Data source
This was a secondary data analysis from the Bangladesh Urban Health Survey (UHS), 2013. The main objective of this survey was to provide detailed information on key health outcomes and service utilization indicators in slums and non-slums areas of City Corporations, and other urban areas in Bangladesh. Fieldwork for this survey started on July 23, 2013 and ended on December 12, 2013. Five types of questionnaires were used in the 2013 UHS survey: a men’s questionnaire, a community questionnaire, a women’s questionnaire, a verbal autopsy questionnaire and a household questionnaire. The UHS, 2013 data were downloaded from the University of North Carolina website at https://dataverse.unc.edu .
Sampling design
The main purpose of sampling design of UHS, 2013 was to find key representative indicators for slum and non-slum populations in the 9 city corporations; and district municipalities and large towns referred as other urban areas. The sample collected in this survey is nationally representative. Under-five mortality and percentage of birth deliveries in health facilities for all births in the last three years were used as the key indicators for determining the sample size. For estimating “under-five mortality”, the sample size for non-slums was estimated to be 27000 households which was expected to obtain an estimate within a margin of error of 4 per 1000 births (95% CI of 19–27) [1]. With similar precision, the sample sizes for other urban areas and slum areas were estimated to be about 11,040 and 15,750 households to obtain estimates within a margin of error of nine (95% CI of 41–59) and eight (95% CI of 55–71) per 1,000 births, respectively. Similarly, for estimating “Percentage of facility deliveries”, the sample sizes were 15750, 11040 and 9000 households within margin of errors 3.4, 5 and 6 percentage points for slums, other urban areas and non-slum areas, respectively.
The sampling frame of this survey was a complete list of urban Mohallas in the 9 city corporation and other areas from the 2011 census. The 9 city corporations were Dhaka metropolitan area, Khulna, Rajshahi, Chittagong, Barisal, Comilla, Rangpur, Narayangang and Sylhet. A 3-stage stratified sampling procedure was used for the collection of data of the UHS, 2013. There were two strata: city corporations and other urban areas. In the first stage, the numbers of randomly selected Mohallas in city corporations and other urban areas were 450 and 184, respectively. A mapping activity was used to separate slum and non-slum clusters in each Mohalla. At second stage, two slum clusters and one non-slum clusters were randomly selected from each Mohalla. In the case of other urban strata, two clusters were randomly selected from each selected Mohalla. In the last stage of selection, households of the selected clusters were randomly selected by using a household listing activity.
Participants
The ever married women who lived in urban slums or non-slums were questioned about their last birth in the preceding three years of the UHS in order to minimize recall bias. This allowed for analysis of 6142 live births. The final sample size was 6137 after excluding the five women who had missing data regarding the number of ANC visits.
Response variable
Women were considered as having complication [11] if they had experienced any of the following five problems during their last pregnancy, during or after delivery: convulsion/fits, severe/heavy bleeding, prolonged labor (> 12 h), high fever with smelly discharge, and oedema face/feet/body. Thus, the response variable, complications variable was coded as 1, if respondent had any of the five complications and 0, if she did not.
Exposure variables
Based on the available literature review, the following socioeconomic, demographic and individual fertility variables were considered as exposure variables in our analysis: wealth index, education level of mothers, access to media, NGO membership, region, number of children ever born, pregnancy intention at last birth, mother’s age at last birth, multiple last birth, delivery by medically trained provider (MTP), sex of last child, place of residence, migration status, at least 4 ANC visits and place of delivery. A woman was considered as migrant if she came to urban area from other city or non-urban area. However, the variables “access to media” and “NGO member” were not directly obtained from the survey data. These two variables were created from the available information in the survey data as follows: a woman was considered as exposed to media if she usually read a newspaper or magazine/ listened to radio/watched TV and she was treated as an NGO member belonging to any of the following organizations: Grameen Bank, BRAC, BRDB, ASHA and PROSHIKA.
Statistical analyses
The UHS, 2013 data were collected from 1718 clusters: 450 in slums, 900 in non-slums and 368 in other urban areas. This clustering might result in the correlated responses. To allow a wide variety of correlation pattern to be explicitly modeled, a generalized linear mixed model (GLMM) was applied. Chi-square test of association was used in bivariate analysis to determine whether there existed significant association between the outcome and exposure variables. In the regression analysis, the crude odds ratios (CORs) and adjusted odds ratios (AORs) were calculated by exponentiating the unadjusted and adjusted effects of the covariates obtained from simple and multiple mixed logistic regression models, respectively.
Suppose, Yik and Xik are the outcome variable and p × 1 vector of exposure variables, respectively taken from the kth (k = 1, 2, …, ni) individual of the ith (i = 1, 2, …, m) cluster. Also assume that β = (β1, β2, …βp) is the p × 1 vector of regression parameters. Then the GLMM for the binary response, Yik can be written as
$$ {\eta}_{ik}=\mathit{\ln}\frac{\pi_{ik}}{1-{\pi}_{ik}}={x}_{ik}^{\prime}\beta +{v}_i $$
where πik = E(Yik| vi) and vi is the random intercept term that is assumed to follow normal distribution with mean 0 and variance \( {\sigma}_v^2 \). Therefore, the conditional log-likelihood function can be written as
$$ l\left(\beta, {\sigma}_v^2|{x}_{ik},{v}_i\right)=\sum \limits_{i=1}^m\sum \limits_{k=1}^{n_i}\ln \left(f\left({y}_{ik}|{x}_{ik},{v}_i\right)\right) $$
The estimates of the parameters β and \( {\sigma}_v^2 \) can be obtained by maximizing the following marginalized likelihood function
$$ L\left(\beta, {\sigma}_v^2\right)={\int}_{-\infty}^{\infty}\left\{\prod \limits_{i=1}^m\prod \limits_{k=1}^{n_i}f\left({y}_{ik}|{x}_{ik},{v}_i\right)\right\}d{v}_i. $$
The intra-cluster correlation coefficient, ρ can be measured by the variance component as [14] \( \rho =\frac{\sigma_v^2}{\sigma_v^2+\frac{\pi^2}{3}} \).