Dietary patterns and lifestyle factors in the Norwegian EPIC cohort: The Norwegian Women and Cancer (NOWAC) study

Objective:

To identify different dietary patterns in Norway using a combination of cluster and factor analysis.

Design:

Cross-sectional study.

Setting:

Nation-wide, population-based study.

Subjects:

The Norwegian EPIC cohort is a subcohort of the Norwegian Women and Cancer study (NOWAC), and consist 37.226 women aged 41–56 y who answered a food frequency questionnaire (FFQ) in 1998.

Interventions:

The associations among 50 food variables were first investigated by using principal component analysis. Five important factors were found. The five principal components were then used as input in the cluster analysis. Different socioeconomic and lifestyle variables were examined.

Results:

Six clusters of dietary patterns were found, and were labelled accordingly: ‘traditional fish eaters’, ‘healthy eaters’, ‘average, less fish, less healthy’, ‘Western’, ‘traditional bread eaters’, and ‘alcohol users’. The traditional fish eaters and the traditional bread eaters were both highly represented in the north and west of Norway and were more likely to be present among persons with lower income and lower education. The healthy and the alcohol drinkers were found mostly in the south and east and were more likely to have higher income. Persons in the alcohol group were more likely to be current smokers. The western group had the highest percentage of three or more persons in the household and the shortest time since last birth, indicating that families with children dominate this group.

Conclusion:

Our data indicate six different dietary patterns in Norway, each with different socio-demographic and lifestyle characteristics.

Sponsorship:

The Norwegian Cancer Society (E 04038/006).

Traditionally, research on nutrition and health has focused on single nutrients or single food items, and not on the entire diet. Since we eat complex food, not single nutrients or food items, the focus has recently turned more towards dietary patterns (Hoffmann et al, 2000; Schulze et al, 2003). Analysing dietary patterns may be a useful tool to explain the complex relationship between diet and health.

Lately, many researchers have used cluster analysis to identify dietary patterns (Wirfalt & Jeffery, 1997; Fraser et al, 2000; Greenwood et al, 2000; Wirfalt et al, 2001); others prefer to use factor analysis (Fung et al, 2003; van Dam et al, 2003). The difference and the similarity between the two methods can be summarized in a few words as follows. Most clustering methods classify subjects into nonoverlapping groups on the basis of similarity in food intake. Factor analysis, on the other hand, summarizes food variables through factor scores. This can, in a current approach, be described as aggregating inter-related variables into possibly overlapping groups, one for each factor score. The two methods should be considered as complementing rather than opposing each other (Hoffmann et al, 2000). In fact, a combination of cluster and factor analysis may give a broader perspective and possibly more convincing pattern definitions than the use of one of the methods only. However, only few studies have tried to combine the two methods (Millen et al, 1996).

One main objective in Norwegian food policy is to reduce health damages related to the diet among the population (Landbruksdepartementet & Informasjonsenheten, 1997). To achieve this goal, it is important to understand what characterizes people's dietary patterns. The traditional Norwegian diet consists mainly of two or three similar cold meals, usually open sandwiches, and one hot meal with either fish or meat served with potatoes and vegetables. (Johansson et al, 1997; Amilien et al, 2000). Carrots are the most preferred vegetable. Fish is important in the traditional Norwegian diet and is often served with melted fat or fatty sauces. But the fish and potato consumption is decreasing in favour of meat and American-style pizzas and pasta dishes (Amilien et al, 2000).

The aim of this study was to identify Norwegian dietary patterns by clustering the subjects into distinct groups representing different dietary patterns. The idea was to do this by using factor analysis to reduce the number of variables as input to clustering procedures. We have also compared different lifestyle characteristics and socio-demographic factors among the patterns.

Study population

The Norwegian Women and Cancer (NOWAC) study is a large population-based cohort study designed to examine cancer-related factors. More than 100 000 Norwegian women are included in the cohort from 1991 to 1997 (Lund et al, 2003). In 1998, 37 226 women from the NOWAC study was included in the large multicentre study EPIC (European Prospective Investigation into Cancer and Nutrition) (Riboli et al, 2002). The Norwegian part of the EPIC study is a subsample of the NOWAC study with women who answered a mailed questionnaire on diet and different lifestyle and health-related issues in 1991/1992. The crude response rate in the first mailing was 57%. In 1998, women who had answered the first questionnaire and who agreed on being contacted again, received a second questionnaire with more detailed questions on diet. Crude response rate on this questionnaire was 82%. Only the women who answered the second questionnaire are part of the Norwegian EPIC cohort. The Norwegian EPIC cohort is described in more detail elsewhere (Brustad et al, 2004).

In the analysis of dietary patterns, we excluded some subjects due to a possible high degree of under- or over-reporting. We estimated the energy intake over energy requirement (calculated from age and bodyweight) and excluded subjects in the lower and upper 1% of this distribution to reduce the effect of extreme values (Ferrari et al, 2002). In addition, we excluded subjects who had answered less than half of the questions in the food frequency questionnaire (FFQ). Owing to the fact that most Norwegians only have one hot meal per day (Kjærnes et al, 2001), subjects reporting more than 60 hot meals per month were also excluded. After the exclusions, 35 554 subjects were included in the analysis.

Data collection

The data were collected by mailing an eight-page questionnaire about different lifestyle habits, which included four pages designed as a semiquantitative FFQ.

In the FFQ, subjects were asked to record how often, on average, they had consumed each food item during the previous year. There were 86 different food items, including alcohol. Four to six fixed frequency choices were given for each item. For fish, the women also had to indicate in what seasons they normally ate different fishes, due to seasonal variation. In addition, there were two variables about use of cod liver oil (liquid and capsules). The portion size per consumption was asked in natural or household units. A Norwegian weight and measurement table (Blaker et al, 1989) was used to calculate the weights in gram for each food item. Daily intake of energy and nutrients was computed based on the Norwegian Food Composition table (Rimestad et al, 2001).

Food items similar in type and/or nutrient content or culinary use were merged to reduce the number of variables and make the factor and cluster analyses easier. The Norwegian food tradition was kept in mind while doing this. The total number of variables ended on 50 and is shown in the Appendix.

The questionnaire also collected information about socioeconomic and lifestyle factors. Food choices are often associated with lifestyle and socioeconomic factors (Hjartåker & Lund, 1998; Johansson et al, 1999). In this paper, we utilized the following variables for the analysis: age, education, income, marital status, number of persons in the household, smoking, dietary supplement use (other than cod liver oil), body mass index (BMI), physical activity, and depression. Physical activity was recorded in the questionnaire on a scale ranging from 1 to 10 and assembled into three categories ‘low’ (1–3), ‘moderate’ (4–7), and ‘high’ (8–10). We also constructed new variables derived from the food information. In particular, we studied total energy intake and some nutrients (total fat, total protein, total carbohydrates, fibre, sugar and vitamin C and D) because these may be predictors of a healthy lifestyle.

Statistical methods

Since some of the variables used in the analysis do not contribute to energy intake (coffee and sugar-free soda), we decided not to adjust for energy. We believe these are important variables in different patterns and energy adjustment would favour these variables. On the other hand, the main analysis was based on the correlation matrix of the 50 variables rather than on the covariance matrix. This is not the same thing as energy intake correction but it does operate a kind of standardization.

The associations among the variables were investigated by using the simplest approach to factor analysis, Principal component analysis (PCA) (PROC FACTOR method=principal, SAS Institute, version 8.02). This choice was motivated by the fact that more sophisticated approaches to factor analysis did not appear to fit the data very well (results not reported here). In contrast to other approaches requiring a model with latent variables, PCA is not necessarily model based and is therefore of a more universal applicability. Several criteria are used in order to choose the number of factors, of which the most popular are the eigenvalue criterion: eigenvalue >1 (Hair, 1995) (the eigenvalues measure the amount of variance explained by each of the factors; the larger the eigenvalue, the more variance is explained by that factor) and the scree-plot (a plot of the eigenvalues against the number of factors with the eigenvalues ordered from the largest to the smallest). We relied on both of these as well as on the interpretability of the principal components. This led to the choice of five principal components, which were the factor scores, used as input to the cluster analysis (see the Results section for details).

In view of the large amount of data, we have used a two-step approach to clustering. The subjects were first clustered into 50 microclusters using k-means clustering (SAS PROC FASTCLUS) (Everitt et al, 2001). This procedure groups subjects represented by the five principal components into clusters based on Euclidian distance between observations. The 50 microclusters were then clustered using WARD's hierarchical clustering (Everitt et al, 2001). The choice of the appropriate number of clusters is a difficult problem. Although no universally valid solution is available, there are a few criteria that have been studied in depth and seem to perform reasonably well. We have chosen here the Calinski and Harabasz index which, according to Milligan and Cooper (1985), is one of the best in given situations. This index is computed from the between and pooled within sum of squares. The optimal number of clusters is the number that maximize this index. Six clusters appeared to be the optimal number of clusters using this criterion, and exploring the six clusters it seems to be possible to interpret six different dietary patterns. These will be described in the Results section.

To study the robustness of the clusters, we divided the data into two arbitrary parts (by assigning a computer-based random number to each subject, sort and divide into two equal sized parts) and following the same strategy to cluster the subjects, we found a similar interpretation choosing a six-cluster solution in the two random sets as in the total set.

As a validation of the six dietary patterns, we investigated the association of the clusters with age and socioeconomic variables. Differences in characteristics between the six clusters were examined using the χ²-test for categorical variables and analysis of variance and Tukey's multiple comparisons test for continuous variables.

Characteristics of the study population

A description of the study population is shown in Table 1.

The mean age of the women was 47.7 years (s.d.=4.3) with a 15-y age distribution. Mean BMI, calculated from self-reported weight and height, was 24.4 kg/m² (s.d.=4.1). Mean energy intake was 7.0 MJ. The women had 12.3 y of education on average.

Dietary patterns

In the PCA, the eigenvalue criterion indicates that there are 16 important factors but the scree-plot indicates that about nine factors should be included. Studying the factors we found that because of large variances in the variables describing liquid consumption, especially nonalcoholic liquids, most of the factors mainly describe associations among these variables. The first five factors, however, describe associations among solid food and alcohol-related variables, and we decided to keep only these factors in the following cluster analyses. Choosing five factors is equivalent to keeping all factors with eigenvalues greater than 1.5 (Handa & Kreiger, 2002; Markaki et al, 2003). These five factors explain 23.7% of the variance.

Considering PCA as a simple version of factor analysis, we can obtain an interpretation of the five factors by looking at the so-called factor loadings, reported in Table 2. In Table 2, we report only loadings greater than 0.30, and these could be used for interpretation. As noticed in the Introduction, this approach can be seen as merging variables into possibly overlapping groups, one for each factor. On the other hand, we have not used the factor scores as above, but have used the PCAs directly as input to the two-step clustering algorithm.

Six clusters of dietary patterns were found (Table 3). As a result of the large sample size, there were statistically significant differences (P<0.001) between the six clusters for all the characteristics listed in Tables 3, 4 and 5, even where the sizes of the differences were of little practical importance. Therefore, attention is focused on the estimates, rather than the P-values.

The first cluster was labelled ‘traditional fish eaters’ because consumption of all the fish items with the traditional garnish were dominant in this cluster. Cluster 2 was labelled ‘healthy’ because it reflected intake of food commonly thought to be healthy, such as skimmed milk, yoghurt, juice, cereals, rice, chicken, fruit, and cod liver oil. Cluster 3 was labelled ‘average, less fish, less healthy’. This was the largest group and there were no dominant foods/food groups. However, this cluster had low consumption of all fish items, cod liver oil, vegetables, juice, and course bread. Cluster 4 was labelled ‘Western’ because meat products, bakery products, deserts and chocolate, and more modern products like pizza, rice and pasta dominated. Cluster 5 was labelled ‘traditional bread eaters’ because this reflected the typical breakfast and lunch habits in Norway with milk, course bread, jam, cheese and fat on bread. Cluster 6 was labelled ‘alcohol users’ because beer, wine, and liquor were dominant here.

After identifying the clusters of dietary patterns, we looked at the distribution of the individuals making up each cluster in terms of their demographic and lifestyle characteristics (Table 4). The traditional fish eaters were highly represented in the north and west of Norway. They were also more likely to be present among persons with lower income and lower education, and there were also a higher percentage of current smokers (39.6%) than former (32.5%) or never smokers (28%) in the fish group. The mean age was highest in the fish group. The BMI was also highest in the fish group (mean BMI=25.1 kg/m²), and it remained highest also after adjusting for age or for age, activity level, and total energy intake (data not shown). Looking at the marital status, the percentage was highest in the fish group for married and cohabiting persons combined. When examining number of persons in the household and age since last birth, the fish group was dominated by two persons household and more than 16 y since last birth.

The healthy eaters were found mostly in the south and east and were more likely to have higher income. They also tended to have higher education. There were more former and never smokers than current smokers in the healthy group. The women in this group were also more likely to take dietary supplements than in the other groups. Among those who reported to take dietary supplements, persons in the healthy group were more likely to take it daily. The highest activity level was reported in this group. When combining married and cohabiting persons, the percentage was lowest in the healthy group. This group had the highest percentage of ‘others’ and the highest percentage of single household.

The persons in the average and the Western pattern were more evenly spread all over the country and more evenly distributed within the income and education groups. Concerning marital status, we found most married women in the Western group. The Western group represented the highest percentage of three or more persons in the household and those with the most recent child births (data not shown), indicating that families with children dominate this group. Highest energy intake was seen in the Western group (8.5 MJ) and lowest in the average group (5.7 MJ).

Like the traditional fish eaters, the traditional bread eaters were highly represented in the north and west of Norway, and also more likely to be present among persons with lower income and lower education. The bread group had the highest percentage of never smokers (40.5%). The lowest BMI was observed for the bread group (mean BMI=23.9 kg/m²).

The alcohol drinkers, like the healthy group, were found mostly in the south and east and were more likely to have higher income, but unlike the healthy group there were a lower percentage of subjects with high education in the alcohol group. The alcohol group had the highest percentage of current smokers (57.2%). There were a higher percentage of persons who reported being depressed (had visited a doctor because of it) in this group. This group had also the lowest reported activity level. Their marital status indicated the smallest percentage of married women in the alcohol group but the highest percentage of cohabiting persons. The alcohol users also had highest percentage of persons without children.

The distribution of fat, protein, carbohydrate, and intake of fibre, sugar and vitamin C and D are shown in Table 5.

We tried adjusting for age, activity level, and total energy intake for all the analysis performed on the final patterns, but this did not change the results markedly (data not shown).

The results of this study suggest that there are different food patterns in Norway and that the patterns in many ways reflect where in the country you live and to what socioeconomic group you belong. This is especially true for the two traditional food groups and for the health conscious group. The north and west of Norway represents the Norwegian coastline with easy access to fish and traditionally high fish consumption. The inner parts of Norway do not have the same access to fresh fish. As for the traditional bread eaters, it is possible that the districts are more likely to hold on to their traditions than people in big cities. The east and south of Norway is dominated by the capital Oslo, the only big city in Norway.

Some studies on fish consumption and health report that fish consumers on average have a healthier lifestyle than nonconsumers (Hu et al, 2002; Oomen et al, 2000). However, in our study we find that there are a higher percentage of overweight and obese persons in the fish group, and the average BMI is higher than for the other groups. This is in accordance with the findings of Jacobsen and Thelle (1987) and may be due to the traditional servings of a Norwegian fishmeal with lots of melted fat or fatty sauces. There were also quite a high percentage of current smokers in the fish group, indicating that fish consumers in our study do not have a healthier lifestyle than average. Not surprisingly, the fish group had the highest intake of vitamin D. Fish and fish products, especially fish liver, are good sources of vitamin D (Rimestad et al, 2001). The fish group had also high consumption of cod-liver oil (Table 3), another important source of vitamin D (Brustad et al, 2004), although use of cod-liver oil were more dominant in the healthy group. This is consistent with the findings of Brustad et al (2004) where consumption of fish and fish products were higher among cod-liver oil users than nonusers.

Many surveys have found that people with higher education are more health conscious (Fraser et al, 2000; Greenwood et al, 2000; Johansson et al, 1999). This is consistent with what we found in the healthy group. The women in the healthy pattern had in average higher education, higher physical activity level, and were more likely to take dietary supplements daily. This is in accordance with what Greenwood et al (2000) found for their healthy group. To support the findings that this is a healthy group, we examined some nutrients normally considered as predictors of a healthy diet and found that the healthy group had the highest percentage of energy from carbohydrates and the lowest from fat. They also had the highest intake of fibre and vitamin C, and second highest intake of vitamin D from the diet, supplements not included (Table 5). The high intake of vitamin D derives most probably from the use of cod-liver oil and relatively high proportion of fish and fish products (Table 3). Fruit and vegetables are important sources of both vitamin C and fibre (Rimestad et al, 2001), and were among the food groups that were dominating in this group (Table 3). Brustad et al (2004) found that cod-liver oil users had higher education, used more dietary supplements and consumed more fruit and vegetables and fish. They also found more daily users in Oslo than in the rest of the country. This correlates well with our findings for the healthy group.

The average group contained about 38% of the total study population but had the lowest estimated energy intake (5.7 MJ). Based on the low energy intake, this group could also be called a low food energy group or low energy reporter (LER) group. Other studies have reported higher BMI in low energy groups (Pryer et al, 1997, 2001; Fraser et al, 2000), but we did not find the same tendency in our LER group. However, we found a somewhat higher percentage of overweight and obese women in this group compared to the other groups, except for the fish group. Still, more than 60% of the women were within the range considered as normal weight (18.5–24.9 kg/m²), and mean BMI (24.5 kg/m², after adjusting for age, activity level, and energy intake) was almost as for the total cohort (24.4 kg/m²) and slightly lower than for the Western group (24.8 kg/m², after adjustment). There can be several reasons why we did not get the same results as reported in other studies. First of all, our data are based on a semiquantitative FFQ and does not cover the entire diet. It is possible that the LER's could not find the food they normally use in the questionnaire. Secondly, the weight and height is self-reported, which means the LER's could have underreported their weight. Spencer et al (2002) demonstrated previously that BMI calculated from self-reported height and weight was underestimated on average by 0.72 kg/m² in women and that heavier people underestimated to a greater extent than leaner people. And third, it is also possible that there is an overreporting in the other groups and that our LER group are the most precise. Some of the women reports up to 60 hot meals per month (approximately two hot meals per day), which is not usual for a traditional Norwegian diet (Kjærnes et al, 2001). However, we decided to set the limit for exclusion on 60 hot meals because many Norwegians are adapting to more continental habits, and today it is also possible to buy hot meals at many work cantinas. This can, however, be a source of overreporting.

The Western group contains most young women and had the highest percentage of more than three persons in the household, which indicates families with children. That may explain the higher consumption of pizza, pasta and meat products, and more sweets and snack products, indicating that the children's wishes and preferences are taken into consideration when preparing a meal. Sanchez-Villegas et al (2003) report younger persons with higher education and more sedentary lifestyle to belong in their Western pattern. The women in our pattern do also have higher education than average, but cannot be called more sedate than average. However, they report a somewhat lower activity level than both the healthy group and the fish group.

The traditional bread eaters had the lowest mean BMI, and the lowest percentage of overweight and obese persons. They had also the highest percentage of persons reporting moderately high activity level, and highest percentage of never smokers. The bread eaters also reported less use of dietary supplements. This may indicate that the bread group are health conscious. It has been a common health advice in Norway to eat more carbohydrates of the type you find in coarse bread. The health authorities have also stated that supplement use other than cod-liver oil is not necessary, because there are adequate amounts of vitamins and minerals in a traditional Norwegian diet. However, when studying Table 3, it does not seem that the bread eaters follow other health advice, like the ‘5-a-day’ advice for instance, where you are advised to eat together at least five fruits and vegetables each day. The consumption of fruit and vegetables was rather low, and they also consumed high fat and semiskimmed milk in favour of skimmed milk. They also had the lowest intake of vitamin C, and along with the Western group, they had the highest intake of added sugar. We have not found other studies that can support our findings on this group.

The alcohol pattern contained the highest percentage of current smokers. Both Slattery et al (1998) and Newby et al (2003) found highest percentage of current smokers among their alcohol patterns. As in our study, Slattery et al (1998) found that persons in the alcohol pattern were more likely to have higher income, but in contrast to what we found Slattery also found that they were more likely to have higher education.

A limitation in our study is that it had an increased response rate by increased education level (Lund et al, 2003). Other studies have found that higher education indicates a healthier lifestyle (Johansson et al, 1999; Fraser et al, 2000; Greenwood et al, 2000). However, in our study we found six different patterns representing different lifestyles. The percentage of persons belonging to each pattern may differ in the general population, but there is no reason to believe that the patterns should be markedly different from what we find in our study. Another limitation is that the study includes only women in a certain age group. Studies including both men and women indicate some differences between genders, but they also find the same patterns to a certain degree (Slattery et al, 1998; Mishra et al, 2002). Most of the women in our study were married or cohabits and it is likely to assume that the rest of their family, including their husbands, eat more or less the same food as the women. The main difference may be the food they eat away from home, which mainly will be the lunch meal (Kjærnes et al, 2001). However, older or younger age groups may have different dietary patterns than the ones found here. Therefore, it seems reasonable to generalize the findings only to the female population in the same age group (40–55 y) in Norway.

The strength of the study is the size of the study and the fact that it is based on a nationwide randomly selected study sample. The external validity of the NOWAC study have recently been examined by Lund et al (2003) with the conclusion that no major selection bias was found that could invalidate the calculation of population attributable risk.

The methods used in this study have been criticized because they may be affected by subjective analytic decisions and results are dependent on the individual sample and decisions about variable input format. Another limitation is lack of stability of dietary pattern and reproducibility difficulties (Jacques & Tucker, 2001). Hu et al (1999) examined reproducibility and validity of dietary patterns defined by factor analysis and they concluded with reasonable reproducibility and validity of the major dietary patterns. There are no data available on reproducibility and stability of dietary patterns derived from cluster analysis, but several studies have examined the validity (Tucker et al, 1992; Quatromoni et al, 2002). The patterns we find in this study are similar to patterns found in several other studies. The Western pattern (Slattery et al, 1998; Fung et al, 2001; Terry et al, 2001) and the healthy pattern (or ‘prudent pattern’) are reported by many others (Slattery et al, 1998; Fung et al, 2001; Pryer et al, 2001; Terry et al, 2001; Chen et al, 2002; Newby et al, 2003), the same is the case for the alcohol pattern (Slattery et al, 1998; Terry et al, 2001; Newby et al, 2003) and the average pattern (or LER pattern) (Fraser et al, 2000; Greenwood et al, 2000; Millen et al, 2002). Several studies also report traditional diet patterns (Greenwood et al, 2000; Pryer et al, 2001; van Dam et al, 2003), but these are naturally different from the ones we find in our study since traditional patterns are specific to each country.

What makes this study different from most other studies on dietary patterns is that we have combined the two methods, factor and cluster analysis. The factors used as input in the cluster algorithm explained 23.7% of the variance. This may be interpreted as saying that 23.7% is the only explainable variation; the rest is ‘noise’. If there are clusters they will likely be in the part of variation that is interpretable and not random. Indeed, the clusters do not and cannot explain all of the food choices of the Norwegian population, but only some of those that can be identified in the midst of the ‘noise’ and interpreted in terms of basic food groups selection. We were looking and finding some similarities of behaviour, some common patterns that do not exclude the fact that every Norwegian has his/her unique food choices (which is what we refer to as ‘noise’). Now these clusters do identify some common patterns. There may be subtler ones hiding in the axes we did not consider, but even if we found them, they would not be clearly supported by the data.

To summarize, our data indicate six different dietary patterns in Norway, each with different socio-demographic and lifestyle characteristics. Many studies have shown the association between certain dietary patterns and different diseases, and this will also be a natural next step to examine for EPIC Norway. Our findings may be of relevance to the development and targeting of Norwegian food politic messages in the future.

Affiliations

1. Institute of Community Medicine, University of Tromsø, Norway

   D Engeset, E Alsaker & E Lund

2. Department of Epidemiology & Biostatistics, McGill University, Montreal, Canada

   A Ciampi

Corresponding author

Correspondence to D Engeset.

Guarantor: D Engeset.

Contributors: DE wrote the paper and contributed to the statistical analysis. EA performed the statistical analysis and contributed to the writing of the paper. AC contributed to the writing of the paper and to the statistical analysis. EL was the principal investigator.

Supplementary Information accompanies the Paper on Euroepan Journal of Clinical Nutrition Website (http://www.nature.com/ejcn).

Reprints and Permissions