Brazilian Longitudinal Study of Adult Health (ELSA-Brasil) participant’s profile regarding self-rated health: a multiple correspondence analysis

This study used baseline data (2008–2010) from the ELSA-Brasil study. The ELSA-Brasil is a longitudinal multicentric cohort study of 15,105 civil servants (35–74 years) conducted at six study research centres in three regions of the country, including the Northeast, South, and Southeast. These research centres are located in five federal universities and the Oswaldo Cruz Foundation [19, 20].

The present study did not use information about retired participants, since they do not have information about occupational characteristics (socio-occupational category and psychosocial work environment). Also, participants that declared their race/ethnicity as Asian or Indigenous were excluded due to the small number of participants in each category (2.4 and 1%, respectively). Furthermore, the information of participants that declared their race/ethnicity as Asian is mainly centered in one of the research centres in São Paulo. The exclusion of Indigenous people was made considering that our participants are urban indigenous in a small number, and they do not represent the indigenous population. Finally, participants who had missing data for any of the study variables were also excluded (Fig. 1).

The variables were age, sex, self-declared race/ethnicity (white, brown and black), marital status (married/united, non-married/non-united – this category includes single, divorced or separated, and widowed people), education (the categories were complete elementary school or less, completed high school, and completed university degree or more. This variable considers the highest level of completed education, with exception of elementary school, i.e. a participant that did not completed high school was considered in the complete elementary school or less category), month per capita household income (low - up to $234, medium - from $234 to $702, and high - more than $702. The cutoff points of this variable were based on the 2008 minimum wage in Brazil and the median income of our population. The low category considers one salary, the medium category considers from 1 to 3 salaries, and the high category considers more than 3 salaries. The median income of our population was $702), and socio-occupational category (manual, middle and higher). This last variable considers different forms of insertion in production (considering the position in the typical occupation), the occupation itself (which non-manual was qualified by the level of formal education required by the occupation and manual was qualified by the sector specialization), and the hierarchy in production. The categories were defined by the Center for Development and Regional Planning (CEDEPLAR), Faculty of Economic Sciences of the Federal University of Minas Gerais (UFMG), based on the literature [21, 22].

The variable representing the psychosocial work environment was job stress (demand–control model). Job stress was accessed using the Swedish demand control support questionnaire (DCSQ). This questionnaire contains 17 items, five items refer to the psychological demand dimension, six items refer to the control dimension, and six items refer to the social support dimension. In our study, the repetitive work item (control dimension) and the social support dimension were not considered since the study about the dimensional structure of the DCSQ in the Brazilian context suggests this item exclusion and a better goodness-of-fit without the social support dimension [23]. The scores of the DCSQ (job demands, 5 items; skill discretion, 3 items; and decision authority, 2 items) were dichotomized into high and low at the median for these dimensions (14, 11, and 6 points, respectively) [24].

The health variables used in analyses were self-rated health and body mass index. Self-rated health (SRH) was measured using the following question: “In general, compared to people of your age, how do you consider your state of health?”. The response options were: “very good, good, fair, poor, or very poor”. For the analyses, the answers were categorized as good self-rated health (very good and good), fair, and poor (poor and very poor). The body mass index (BMI) cutpoints were considered as: ≤ 24.9 kg/m² for underweight and normal weight (the categories of underweight, ≤ 18.5 kg/m², and normal weight were grouped due to the small number of participants who were underweight, < 1%), between 25 and 29.9 kg/m² for overweight, and ≥ 30 kg/m² for obesity.

Proportions were used to describe population characteristics regarding self-rated health. Self-rated health, sex, self-declared race/ethnicity, marital status, education, per capita household income, body mass index, socio-occupational category, and job strain were analyzed by multiple correspondence analysis (MCA). Stratified analyses by age were conducted due to our consideration of aging as an effect modifier [7, 12, 25]. Since our average population age is 49.14 years, two age groups were created to stratify the analyses (up to 49 years old and 50 years old or more).

Correspondence analysis is an exploratory technique applied to categorical data. This analysis graphically illustrates the relationship within one set of variables, and the proximity of categories, in space, indicates a relationship or correspondence between them [26, 27]. The advantage of this statistical method is the absence of any assumption about probability distributions and the lack of a need to establish predetermined relations among the variables, such as the unidirectional relationships estimated by regression models. This type of analysis provides total inertia, which means the percentage of variability explained by each dimension, and in this paper, the number of dimensions was chosen by analyzing the decline of adjusted inertias (eigenvalues) [27].

Scatterplots (formed by the coordinates of each category in each dimension) were analyzed with regard to dimensions, and clusters of categories were created to delineate different profiles in the sample. The results based on hierarchical cluster analysis (dendrogram) of the standard coordinates obtained in the correspondence analysis were confronted with the resulting clusters visualized in the multiple correspondence plot. The dendrogram provided a clear visualization of the categories of the variables in each group, and it is more useful in the case of many dimensions in the MCA (which become hard to use biplots).

The x-axis of the scatterplots represents the data variability explained by the first dimension, while the y-axis represents the data variability explained by the second dimension. The dots represent each variables categories.

The analyses were performed in the R software [28], version 3.5.1, library “ca”, “ggplot2”, “dendextend”, and “factoextra”.