Focused Principal Component Analysis: a graphical method for exploring dietary patterns Análise de Componente Principal Focada: um método gráfi co para explorar padrões alimentares

The aim of the present study was to introduce Focused Principal Component Analysis (FPCA) as a novel exploratory method for providing insight into dietary patterns that emerge based on a given characteristic of the sample. To demonstrate the use of FPCA, we used a database of 1,968 adults. Food intake was obtained using a food frequency questionnaire covering 26 food items. The focus variables used for analysis were age, income, and schooling. All analyses were carried out using R software. The graphs generated show evidence of socioeconomic inequities in dietary patterns. Intake of whole-wheat foods, fruit, and vegetables was positively correlated with income and schooling, whereas for refined cereals, animal fats (lard), and white bread this correlation was negative. Age was inversely associated with intake of fast-food and processed foods and directly associated with a pattern that included fruit, green salads, and other vegetables. In an easy and direct fashion, FPCA allowed us to visualize dietary patterns based on a given focus variable. Food Consumption; Principal Component Analysis; Nutritional Epidemiology Introduction While dietary patterns reflect an individual’s food preferences, they are also influenced by other characteristics, such as economic history, income, schooling and demographic characteristics (sex and age). The statistical methods most commonly used for identifying dietary patterns among populations, or among specific population groups, include data reductions based on a posteriori models, such as cluster analysis and principal component analysis (PCA) 1. Both clustering and PCA are able to identify underlying structures among different food items, i.e. patterns of reduction and clustering of the dataset. However, investigating the relationship between the dietary patterns identified by these methods and population characteristics requires subsequent dependence analysis, which entails resorting to multivariate regression models that include both the dietary pattern and other characteristics of the sample. In Brazil, a limited number of studies have attempted to identify dietary patterns and their association with population characteristics. The findings from these studies are consistent in that they indicate the existence of socioeconomic inequities in the dietary patterns of the population. Lenz et al. 2 identified five dietary patterns by means of PCA. Of these, three displayed the characteristics of a healthy diet, such as the presARTIGO ARTICLE Canuto R et al. 2150 Cad. Saúde Pública, Rio de Janeiro, 26(11):2149-2156, nov, 2010 ence of fruits, vegetables, whole-wheat/wholemeal bread, brown rice and nuts, and two were composed of unhealthy forms of carbohydrates and fats. Subsequent multivariate analysis using Poisson regression showed that healthy patterns were more likely followed by women with higher income and schooling, and by older women 2. Olinto et al. 3, also using PCA followed by Poisson regression, identified dietary patterns among young adults and found that healthy patterns were associated with female sex and higher socioeconomic status. Focused Principal Component Analysis (FPCA) has emerged as a novel method for a posteriori exploratory analysis that is appropriate for scenarios in which explanations are sought for the relationships among a group of variables based on a given characteristic of the sample. Applying FPCA to food data makes it possible to view the correlation between diet and a given variable of interest, at the same time as enabling detection of correlations between the different food items themselves. In FPCA, unlike in PCA, dietary patterns focusing on a particular variable of interest are formed, and are presented exclusively in graphical format 4. With this in mind, the aim of the present study was to present FPCA analysis by applying this methodology to one food intake dataset. As an example, we will use a demographic variable (age) and two socioeconomic variables (income and schooling) as the focus variables.