Data
For objective 1: For assessing the first objective, we used the third round of the Bangladesh Multiple Indicator Cluster Survey (BMICS) 2019 data- the most recent publicly available nationwide dataset. The key variables collected in BMICS 2019 included housing characteristics, women and child’s sociodemographic characteristics, reproductive and maternal health indicators (ANC, facility birth, postnatal care), child health, nutrition and development (malnutrition, disease episodes, early childhood development) and etc. BMICS 2019 adopted a two-stage stratified cluster sampling technique making the data representative at national, division, district, and rural and urban level [17]. The urban and rural areas within each district formed the 128 strata from which 3220 clusters were chosen with probability proportional to size. Then a systematic sample of 20 households was selected from each cluster. BMICS 2019 was designed to provide estimates for a large number of indicators on the situation of children and women. BMICS interviewed 64,378 women from 61,242 households with a response rate of 93.7%. The key indicator used for calculating the sample size was the proportion of women with at least four ANC visits among women aged 15–49 years with a live birth in 2 years preceding the survey. BMICS considered the proportion to be 35% and survey estimate was 37% from a sample size of 9,285 which yields an estimated power to be 98%. This also yields a sufficient sample size for estimating other MHC indicators like facility birth, and post-natal care. The detailed sampling procedures of BMICS 2019 can be found in the survey report [17].
For objective 2: We explored the schooling status of male adolescents using data from the Bangladesh Adolescent Health and Wellbeing Survey (BAHWS) 2019–20- the first national-level survey on adolescents aged 15 to 19 in Bangladesh [27]. We included all 5,523 male adolescents aged 15 to 19 years interviewed in BAHWS 2019–20.
Participants
For objective 1: The BMICS 2019 collected ANC and facility birth information from 9,285 women who had their last birth two years preceding the survey. As we aimed examining the role of HHs’ education on women’s MHC usage, women living in households where there are HHs who may influence women’s MHC uptake will be the target population. Women who themselves are the HHs will be independent from the influence of HHs and thus will not be in our target population. So, we excluded 333 women who were the head of their households. This exclusion will not introduce any biases to our estimates, because these women did not meet the criterion of target population.
ANC, facility birth, and education of HH were missing respectively for 5,1, and 3 women. As the sample size was sufficient enough for estimating ANC and facility birth, we excluded these 9 women from the analyses. Exclusion did not introduce potential bias in the estimates. If all these 9 women had 4 ANC visits, and facility birth, the estimated prevalences would change only at the second decimal (for ANC: without exclusion 36.96% vs after exclusion 36.89% and for facility birth: 53.54% vs 53.50%).
Finally, the analytical sample includes 8,943 ever-married women aged 15–49 years who gave birth two years preceding the survey.
Outcome measures
For objective 1: The two primary outcomes of interest were at least four ANC uptake and IDS usage for the most recent birth, which we dichotomized in the following way:
-
1.
At least four ANC visits: yes (if a woman had at least four ANC visits during her last childbirth), no (if a woman had no or at most three ANC visits during her last childbirth).
-
2.
IDS usage: (if a woman gave her last birth in institutional settings other than home), no (if a woman gave her last birth at her, relatives’ or others’ home).
For objective 2: To explore the educational status of male adolescents we looked at three indicators: school dropout, educational level completed before school dropout, and reasons for dropout.
Covariates
Covariates of interest: A pregnant woman is not the decision maker of MHC usage solely. Usually, HH controls the power dynamic of the household. Thus, if HHs are unaware of the benefits of MHC usage, they may not allocate the financial or human resources of the households for MHC usage. Educated HHs may have greater health literacy [28] and more egalitarian attitudes toward women’s autonomy in decision-making [25]. Earlier studies from Uganda and Bangladesh showed that HHs’ education helps improve MHC usage [25, 29]. Hence, keeping in mind the study objectives, the main covariate of interest was the education level of HH which we categorized as follows: none or pre-primary, primary, secondary, and above secondary.
Other covariates: We used Penchansky and Thomas’s Theory of Healthcare Access (ToHA) modified by Saurman [30] for other covariate selection. The six domains of the ToHA are accessibility, availability, acceptability, affordability, adequacy, and awareness. We conceptualized four broad groups of factors- household characteristics, women’s individual-level factors, women’s birth history, and contextual factors that may influence the six domains of the ToHA. Here one group of factors may influence more than one dimension of the ToHA. For example: the contextual factor “type of residence” can influence accessibility, availability, and acceptability domains respectively through transportation system, number of facilities, and cultural norms of using the facilities. Factors under each of the four broad groups were selected based on the earlier literature from Bangladesh [21, 24, 25, 30], India [31,32,33,34], Pakistan [35,36,37,38], and Nepal [35, 39,40,41]. Household characteristics include sex, age, religion of the HH; relationship with the HH; and household wealth status. Women’s individual-level factors comprise educational status, media exposure, and age at index birth. Women’s birth history contains the birth order and sex of the index child, and history of child death. Lastly, contextual factors include the place of residence and administrative division. We presented a detailed description of all the covariates in Appendix Table A.2.
Confounders: HHs’ gender; religion, wealth, and residence of the household; and women’s education have contextual rationale to be confounders. In Bangladesh context, male HHs are more educated than female HHs, but women from male-headed households are less likely to use MHC probably because male heads recognize women’s needs less than female heads do. Women from non-muslim, wealthy, urban households are more likely to have educated HHs and higher chances of MHC usage than the rest. Women’s education can also determine HHs’ education because educated women are more likely to be married in educated households. The positive role of women’s education in MHC usage is also recognized.
Statistical analysis
Univariate statistics were used to understand the sociodemographic context of the study participants. Sampling weights and survey design characteristics of the corresponding survey were incorporated to reduce the bias from the estimates and produce robust standard error of the estimates. All the analyses were done using Stata version 14.0 (Stata SE 14, Stata Corp, College Station, TX, USA) and R (V.4.2.2, RStudio2023.09.0).
Analysis plan under objective 1
Model specification
BMICS 2019 used a cluster sampling design. Due to cluster-level heterogeneity in customs, beliefs, and infrastructural settings, MHC usage within a cluster can be correlated and so different clusters may start from different baseline levels. Hence, we used a mixed-effect logistic regression model considering random intercept at the cluster level to examine the association of HHs’ educational status with at least four ANC uptake and facility birth. This can deal with the hierarchical nature of the clustered data by incorporating both fixed effects and random effects at the cluster level. These random effects capture the variation across clusters by explicitly modeling the clustering structure and fixed effects represent the average relationship between the predictors and the response variable across all clusters.
If there is enough sample size, multivariable regression is a mathematical model that estimates an adjusted coefficient that controls for multiple confounders and covariates simultaneously [42]. Hence, we used multivariable regression models to adjust for potential confounders. The first model included only the main covariate of interest (HHs’ education). In the 2nd model, we added the confounders, and the 3rd model included other covariates as well. We performed sensitivity analyses for unmeasured confounding using the methodology proposed by VanderWeele and Ding in 2017 [43, 44]. We estimated the E-value, defined as the minimum strength of association, on the odds-ratio scale that an unmeasured confounder would need to have with both the exposure and the outcome to explain away the exposure-outcome association [45].
Marginal effect estimation
Further, we examined whether less educated HHs impede achieving the maximum benefits of ANC in facilitating institutional birth. For this, we estimated the marginal effects of HHs’ education on facility birth. In general, the marginal effect of X on the outcome (Y) is the instantaneous rate of change of Y with respect to X. For binary outcome with categorical covariate, the average marginal effect is the average change in outcome probability when a covariate changes from the reference category. For example: To estimate the marginal effects of a binary variable X, firstly one has to fit the regression model and estimate the predicted probability of the outcome for all the observed observations at the reference category (X = 0). Then subtract these from the predicted probability of outcome for all observed observations at the other category of X (X = 1). The average of these differences is the marginal effect of X on the outcome probability. So, the marginal effect equals 10 means the probability of the outcome will increase by 10% points among respondents with X = 1 than those with X = 0.
In our examination, we first constructed another mixed-effect multiple logistic regression considering random intercept at the cluster level with an interaction of HH’s education and the number of ANC visits. We estimated multiple logit-based marginal probabilities of institutional birth for each education level of HHs across the three groups of comparable number of ANC visits (No ANC visits, 1–3 ANC visits, 4 + ANC visits). Then, we estimated the marginal effects of HH’s education on institutional birth for the comparable number of ANC visits. These marginal effects allow concluding whether low education of HH diminishes the full-length benefits of ANC uptake in elevating institutional birth.
Spatial mapping
Lastly, we estimated the proportion of HHs with no education or primary education for each district and drew the spatial map of these district-wise percentages which allows for identifying the regions that may require immediate involvement of HHs in improving MHC usage.
Analysis plan under objective 2
Using BAHWS 2019–20 we estimated the percentage of school dropouts among male adolescents for each administrative division. Further, we explored the completed years of schooling of these dropouts by division which helps to understand the timing of quitting school. Lastly, we explored the reasons for quitting school.