectal cancer in two prospective cohort studies of Swedish women and men.Am J Epidemiol2006;163:638–644.
Levi F, Pasche C, Lucchini F, La Vecchia C.Macronutrients and colorectal cancer: a Swiss case–control study.Ann Oncol2002;13:369–373.
Ma E, Sasazuki S, Inoue M, Iwasaki M, Sawada N, Takachi R, et al..High dietary intake of magnesium may decrease risk of colorectal cancer in Japanese men.J Nutr2010;140:779–785.
Marques-Vidal P, Ravasco P, Ermelinda CM.Foodstuffs and colorectal cancer risk: a review.Clin Nutr2006;25:14–36.
Masson LF, McNeill G, Tomany JO, Simpson JA, Peace HS, Wei L, et al..Statistical approaches for assessing the relative validity of a food-frequency questionnaire: use of correlation coefficients and the kappa statistic.Public Health Nutr2003;6:313–321.
Mizoue T, Yamaji T, Tabata S, Yamaguchi K, Shimizu E, Mineshita M, et al..Dietary patterns and colorectal adenomas in Japanese men: the Self-Defense Forces Health Study.Am J Epidemiol2005;161:338–345.
Moghaddam AA, Woodward M, Huxley R.Obesity and risk of colorectal cancer: a meta-analysis of 31 studies with 70 000 events.Cancer Epidemiol Biomarkers Prev2007;16:2533–2547.
Moskal A, Norat T, Ferrari P, Riboli E.Alcohol intake and colorectal cancer risk: a dose–response meta-analysis of published cohort studies.Int J Cancer2007;120:664–671.
Ocke MC, Larranaga N, Grioni S, van den Berg SW, Ferrari P, Salvini S, et al..Energy intake and sources of energy intake in the European Prospective Investigation into Cancer and Nutrition.Eur J Clin Nutr2009;63Suppl 4S3–S5.
Peto R, Gray R, Collins R, Wheatley K, Hennekens C, Jamrozik K, et al..Randomised trial of prophylactic daily aspirin in British male doctors.Br Med J (Clin Res Ed)1988;296:313–316.
Renehan AG, Tyson M, Egger M, Heller RF, Zwahlen M.Body-mass index and incidence of cancer: a systematic review and meta-analysis of prospective observational studies.Lancet2008;371:569–578.
Satia-Abouta J, Galanko JA, Potter JD, Ammerman A, Martin CF, Sandler RS.Associations of total energy and macronutrients with colon cancer risk in African Americans and Whites: results from the North Carolina colon cancer study.Am J Epidemiol2003;158:951–962.
Slattery ML, Caan BJ, Potter JD, Berry TD, Coates A, Duncan D, Edwards SL.Dietary energy sources and colon cancer risk.Am J Epidemiol1997;145:199–210.
Slattery ML, Boucher KM, Caan BJ, Potter JD, Ma KN.Eating patterns and risk of colon cancer.Am J Epidemiol1998;148:4–16.
Smith-Warner SA, Elmer PJ, Fosdick L, Randall B, Bostick RM, Grandits G, et al..Fruits, vegetables, and adenomatous polyps: the Minnesota Cancer Prevention Research Unit case–control study.Am J Epidemiol2002;155:1104–1113.
Stemmermann GN, Nomura AM, Heilbrun LK.Cancer risk in relation to fat and energy intake among Hawaii Japanese: a prospective study.Princess Takamatsu Symp1985;16:265–274.
Tavani A, La Vecchia C.Coffee, decaffeinated coffee, tea and cancer of the colon and rectum: a review of epidemiological studies, 1990–2003.Cancer Causes Control2004;15:743–757.
Terry P, Hu FB, Hansen H, Wolk A.Prospective study of major dietary patterns and colorectal cancer risk in women.Am J Epidemiol2001;154:1143–1149.
Theodoratou E, Kyle J, Cetnarskyj R, Farrington SM, Tenesa A, Barnetson R, et al..Dietary flavonoids and the risk of colorectal cancer.Cancer Epidemiol Biomarkers Prev2007;16:684–693.
Theodoratou E, Campbell H, Tenesa A, McNeill G, Cetnarskyj R, Barnetson RA, et al..Modification of the associations between lifestyle, dietary factors and colorectal cancer risk by APC variants.Carcinogenesis2008;29:1774–1780.
Van den Brandt PA, Smits KM, Goldbohm RA, Weijenberg MP.Magnesium intake and colorectal cancer risk in the Netherlands Cohort Study.Br J Cancer2007;96:510–513.
West DW, Slattery ML, Robison LM, Schuman KL, Ford MH, Mahoney AW, et al..Dietary intake and colon cancer: sex- and anatomic site-specific associations.Am J Epidemiol1989;130:883–894.
Wirfalt E, Midthune D, Reedy J, Mitrou P, Flood A, Subar AF, et al..Associations between food patterns defined by cluster analysis and colorectal cancer incidence in the NIH-AARP diet and health study.Eur J Clin Nutr2008;63:707–717.
Xie J, Itzkowitz SH.Cancer in inflammatory bowel disease.World J Gastroenterol2008;14:378–389.
