This study was approved by the Committee of Ethics and Research of the Federal University of Rio de Janeiro and informed consent was obtained from all participants.
Study design and settings
A population-based cohort multilevel study on social capital, health behaviours and health outcomes in a representative sample of pregnant and postpartum women was conducted in two cities in the state of Rio de Janeiro, Brazil. The cities were deliberately selected based on proxy measures for social capital that included violence rates and per capita income [23] and according to demographic indicators. Per capita income for the selected cities with high and low social capital was U$222 and U$101 for city 1 and city 2, while homicides rates was 0.8 (city 1) and 7.4 (city 2) per thousand for the period of study. The population size of the cities was similar (<300,000 inhabitants) and the birth rates were between 130 and 170 births per 100,000 inhabitants [24, 25]. The coverage of antenatal care was greater than 90 % in both cities and antenatal care was provided in few health care units facilitating the recruitment of a representative sample of pregnant women in both cities.
The selection of the sample was conducted at the four main public prenatal care units of the cities where 95 % of the antenatal care in both cities are provided [26].
Pregnant women attending prenatal care in public units are predominantly from moderate and lower social classes. Based on information from the Department of Health of the two cities, pregnant women from 46 neighbourhoods used the antenatal care units selected for this study. Therefore, each woman was allocated to a neighbourhood area, according to residential zip code, which is a good reflection of a neighbourhood geographic area [27].
Study participants and periods of study
The participants were pregnant women who had sought prenatal care at public health care units administered by the Brazilian National Health Care System. Individual face-to-face interviews were carried out to collect neighbourhood and individual-level primary data between October 2008 and December 2009 [26, 28]. The baseline study was conducted at the antenatal care units during the first trimester of pregnancy and the follow-up data was collected at 30 days postpartum and related to the last 6 months of pregnancy at women’s houses.
Study power calculation
The formulae for the study power calculation using the method for proportions with cluster randomization [29] was used to estimate the minimum statistical difference between groups considering the observed sample size of 1046 women in 46 neighbourhoods and 23 as the observed average number of women per neighbourhood. The sample intracluster correlation was based on intraclass correlation coefficient of 0.117. The minimum differences of prevalence of cigarette smoking during pregnancy, inadequate diet and alcohol consumption at baseline to be detected between areas with low and high social capital were calculated considering the respective observed prevalences (18.1 %, 53.4 %, 7.6 %). Considering a 80 % power and a significance level of 5 %, the sample size used in this study was able to detect at least 20 % of the differences of the health compromising behaviours.
Inclusion criteria
The inclusion criteria were women in the first trimester of pregnancy, living at their current address for at least 12 months and who did not change the address by the follow-up study. The two latter criteria were used to assess the effect of social capital on health-compromising behaviours since the neighbourhood effects and individual social capital tend to be stable after some months living in the same place. The interviewers inspected the medical records and all eligible pregnant women were invited to participate. The women were informed about the objectives of the study and their participation was requested before interview. Women who had a miscarriage or abortion were excluded.
Reliability and consistency
Reliability and internal consistency of the social capital scales were assessed through intraclass correlation coefficient and Cronbach’s α in the test–retest study at 15-day interval. A pilot study (N = 130) was performed to test the questionnaires.
The Cronbach’s α coefficient of the social capital questionnaire was 0.684. In the confirmatory factorial analysis, all items of social capital questionnaire loaded coefficients higher than 0.30. Psychometric characteristics of the social capital scales were published elsewhere [28].
Health-compromising behaviours
The investigated health-compromising behaviours were cigarette smoking, alcohol consumption and inadequate diet. The participants reported current smoking, current alcohol consumption and fruits and vegetable intake at baseline. Frequency of fruits and vegetable intake, smoking and alcoholic beverage consumption during the follow-up period were also recorded. Current smoking was assessed with the following question: ‘Are you a current smoker? (response options: Yes/No)’. The question assessing alcohol consumption was ‘Do you drink alcoholic beverages? (response options: Yes/No)’. These items were tested and used in a study in the same population and shown to be valid [30].
Inadequate diet was evaluated based on the weekly frequency of intake of fruit and vegetables. The items were from the Brazilian national survey on the prevalence of risk factors for chronic diseases in adolescents conducted in 2009 and are considered valid measures of a healthy diet [31]. Pregnant women reporting fruit and vegetable intake lower than five times a week were considered to have inadequate diet [31, 32].
Participants were divided into four groups according to the pattern of each health-compromising behaviour at baseline and follow-up periods. Women who maintained the health-compromising behaviours in both the baseline and follow-up periods were the “stable risk behaviour group”. The “stable healthy behaviour group” included women who did not have the health-compromising behaviours at baseline or follow-up. For example, women who did not smoke throughout gestation. Women who changed their behaviour were classified as “positive behavioural change group” when they changed the behaviour between baseline and follow-up in a positive way, i.e. they stopped smoking during pregnancy. The “negative behavioural change group” were those who adopted the health-compromising behaviour after baseline. For example, adopted a inadequate diet during pregnancy.
Neighbourhood and individual social capital
Neighbourhood and individual social capital measures were calculated using valid instruments at baseline [33, 34]. Neighbourhood social capital refers to the relationships between social groups and their neighbourhood and is largely based on day-to-day interaction between neighbours [15]. The neighbourhood social capital questionnaire was adapted from a previous study in Brazil and included four dimensions confirmed by factorial analysis: social trust, social control, neighbourhood security and political efficacy [34]. Two core sets of questions from Sampson’s seminal paper on collective efficacy were employed to measure social trust and social control [35] and from Stafford et al. [36]. The social trust measures included from the latter study were if people were comfortable asking a neighbour to collect prescription if ill in bed, to lend a small amount of money or confiding about a personal problem. Included items related to social control were related to people’s reaction if they see children ditching classes, people vandalising things or fighting and treating each other with respect [36]. Items relating to political efficacy were from the American and British Political Action Surveys [37], and frequency of violent episodes in the neighbourhood was used to assess neighbourhood security [35]. As each subscale of the social capital questionnaire consisted of different numbers of items, the final scores for each subscale were standardized from 0 to 100. In this way, subscales were comparable and could be added up to form the neighbourhood social capital variable. The score of social capital was computed at the individual level and then aggregated at neighbourhood level. Participants were grouped into 46 neighbourhood areas: 28 neighbourhoods in City 1 and 18 in City 2. The neighbourhoods were then categorized into three equal groups according to tertiles of the social capital score [29] as follows: low (from 32.08 to 41.63), moderate (from 41.64 to 45.93) and high (from 45.94 to 58.51) neighbourhood social capital. A higher proportion of neighbourhoods in the city with greater violence and low income (city 2) were categorised as low in social capital.
Individual social capital was assessed by the levels of social support and social networks. Social support was measured using a social support scale, which consists of 19 items comprising five dimensions of functional social support: material, affective, emotional, positive social interaction and information [33, 38]. Higher score of social support indicates more support. Social support score was multiplied by 10 on the log scale, so that they indicate a change in the outcome variable for every increase of 10 points in the scale. Social networks are considered as the ‘web’ of social relationships surrounding the individual as well as their characteristics, or groups of people they have contact with [39]. Social networks were assessed based on the number of friends (0–1, 2+) and family members (0–1, 2+) that participants reported that they could talk to openly about any topic.
Covariates
The covariates were demographic and socioeconomic data collected at baseline.
Contextual covariate
Social class was evaluated using an economic classification commonly used in Brazil that comprises a group of indicators based on market power and level of education of the head of household [40]. A final score was obtained using a set of points assigned to these indicators which defines the socioeconomic groups; A (highest), B, C, D, and E (lowest). Those with the highest scores represented the highest socioeconomic groups. Because of the small number of observations in classes A and E, data were categorized into three groups: high (A + B); moderate (C); low (D + E). Social class was also aggregated at the neighbourhood level. Neighbourhoods were categorized as low, moderate and high socioeconomic status, based on the tertiles of the distribution of subjects into high social class.
Individual covariates
Individual maternal socioeconomic and demographic characteristics included marital status (married, living with a partner; has a partner, not living with him; single without a partner), number of children (1 child; 2 children; 3 or more children), years of schooling (0–4; 5–8; 9 or more years), family monthly income (<1 Brazilian Minimal Wage (BMW. One BMW was US$ 178.00 at the time of data collection); 1 or more BMW), occupational context (no paid work – women with no paid work, housewives or unemployed women; paid work – employed women with paid work), age (13–19; 20–30; more than 30 years) and ethnicity. The latter variable was assessed through the self-reported skin colour method as proposed by the Brazilian Institute of Geography and Statistics. Participants were asked to describe their skin colour using the following options: white, brown and black [41].
Statistical methods
Clustering analysis
The clustering of the three health-compromising behaviours before and during pregnancy was examined. Clustering of behaviours existed when the observed combination of behaviours exceeds the expected prevalence of the combination. The expected prevalence of a specific combination of behaviours was calculated on the basis of the individual probabilities of each behaviour based on their occurrence in the study population [19]. The observed/expected ratios were examined by calculating the prevalence odds ratios (POR) and the 95 % confidence interval based on a Poisson distribution [42].
Multilevel analysis
This study investigated the association of neighbourhood and individual social capital on health-compromising behaviours before and during pregnancy. The multilevel structure of analysis included 1057 (baseline) and 1046 (follow-up) women (level 1) grouped into 46 neighbourhoods (level 2). A two-level random intercepts and fixed-slopes model structure with individuals nested within neighbourhoods was fitted.
Five outcomes of health-compromising behaviours were considered as follows: (i) number of behaviours at baseline, (ii) number of behaviours at follow-up, (iii) pattern of smoking between baseline and follow-up, (iv) pattern of alcohol consumption between baseline and follow-up, and (v) pattern of diet between baseline and follow-up. Multilevel nested ordered (e.g., number of health related behaviours) and unordered (e.g., pattern of health related behaviours during pregnancy) multinomial logistic regressions, adjusted for confounders were carried out.
Number of health-compromising behaviours was a four-level ordinal outcome, namely none, 1, 2 and 3 risk behaviours, and ordered logit models were used to estimate the cumulative distribution probabilities of the response categories. The reference group was “no health-compromising behaviour”. Coefficients estimated in these models indicated the likelihood of moving into a higher category of number of behaviours. The cumulative response probabilities were modelled, and the proportional odds (cumulative logits) for the three categories presented in relation to independent variables.
“Stable risk behaviour” and “positive behavioural change” were nominal outcomes investigated concerning the patterns of health-related behaviours during pregnancy. They were compared with “stable healthy behaviour” and “stable risk behaviour”, respectively. Unordered logit models were used to estimate the distribution probabilities of each of the response categories.
Fixed- and random parameter estimates for the two-level ordered logit models were calculated by marginal quasi-likelihood (MQL) procedures with first-order Taylor series expansion, RIGLS (restricted iterative generalized least squares) estimation method, as implemented within MLWIN software version 2.24.
The results of multilevel analyses are presented as odds ratios (ORs) with 95 % confidence intervals (95 % CI). In these analyses, variables that presented P ≤ 0.10 in bivariate analysis were considered for multivariate analysis. Four models were tested for each outcome. The association between neighbourhood social capital and neighbourhood socioeconomic status (social class) and the health-related behaviours outcomes was tested in Model 1. Individual-level social capital measures (social support and social network) was added in Model 2. Individual-level sociodemographic confounders described in the theoretical model (Fig. 1) were identified from previous studies in pregnant women [9–12]. Socioeconomic factors (marital status, number of children, years of schooling, family monthly income and occupational context) were inserted in Model 3 and demographic characteristics (age and ethnicity) inserted in Model 4. Independent variables of each block were adjusted for each other using backward selection method. Those that remained significant at 5 % (P ≤ 0.05) were retained in the analysis for adjustment in the next model.