Data analysis
The data was analysed in STATA v12.1 using Pearson’s chi-square test to describe associations between categorical variables for each site. Bivariate and multivariate logistic regression analyses were used to test for associations between the outcome variables: (i) timing of ANC initiation: before 16 weeks (early) versus after 16 weeks (late) gestational age and (ii) frequency of attendance: at least 4 visits (adequate) versus less than 4 visits (inadequate) against the psychosocial variables (substance use, feelings about pregnancy, social capital, social support, cultural beliefs, mental health perceptions, self-esteem and partner characteristics).
Multiple logistic regressions were used to model the effects of the psychosocial variables on the outcome variables. A predictive modelling strategy was used as follows: Variables other than psychosocial variables that were potentially associated with ANC attendance such as demographic factors (age, level of education, marital status, employment status and race), parity and previous miscarriage, were selected by looking at the reduction in the deviance due to adding each potential ‘confounder’ to a model with only the outcome variable to form a baseline model with all relevant ‘confounders’.
Psychosocial variables were then added one at a time to the baseline model and their significance was assessed by looking at the reduction in the deviance. Interactions between the independent variables and confounders were assessed for any change in the effect on the outcome variable. The models were then compared using likelihood ratio chi-square statistics and the model with the lowest Aikakes Information criterion (AIC) was selected as the best model. The urban and rural sites were treated as separate strata in analysis rather than combined in one multivariate analysis because previous research had suggested substantial cultural and social differences in these two populations [17]. To adjust for clustering at the rural site, the multivariate analysis was repeated using the survey command ‘svy’ in STATA. Modeling was done separately for the urban and rural sites. Two final models were therefore selected per site, for (i) timing of ANC initiation and (ii) frequency of ANC attendance for the urban and rural sites.