Introduction

Colorectal cancer (CRC) is the fourth most common cancer and the third leading cause of cancer-related deaths both among male and female individuals worldwide in 2020. In recent years, its prevalence has been rising at an alarming rate globally particularly in low- and middle-income countries, with the highest incidence and mortality remaining in high-income countries1. In Iran, CRC is the fourth most common cancer and the fifth leading cause of cancer death, with a predicted incidence increase of 54.1% from 11,558 cases in 2016 to 17,812 new cases by 20252.

Besides genetics, several modifiable factors may affect CRC risk3, the most important ones including tobacco smoking, unhealthy diet, high alcohol consumption, physical inactivity, and excess body weight. As for diet, several previous epidemiological studies reported on the potentially protective (e.g., folic acid, calcium, vitamin D, vegetables, fruits, and fiber) or harmful effects (e.g., red and processed meat, trans-and saturated fatty acids) of macro- and micronutrients regarding their association with different health conditions, such as CRC cancer4.

While evaluation of individual nutrients, foods, or food groups provide relevant information to understand which dietary factors may increase or decrease cancer risk, it does not provide information of the overall diet. Measures of total diet quality have become more prevalent in recent years, as they can take into account undiscovered characteristics of foods and the interaction between all nutrients in different populations5. Among them, there are a priori and a posteriori dietary patterns, the former being based on scientific knowledge on the relationship between diet and disease and the latter deriving solely from the data. Statistical methods for the identification of a posteriori dietary patterns (DPs) allow to evaluate different dietary factors at the same time, even in the same regression model, but without issues related to multicollinearity that would instead be easy to encounter if dietary factors were considered as separate factors.

The authors of a recent systematic review concluded that two common patterns are associated with CRC risk6: 1) a healthy eating pattern that includes a large amount of fruit and vegetables, whole grains, nuts, and legumes, fish, seafood, and dairy products; 2) an unhealthy pattern, with high intakes of red meat, processed meat, sugar-sweetened beverages, refined grains, desserts, and potatoes, although maybe there are some specific differences between different populations. The World Cancer Research Fund indicated that the evidence for the relationship between DPs and CRC risk is limited and inconclusive due to the limited number of available studies, most of which emphasized DPs based on specific definitions and did not consider international population diversity3. Therefore, further research is needed to confirm these findings.

Additionally, previous meta-analyses reported mixed results regarding the association between dietary fats and CRC7,8,9, while no previous studies focused specifically on data-derived DPs high in fats, especially unhealthy ones. Consumption of trans fatty acids is particularly high in the Eastern Mediterranean Region, thus warranting more evidence on the topic to provide recommendations aimed at improving local population’s health10,11.

For these reasons, we aimed to investigate the association between different a posteriori DPs and CRC risk in an Iranian population (IROPICAN study) and to evaluate whether observed associations change according to CRC anatomical subsite and participants’ characteristics.

Results

Participants’ characteristics

Study participants’ demographic and lifestyle characteristics are reported in Table 1. Our study included 865 CRC cases (434 with colon cancer, 404 with rectal cancer, and 27 with missing information on cancer subsite) and 3,204 controls. Cases were older and a larger proportion of them were female, overweight or obese, had family history of CRC, or reported low SES compared with controls. In addition, waterpipe smoking was more common among cases, while the opposite was observed for aspirin intake (Table 1).

Table 1 Main study participants’ characteristics.
Full size table

Factor analysis

We retained three DPs from the principal component factor analysis, and their factor loading matrix is reported in Table 2. Overall, they explained 24.7% of the variance of intake of food groups. In particular, the first pattern was named “healthy diet”, having high factor loadings for fruit and vegetables, legumes, chicken meat, fish, egg, and condiments, such as olive and other vegetable oil. The second DP was named “Western-type diet” due to high loadings for nonalcoholic beverages and fruit juice, pizza, red and organ meat, cakes and desserts, dairy products, and snacks. The last one was named “other” and showed high contribution of beverages, especially coffee and tea, as well as sugar and candies, processed meat, salt, and hydrogenated fats.

Table 2 Factor loading matrix for dietary patterns identified through factor analysis.
Full size table

Participants’ characteristics and daily intakes for each food group by quartiles of DPs are reported in Supplementary Table 2 and 3, respectively.

ORs of CRC across quartiles of each identified DPs are presented in (Table 3). In particular, we found that both the healthy diet (OR Q4 VS Q1 = 1.48; 95% CI = 1.12, 1.97, ptrend = 0.001) and the Western-type diet (OR Q4 VS Q1 = 4.15; 95% CI = 3.15, 5.46, ptrend < 0.001) had a positive association with CRC, while the “other” DP was not associated with CRC. Results were robust to FDR correction for the Western-type diet (p-value Q4 VS Q1 < 0.001), but not for the healthy diet (p-value Q4 VS Q1 = 0.144). Results were similar when considering individual subsites of CRC (Table 3).

Table 3 Odds ratios (ORs) and corresponding 95% confidence intervals (CIs) for the association between dietary patterns identified from the factor analysis and colorectal cancer, overall and by cancer subsite.
Full size table

Results of the stratified analysis are reported in Supplementary Table 4. After FDR correction, the Western-type diet showed consistent associations with CRC among most investigated strata. In general, no substantial differences across strata of considered participants’ characteristics, including cigarette smoking, waterpipe smoking, and opium use, were observed for all DPs (Supplementary Table 4).

Reduced rank regression analysis

Factor loadings of food groups for the DP identified through reduced rank regression are reported in (Table 4). This DP high in unhealthy fats had animal products as the major contributors, with the highest factors loadings from eggs, dairy products, red meat, organ meat, margarine and hydrogenated fats, and butter. Instead, high negative loadings were found for fruit, pasta and rice, and bread.

Table 4 Factor loading matrix for the dietary pattern high in unhealthy fats identified through reduced rank regression.
Full size table

Detailed participants’ characteristics according to quartiles of this DP are reported in Supplementary Table 2. Intakes for each food group by quartiles of the DP high in unhealthy fats are described in Supplementary (Table 3).

Findings regarding the association between the DP high in unhealthy fats and CRC are presented in Table 5, showing higher odds of CRC among individuals in the highest quartile compared with the lowest (OR Q4 VS Q1 = 2.14; 95% CI = 1.70, 2.69, ptrend < 0.001). Results were robust to FDR correction, both for CRC overall (p-value Q4 VS Q1 < 0.001) and for individual subsites of CRC (Table 5). Similarly, we observed no differences across considered strata of participants’ characteristics (Supplementary Table 5), with p-values for heterogeneity of 0.303, 0.874, and 0.895 for opium use, cigarette smoking, and waterpipe smoking, respectively.

Table 5 Odds ratios (ORs) and corresponding 95% confidence intervals (CIs) for the association between a dietary pattern high in unhealthy fats identified using reduced rank regression and colorectal cancer, overall and by cancer subsite.
Full size table

The DP high in unhealthy fats was correlated with the Western-type diet identified through factor analysis (ρ = 0.30, p < 0.0001). A weak negative correlation was also observed with the healthy DP (ρ = -0.04, p = 0.035), while no correlation was found with the “other” DP (ρ = 0.00, p = 0.854).

Discussion

The analysis carried out in the IROPICAN case-control study and presented herein led to the identification of three different a posteriori DPs using factor analysis. The findings of the study showed that both the healthy diet and the Western-type diet were positively associated with CRC, although results were robust to FDR correction only for the Western-type diet, suggesting that observed associations for the healthy diet might be due to chance. Furthermore, the application of reduced rank regression allowed the identification of an DP high in unhealthy fats, weakly correlated with both the Western-type diet and the healthy diet, which was also associated with CRC. For all DPs, findings were consistent among investigated strata of study population’s characteristics.

As reported by a recent meta-analysis12, a few previous studies evaluated the association between a posteriori DPs and CRC. In particular, the meta-analytic estimates showed that a Western DP was positively associated with CRC, while the opposite was true for a prudent pattern entailing high consumption of fruit and vegetables and similar to the healthy diet that we identified in our study. Thus, while previous results can be considered in line with our findings for the Western-type diet, this does not completely hold true for the healthy diet, which instead was not associated with CRC after FDR correction in our analyses. It should be noted, however, that given that factor analysis allows the identification of patterns that are uncorrelated with each other13,14, according to our results reducing consumption of food groups from the Western pattern and replacing them with those from the healthy pattern would still be expected to reduce CRC risk, given the association with CRC of the former pattern but not of the latter. As for the Western-type diet, our estimates are in agreement with those from previous studies regarding the direction of the association12, although we observed a much stronger association with CRC. Even though this may be due to different sources of bias, as discussed in detail below, another possible explanation is heterogeneity in dietary habits in our study population. Indeed, associations might be stronger in Iran or other low- and middle-income countries, where CRC incidence rate is still increasing2, because there could be a greater range in diet, while in high-income countries, where the rates are high but stable1, heterogeneity of dietary habits among the population might be lower, hence showing weaker associations15. Indeed, individuals in the highest quartile of the Western-type diet in our study showed mean daily intakes of all food groups with high factor loadings for this DP at least ~ 3.5 (except dairy products, ~ 1.9) times higher than those in the lowest quartile, although with large dispersion in some cases.

While existing evidence from other world regions can be considered substantial, the number of studies on this issue from the Eastern Mediterranean Region is limited16,17,18,19,20, although they mostly reported results in agreement with those of the abovementioned meta-analysis of observational studies12. Some previous studies also applied reduced rank regression to derive DPs and evaluated their association with CRC21, although none of them, to our knowledge, adopted dietary fat-related variables as response variables to derive DPs or were conducted in the Eastern Mediterranean Region. In this regard, high dietary intake of trans fatty acids has been reported to be consistently associated with CRC in recent meta-analyses7,8, which is in line with our findings, while no association for total fat and saturated fatty acids was observed9, warranting further research on this topic. In this context, it should be noted that our findings are particularly relevant for population’s health in the Eastern Mediterranean Region, since it has been reported by the World Health Organization to have one of the highest levels of trans fatty acids in its food supply compared with other world regions10,11. Previous reports suggested that the average per-person intake of trans fats in Iran be higher than 10 g/day22 and much higher than that of Western countries23. Also, although the content of trans fats in various food products has declined over recent years in Iran, it is still relevant24. However, the observed association between the DP high in unhealthy fats and CRC may be not only due to the dietary intake of unhealthy fats, but also to animal proteins, given the substantial contributions of animal products to this identified DP. Indeed, animal proteins favor insulin resistance and hyperinsulinemia25, which in turn have been shown to be associated with CRC26.

To the best of our knowledge, this is the largest study investigating the association between a posteriori DPs derived using factor analysis and CRC carried out so far in the Eastern Mediterranean Region, as well as the first study to adopt reduced rank regression to identify a DP high in unhealthy fats and to assess whether it is associated with CRC. In addition, we were able to adjust the analyses for main known risk factors of CRC and the large numbers of cases and controls allowed us to evaluate whether associations were constant across strata of participants’ characteristics and for different cancer subsites. Nevertheless, our analyses did not take into account participants’ physical activity due to lack of related data, we may bias estimates away from the null. Also, our study may suffer from selection bias, a type of bias typically affecting case-control studies, especially hospital-based ones in which the distribution of risk factors might not accurately reflect that of the base population where cases are drawn from27,28. Nevertheless, both cases and controls were enrolled at the same university hospitals and matched for residential city to mitigate the referral pattern and selection bias. Also, the high participation rate in our study suggests that potential selection bias related to low response rates sometimes observed in observational studies be limited. Dietary information was self-reported through FFQs29,30, which might lead to recall bias and thus affect the results of our study, although misclassification of exposure has been reported to be non-differential in this context31. Similarly, recall bias might also affect other information regarding lifestyle factors, which were included as adjustment factors in the analyses. Additionally, the methodology used for data collection did not allow to evaluate changes of dietary habits during the lifespan, possibly contributing to non-differential misclassification of exposure mentioned above. Also, DPs from the factor analysis explained a low proportion of variance of food groups, which might limit their validity.

In summary, the results of our study suggest that both a Western-type diet and a DP high in unhealthy fats were associated with CRC, suggesting the need for a decrease of the consumption of unhealthy foods, including those high in trans and saturated fatty acids, in order to improve population’s health. Also, the observed positive association between a healthy diet high in fruit and vegetables and CRC warrants further investigation.

Methods

Study design and participants

IROPICAN is a multicenter hospital-based case-control study carried out in ten provinces in Iran, between May 2017 and July 2020, with the main objective of investigating the association between opium consumption and risk of lung, bladder, head and neck, and colorectal cancer32. Seven provinces including Tehran, Fars, Mazandaran, Kerman, Golestan, Kermanshah, Khorasan-Razavi were involved in the CRC study. Individuals aged 30–75 years old, who had resided in the study regions for at least two years, were of Iranian nationality, were able to speak and understand Farsi language, and were able to take part in the interview (duration of ~ 80 min) were eligible to participate. Cases were 906 individuals with histologically confirmed incident primary CRC, including all histological types of CRC except melanoma and sarcoma, admitted to university hospitals in the study area and diagnosed with CRC within one year before enrolling in the study. They were categorized according to the International Classification of Diseases (ICD-O-3 codes) as cases of cancer of the colon (C18), proximal colon (from cecum to the splenic flexure, ICD-O codes: C18.0 through C18.5), distal colon (from descending colon to sigmoid colon, ICD-O codes: C18.6 through C18.7), and rectum (from the recto-sigmoid junction (C19) down to the rectum). Cancers involving two different subsites of the colon (e.g., proximal and distal colon) were classified as “overlapping lesions”. Controls were 3,243 healthy and cancer-free individuals enrolled concurrently and at the same hospitals as the cases. Controls were enrolled among relatives or friends of patients in non-oncology wards or among healthy visitors who were at the hospital for reasons other than receiving treatment. For each case of the four cancer types included in the study, one control, was enrolled in the study, frequency-matched by sex, five-year age group, and place of residence with the corresponding cancer case. In the current analysis we included controls recruited for all cancer sites in the original IROPICAN study, and the matches were broken for the purpose of the analysis. Overall, 1% and 11% of initially contacted cases and controls refused to participate, and the main reasons for non-participation were inability to take part in the interview due to sickness and lethargy among cancer patients and reported lack of time or unwillingness to donate biological samples among controls. In addition, individuals with no histological confirmation of diagnosis (n = 18), those with no information regarding their case-control status (n = 2), and those who did not complete the food-frequency questionnaire (FFQ) or who were in the highest and lowest 1% of the ratio between total energy intake and estimated energy requirement (n = 60) were excluded (Supplementary Fig. 1).

The study was approved by the Ethics Committee of the National Institute for Medical Research Development (NIMAD) (Code: IR.NIMAD.REC.1394.027). All participants provided written informed consent. The study was conducted in accordance with the Declaration of Helsinki.

Data collection

Participants were interviewed using a structured questionnaire by trained interviewers in order to collect information on sociodemographic characteristics and lifestyle factors, such as age, sex, education, cigarette and waterpipe smoking, opium use, alcohol drinking, socioeconomic status (SES), previous illness, family history of cancer, and use of nonsteroidal anti-inflammatory drugs. The same team of trained interviewers generally measured the standing height of participants (cm) at enrollment. In addition, cases were asked to report their body weight before cancer diagnosis, while controls were measured their body weight at the time of the interview.

Information on diet during the year before diagnosis (or interview for controls) was collected by trained interviewers using the Persian Cohort food frequency questionnaire (FFQ), a validated semi-quantitative FFQ including 131 items (113 food items and 17 supplements)29,30. Daily intakes of foods in grams were obtained by converting the reported frequency of intake (daily, weekly, monthly, or yearly) to frequency per day and by multiplying it by the standard portion size (grams) using household measures. Thus, total energy intake and single nutrients intakes were computed using the food composition database developed for the Iranian population, which was based on the US Department of Agriculture (USDA)33, Near-East34 and Bahrain35 food composition tables. Food items from the FFQ were categorized into the following 29 food groups based on their similarities of both culinary practices and nutritional profiles: processed meat, organ meat, total fish, margarine and hydrogenated fats, dairy products, coffee and tea, fruit, fruit juice, vegetables, bread, pasta and rice, cereals, snacks, nuts, nonalcoholic beverages, tomatoes, legumes, cakes and desserts, sugar and candies, olives and olive oil, other oils and condiments/dressings, pickles, red meat, chicken meat, eggs, butter, potatoes, pizza, salt (Supplementary Table 1).

Dietary pattern analysis

The DP analysis was carried out on the 29 different food groups (g/day) described above.

The first part of the analysis is represented by a factor analysis used to derive a posteriori DPs. We assessed the factorability of the correlation matrix by visually inspecting it and by using statistical procedures. Both Bartlett’s test of sphericity (p < 0.001) and Kaiser–Meyer–Olkin statistic (KMO = 0.768) showed satisfactory results and suggested that factor analysis was appropriate with our data. Thus, we applied a principal component factor analysis to derive a posteriori DPs among controls13. The number of factors to retain was chosen based on the following criteria: factor eigenvalue > 1, scree-plot visual inspection (Supplementary Fig. 2), and interpretability of factors. We applied a varimax rotation to the factor loading matrix to derive uncorrelated factors and to improve interpretability14. Derived DPs were interpreted according to food groups with factor loadings ≥ 0.3 in absolute value. Hence, we computed factor scores, which are aimed at representing the degree of adherence of each study participant’s diet to each identified DP, among both cases and controls and for each DP using Bartlett’s method36,37. We also repeated the analysis by including both cases and controls, but since results did not substantially change, we report here only those deriving from the factor analysis applied to controls.

Subsequently, we adopted reduced rank regression38 to identify a DP high in fats (excluding healthy unsaturated fats). Thus, we used the abovementioned 29 food groups as predictor variables and the natural logarithm of energy-adjusted (using the residual method) daily intakes of trans fatty acids, saturated fatty acids, and cholesterol from natural sources as response variables in the analysis. We retained only the first identified factor, which is the one explaining the largest part of the variance of response variables, and we thus derived its related factor score. To check robustness of the derived DP, we repeated the analysis on each half of the original study population derived from splitting it randomly, which confirmed results from the main analysis.

Statistical analysis

For each DP, identified using either factor analysis or reduced rank regression, factor scores were categorized into quartiles according to their distribution among controls. Thus, we estimated odds ratios (ORs) and corresponding 95% confidence intervals (CIs) for the association between quartiles of each DP and CRC, overall and by anatomical subsite, by including all DPs at the same time in the unconditional logistic regression model (for DPs from the factor analysis only), which was also adjusted for province of residence (Tehran, Fars, Kerman, Golestan, Mazandaran, Kermanshah, Khorasan-Razavi), age (< 30, 30–39, 40–49, 50–59, 60–69, ≥ 70), sex (male, female), opium use (never user, non-regular user, regular user), cigarette smoking (never, ever), regular waterpipe smoking (no, yes), SES (low, intermediate, high), family history of CRC, aspirin intake, body mass index (BMI) (underweight [BMI < 18.5 kg/m2], normal weight [18.5 kg/m2 ≤ BMI < 25 kg/m2], overweight [25 kg/m2 ≤ BMI < 30 kg/m2], obese [BMI ≥ 30 kg/m2]), total energy intake (quartiles according to the distribution among controls). Alcohol consumption was not taken into account in the analysis because the proportion of drinkers among the study population was negligible. SES was calculated for each participant, using principal component analysis, based on the number of years of education and ownership of any assets like vacuum cleaners, clothes washers, dishwashers, freezers, internet access, microwaves, laptops, mobile phones, cars, and shops39. Additionally, we tested for linear trends by replacing quartiles of DPs with quartile-specific median values of factor scores among controls, which were thus included as continuous variables in the regression models. Eventually, we conducted stratified analyses by age, sex, opium use, cigarette smoking, regular waterpipe smoking, and BMI, and evaluated heterogeneity between strata using likelihood ratio tests comparing the models with or without the interaction term between the stratification variable and the DP of interest.

Furthermore, we also took into account the large number of tests carried out by estimating False Discovery Rate (FDR)-corrected p-values according to Benjamini-Yekutieli procedure40,41.

As a secondary analysis, correlation between factor scores of the DPs retained from the factor analysis and the DP identified using reduced rank regression was evaluated by computing Spearman’s rank correlation coefficients.

Statistical analyses were carried out using Stata software, version 17 (StataCorp LP. College Station. TX).