Dietary pattern analysis has been of growing interest in nutritional epidemiology as a multidimensional approach that allows the investigation of the overall effects of diet on human health, by taking into account the complex interactions among foods and nutrients consumed( Reference Hu 1 – Reference Jacob and Steffen 3 ). Two major approaches have long been applied to investigate dietary patterns. In the hypothesis-driven approach, scientific evidence or dietary recommendations are used to construct dietary indices or scores that evaluate the quality of the diet or the adherence to a particular pre-defined diet, as the Diet Quality Index, the Healthy Eating Index and the Mediterranean diet score( Reference Ocké 4 ). In the data-driven approach, statistical methods such as exploratory factor analysis (EFA), principal component analysis and cluster analysis are used to empirically derive dietary patterns based on dietary data collected( Reference Ocké 4 ).
Over the past few decades, a number of evidences have been emerged towards the association between empirically derived dietary patterns and CVD risk( Reference Hu, Rimm and Stampfer 5 , Reference Fung, Willett and Stampfer 6 ) and cardiovascular mortality( Reference Heidemann, Schulze and Franco 7 ). The ‘Prudent pattern’, characterised by fruits, vegetables, legumes, fish and whole grains has been associated with lower risk of CHD( Reference Hu, Rimm and Stampfer 5 ) and lower cardiovascular mortality among women( Reference Heidemann, Schulze and Franco 7 ), whereas the ‘Western pattern’, characterised by red and processed meats, refined grains, sugar, sweets and dairy products, has been associated with a greater risk of CHD and cardiovascular mortality( Reference Hu, Rimm and Stampfer 5 , Reference Heidemann, Schulze and Franco 7 ).
The association of empirically derived dietary patterns with metabolic CVD risk factors has also been investigated( Reference Fung, Rimm and Spiegelman 8 – Reference Olinto, Gigante and Horta 14 ), but no studies have proposed and tested a conceptual model, that is, a theoretic model that represents the hypothesised relationships between the variables under investigation, such as the biological inter-relationships of the CVD risk factors, and the effects of dietary patterns on these factors.
To investigate this association under a conceptual model, the structural equation model (SEM) is a suitable statistical method, allowing investigators to test the validity of the model based on a set of measured variables in an attempt to explain their observed variances and covariances( Reference Asparouhov and Muthén 15 ). Moreover, SEM allows for multiple linear equations, including direct and indirect effects, and latent variables, features not allowed by traditional regression methods.
Therefore, the aim of the present study was to derive dietary patterns and examine their associations with metabolic CVD risk factors under a conceptual model tested by SEM.
Methods
Study population
Data were obtained from the Health Survey of the City of São Paulo, a cross-sectional population-based survey, designed to collect health and nutrition information as well as life conditions on a probabilistic sample of residents of the city of São Paulo, Brazil, between September 2008 and August 2011.
A two-stage cluster sampling of census tracts and households was performed. In the first stage, seventy census tracts were randomly selected from the totality of urban census tracts in the city of São Paulo as the primary sampling units. In the second stage, households were randomly selected within census tracts, in order to allow data collection from about 600 adults (20–59 years) and 600 elderly (60 years and over) of both sexes.
A total of 1102 individuals aged 20 years and above were evaluated in the first stage of data collection, which occurred between 2008 and 2009. In this stage, a structured questionnaire enquiring about socio-economic (per capita family income; educational level), demographic (skin colour, age) and lifestyle characteristics (smoking status; alcohol use; physical activity), as well as food consumption (the first 24-h dietary recall (24-HDR)), was administered by qualified interviewers in individuals’ homes. In the second stage of the study, which occurred between 2010 and 2011, a total of six telephone contacts and one home visit per individual were performed in order to find and invite the same individuals to participate. A total of 642 individuals agreed to participate, and 548 of them underwent a second dietary assessment, blood sample analysis and blood pressure (BP) assessment, as well as anthropometric measurements. In order to evaluate long-term changes in diet, that is, changes occurring between the first and the second dietary data collection, individuals were enquired by telephone about modifications in amounts and types of foods consumed after concluding the second dietary data collection. For the present study, only individuals with no missing data for each variable evaluated and with no changes in their diet were included (n 462).
This study was conducted according to the guidelines established in the Declaration of Helsinki. All procedures involving human subjects in this study were approved by the Research Ethics Committees at the School of Public Health, University of São Paulo (study protocol number 2001; authorisation 275/09). All participants provided their written informed consent before data collection in each stage of the study.
Dietary data collection
The first 24-HDR data were collected by face-to-face interviews according to procedures described in the United States Department of Agriculture (USDA) five-step multiple-pass method( Reference Raper, Perloff and Ingwersen 16 ). This method guides the individual through a 24-h reference period of food intake (more commonly, the day before interview) and provides different opportunities for he/she to remember and describe all foods and beverages consumed( Reference Raper, Perloff and Ingwersen 16 ).
The second 24-HDR data were collected from the same individuals by telephone interviews according to the interviewing system incorporated into the University of Minnesota’s Nutrition Data System for Research (NDS-R). This interviewing system enhances data quality as it standardises the probes about foods and portions consumed( Reference Thompson, Subar and Loria 17 ). All individuals were advised to report food consumption in household measures as well as to mention the eating occasions, meal time, cooking methods, seasonings used and brand names. Quality control of the 24-HDR was conducted during data collection in order to identify and correct reporting errors. Dietary data collection occurred on non-consecutive days throughout all seasons and days of the week.
After dietary data collection, all household measures reported in each 24-HDR were converted into grams and millilitres according to Brazilian publications, which were also consulted to obtain standard recipes of regional food preparations( Reference Pinheiro, Lacerda and Benzecry 18 , Reference Fisberg and Villar 19 ). The NDS-R, version 2007, was used to determine the nutrient content of each food and beverage consumed. This programme was developed by the Nutrition Coordinating Center at the University of Minnesota, Minneapolis, MN, USA and has the USDA Food Composition Table as the primary database source.
Implausible energy intake
In order to avoid biased estimates for the relationship between dietary patterns and metabolic CVD risk factors, individuals who reported implausible energy intakes (i.e. under- and over-reporters) were removed from the analysis. The misreporting of energy intake was defined following procedures described in the study by McCrory et al. ( Reference McCrory, Hajduk and Roberts 20 ) and applied by Avelino et al. ( Reference Avelino, Previdelli and Castro 21 ) in a sample of Brazilian adults. In brief, the agreement between the average reported energy intake (rEI) and the predicted total energy expenditure (pTEE) was evaluated for each individual. In this study, the pTEE was obtained by the equation proposed and validated by Vinken et al. ( Reference Vinken, Bathalon and Sawaya 22 ). This equation was developed based on data from free-living individuals aged 18–81 years who were evaluated by the double-labelled water (DLW) method.
To identify under- and over-reporters, a cut-off point of ±2 sd of the agreement between rEI and pTEE was applied based on principles proposed by Black( Reference Black 23 ). This cut-off point takes into account the within-person CV of energy intake on the days of dietary assessment (in this case, 34·36 %), the CV of the technical measurement error of the DLW plus the biological variation in total energy expenditure (TEE) (8·2 %) and the CV of prediction error of the TEE by Vinken’s equation (17·7 %)( Reference McCrory, Hajduk and Roberts 20 ).
In the present study, the ±2 sd cut-off was equal to ±62 %. Therefore, individuals were considered under-reporters of energy intake if the agreement between rEI and pTEE was <38 % (n 26; 54·9 % BMI<30 kg/m2; 68·6 % female), and individuals were considered over-reporters if the agreement between rEI and pTEE was >162 % (n 19; 66·7 % BMI<30 kg/m2; 70·4 % female).
After removing under- and over-reporters (n 45), the final sample comprised 417 adults.
Foods grouping
In both 24-HDR, a total of 758 different foods were reported, of which 547 were consumed by at least 5 % of the sample, and then collapsed into thirty-four food groups as performed by Selem et al. ( Reference Selem, Castro and César 24 ) and by Castro et al. ( Reference Castro, Baltar and Selem 25 ). In the first step, the foods were combined according to the similarity of the nutrient profile( Reference de Souza, Madruga and Gigante 26 – Reference Cunha, de Almeida and Sichieri 28 ) (e.g. all types of coffees were combined into the ‘coffee’ group) and the particular dietary habits and culinary usage of the Southeastern Brazilian population( Reference Sichieri, Castro and Moura 29 ) (e.g. ‘beans’ group included brown and black beans because they are cooked legumes usually eaten with rice, whereas the ‘legumes’ group included soyabeans, lentils, chickpeas and snow peas because they are usually consumed in different preparations such as soups, creams and salads).
In the second step, food groups were analysed by the correlation matrix in order to verify how they were correlated to each other. Similar food groups with positive and significant correlations were aggregated into a single food group (e.g. leafy and non-leafy vegetables), and similar food groups with negative and significant correlations were maintained in different groups (e.g. whole milk and reduced-fat/skimmed milk). Finally, twenty-eight food groups were available for analysis (a detailed description of their composition is provided in Table 1).
Factor-loading matrix for dietary patterns derived through exploratory structural equation modelling according to dietary data from two non-consecutive 24-h dietary recall, São Paulo, Brazil, 2008–2011 (Factor loadings with their standard errors)Footnote *
RMSEA, residual mean square error of approximation; CFI, comparative fit index; TLI, Tucker–Lewis index; SRMR, standardised root mean square residual.
* Inter-factor correlations: factor 1 (Traditional pattern)×factor 2 (Prudent pattern): r 0·04; P=0·588. Factor 1 (Traditional pattern)×factor 3 (Modern pattern): r −0·14; P=0·193. Factor 2 (Prudent pattern)×factor 3 (Modern pattern): r 0·30; P=0·157. Goodness-of-fit indices: RMSEA: 0·031; 90 % CI 0·026, 0·035; CFI: 0·92; TLI: 0·90; SRMR: 0·043.
Before dietary pattern analysis, the food group intakes (g) were adjusted for the within-person variation through the web-based statistical modelling technique Multiple Source Method (MSM). The MSM was developed within the European Food Consumption and Validation Project as a suitable technique for estimating the usual nutrient and food intakes (including those episodically consumed) based on two or more short-term dietary methods per individual such as 24-HDR( Reference Harttig, Haubrock and Knüppel 30 ).
Anthropometric measurements
Anthropometric measurements, BP and blood samples were obtained in participant homes by a trained nursing assistant following standardised protocols developed for the study. Body weight (BW, kg), height (cm) and waist circumference (WC) (cm) were measured for each individual according to the procedures recommended by the World Health Organisation( 31 , 32 ). BW, height and WC were measured in duplicate using a calibrated digital scale (Tanita®, HD-313; Tanita, accuracy: 100 g), a portable wall-mounted stadiometer (Seca®, model 208; Seca, accuracy: 1 mm) and a flexible anthropometric tape, respectively. The measurement of BW and height was performed with the individual standing firmly on a levelled surface, wearing light clothes and no shoes. WC was measured midway between the lowest rib and the iliac crest at the end of expiration( 32 ). The BW reading was recorded to the nearest 0·1 kg, whereas height and WC were recorded to the nearest 0·1 cm. The arithmetic means of BW, height and WC were calculated.
Blood pressure
BP was measured according to the 5th Brazilian Guidelines for Hypertension( 33 ) using a validated oscillometer (Omron®, model HEM-712C; Omron Healthcare Inc.). Participants rested for 5 min while sitting before measurements, with the arm supported at heart level. Systolic blood pressure (SBP) and diastolic blood pressure (DBP) were initially measured in the right arm; 1 min after the first measurement, SBP and DBP were measured in the left arm. Two additional BP measurements were obtained in the arm with the largest values of SBP and DBP, with a 1 min interval between them. The arithmetic means of the last two measurements of SBP and DBP were calculated and considered for analysis.
Blood samples
Blood samples were collected by venepuncture in two 10-ml vacutainers after 12 h of overnight fasting. Blood samples were maintained in a polystyrene box with ice packs and transported within 2 h to the laboratory for immediate centrifugation at 3000 rpm for 15 min at room temperature. After centrifugation, serum samples were stored at −80°C until analysis. Serum total cholesterol (TC) and fractions (VLDL, LDL, HDL) as well as serum TAG were determined by enzymatic-colourimetric method. TC:HDL ratio and TAG:HDL ratio were also calculated. Serum high-sensitivity C-reactive protein (hs-CRP) was determined by an automated immunoturbidimetric assay system (IMMAGE Immunochemistry System; Beckman Coulter) with a minimum detectable dose of 0·06 mg/l. Serum leptin (LEP) concentrations were determined using a human RIA kit (HL-81K; Millipore) with a minimum detectable dose of 0·437 ng/ml. Fasting plasma glucose (FPG) was measured by enzymatic-colorimetric glucose oxidase procedure (GOD-Trinder; Labtest Diagsnóstica) with a minimum detectable dose of 20 mg/l.
Statistical analysis
Dietary patterns modelling
The exploratory structural equation modelling (ESEM) using the robust maximum likelihood (MLR) parameter estimation and the oblique Geomin rotation was applied to food group variables in order to empirically derive dietary patterns. ESEM is a multivariate statistical technique that can be interpreted as a combination of EFA and SEM. It relies on the covariance structure of the observed variables and is indicated when the researcher has a weak hypothesis about how multiple-observed variables load on the factors( Reference Kline 34 ), a common situation in dietary pattern analyses. In addition, ESEM allows investigators to test the significance of factor loadings( Reference Asparouhov and Muthén 15 ), which contributes to reduce the subjectivity during modelling and interpreting dietary patterns. More details about this method can be found in the study by Asparouhov & Muthén( Reference Asparouhov and Muthén 15 ).
The MLR estimation method was chosen because it is an iterative estimation procedure that leads to more robust standard error estimates for continuous data following a non-normal multivariate distribution( Reference Kline 34 , 35 ). The oblique Geomin rotation was used to test the inter-factor correlations and to obtain a simple factor structure (i.e. a structure with variables loading high on as few factors as possible and with a minimal number of cross-loadings)( Reference Saas and Schmitt 36 ).
Similar to confirmatory factor analysis, the ESEM requires the previous specification of the number of factors under investigation. For this, models with two, three and four factors were sequentially fitted and compared by the log-likelihood ratio test (LRT) and the χ 2 difference test, both adjusted by the Satorra–Bentler scaled correction. The number of factors in each model was based on the common number of previously reported dietary patterns derived through data-driven methods in national studies involving adults( Reference Marsola, Rinaldi and Siqueira 13 , Reference Olinto, Gigante and Horta 14 , Reference Selem, Castro and César 24 , Reference de Souza, Madruga and Gigante 26 – Reference Sichieri, Castro and Moura 29 , Reference Sichieri 37 – Reference da Silva, Neutzling and Camey 43 ). The Satorra–Bentler scaled correction was used in order to adjust the χ 2 and the log-likelihood estimates obtained by the MLR method( Reference Asparouhov and Muthén 15 , Reference Satorra and Bentler 44 ). The LRT and the χ 2 difference test provide a comparative evaluation of the goodness-of-fit of two models with different number of factors and/or free parameters, guiding researchers to choose the model with the smallest number of factors that fits well with the data( Reference Asparouhov and Muthén 15 ). Additional goodness-of-fit indices – namely, comparative fit index (CFI), Tucker–Lewis index (TLI), residual mean square error of approximation (RMSEA) and its 90 % CI and standardised root mean square residual (SRMR) – were also considered. They provide different information about model fit allowing for a more conservative and reliable evaluation of the model( Reference Kline 34 ). Acceptable model fit was defined according to the following criteria: CFI (>0·90)( Reference Bentler 45 ), TLI (>0·90)( Reference Bentler and Bonett 46 ), RMSEA (≤0·06, 90 % CI<0·08)( Reference Browne and Cudeck 47 ) and SRMR (≤0·08)( Reference Hu and Bentler 48 ).
The interpretability of the factors was also evaluated and considered during comparison of the models. Food groups with positive and significant factor loadings can be interpreted as contributing directly to the factor, whereas food groups with negative and significant factor loadings can be interpreted to be inversely correlated with the factor.
Modification index test was evaluated in order to identify significant correlations between measurement error terms (i.e. residuals) of food group variables that would improve model fit. Only plausible correlations were allowed (e.g. correlation between error terms of breads and of butter and margarine), considering that reporting errors for two or more foods may be correlated. All the analyses were executed in Mplus software (version 6.12; Muthén & Muthén). A P value<0·05 was considered to be significant in two-sided tests.
Conceptual model for the association between dietary patterns and metabolic CVD risk factors
The conceptual model for the association between dietary patterns and metabolic CVD risk factors was tested using the SEM analysis. Dietary patterns were expected to be directly associated with each metabolic CVD risk factor, namely, FPG, SBP, DBP, TAG:HDL, TC:HDL and hs-CRP, as well as with serum LEP, BW and WC.
Considering that serum LEP and WC represent obesity indicators that are positively correlated to BW, a latent variable named ‘obesity’ was constructed. A latent variable is a non-observed random variable, such as a factor, which comprises two or more correlated measured variables. The obesity latent variable allowed the estimation of the combined effects of serum LEP, BW and WC on each CVD risk factor.
The indirect effects of dietary patterns on metabolic CVD risk factors mediated by the obesity latent variable were estimated, under the hypothesis that dietary patterns can affect metabolic CVD risk factors also by leading to changes in total BW as well as in amounts and distribution of body fat.
Residual correlations between the outcome variables SBP and DBP and between TAG:HDL and TC:HDL were estimated based on a previous model for metabolic CVD risk factors( Reference de Castro, Baltar and Marchioni 49 ). In addition, residual correlations of FPG with TAG:HDL, TC:HDL and hs-CRP were expected, owing to evidences towards the association of glucose metabolic disturbances with atherogenic dyslipidaemia and inflammation( Reference Yudkin, Stehouwer and Emeis 50 ).
The variables sex, age and race/ethnicity were controlled in all regression models. In addition, the variable energy was added to the obesity regression model. Antihypertensive, hypocholesterolaemic and antidiabetic medications were also controlled in regression models of BP, atherogenic dyslipidaemia and FPG, respectively.
The conceptual model was tested in Mplus software version 6.12 using the MLR estimation method and the oblique Geomin rotation. Delta method was used to estimate the standard errors for indirect effects. Standardised estimates of regression coefficients, correlation coefficients and indirect effects were reported. The acceptable goodness-of-fit of the model was also evaluated by the indices CFI (>0·90)( Reference Bentler 45 ), TLI (>0·90)( Reference Bentler and Bonett 46 ), RMSEA (≤0·06, 90 % CI<0·08)( Reference Browne and Cudeck 47 ) and SRMR (≤0·08)( Reference Hu and Bentler 48 ).
Results
Participant characteristics
Participants comprised 164 men (39 %) and 253 women (61 %), with an average age of 54 (sd 19) years. About 61 % of the participants were of white race. According to the National Cholesterol Education Program/Adult Treatment Panel III criterion( 51 ), 55 % of the participants had abdominal obesity, 38 % high LDL-cholesterol, 37 % low HDL-cholesterol, 32 % hypertriacylglycerolaemia, 13 % elevated FPG and 44 % elevated BP. The overall prevalence of the metabolic syndrome was 26 % (data not shown). The usual energy intake of participants averaged 7255 (sd 1912) kJ (1734 (sd 457) kcal) with men exhibiting a higher average intake than women 8104 v. 6708 kJ (1937 v. 1602 kcal, respectively, P<0·0001) (data not shown).
Dietary patterns modelling
The LRT and the χ 2 difference test adjusted by the Satorra–Bentler scaled correction were significant for comparison of models with two and three factors, suggesting a better fitting for the latter (LRT=133·34 (df 26), P<0·001; χ 2=123·01 (df 26), P<0·001). Moreover, the goodness-of-fit indices of the model with two factors indicated an unacceptable fit (RMSEA: 0·037; 90 % CI 0·031, 0·044; CFI: 0·85; TLI: 0·80, SRMR: 0·05). In contrast, the model with three factors attained the criteria for an acceptable fit (RMSEA: 0·022; 90 % CI 0·010, 0·030; CFI: 0·95; TLI: 0·93, SRMR: 0·04).
Comparing models with three and four factors, a better fitting was observed for the model with four factors based on the LRT and the χ 2 difference test (LRT=53·14 (df 25), P<0·001; χ 2=59·17 (df 25), P<0·001). The goodness-of-fit indices confirmed a better fit for the model with four factors in comparison with the model with three factors (RMSEA: 0·011; 90 % CI 0·000, 0·023; CFI: 0·99; TLI: 0·98, SRMR: 0·03). However, only two food groups (red meats and alcoholic beverages) showed significant loadings on the fourth factor, which made the interpretation difficult. Therefore, the model with three factors, more parsimonious and interpretable, was selected for the subsequent analysis.
The ESEM diagram with three factors is displayed in Fig. 1. The ellipses represent the dietary patterns (i.e. the factors) and the rectangles denote the food groups (i.e. the measured variables). The two-sided arrows indicate the allowed correlation between dietary patterns (arrows between ellipses) and between the residual variances of the food group variables (arrows between rectangles). The one-sided arrows from the ellipses represent the factor loadings, whereas the one-sided arrows directed to food groups represent the residual variances of each variable.
Diagram of exploratory structural equation modelling with three factors (dietary patterns) and their indicator variables (food groups), Health Survey of São Paulo, Brazil, 2008–2011.
Dietary patterns composition and associations with metabolic CVD risk factors
The factor-loading matrix of dietary patterns derived by ESEM analysis is presented in Table 1. The first dietary pattern labelled ‘Traditional’ had high positive loadings on rice and beans and low-to-moderate positive loadings on red meats, eggs, whole milk, butter and margarine, and sugar. Negative loadings were observed for legumes, low-fat and skimmed milk, cheese, whole breads, sweets, pasta, cold cuts, mayonnaise, sandwiches and alcoholic beverages. The second dietary pattern labelled ‘Prudent’ included vegetables, fruits, juices, fish and poultry, low-fat and skimmed milk, whole breads, olive oil and seasonings, all with positive loadings. Negative loadings were observed for sandwiches and alcoholic beverages. The third dietary pattern labelled ‘Modern’ included soda pop, red meats and alcoholic beverages with the highest positive loadings, followed by sandwiches, cold cuts, seasonings, mayonnaise and salty snacks. Negative loadings were observed for fruits, whole milk, coffee and tea, and sugar. No correlations were found among the three dietary patterns derived.
Fig. 2 shows the SEM diagram with standardised estimates for the relationship between dietary patterns and metabolic CVD risk factors. The one-sided arrows from dietary patterns to the obesity latent variable and to metabolic CVD risk factors represent the standardised regression coefficients that can be interpreted as the change in standard deviation units in the outcome variable associated with a standard deviation change in the predictor variable( Reference Hendrie, Coveney and Cox 52 ). The one-sided arrows from the obesity latent variable to metabolic CVD risk factors represent the regression coefficients, whereas the arrows from obesity to serum LEP, BW and WC indicate the standardised factor loadings of the measured variables. The two-sided arrows represent the correlation coefficients between dietary patterns. The residual correlation coefficients between outcome variables were omitted for simplicity.
Structural equation model diagram with standardised estimates for the relationship between dietary patterns and metabolic CVD risk factors, Health Survey of São Paulo, Brazil, 2008–2011. * Standardised coefficients significant at the critical value of 0·05. SBP, systolic blood pressure; DBP, diastolic blood pressure; FPG, fasting plasma glucose; hs-CRP, high-sensitivity C-reactive protein; LEP, leptin; WC, waist circumference; BW, body weight.
The ‘Traditional’ dietary pattern had negative direct effects on the obesity latent variable, whereas the ‘Prudent’ pattern had negative direct effects on SBP. No association was observed between the ‘Modern’ pattern and obesity and neither between the pattern and metabolic CVD risk factors (Table 2). Moreover, the ‘Traditional’ pattern had small negative indirect effects on TAG:HDL, TC:HDL and FPG, through mediation effects of obesity (Table 3).
Parameter estimates and model fitness of the sem analysis of dietary patterns and metabolic CVD risk factors, São Paulo, Brazil, 2008–2011
SBP, systolic blood pressure; DBP, diastolic blood pressure; hs-CRP, high-sensitivity C-reactive protein; FPG, fasting plasma glucose; LEP, leptin; BW, body weight; WC, waist circumference; TC, total cholesterol; AIC, Akaike criterion; BIC, sample size-adjusted Bayesian criterion; χ 2/df, adjusted χ 2 test; RMSEA, residual mean square error of approximation; CFI, comparative fit index; TLI, Tucker–Lewis index; SRMR, standardised root mean square residual.
Total and indirect effects of dietary patterns on metabolic CVD risk factors, Health Survey of São Paulo, Brazil, 2008–2011 (Standardised coefficients with their standard errors)
SBP, systolic blood pressure; DBP, diastolic blood pressure; TC, total cholesterol; FPG, fasting plasma glucose; hs-CRP, high-sensitivity C-reactive protein.
* Total effects are the sum of direct and indirect effects of dietary patterns on metabolic CVD risk factors.
† Indirect effects of the dietary patterns on metabolic CVD risk factors mediated by obesity latent variable.
The obesity latent variable showed positive effects on all the CVD risk factors evaluated. Positive residual correlations were found between BP variables, TAG:HDL and TC:HDL, FPG and TAG:HDL as well as between FPG and TC:HDL (Table 2).
The model adjustments revealed positive associations of age with the ‘Prudent’ pattern (β=0·29 (se 0·09); P=0·001), SBP (β=0·55 (se 0·07); P<0·001) and DBP (β=0·28 (se 0·08); P=0·001), and a negative association with the ‘Modern’ dietary pattern (β=−0·56 (se 0·14); P<0·001). Female sex was negatively associated with the ‘Modern’ dietary pattern (β=−0·34 (se 0·07); P<0·001), obesity (β=−0·22 (se 0·08); P=0·004), TAG:HDL-cholesterol (β=−0·18 (se 0·07); P=0·008) and TC:HDL (β=−0·19 (se 0·08); P=0·012), and was positively associated with the ‘Traditional’ pattern (β=0·38 (se 0·05); P<0·001). Antihypertensive, hypocholesterolaemic and antidiabetic medications were positively associated with BP variables (SBP: β=0·19 (se 0·05); P<0·001; DBP: β=0·13 (se 0·05); P=0·011), TC:HDL (β=0·11 (se 0·04); P=0·007) and FPG (β=0·41 (se 0·07); P<0·001), respectively. No associations of race/ethnicity were observed. Energy intake was positively associated with the ‘Modern’ dietary pattern (β=0·34 (se 0·13); P=0·009) but not with the obesity latent variable (data not shown).
Discussion
This is an original study that proposed the application of the ESEM to empirically derive dietary patterns and used the SEM to examine their associations with metabolic CVD risk factors following a conceptual model for adults. In addition, to our knowledge, this is the first study to report an inverse association between the ‘Traditional’ dietary pattern and a combination of obesity markers, namely, BW, WC and serum LEP.
In the present study, the ‘Traditional’ pattern, characterised by high loadings on typical Brazilian staple foods such as rice and beans and by low-to-moderate loadings on red meats, eggs, whole milk, butter and margarine, and sugar, showed a direct negative effect on the obesity latent variable. This finding suggests a protective role of the ‘Traditional’ pattern against weight gain, increased fat mass and abdominal fat accumulation, which is in agreement with a previous report of 13–14 % lower risk of overweight and obesity in Brazilian subjects following a ‘Traditional’ diet( Reference Sichieri 37 ).
Among the food groups pertaining to the Brazilian ‘Traditional’ pattern, beans may be the major responsible food group for the protective effects observed. Beans are a subgroup of legumes with low glycaemic index that are important sources of phytochemicals and nutrients, such as insoluble and soluble dietary fibre, K and vegetable proteins( Reference McCrory, Hamaker and Lovejoy 53 ). In Brazil, this food group is the most consumed vegetable protein-rich food and the main dietary source of fibre among adults( Reference de Mattos and Martins 54 ), with its average usual consumption in the population evaluated to be about 87 g (equivalent to one serving) (data not shown).
Evidences towards the association between beans consumption and lower BW and WC were previously described. In a cross-sectional study with more than 8000 adult participants in the National Health and Nutrition Examination Survey (1999–2002), Papanikolaou & Fulgoni( Reference Papanikolaou and Fulgoni 55 ) reported lower BW, smaller WC and reduced risk of overall and abdominal obesity among beans consumers compared with non-consumers. The authors argued that the favourable weight associations observed among beans consumers may be related to the fibre content of this food, owing to the effects of fibres on increased satiety, and consequently on reduced energy intake( Reference Howarth, Saltzman and Roberts 56 ).
Moreover, an independent negative association of legumes and total fibre intake with circulating LEP was reported in two population-based studies with Japanese adults( Reference Kuroda, Ohta and Okufuji 57 , Reference Murakami, Satoshi and Takahashi 58 ). Although the biological mechanisms underlying this association require further investigation, it can be interpreted that the ‘Traditional’ pattern may contribute to prevention of excessive BW and abdominal obesity by reducing LEP, a key adipocyte-derived hormone that plays an important role in regulating appetite and BW( Reference de Castro, Baltar and Marchioni 49 ).
In contrast with previous studies that have not detected associations of ‘Traditional’ dietary pattern with metabolic CVD risk factors( Reference Marsola, Rinaldi and Siqueira 13 ) or detected positive effects of this pattern on one or more CVD risk factors( Reference Vilela, Sichieri and Pereira 27 , Reference Neumann, Martins and Marcopito 38 , Reference Noel, Newby and Ordovas 59 ), the present study identified small negative effects of the ‘Traditional’ dietary pattern on TAG:HDL, TC:HDL and FPG, through mediation of obesity markers. These results indicate the suitability of the Brazilian ‘Traditional’ dietary pattern as an adjuvant for BW control and prevention of atherogenic dyslipidaemia and hyperglycaemia.
The second dietary pattern labelled ‘Prudent’ was quite similar to the Dietary Approach to Stop Hypertension (DASH) diet( Reference Appel, Moore and Obarzanek 60 ) and resembled other dietary patterns labelled ‘Healthy’( Reference Berg, Lappas and Strandhagen 9 , Reference Marsola, Rinaldi and Siqueira 13 , Reference Arruda, da Silva and Kac 42 ), ‘Healthy conscious’( Reference Heidemann, Scheidt-Nave and Richter 12 , Reference Mikkilä, Räsänen and Laaksonen 61 ), ‘Health aware’( Reference Hamer and Mishra 11 ) and, more recently, ‘Fruits, vegetables, nuts and legumes’( Reference Gadgil, Anderson and Kandula 62 ), by including fruits, vegetables, juices, fish and poultry, low-fat and skimmed milk, whole breads, olive oil, and seasonings. The ‘Prudent’ dietary pattern was negatively associated with SBP, but not with DBP, suggesting its effectiveness in lowering BP of individuals with isolated systolic hypertension (ISH), an important risk factor for coronary events( Reference Sesso, Stampfer and Rosner 63 ). ISH is more frequent among middle-aged and elderly people owing to the rise in SBP accompanied by the fall in DBP with age( Reference Chobanian 64 ), and was the most frequent subtype of uncontrolled hypertension among US middle-aged adults( Reference Franklin, Jacobs and Wong 65 ). Other beneficial effects ascribed to the ‘Prudent’ dietary pattern include reduction in risk of stroke (−32 %)( Reference Stricker, Onland-Moret and Boer 66 ), CHD (up to −30 %)( Reference Hu, Rimm and Stampfer 5 , Reference Fung, Rimm and Spiegelman 8 , Reference Stricker, Onland-Moret and Boer 66 ) and cardiovascular mortality (−28 %)( Reference Heidemann, Schulze and Franco 7 ). The synergistic effect of multiple nutrients and phytochemicals found in fruits and vegetables (e.g. K, carotenoids, vitamin C and flavonoids)( Reference Bhupathiraju and Tucker 67 ), in whole grains (e.g. total fibre, resistant starch, oligosaccharides, B-vitamins, antioxidants)( Reference Jacobs and Gallaher 68 ), in low-fat dairy products (e.g. Ca, vitamin D and conjugated linolenic acid)( Reference Ruidavets, Bongard and Simon 69 , Reference Rice, Cifelli and Pikosky 70 ) and in olive oil (e.g. oleic acid and phenolic compounds)( Reference Covas and Konstantinidou 71 , Reference Terés, Barceló-Coblijn and Benet 72 ) may explain the cardioprotective role of the ‘Prudent’ pattern.
An unexpected finding was the lack of association between the ‘Modern’ dietary pattern and metabolic CVD risk factors. This pattern, composed primarily by soda pop, alcoholic beverages, red meats, cold cuts, mayonnaise, sandwiches, seasonings and salty snacks, resembles the ‘Western’ and the ‘Processed Foods’ dietary patterns identified in national( Reference Olinto, Gigante and Horta 14 , Reference Vilela, Sichieri and Pereira 27 , Reference Cunha, de Almeida and Sichieri 28 , Reference Sichieri 37 ) and international studies( Reference Hu, Rimm and Stampfer 5 , Reference Fung, Willett and Stampfer 6 , Reference Fung, Rimm and Spiegelman 8 , Reference Stricker, Onland-Moret and Boer 66 , Reference Zuo, Shi and Dai 73 ) as playing harmful effects on cardiovascular health. Elevated insulin concentration( Reference Fung, Rimm and Spiegelman 8 ), increased abdominal obesity( Reference Olinto, Gigante and Horta 14 , Reference Vilela, Sichieri and Pereira 27 , Reference Cunha, de Almeida and Sichieri 28 ), elevated BP, LDL and TC( Reference Olinto, Gigante and Horta 14 ), and hypertriacylglycerolaemia( Reference Heidemann, Scheidt-Nave and Richter 12 ) are the most frequently metabolic disturbances related to the ‘Western’ dietary pattern. Moreover, a 28 % higher risk of mortality from CVD was identified in women at the upper quintile of adherence to the ‘Western’ pattern in comparison with those at the lowest quintile( Reference Heidemann, Schulze and Franco 7 ). Considering that the ‘Modern’ dietary pattern was markedly related to lower age, it is possible that our study did not have enough statistical power to detect this association in a subgroup at lower risk of CVD disease.
The present study has some methodological features that must be outlined. Applying the ESEM analysis to empirically derive dietary patterns allowed testing the significance of the factor loadings in lieu of applying pre-determined cut-off points to the factor-loading matrix, as performed in EFA and principal component analysis. Taking the present data, a great number of significant food groups would not be considered if a factor loading cut-off ≥|0·25| or ≥|0·30| was applied, for example. Therefore, the ESEM analysis has the advantage to ensure methodological strictness to dietary pattern investigation by reducing the subjectivity during modelling.
In addition, the SEM allowed the estimation of indirect effects of the dietary patterns on metabolic CVD risk factors, which contributed to expand the body of evidences towards the association of the dietary patterns with metabolic CVD risk modulation. It should be mentioned, however, that, although the conceptual model of this study showed an acceptable fit, it does not rule out other models with equal goodness-of-fit that underline a different pathway for the relationship between dietary patterns and metabolic CVD risk factors.
Limitations of this study include the cross-sectional design that does not allow inferring the causal relationships between dietary patterns and metabolic CVD risk factors owing to the chance of reverse causality. Moreover, other variables not evaluated in this study may exert some effects on both dietary patterns and metabolic CVD risk factors (e.g. socio-economic status, educational level, physical activity, smoking habits, genetic variations). Therefore, further studies in longitudinal settings that control for other variables should be performed to confirm the present findings.
Furthermore, the dietary patterns derived were based on data collected by two non-consecutive 24-HDR, which are prone to measurement error, memory bias and to a large within-person variation of the dietary estimates( Reference Tucker 74 ). To overcome these aspects and to enhance data reliability, procedures were applied to structure data collection following the multiple-pass method, to control for the quality of the 24-HDR during data collection, to remove under- and over-reporters of energy intake and to adjust for the within-person variation of the food group intakes.
In conclusion, the ‘Traditional’ and ‘Prudent’ dietary patterns, derived by ESEM, were negatively associated with metabolic CVD risk factors among Brazilian adults. Their apparent protective effects against obesity and high BP may be important non-pharmacological strategies for the prevention and control of obesity-related metabolic disorders and CVD.
Acknowledgements
The authors are grateful to the funding bodies for the financial support that enabled the development of this study. The data collection of the Health Survey of the City of São Paulo was supported by the following public funding institutions: National Council of Scientific and Technological Development (CNPq; grant number 473100/2009-6), State of São Paulo Research Foundation (FAPESP; grant number 2009/15831-0) and Department of Health of São Paulo (no grant number). CNPq also granted the scholarship to the first author (grant number 143115/2011-1).
All the authors participated sufficiently during manuscript preparation to take public responsibility for the article content. M. A. C. designed the study’s analytic strategy, performed all statistical analyses and drafted the manuscript. V. T. B. provided statistical expertise and collaborated with manuscript preparation. R. M. F. coordinated data collection. D. M. M. and R. M. F. supervised statistical analysis and manuscript preparation and performed critical revision of the manuscript.
The authors declare that there are no conflicts of interest.
