Relationships between diet and health have already been strongly established in the literature( Reference Willett and Stampfer 1 , Reference Martinez-Gonzalez and Bes-Rastrollo 2 ). Accordingly, the consumption of some foods (e.g. red meat, fruits, fish, alcohol, etc.) or some nutrients (e.g. SFA, Na, K, etc.) is generally associated with an increased or decreased risk of many chronic diseases such as obesity( Reference de la Fuente-Arrillaga, Martinez-Gonzalez and Zazpe 3 – 5 ), hypertension( Reference Reddy and Katan 6 – Reference McCartney, Byrne and Turner 8 ), CVD( Reference Reddy and Katan 6 , Reference de Oliveira Otto, Mozaffarian and Kromhout 9 – Reference Mente, de Koning and Shannon 11 ) or certain cancers( Reference Liu and Ma 12 – Reference Glade 14 ).
The standard approach for exploring these risk–benefit relationships has been to focus on one specific food or one nutrient without considering the diet as a whole( Reference Jacobs and Tapsell 15 ). However, it is necessary to consider the potential interactions or synergistic effects between foods or nutrients in order to depict the overall effect of diet on health( Reference Jacobs and Tapsell 15 – Reference Jacobs and Steffen 17 ). Multidimensional approaches, such as the identification of dietary patterns, thus allow the estimation of more reliable associations between diet and health, taking into account the overall diet and its complexity( Reference Hu 18 – Reference Moeller, Reedy and Millen 20 ). Moreover, the nutritional and socio-economic characterisation of dietary patterns can be used to define practical public health policies to promote better food behaviours in specific groups of consumers( 21 , Reference Tucker 22 ). From an environmental health perspective, dietary patterns can finally contribute to identify the most exposed consumers to a series of food contaminants( Reference Béchaux, Zetlaoui and Tressou 23 ).
In recent years, there has been increasing interest in studying national diets using a multidimensional approach( Reference Moeller, Reedy and Millen 20 , Reference Kant 24 ). The standard approaches applied were principal factor analyses, such as principal component analysis (PCA), multi-component analysis or cluster analysis (e.g. hierarchical cluster analysis (HCA))( Reference Hu 18 , Reference Schulze and Hoffmann 25 , Reference Wirfält, Drake and Wallström 26 ). In France, only a few studies have investigated dietary patterns at a national level using multifactorial methods( Reference Mahe, Francou and Colin 27 – Reference Escalon, Bossard and Beck 30 ), and only one, to our knowledge, from a representative sample of the French population( Reference Bertin, Touvier and Dubuisson 29 ). Although these studies have provided comparable results, the lack of homogeneity of the methods performed and differences in the study population make comparisons difficult. Besides, standard approaches are not really suitable for constructing dietary patterns because of the inherent structure of the data. For instance, food consumption data include a significant number of zeros because of non-consumption of certain categories of foods and only have positive values. Hence, the assumption of a Gaussian distribution may not be valid( Reference Kipnis, Midthune and Buckman 31 ). Moreover, the approaches usually used show poor fit because of non-negative data and the excess of zero values, generally termed ‘sparse data’( Reference Shlens 32 ). Lee & Seung( Reference Lee and Seung 33 ) proposed a new latent-variable-based method, the negative matrix factorisation (NMF) method, specifically adapted to sparse and non-negative data. This method has already been proven to be effective in food risk assessment to identify dietary patterns or chemical mixtures( Reference Béchaux, Zetlaoui and Tressou 23 , Reference Zetlaoui, Feinberg and Verger 34 , Reference Sy, Feinberg and Verger 35 ).
The major aim of this study was thus to identify the main dietary patterns in the French adult population using the NMF approach and the food consumption data of a nationally representative survey (the second French national cross-sectional dietary survey (INCA2)). Next, we determined their main demographic and socio-economic characteristics and assessed their nutritional and environmental profiles in order to highlight their specific features. The dietary patterns revealed in this work will thus give an overview of the different food consumption behaviours in the French adult population, according to distinct dimensions.
Methods
Study population
The French INCA2 survey was carried out between December 2005 and May 2007 by the French Food Safety Agency( 36 ). This cross-sectional survey was initially designed to assess food intake in a nationally representative sample of the French population. Two independent random samples of 3- to 17-year-old children and 18- to 79-year-old adults were drawn using a multistage cluster sampling technique. The complex sampling frame was established from the national census, published by the French National Institute of Statistics and Economic Studies (INSEE), and it has been described elsewhere( Reference Dubuisson, Lioret and Touvier 37 , Reference Lioret, Touvier and Dubuisson 38 ). In brief, 181 geographical units, stratified by region of residence and size of urban area, were first randomly selected with a probability proportional to size. Then, households were randomly drawn within each primary sampling unit, and two independent sampling frames were set up: one restricted to households including at least one child and the other including households with or without children. Last, within each household, either a child or an adult was randomly selected. Participation rates were 63 % for adults and 69 % for children, yielding samples of 2624 adults and 1455 children, respectively. To ensure the national representativeness of each sample, a weighting factor for unequal sampling probabilities for differential non-responses by region, agglomeration size, age, sex, occupation of the household head, size of the household and season has been assigned to each individual. These variables were selected for adjustment because of high discrepancy between their distribution among the INCA2 sample and among the French population, using an external source (Labour force survey 2005-INSEE)( 36 , 39 ) (distribution among the adult sample is presented in the online Supplementary Table S1). The low variability of the weighting factor for adults (mean of 1 and a sd of 0·7) demonstrated the good representativeness of the INCA2 adult sample compared with the French general adult population.
Only adults were considered in this study. As recommended by the European Food Safety Authority, under-reporting subjects (i.e. those who, voluntarily or not, under-reported amounts consumed; 26·9 % of adult sample) were identified and included in the statistical analyses(
40
). Besides, twenty-four subjects (0·9 % of the adult sample) with an extremely low total energy intake (TEI) were excluded from the final sample (estimated from the following formula:
$${\rm log}\,{\rm (TEI)\,\lt\,mean} {\left( {{\rm log}\,{\rm (TEI}} \right)\,{\minus}\,{\rm 3\, \scale 70%{SD}}\,\left( {{\rm log}\,{\rm (TEI)}} \right)\right)}$$
(
Reference Lioret, Dubuisson and Touvier
41
).
The INCA2 survey was approved by the French Data Protection Authority (Commission Nationale de l’Informatique et des Libertés) and the French National Council for Statistical Information (Conseil National de l’Information Statistique).
Data
Collection of data on food consumption
Dietary intake was assessed using a 7-d food record. A trained and certified investigator delivered at home the food record with a self-administered questionnaire and explained to the subjects how to complete them. The investigator returned to the home immediately after the week to check the accuracy of the information reported in both documents. Each day of the food record was divided into three main meals (breakfast, lunch and dinner) and three between-meal snacks. The subjects were asked to describe as precisely as possible the nature and the amount of all foods and beverages consumed during the survey week. Consumed quantities were estimated using the SU.VI.MAX (SUpplémentation en VItamines et en MInéraux AntioXydants) photographic booklet( Reference Hercberg, Deheeger and Preziosi 42 ) or expressed directly in weight or in household measures (e.g. spoon).
Foods and beverages declared were subsequently allocated a food code including 1280 food items and were categorised into forty-three food groups and 121 subgroups. McCann et al.( Reference McCann, Marshall and Brasure 43 ) and preliminary analyses (data not shown) showed that the quality of the description of dietary patterns is strongly affected by the level of aggregation of foods. To obtain a satisfactory trade-off between the level of detail to discriminate individuals according to their food consumption and the difficulty in exploring a large data set by factorial analysis, the nomenclature was modified step-by-step for this study and the 1280 food items were finally reclassified into seventy-four new food groups (Table 1). This classification was based on the foods’ nutritional composition and results of previous analyses (data not shown). Eight food groups (i.e. wholegrain pasta/rice/wheat, whole milk, skimmed milk, sweetened milk, low-fat cheese, dried fruit, nectar, soft drinks with fruit) with a consumption rate <10 % were excluded to avoid excessive noise in the data, which could lead to underline too particular and isolated dietary behaviours( Reference Bailey, Gutschall and Mitchell 44 , Reference Grieger, Scott and Cobiac 45 ).
Table 1 Nomenclature (food groups and consumption rate among the 2600 individuals)
Individual characteristics
Individual demographic and socio-economic variables were collected using face-to-face questionnaires and self-reported data. Questionnaires provided information on individual occupational status, education level and household wealth. Household wealth was defined through questions on the household income and other related variables such as ‘having gone away on holiday for more than 4 d within the last 12 months’, ‘the number of cars in the household’, ‘the number of domestic electrical appliances’, ‘how the financial situation is perceived’, ‘financial access to desired food products’, ‘whether the idea of lacking food would be a concern’, ‘giving up health care for financial reasons’ and ‘housing occupancy status’. A wealth index was derived from a correspondence analysis as already done by Fillol et al.( Reference Fillol, Dubuisson and Lafay 46 ) on variables describing household wealth (cited above). From the correspondence analysis, the score of each subject on the first principal component was used as the summary wealth index, which was divided into tertiles. In addition, for this study and according to Darmon et al.( Reference Darmon, Bocquier and Vieux 47 ), an individual was considered as living in a household experiencing food insecurity for financial reasons if she/he declared not having enough to eat (often or sometimes) because of economic reasons. Respondents were also asked to report other information such as age, sex, household composition, region and size of municipality in which the household was located. The variables and associated categories are described in the supporting information (online Supplementary Table S2).
Nutritional composition data
Nutritional intake was estimated by matching the French Food Composition database for the year 2008( Reference Feinberg, Favier and Laussucq 48 , 49 ) to the individual food consumption data. The individual average daily intake of macronutrients (i.e. total energy content, total carbohydrates, simple carbohydrates, total fats, SFA, proteins, alcohol, fibres and salt), minerals (i.e. Ca, Fe, Na, Mg, K) and vitamins (i.e. vitamins A, C, E, B1, B6, B9) was thus determined.
Food contamination data
Food contamination data were provided by the Second French Total Diet Study (TDS2). The TDS2 was conducted between 2006 and 2010 to evaluate the exposure of the French population to various substances that are likely to be found in foods ‘as consumed’. This study collected 20 000 food products, representing 212 types of food, for which 445 substances of interest were investigated. Food sampling was based on the data from the INCA2 survey, covering about 90 % of dietary consumption in the adult and child populations( Reference Sirot, Volatier and Calamassi-Tran 50 ). The 212 foods selected were linked to the INCA2 nomenclature. Of the 445 substances analysed, ten chemical substances, for which toxicological risk could not be excluded, were considered in this study( Reference Sirot, Volatier and Calamassi-Tran 50 ): trace elements (i.e. Pb, Al, Cd, inorganic As, organic Hg), acrylamide, one mycotoxin (i.e. deoxynivalenol (DON) and its acetylated derivatives), polychlorinated biphenyls (PCB)/dioxins (i.e. non-dioxin-like polychlorinated biphenyls (NDL-PCB), polychlorinated dibenzo-p-dioxins and dibenzofurans and dioxin-like polychlorinated biphenyl) and one additive (sulphites). The individual average daily exposure levels to the ten substances were estimated by combining individual food consumption data and contamination data from the food sample analysis, considering the same hypotheses as those described in the TDS2 report( 51 , 52 ).
Statistical analyses
Identification of dietary patterns
The NMF method was applied to the data set composed of the 2600 individual daily intake (g/d) of the sixty-six food groups. The analysis was performed on the overall adult population because similar dietary patterns were identified separately in men and women (data not shown). To account for individual weight in pattern identification, the iterative least squares (LS)-NMF algorithm developed by Wang et al.( Reference Wang, Kossenkov and Ochs 53 ) and based on that described by Lee & Seung( Reference Lee and Seung 33 ) was used. The goal of this factorial analysis is to summarise the information available in food consumption data into an optimal number k of consumption systems (CS)( Reference Béchaux, Zetlaoui and Tressou 23 , Reference Zetlaoui, Feinberg and Verger 34 , Reference Sy, Feinberg and Verger 35 ). In contrast to the PCA technique, each CS k in the NMF is defined as a positive linear combination of foods, which are generally associated in the same diet. Thus, all CS k describe the different associations of foods within the population. For each CS k , each food group had a coefficient that can be interpreted as the contribution of this food group to the construction of the system CS k . The weight of each CS k in each individual’s total diet was also determined. The diet of an individual is thus represented by a combination of different CS k .
To implement the NMF method, an optimal number of CS must be chosen. In this study, it was selected according to the quality of the interpretation of the CS (relevancy and ease of interpretation) and a graphical approach as done in Béchaux et al.( Reference Béchaux, Zetlaoui and Tressou 23 ) and Sy et al.( Reference Sy, Feinberg and Verger 35 ). Finally, a HCA was conducted to identify individuals with similar combinations of CS, defining a dietary pattern. The scores of each individual on the CS selected were used as input to the HCA. This classical clustering method consists of a step-by-step aggregation of individuals or groups of individuals who combined the CS in a similar way( Reference Kaufman 54 ), leading to one single class that includes the entire population. The number of clusters to retain was based on the inter-cluster inertia:total inertia ratio and the interpretability of the different clusters.
For each dietary pattern, the relative contribution (%) of each CS k was calculated (i.e. among individuals within the same dietary pattern, the contribution of the CS k is the ratio between the sum of weights of the CS k and the sum of the weights of all the CS). The CS that best describes each pattern was identified according to the V test indicator, which compares the average weight of the CS k in one dietary pattern with the average weight of the CS k in the whole population( Reference Lê, Josse and Husson 55 , Reference Lebart, Morineau and Piron 56 ). The CS k with significant and positive V tests were used to describe dietary patterns.
Characterisation of the dietary patterns
Demographic and socio-economic characteristics of each dietary pattern were investigated using binomial logistic regression. Each tested model identified the main demographic and socio-economic determinants of each dietary pattern independently of the others, by comparing with the overall population. Variables considered were age, level of education, wealth index, household size, household composition, occupational status, region, food insecurity and municipality size. These factors were selected because of their significant associations with the dietary patterns in univariate analysis (data not shown), as well as the consistent associations between dietary intake and these demographic and socio-economic determinants( Reference Novaković, Cavelaars and Geelen 57 – Reference Cappuccio, Ji and Donfrancesco 61 ). All analyses were performed among men and women separately in order to take into account the significant interaction observed between sex and other factors (data not shown).
The mean nutrient intake was calculated for each dietary pattern. The association between nutritional intake and dietary patterns was assessed using ANOVA, and specific nutrient intake was identified by comparing the mean of each dietary pattern with the overall mean. All models were controlled for age, sex, season, TEI, level of education, wealth index, occupational status, household size, food insecurity, household composition, municipality size and region. As previously mentioned, these covariates were selected on preliminary analyses and previous studies( Reference Novaković, Cavelaars and Geelen 57 – Reference Cappuccio, Ji and Donfrancesco 61 ).
Diet quality indices can evaluate the overall diet of an individual based on the following: (i) nutrient indicators, which reflect the adequacy to nutritional requirements; and (ii) foods to assess the variety of food intake(
Reference Wirt and Collins
62
,
Reference Alkerwi
63
). Three scores were selected to illustrate the overall quality of the diet: the energy density (ED) of the diet(
Reference Kant and Graubard
64
), the mean adequacy ratio (MAR)(
Reference Vieux, Soler and Touazi
65
) and the dietary diversity score (DDS)(
Reference Kant
66
). The ED was used as an indicator of bad nutritional quality. Low ED has been shown to have a good nutritional quality(
Reference Ledikwe, Blanck and Khan
67
), and a decrease of ED of the diet is recommended by several public health authorities to prevent obesity(
68
,
Reference Swinburn, Caterson and Seidell
69
). For this study, ED was calculated for each individual with respect to the energy content (kJ/g (kcal/g)) of all foods consumed (except beverages such as water, soft drinks, alcohol, milk, coffee, tea). The mean ED was assessed for each dietary pattern. MAR was used as an indicator of good nutritional quality. The MAR represents the nutritional adequacy of the diet. Multiple versions of this index have been related to health indicators(
Reference Keller, Ostbye and Bright-See
70
), as well to other diet quality indexes(
Reference Krebs-Smith, Smiciklas-Wright and Guthrie
71
–
Reference Torheim, Ouattara and Diarra
73
). It was calculated as the mean percentage of the French daily recommended intake for twenty keys nutrients (namely proteins, fibres, vitamins A, C, E, D, B1, B2, B3, B6, B9, B12, Ca, K, Fe, Mg, Zn, Cu, I and Se). Each ratio was truncated at 100, so that a high intake of one nutrient could not compensate for the low intake of another:
$MAR_{i} {\equals}{1 \over {20}}{\times}\mathop \sum\nolimits_{n\,\,1}^{n\,\,20} {{intake_{{i,\,n}} } \over {RDA_{n} }}{\times}100$
where
$intake_{{i,n}} $
is the individual nutrient intake of the nutrient n and
$RDA_{n} $
is the French RDA for the nutrient, taking into account the age and the sex of the individual(
Reference Vieux, Soler and Touazi
65
). Besides, the diet diversity is also a key element of the high quality of diets. A diverse diet increased the probability of nutrient adequacy(
Reference Ruel
74
), and it has been associated with positive health outcomes(
Reference Lucenteforte, Garavello and Bosetti
75
,
Reference Garavello, Giordano and Bosetti
76
). DDS is defined as the number of specific food groups consumed over a specific period(
Reference Kant
66
,
Reference Drewnowski, Henderson and Shore
77
). In this study, 3 d were randomly chosen for each subject: 2 weekdays and 1 weekend day. Five food groups were considered: dairy products (milk, yogurt, cheese), meat (red meat, poultry, fish and crustaceans), cereals (rice, pasta, wheat), fruits (fresh fruit, processed fruit and dry fruit) and vegetables (fresh vegetables and prepared vegetables). A food group was considered to have been consumed if at least 30 g was ingested during the 3 d. A DDS score was calculated for each individual, and it varied from 0 to 5. The mean DDS score was calculated for each pattern. Associations between dietary patterns and diet quality scores were also assessed using ANOVA adjusted for covariates, as described above. The mean of quality scores of each dietary pattern was thus compared with the overall mean.
Finally, mean contaminant exposure levels were calculated for each dietary pattern. Associations between dietary patterns and exposure levels were assessed using ANOVA-adjusted covariates described above. On the basis of the ANOVA model, specific exposure levels were identified by comparing the mean contaminant levels of each dietary pattern with the overall mean.
All values were survey-weighted means. A P value of 0·05 was used as the threshold of significance. All analyses were implemented in the software R version 3.0.2. The LS-NMF algorithm was implemented using the R package ‘NMF’( Reference Gaujoux and Seoighe 78 ). The package ‘Factominer’ was used to run the clustering( Reference Lê, Josse and Husson 55 ). The package ‘Survey’ was used to account for the complex INCA2 sampling frame design( Reference Lumley 79 ).
Results
Identification of dietary patterns
By combining graphical and interpretability criteria, seven distinct CS summarised the consumption behaviours of the 2600 individuals with respect to the sixty-six food groups. The inclusion of additional CS did not provide any further useful information for the interpretation of the dietary patterns. Moreover, additional CS were difficult to interpret, as they were composed of very few food groups (data not shown). Food groups with a score ≥2·5 % were considered as main contributors to a CS. Table 2 shows the relative contribution of the main food groups associated with each of the seven CS, designated as ‘Tradition’, ‘Snacking’, ‘Mediterranean’, ‘Simplicity’, ‘Dietetic’, ‘High-fat/sugar/salt’ and ‘Pleasant-and-convenient’ food behaviours. No strong Pearson’s correlations (<0·2715) were found between the different CS, suggesting that food behaviours related to each CS were independent of each other.
Table 2 Food consumption characteristics of each dietary pattern
CS, consumption system.
* CS contributing significantly more than the overall population (name and % of contribution).
† Foods contributing >2·5 % to the construction of the CS (name and % of contribution).
‡ Individuals in the population (%).
Then, seven dietary patterns with homogeneous CS combinations were identified and named according to their food consumption patterns. The major CS that best described each dietary pattern were identified and presented in Table 2. In brief, the first dietary pattern called ‘Small eater’ represented 23·0 % of the population. It consisted of consumers who used all the CS but to a lesser extent than the overall population, which means that they consumed all foods but in a lower quantity than the overall population. The second dietary pattern called ‘Health conscious’ grouped 12·6 % of the population and was characterised by individuals who used the dietetic CS significantly more than the overall population, which was mainly associated with low-fat or light foods, soups, fruits, tea and herbal tea and, paradoxically, cakes and pastries. The third dietary pattern, named ‘Mediterranean’, grouped 13·0 % of the population and was represented by individuals who used the Mediterranean CS significantly more than the overall population, which was characterised by unprocessed foods (vegetables, oil, herbs and spices, unprocessed fish, unprocessed fruit, etc.) and dairy products (condiments and cold dips (not low-fat), yogurt and cottage cheese (30–40 % fat)). Individuals in the fourth dietary pattern called ‘Sweet and processed’ grouped 13·5 % of the population. This pattern was characterised by food behaviour represented by the Pleasant-and-convenient CS characterised by an association of sweetened products such as breakfast cereals, fruit juices, chocolate bars/confectionery, dairy desserts and meals easy to prepare such as puff pastries, quiches, warm sauces, cereal-based mixed dishes, etc. The fifth dietary pattern identified as ‘Traditional’ accounted for 16·5 % of the population and was represented by individuals who followed the Tradition CS significantly more than the overall population and the High-fat/sugar/salt CS. Individuals in this pattern were therefore characterised by a consumption of foods such as alcohol (in particular wine), processed meat, cheese, bread products with wheat flour, coffee, red meat, but also crackers, confectionery without chocolate, grains and nuts, cakes and pastries, and sweetened biscuits, which characterised the High-fat/sugar/salt CS. The sixth pattern, identified as ‘Snacker’, was represented by 11·5 % of the population and was characterised by individuals who followed the Snacking CS, mainly represented by take-away products such as sandwiches, pizza, sodas and colas, puff pastries (such as ham puff pastry, ‘bouchée à la reine’, etc.) and processed foods such as processed potato products and cereal-based mixed dishes (as spaghetti carbonara, pasta gratin, etc.). This pattern also followed the High-fat/sugar/salt CS more than the overall population. The last dietary pattern called ‘Basic consumer’ accounted for 10·0 % of the population and was characterised by individuals who followed the Simplicity CS, which associated mostly simple foods such as butter/other animal fat, refined pasta/rice/wheat, unprocessed potatoes, yogurt and cottage cheese (20 % fat), bread and bread products (including bread, loaf and rusk).
Characterisation of dietary patterns
OR and 95 % from logistic regressions are detailed in the Table 3 for men and women separately. Regardless of sex, the probability of belonging to the Health-conscious and Mediterranean dietary patterns (only for men in Traditional pattern) increased with age, conversely to the probability of belonging to the Sweet-and-processed and Snacker dietary patterns. In addition, both women and men in the Health-conscious pattern were more likely to have a higher wealth index, as well as women from the Mediterranean pattern. In contrast, men in the Snacker pattern were more likely to have a relatively low wealth index. Women from the Traditional and Small-eater patterns were more likely to have a low educational level conversely to women from the Mediterranean pattern. Women belonging to the Traditional, Snacker or Health-conscious dietary patterns were more likely to live in households experiencing food insecurity, compared with women from the Small-eater, Mediterranean and Sweet-and-processed dietary patterns. Among men, individuals from the Sweet-and-processed dietary pattern were more likely to live in households experiencing food insecurity, conversely to men belonging to the Traditional and Small-eater dietary patterns. The Mediterranean and Snacker dietary patterns had a higher probability of living in large towns or cities.
Table 3 Demographic and socio-economic determinants of each dietary pattern by sex (Odds ratios and 95 % confidence intervals)
Nutritional intake for each dietary pattern is shown in Table 4. The energy intake was lower than the overall population for the Small-eater but higher for the Sweet-and-processed, Traditional and Basic-consumer dietary patterns. These three latter dietary patterns were also characterised by higher intake of SFA, mainly because of a higher consumption of savoury or sweet pastries, chocolate for Sweet-and-processed pattern and higher consumption of animal products (i.e. butter, cream, cheese or red meat) for Traditional and Basic-consumer patterns. The Health-conscious and Mediterranean dietary patterns had higher intake of fibres than the overall population, primarily because of a higher consumption of fruits, vegetables and wholemeal bread (for Health-conscious pattern only), leading also to higher intake of many minerals and vitamins than the overall population. In addition, Sweet-and-processed pattern showed higher intake of some minerals and vitamins, probably because of a higher consumption of fruits juice and breakfast cereals (which are, for most of them, fortified). Conversely, the Small-eater, Snacker, Traditional and Basic-consumer dietary patterns showed intake of almost all mineral and vitamins studied, which was lower than the overall population. Only the Traditional and Health-conscious dietary patterns had higher intake of Na than the overall population, primarily because of a high consumption of cheese and processed meat and a high consumption of wholemeal bread and bottled water for each pattern, respectively.
Scores of nutritional quality (DDS, MAR, ED) were significantly different across dietary patterns (Table 4). Mostly because of an insufficient intake of fruits and vegetables, the Traditional and Snacker dietary patterns showed significantly lower DDS values than the overall population; 20·4 and 30·7 % of individuals from the Traditional and Snacker dietary patterns, respectively, had a DDS value of 4, and 13·5 and 6·3 %, respectively, had a DDS value of 3 (data not shown). Conversely, the Health-conscious and Mediterranean dietary patterns consumed at least 30 g of dairy products, meat, cereals, fruits and vegetables over 3 d, leading to higher DDS values than the overal population; 95 and 92 % of consumers, respectively, had a DDS value of 5 (data not shown). The MAR, a composite indicator for nutrient adequacy, was higher than the mean in the overall population for individuals from the Health-conscious and Mediterranean dietary patterns, as well as for Sweet-and-processed and Basic-consumer dietary patterns. Individuals from the Health-conscious and Mediterranean patterns, who consumed higher amounts of foods with high nutritional density and low ED, such as fruits, vegetables and unprocessed fish, had also a lower ED than the overall population. ED was higher than the mean in the overall population for the Small-eater, Traditional and Snacker dietary patterns, patterns for which the MAR was significantly lower than the overall population.
Table 4 Nutrient intake and diet quality indicators of each by dietary pattern (Survey-weighted mean values and standard deviations)
Vit, vitamins; carbo, carbohydrates; pop., population; MAR, mean adequacy ratio.
* ANOVA adjusted for sex, season, level of education, wealth index, occupational status, household size, food insecurity, household composition, municipality size and region, and total energy intake (except for the variable energy), significant at P<0·05.
† Nutritional intake significantly lower than the overall population; significant at P<0·05.
‡ Nutritional intake significantly higher than the overall population.
For the ten substances considered in this study, Table 5 gives the mean exposure levels for each dietary pattern. Except for acrylamide and DON and its derivatives, the Snacker dietary pattern was significantly less exposed than the overall population for all substances studied. This result can be attributed to relatively low consumption of foods that are recognised as contributors to substance exposure. On the contrary, the Health-conscious and Mediterranean dietary patterns were more exposed than the overall population to numerous substances. For instance, these patterns showed the highest exposure level to Pb, primarily because of higher consumption of water and hot drinks. Furthermore, as a result of their higher consumption of vegetables, individuals from the Health-conscious and Mediterranean dietary patterns were more exposed to Al than the overall population. The Health-conscious dietary pattern was also more exposed to PCB-NDL, primarily because of higher consumption of fish and fish products. The Basic-consumer dietary pattern was also significantly more exposed than the overall population to Cd, because of higher consumption of bread products, and to PCB-NDL, mostly because of high consumption of butter and other dairy products. Because of their high consumption of alcohol (mainly wine), individuals belonging to the Traditional dietary pattern were more exposed to sulphites than the overall population.
Table 5 Contaminant exposure levels of each dietary pattern (Survey-weighted means and standard deviations)
bw, Body weight; TEQ, toxicity equivalent quantity.
* DON, deoxynivalenol; NDL-PCB, non-dioxin-like polychlorinated biphenyls; PCDD, polychlorinated dibenzo-p-dioxins and dibenzofurans; DL-PCB, dioxin-like polychlorinated biphenyl ANOVA adjusted for sex, season, level of education, wealth index, occupational status, household size, food insecurity, household composition, municipality size and region and total energy intake, significant at P<0·05.
† Contaminant exposure level significantly higher than the overall population.
‡ Contaminant exposure level significantly lower than the overall population; significant at P<0·05.
Discussion
This study identified seven main dietary patterns in the adult population in France, with very distinct food consumption behaviours. These patterns reflected specific nutritional intake and food contaminant exposure levels, as well as particular demographic and socio-economic determinants. According to their CS composition, these patterns were named Small eater, Health conscious, Mediterranean, Sweet and processed, Traditional, Snacker and Basic consumer. The results of this study were consistent with other studies, both national and international. Indeed, the patterns reported as reproducible in the review of Newby & Tucker( Reference Newby and Tucker 80 ) (Healthy, Western, Alcohol/Drinker, and Sweets/Dessert) are similar to some patterns we observed here. Nevertheless, although some patterns were comparable across populations (in many diverse countries and continent), there was natural variation in food consumption, which can be partly attributed to the specificity of French food culture.
First of all, two dietary patterns in particular are consistently reported in industrialised countries: one is less healthful and designated as a ‘Western-style’ pattern, and the other is more healthful and called the ‘Prudent’ pattern( Reference Hu 18 , Reference Kant 24 , Reference Wu 81 , Reference Sofianou, Fung and Tucker 82 ). First, the Western-style pattern generally features high consumption of bread, red and processed meat, starchy foods and high-fat products and is relatively similar to the patterns described as Traditional and Basic-consumer in this study. However, some disparities remained. On one hand, the Basic-consumer pattern was also characterised by a higher consumption of basic and unprocessed foods (egg, unprocessed potatoes, pulses) than the overall population with relatively high consumption of dairy products (cream, yogurt and butter), which may specifically reflect an older French model( Reference Mahe, Francou and Colin 27 , Reference Mahe, Tavoularis and Pilorin 83 ). On the other hand, high consumption of alcoholic drinks (in particular wine), observed in our Traditional dietary pattern, is not particularly noticed for the ‘Western’ diet. Other French studies have revealed an Alcohol/meat dietary pattern, but distinctive only in its amount of alcohol and meat consumed( Reference Kesse-Guyot, Bertrais and Péneau 28 , Reference Charreire, Kesse-Guyot and Bertrais 84 , Reference Bessaoud, Tretarre and Daurès 85 ). Our Traditional pattern seems to reflect at least one aspect of the French culinary culture, with its strong attachment for conviviality, and pleasure of eating( Reference Mahe, Tavoularis and Pilorin 83 , Reference Poulain 86 ). The dietary behaviours of these two ‘Western-like’ dietary patterns led to less healthy nutritional intake, with high energy and SFA intake and low vitamins and minerals intake. Individuals from these patterns were likely to have a lower socio-economic status. The results tend to support the assumption, often reported in the literature, that consumption is strongly influenced by socio-economic status and notably confirm a strong relationship between a higher consumption of energy-dense foods (such as fried products, cereals, potatoes, meat and meat products) and a lower socio-economic status( Reference Darmon and Drewnowski 87 – Reference Trichopoulou, Naska and Costacou 89 ). Second, the Mediterranean and Health-conscious patterns were comparable to the ‘Prudent’ pattern, commonly identified in the literature. The name Mediterranean was chosen following the definition of a Mediterranean diet in the literature, such as a high consumption of whole grains and carbohydrates, fruits, vegetables, fish, olive oil, legumes and low to moderate amounts of saturated animal fats, red meat and wine( Reference Sofi 90 , Reference Georgoulis, Kontogianni and Yiannakouris 91 ). Effectively, our Mediterranean pattern was characterised by a high consumption of fruits, vegetables, fish and oil, of which 56 % was olive oil (against 52 % in the overall population). Moreover, the Mediterranean Diet Score proposed by Trichopoulou et al.( Reference Trichopoulou, Costacou and Bamia 92 ) has been calculated for each individual and confirmed the existence of this Mediterranean pattern among the French adult population (data not shown). Nevertheless, our Mediterranean pattern was not characterised by a high consumption of legumes and whole-grain products, as described in the literature( Reference Sofi 90 , Reference Georgoulis, Kontogianni and Yiannakouris 91 ). Similar patterns have also been identified in other French studies( Reference Kesse-Guyot, Bertrais and Péneau 28 , Reference Bertin, Touvier and Dubuisson 29 , Reference Charreire, Kesse-Guyot and Bertrais 84 ), but which were also characterised by high consumption of breakfast cereals, which was not observed in this study. The Health-conscious pattern describes individuals who ate more dietetic products. Few studies have identified a group of consumers characterised by high consumption of dietary products( Reference Hearty and Gibney 93 , Reference Ax, Warensjö Lemming and Becker 94 ). The consumption of diet products appeared long before the INCA2 study – that is in the 1980s( 95 ); thus, the identification of such a pattern was probably because of the level of aggregation of foods chosen in this study, which identified diet products separately. Consumers in both these dietary patterns seemed to have the most nutritious dietary behaviour with a nutrient-dense diet, a higher MAR and higher consumption of foods with low ED. These dietary patterns were associated with higher socio-economic status, which support the association between a higher socio-economic status and so-called healthy foods, such as wholemeal cereal-based products, fruits and vegetables or fish already identified in the literature( Reference Darmon and Drewnowski 87 , Reference Vlismas and Stavrinos 88 ).
In addition, we identified two patterns (Snacker and Sweet and processed) characterised by a high consumption of processed and modern foods (i.e. easy to prepare and to eat). Only one such pattern per study has generally been reported in the literature, either under the name of Processed/Unhealthy foods, characterised by the high consumption of high-energy beverages and savoury snacks( Reference Bertin, Touvier and Dubuisson 29 , Reference Hearty and Gibney 93 ), or Sweets( Reference Venkaiah, Brahman and Vijayaraghavan 96 – Reference Newby, Muller and Hallfrisch 98 ), with a high consumption of dairy desserts and sweet products. These two profiles were both characterised by high energy intake and SFA intake. Conversely to the description of the pattern ‘Sweet foods and breakfast cereal’ identified by Hearty et al.( Reference Hearty and Gibney 93 ) among Irish adults, individuals in the Sweet-and-processed dietary pattern also had higher intake in some vitamins and minerals than the overall population, probably because of a higher consumption of fruits juices and fortified breakfast cereals. In both these dietary patterns, they were more likely to be younger, which confirms the negative association observed by Adams & White( Reference Adams and White 99 ) between age and energy from ultra-processed foods (i.e. ready-to-eat, convenient and accessible foods such as breakfast cereals, biscuits, mixed dishes, pizza, etc.).
Finally, the Small-eater dietary pattern in our study was characterised by a significantly lower consumption of all foods compared with the overall population, with lower intake of micronutrients. To our knowledge, only two studies identified a similar pattern, but these studies were performed among an elderly population( Reference Corrêa Leite, Nicolosi and Cristina 100 , Reference Schroll, Carbajal and Decarli 101 ). In our study, although no association was observed between this dietary pattern and age, individuals belonging to the Small-eater pattern had a tendency to be older (11, 17, 32, 21 and 19 % of individuals from the Small-eater pattern were 18–24, 25–34, 34–49, 50–64 and >64 years old, respectively; data not shown). Because the presence of under-reporters might have suggested potential bias, their distribution was studied. In fact, under-reporters did not represent the majority of individuals from this pattern (48 % of individuals from the Small-eater pattern were identified as under-reporters), and, consequently, those individuals who had lower energy intake than the overall population could be considered as real small consumers.
As reported in other studies, dietary patterns can highlight the specific food habits, preferences and availability of the countries(80,102). The multiplicity of dietary patterns identified in this work clearly reflects the contradictory attitudes of the French population toward food, such as health awareness, indulgence, pleasure, conviviality, but also convenience and practicality( Reference Laisney 103 , Reference Laisney 104 ). These different food consumption behaviours were also noticeable in the BMI, as defined by the World Health Organization( 5 ). For instance, profiles with ‘healthy’ food behaviour (i.e. the Health-conscious and Mediterranean patterns) had a lower proportion of individuals considered as overweight (32·8 and 29·1 %, respectively) or obese (13·0 and 10·9 %) than the more ‘unhealthy’ profiles, such as the Traditional pattern (44 % of overweight and 15·5 % of obese individuals).
One original aspect of our work was to focus, in addition to the nutritional quality of the diet, on food contaminant exposure levels of each dietary pattern. Whereas Health-conscious and Mediterranean dietary patterns seemed to have healthy dietary behaviours, these two groups of consumers seemed to be more at risk for exposure to some chemical substances. In comparison with health-based guidance values (HBGV)( 51 ), the Health-conscious pattern was considered to be at risk for its exposure to Pb, Cd, inorganic As and Al, and the Mediterranean pattern was identified to be at risk for its exposure to Pb, inorganic As, organic Hg and NDL-PCB. Conversely, the Snacker pattern had a higher ED, a lower MAR and, in comparison with recommended nutrient intake values, the highest prevalence of inadequate nutrient intake (data not shown). However, according to the HBGV, its exposure to the ten substances studied was not considered to match at-risk levels (except for acrylamide exposure). Finally, our results suggest that diets should be analysed further according to a risk:benefit ratio. Unfortunately, no comparison can be made here because, to our knowledge, to date, no other study in the literature has characterised the dietary patterns by levels of contaminant exposure.
Otherwise, this study shows that the novel factorial analysis used, the NMF, was well adapted to determining dietary patterns and successfully summarised the precise variability of food consumption in a given population. Moreover, by using an appropriate algorithm, this is the first study on this topic for which individual sampling weight was taken into account in the NMF to be representative of the French adult population( Reference Béchaux, Zetlaoui and Tressou 23 , Reference Zetlaoui, Feinberg and Verger 34 ). In contrast to PCA, for which dietary patterns are constructed based on an opposition of ‘foods consumed’ and ‘non-consumed’, the NMF constructs food behaviour patterns using only a positive association of foods, which may better reflect reality. In addition, although it is well known that food consumption is a multidimensional phenomenon, classical factorial analysis approaches mean that one dietary pattern corresponds to one common underlying dimension (factor) of food consumption( Reference Hu 18 , Reference Kesse-Guyot, Bertrais and Péneau 28 ). With NMF, one dietary pattern can be represented by different CS. Thus, as an example, consumers from the Traditional dietary pattern were characterised by foods composing the Traditional CS (processed meat, alcoholic drinks, coffee, etc.) but also the High-fat/sugar/salt CS (grains and nuts, crackers, etc.). Moreover, our study used different levels of aggregation of foods and distinguished some dietary patterns that were previously confounded in other studies and provided a better characterisation of those patterns. For instance, Bertin et al.( Reference Bertin, Touvier and Dubuisson 29 ) identified five dietary patterns named Traditional, Diversified, Processed, Prudent and Sandwiches using a PCA based on the average frequency of consumption of forty-three food groups from the INCA2 survey. According to the foods that characterised the dietary patterns, the Processed pattern was similar to the Snacker and Sweet-and-processed patterns of this study, and the Prudent pattern was similar to the Health-conscious and Mediterranean dietary patterns. Furthermore, the Processed pattern identified by Bertin et al.( Reference Bertin, Touvier and Dubuisson 29 ) differed from the overall population only by the higher consumption of sandwiches and lower consumption of other foods, whereas our Snacker dietary pattern included individuals who consumed higher quantities of several foods (sandwiches, pizza, sodas and colas, processed potato products, etc.) than the overall population. These differences further demonstrate that the NMF provides a better characterisation of the different food consumption behaviours.
Another major strength of this study was that it was based on two robust national studies. First, the INCA2 survey was conducted on a large and representative sample of the French adult population using a complex sampling frame design, with a robust collection of dietary intake using a 7-d food record, as well as numerous variables relative to demographic and socio-economic status( Reference Dubuisson, Lioret and Touvier 37 ). For the TDS2 study, a complex food sampling plan covering 90 % of the French diet was designed, taking into account the seasonal nature of products and the regional variations, leading to an accurate assessment of the population exposure at the national level( Reference Sirot, Volatier and Calamassi-Tran 50 ). In addition, the use of factorial analysis raises some concerns about the degree of subjectivity involved in the analytical process (e.g. the determination of the number of CS, the level of aggregation of foods, the determination of the number of patterns identified). However, as highlighted in Newby & Tucker( Reference Newby and Tucker 80 ), the consistency and reproducibility with regard to other national and international studies help to confirm the validity of our findings.
In conclusion, from the INCA2 survey, we identified seven distinct dietary patterns in the French adult population, with specific demographic, socio-economic, nutritional and environmental characteristics. These findings provide new information on the diversity of food consumption in France and give an overview of the nutritional quality of the different food consumption behaviours. From a public health perspective, our results provide interesting insights for developing behaviourally targeted policies. In addition, because of contradictory results for a given dietary pattern between high-quality nutritional intake and high contaminant exposure levels (and vice versa), this study also demonstrates the necessity to analyse the risks and the benefits of food consumption behaviours, particularly in a public health context. Finally, the food consumption data were collected several years ago (i.e. 2006–2007) and the third INCA survey is currently underway, potentially providing the opportunity to assess the trends in dietary patterns at the national level.
Acknowledgements
The authors wish to thank the CIQUAL for providing the national food composition tables, the Institut de Sondage Lavialle (ISL) team for the collection of data, and all the families for their cooperation.
This research received institutional support from the French agency for food, environmental and occupational health safety (ANSES).
R. G. designed the study, analysed and interpreted the data, wrote the manuscript and had primary responsibility for final content. C. B. and S. H. contributed to the analytical approach, the interpretation of the results and revised each draft. S. H., P. D.-P., C. D., V. S. and A. C. contributed to the design of the surveys (INCA2 and EAT2), to the data collection and help to write the paper.
The authors have no financial or personal conflicts of interest to declare.
Supplementary Material
For supplementary material/s referred to in this article, please visit http://dx.doi.org/10.1017/S0007114516001549
Open access
