Socioeconomic and geographic disparities in institutional delivery in Bangladesh: a Bayesian multilevel modelling framework

Abstract

Background

Despite substantial improvements in institutional delivery rates, significant socioeconomic and regional disparities persist in Bangladesh. Addressing these inequalities remains crucial to ensuring equitable maternal healthcare and achieving the Sustainable Development Goals (SDGs).

Methods

This study used nationally representative data from the 2022 Bangladesh Demographic and Health Survey (BDHS), comprising 4,433 women aged 15–49 years. Multilevel binary logistic regression models were estimated within a Bayesian framework via the brms package in R. We specified five hierarchical models—null, individual, household, community, and full—to identify factors associated with institutional delivery. Adjusted odds ratios (AORs) with 95% credible intervals (CrIs) are reported. Model adequacy was evaluated via the widely applicable information criterion (WAIC) and leave-one-out information criterion (LOOIC). Weakly informative priors were applied, and sensitivity analyses using alternative prior specifications were conducted to evaluate the robustness of posterior estimates.

Results

The full model showed a proportional change in variance (PCV) of approximately 59.9%, indicating a substantial reduction in between-cluster variance after adjustment for socioeconomic and demographic characteristics. The overall prevalence of institutional delivery was 63.5%. Antenatal care (ANC) visits and household wealth were the strongest predictors. Women who attended 4 + ANC visits (AOR = 7.83; 95% CrI: 5.51–11.26) and those in the richest wealth quintile (AOR = 2.76; 95% CrI: 1.94–3.89) had higher odds of delivering in a health facility. Higher maternal education (AOR = 2.87) and lower birth order (AOR = 0.57) were significant predictors. Conversely, women in paid employment had 17% lower odds of institutional delivery (AOR = 0.83). Pronounced divisional (e.g., Khulna vs. Sylhet) and significant religious disparities were also observed.

Conclusions

Institutional delivery in Bangladesh is associated with a complex interplay of individual-, household-, and community-level factors, with antenatal care (ANC) visits and household wealth demonstrating the strongest associations. Persistent socioeconomic and geographic disparities, particularly in Barishal, Sylhet, and Mymensingh, highlight the need for targeted, multilevel interventions. MOR = 2.09 and PCV = 59.9% indicate the need for policies targeting cluster-level inequalities, household wealth, and education through combined efforts in poverty reduction and health promotion.

Supplementary Information

The online version contains supplementary material available at 10.1186/s12889-026-26297-5.

Background

Maternal healthcare services play a crucial role in reducing maternal and neonatal mortality worldwide [1–3] and has been a key focus of global development initiatives, first under the Millennium Development Goals (2000–2015) and continuing under the Sustainable Development Goals (SDGs) for 2015–2030 [4]. Institutional delivery remains the most effective form of facility-based maternal care, ensuring the presence of skilled medical personnel to manage complications and safeguard maternal and neonatal health [3, 5]. Evidence suggests that delivery in health facilities substantially reduces maternal and neonatal mortality by enabling safe and timely postnatal care [3, 5]. Despite global progress, an estimated 287,000 women die each year from maternal causes, with the vast majority of deaths occurring in low- and middle-income countries (LMICs), particularly in Southern Asia and Sub-Saharan Africa [6–8].

In Bangladesh, the maternal mortality ratio (MMR) declined markedly, from 437 deaths per 100,000 live births in 2001 to 123 in 2020 [6, 8]. Nevertheless, achieving the SDG target of reducing the MMR to 70 per 100,000 live births by 2030 will require sustained and targeted interventions [9, 10]. Among recent live births, 65% occurred in health facilities [11]. The institutional delivery rate has increased from 15% in 2007 to 65% in 2022, with significant regional disparities [11, 12]. This remarkable progress underscores the influence of socioeconomic development, health service expansion, and targeted maternal health initiatives [3, 9, 13]. However, substantial regional disparities persist, with facility delivery rates ranging from 82% in Khulna to just 51%, 52%, and 54% in Barishal, Sylhet, and Mymensingh, respectively [11, 14]. Rural-urban disparities in maternal healthcare persist due to concentrated shortages of facilities and professionals in rural areas [15]. Socioeconomic inequalities, limited awareness, and financial constraints drive disparities in maternal healthcare utilization in Bangladesh [10]. Economic disparities in maternal healthcare persist, resulting in poorer women relying on inadequately resourced traditional birth attendants, while wealthier women access safer, skilled maternity services [16]. Such variations highlight that, despite national gains, progress has been uneven, and inequities remain entrenched [10, 17].

In 2015, approximately 5,500 maternal deaths were reported in Bangladesh [7]. Despite efforts to increase safe delivery, most mothers in LMICs still deliver at home without professional assistance [18–20]. Covariates at the individual, household, and community levels strongly affect the utilization of healthcare services. Previous studies have examined individual-level variables (e.g., ANC visits [8, 21], maternal age [8], education [8, 22, 23], birth order [7, 9], and working status [7, 24]), household-level characteristics (e.g., wealth index [8, 22] and religion [25]), and community factors (e.g., division [26] and residence type [22, 23]), which collectively influence the use of institutional delivery care.

However, existing studies often examine these determinants using traditional frequentist statistical techniques that overlook the hierarchical structure of survey data [27]. Failing to account for this structure may result in biased estimates due to the assumption of independent observations [28]. Although DHS data are generated using complex survey designs with sampling weights, most existing studies have relied on conventional multilevel models that do not adequately account for these weights, largely due to methodological challenges. This limitation may lead to biased or inefficient estimates. To address this methodological limitation, the present study applies a survey-weighted Bayesian multilevel logistic regression model, which provides more accurate and robust estimates than conventional frequentist approaches in hierarchical data settings [29]. Moreover, the Bayesian approach incorporates prior information, offers greater flexibility in handling uncertainty, and yields more credible interval-based inference [30, 31].

Therefore, this study aims to identify the key determinants of institutional delivery at the individual, household, and community levels, and to examine socioeconomic and regional disparities in Bangladesh using an advanced Bayesian multilevel modelling framework.

Methods

Data sources

This study used secondary data from the 2022 Bangladesh Demographic and Health Survey (BDHS), the ninth nationwide survey on population and health in Bangladesh [11]. The BDHS 2022 aimed to collect comprehensive data on reproductive, maternal, and child health; nutrition; family planning; women’s empowerment; and access to healthcare services. We adopted a two-stage stratified sampling design. In the first stage, 237 urban and 438 rural clusters (enumeration areas, EAs) were selected using probability proportional to size (PPS) sampling. In the second stage, 45 households were systematically selected from each of the EAs. Of these, 30 households were surveyed via the full questionnaire, whereas 15 households received a shorter version focused primarily on biomarker measurement. Women aged 15–49 years with prior marital experience were eligible for individual interviews.

In total, 30,330 households across 674 clusters from all administrative divisions were successfully surveyed. After excluding observations with missing information on key variables, 4,433 women were included in the final analysis. A consort flow chart, showing the process of obtaining 4,433 observations, is included in the supplementary file (Figure S1). Further details about the BDHS methodology and data access are available from The DHS Program.

Study variables

Dependent variable

The dependent variable was the place of delivery, categorized as binary. Deliveries occurring in a medical facility, such as a government or private hospital, an NGO-operated health centre, or a clinic, were coded as 1 (institutional delivery), whereas home or other non-institutional deliveries were coded as 0.

Independent variables

Independent variables were classified into three hierarchical levels: Individual level: maternal age in 5-year groups, employment status, highest education level, number of antenatal care (ANC) visits, birth order, number of living children, media exposure (newspaper, radio, television, internet use), and literacy. Household level: sex of household head, household electricity, ownership of a television, distance to health facility, religion, and wealth index. Community level: administrative division and place of residence (urban/rural).

The variables were regrouped into analytically meaningful categories, e.g., ANC visits, birth order, number of living children, media exposure, literacy, and place of delivery. Earlier guidelines and common practice in Bangladesh have emphasized a minimum of four visits [32]. Therefore, to capture the non-linear relationship between ANC utilization and institutional delivery, this variable was categorized into three groups: no visit, 1–4 visits, and 4 + visits. According to a previous study, birth order was classified as low (1–2) versus high 3+, recognizing that mothers with fewer children may seek care differently than those with many children [10]. Media exposure was dichotomized to reflect any versus no access to health information, consistent with other studies [7, 10, 33]. Literacy was defined as literate versus illiterate, following the standard BDHS definition, ensuring consistency with prior DHS-based studies [34].

Conceptual framework

This study employs a multilevel socioeconomic and regional framework, acknowledging that institutional delivery is influenced by a combination of individual, household, and community-level factors. All three levels are consistent with the socioecological model of health behaviour. Individual factors (e.g., education, use of ANC) interact with household characteristics (wealth index, religion), all within the wider community setting (division, place of residence) [35, 36]. Our Bayesian multilevel model allows us to examine associations across all three levels, while acknowledging the observational nature of the data. It also accounts for differences between clusters, illustrating how these interconnected factors are collectively linked to disparities in institutional delivery.

Statistical analysis

The initial stage involved cleaning the data and handling missing values by selecting cases and deleting those with incomplete data. Data cleaning and initial descriptive analyses were conducted using IBM SPSS Statistics version 27. Bayesian modelling was performed in R version 4.4.3.

Bivariate analysis was first conducted to explore the associations between institutional delivery and each explanatory variable. Cross-tabulations were used to estimate frequencies and percentages [37], and Pearson’s tests were applied to assess statistical significance at [9, 38]. To visualize the relative influence of factors at different levels, a Sankey diagram [31] was generated.

Given the hierarchical structure of the BDHS data, where individuals are nested within households and households within 674 clusters, a Bayesian multilevel binary logistic regression model was employed to account for intra-cluster correlation and unobserved heterogeneity [39].

Bayesian modelling framework

A Bayesian framework was adopted for this study, treating all unknown parameters as random variables. This approach quantifies uncertainty through probability distributions and provides full posterior inference for both fixed and random effects by combining prior information with the likelihood. It is particularly well-suited for quantifying cluster-level heterogeneity using measures such as the intraclass correlation coefficient (ICC), median odds ratio (MOR), and proportional change in variance (PCV), which are not naturally probabilistic under frequentist inference. Variance components and heterogeneity measures are also more straightforward to estimate and interpret within a hierarchical Bayesian setting. In addition, Bayesian models provide a flexible approach for incorporating survey weights in complex datasets, such as those from the DHS. Given the clustered structure of the survey data and emphasis on between-cluster variation, this framework is particularly suitable for identifying actionable determinants and prioritizing interventions to reduce disparities.

The binary response variable representing institutional delivery was modelled via a Bernoulli distribution [27]. The association between the outcome and the predictors was specified through a logit link function as follows:

where is the binary outcome for the individual in cluster , is the intercept, are fixed-effect coefficients for covariates , and is the random intercept for the cluster , which is assumed to follow a normal distribution with mean 0 and variance [27, 29, 40].

To account for the complex survey design, we used a frequency weight expansion approach. The sampling weight (V005) was first normalized by dividing it by 1,000,000 to convert it from a 6-decimal format and then scaled to have a mean of 1. Each observation was repeated according to its rounded, normalized weight, producing an expanded dataset that reflects the national population distribution. This method is statistically equivalent to weighted likelihood estimation, allowing our Bayesian multilevel models to generate estimates that are representative of the target population.

The model was fitted in R version 4.4.3 via the brms package with the CmdStanR backend [41], which implements Bayesian multilevel modelling via the Hamiltonian Monte Carlo (HMC) algorithm with a no-U-turn sampler (NUTS). Four Markov chains were run, each with 2,000 iterations (1,000 warm-up iterations), yielding 4,000 post-warm-up samples. To ensure efficient sampling, the adapt_delta parameter was set to 0.95.

Weakly informative priors were used to regularize the model, allowing the data to drive the estimation. Fixed effect coefficients were assigned normal priors with a mean of 0 and a standard deviation of 2.5, allowing the data to guide the estimation. The intercept and cluster-level standard deviations were assigned Student-t priors with 3 degrees of freedom, location 0, and scale 2.5, providing stability while accommodating variability. Posterior distributions were obtained by combining these priors with the likelihood. Sensitivity analyses were performed to assess the robustness of the estimates by comparing models with different prior specifications, N[0, 5] vs. default priors, confirming that the associations remained stable across models. Results showed minimal differences in effect estimates (AORs), with all variables maintaining the same direction of association and overlapping credible intervals. Convergence diagnostics were good, indicating robust model performance.

Model convergence was assessed via R-hat, Bulk effective sample sizes (ESS), Tail effective sample sizes (ESS), and trace plots. Convergence of the Markov chains was indicated by R-hat ≈ 1, evaluated both bulk and tail ESS ≥ 1000, and visual inspection of the trace plot showed well-mixed chains, indicating good convergence and sufficient for reliable posterior estimates [39, 42]. Posterior predictive checks were performed to evaluate model adequacy and assess model fit by comparing observed data with posteriorly replicated data.

Variance components were estimated via the cluster variance, intraclass correlation (ICC), median odds ratio (MOR), and proportional change in variance (PCV). The size of cluster-level variation and the ICC will be interpreted to understand how much geographic or contextual differences influence institutional delivery. This will help identify whether resource allocation or intervention strategies should be prioritized for specific clusters or regions. Model fit was compared using the Watanabe–Akaike information criterion (WAIC), and the leave–one–out information criterion (LOOIC) was calculated using the LOO package [43]. Inferences were based on the model with the lowest WAIC and LOOIC, indicating the best fit to the data [39]. HMC and NUTS have been used to achieve the posterior distribution of WAIC and LOOIC, specifically for model selection in the Bayesian setting [39, 44]. The results are presented as adjusted odds ratios (AORs) with 95% credible intervals (CrIs) for five hierarchical models: null, individual, household, community, and full models. To visualize the complex, multidimensional nature of disparities in institutional delivery, we constructed a Sankey diagram using the plotly package [45] in R. The width of each node reflects the size of the group, while the links indicate how women move between categories, highlighting the pathways that are most at risk. To illustrate the most influential predictors, a forest plot was constructed to visualize the odds ratios from the Bayesian multilevel logistic regression model [46]. The R code used for data preparation, Bayesian modelling, and generating figures is available in the supplementary materials (Code S1).

Results

A total of 4,433 ever-married women aged 15–49 years were included in this study to identify the determinants of institutional delivery via a Bayesian multilevel approach. Among them, 63.5% delivered in a healthcare facility, whereas 36.5% had non-institutional births.

Bivariate analysis of institutional delivery and associated factors

This section presents the findings of a bivariate analysis examining how institutional delivery varies by individual, household, and community-level factors, using data from the BDHS 2022. The analysis revealed substantial inequalities across socioeconomic, demographic, service-use, and geographic dimensions presented in Table 1.

Table 1.

Associations between respondent characteristics and institutional delivery

Variable name	Category	Frequency	Institutional Delivery (%)	Chi-square	P value
Age in 5-year groups	15-19	549	62.5	20.02	0.003
	20-24	1365	65.4
	25-29	1265	64.9
	30-34	829	63.0
	35-39	348	55.2
	40-44	69	58.0
	45-49	8	25.0
Respondent working	No	3408	65.0	14.75	<0.001
	Yes	1025	58.3
Highest educational level	No education	255	38.8	393.83	<0.001
	Primary	1075	46.5
	Secondary	2314	65.9
	Higher	789	87.2
ANC visits	No visit	381	19.2	512.67	<0.001
	1-4 visits	2995	61.9
	4+ visits	1057	83.7
Birth order	Low	3094	70.1	196.15	<0.001
	High	1339	48.0
Number of living children	Low	3247	69.3	175.61	<0.001
	High	1186	47.6
Reading newspaper	Not at all	4115	62.0	53.84	<0.001
	Sometimes	318	82.7
Listening to the radio	Not at all	4273	63.1	7.13	<0.001
	Sometimes	160	73.8
Watching television	Not at all	2076	51.8	226.62	<0.001
	Sometimes	2357	73.7
Use of the internet	Never	3042	56.6	195.13	<0.001
	Often use	1391	78.4
Literacy	Illiterate	539	39.3	152.81	<0.001
	Literate	3894	66.8
Household electricity	No	66	25.8	39.43	<0.001
	Yes	4367	64.0
Household television	No	2265	52.1	256.68	<0.001
	Yes	2168	75.3
Distance to health facility	Not a big problem	2445	66.3	18.73	<0.001
	Big problem	1988	60.0
Sex of household head	Male	3952	63.6	0.22	<0.001
	Female	481	62.4
Religion	Islam	4052	61.6	79.39	<0.001
	Hinduism	346	85.0
	Buddhist	25	52.0
	Christianity	10	90.0
Wealth index	Poorest	938	38.8	543.15	<0.001
	Poorer	883	54.0
	Middle	866	63.9
	Richer	889	74.9
	Richest	857	87.5
Division	Barishal	475	57.3	179.47	<0.001
	Chattogram	768	60.3
	Dhaka	668	69.2
	Khulna	516	83.3
	Mymensingh	532	51.1
	Rajshahi	439	71.8
	Rangpur	521	62.4
	Sylhet	514	53.3
Residence type	Urban	1463	74.9	122.9	<0.001
	Rural	2970	57.8

Utilization of health services showed the strongest association. Women with no ANC visits had the lowest institutional delivery rate (19.2%), a 64.5%-point difference, whereas women with 4 + visits achieved an 83.7% rate. Institutional delivery was more than 4.3 times higher among women with 4 + visits than among those with no visits. Distance to a health facility was also an important determinant. Those who considered distance “not a problem” accounted for 66.3%, whereas those who cited distance as a “big problem” had a rate of 60.0%, representing a modest but significant gap.

Wealth-related disparities were observed in institutional delivery, showing the steepest socioeconomic gradient. Among them, the rate of richest women was 87.5%, whereas it was only 38.8% in the poorest households, representing a 48.7% gap. Similarly, maternal education was strongly associated with institutional delivery. A total of 38.8% of women with no education received health facility delivery, and 87.2% had higher education. Regional disparities were evident. The Khulna division recorded the highest facility delivery rate (83.3%), whereas Mymensingh lagged, with only 51.1%. The urban–rural divide also persisted, with a significantly higher proportion of institutional deliveries in urban areas (74.9%) compared to rural areas (57.8%).

Internet usage and television viewership were strongly associated with the use of health facilities. Those who used the internet or viewed television had health facility rates of 73.7% and 78.4%, respectively. Furthermore, the non-use rate was 56.6%, and the no-watching rate was 51.8%.

A Sankey diagram (Fig. 1) illustrates multidimensional inequalities by residence, wealth, and ANC visits.

Fig. 1.

Sankey diagram illustrating disparities in institutional delivery by residence, wealth, and antenatal care (ANC) visits

Although the proportion of individuals in rural areas was greater than that in urban areas, access to healthcare delivery was notably greater in urban areas. The poorest women experienced the lowest rate of institutional deliveries, whereas those from the richest households had the highest rate. Similarly, women with 4 + ANC visits had significantly higher institutional delivery rates than those with no ANC visits, highlighting disparities in community-, household-, and individual-level factors.

Survey-weighted bayesian multilevel binary logistic regression

Five hierarchical Bayesian logistic regression models were estimated: null, individual, household, community, and full models. The report and final summary relied on the model with the smallest information criterion values, indicating the most appropriate fit. The full model was selected, LOOIC (4468) and WAIC (4464), reflecting the best fit and highest predictive accuracy. This model achieved the greatest reduction in between-cluster variance (PCV = 59.9%). This finding indicates that nearly 60% of the unexplained cluster-level variance from the null model is accounted for by the combined influences of individual-, household-, and community-level factors (Table 2). The ICC in the full model was 0.15, indicating that approximately 15% of the total variation in institutional delivery was attributable to differences between clusters. The remaining 85% of the variation was attributed to within-cluster differences. Furthermore, MOR = 2.09 indicates that if a woman moved from a lower-risk to a higher-risk cluster (with identical individual, household, and community characteristics), her odds of institutional delivery would increase by a median factor of 2.09. The between-cluster variance decreased from 1.49 in the null model to 0.60 in the full model, while the within-cluster variance remained constant at 3.29 across all the models, consistent with the assumption of multilevel logistic regression for binary outcomes.

Table 2.

Adjusted odds ratios (AORs) and 95% credible intervals from bayesian multilevel logistic regression model

Variable & Category	Null model AOR [95% CrI]	Individual model AOR [95% CrI]	Household model AOR [95% CrI]	Community model AOR [95% CrI]	Full model AOR [95% CrI]
Fixed effects
Intercept	1.86 [1.65, 2.13]	0.18 [0.11, 0.29]	0.26 [0.11, 0.60]	2.28 [1.57, 3.35]	0.06 [0.02, 0.16]
Age in 5-year groups
15–19 (Reference)	-	1.00 (Ref)	-	-	1.00 (Ref)
20–24	-	1.80 [0.83, 1.40]	-	-	1.10 [0.85, 1.40]
25–29	-	1.33 [1.01, 1.76]	-	-	1.30 [1.00, 1.72]
30–34	-	1.55 [1.26, 2.13]	-	-	1.43 [1.04, 1.96]
35–39	-	1.32 [0.90, 1.97]	-	-	1.15 [0.77, 1.72]
40–44	-	1.81 [0.91, 3.67]	-	-	1.65 [0.83, 3.38]
45–49	-	0.68 [0.09, 4.02]	-	-	0.63 [0.08, 3.68]
Respondent working
No (Reference)	-	1.00 (Ref)	-	-	1.00 (Ref)
Yes	-	0.82 [0.67, 0.99]	-	-	0.83 [0.68, 1.00]
Highest education level
No education (Reference)	-	1.00 (Ref)	-	-	1.00 (Ref)
Primary	-	0.98 [0.66, 1.46]	-	-	0.98 [0.65, 1.47]
Secondary	-	1.56 [1.00, 2.45]	-	-	1.37 [0.88, 2.15]
Higher	-	3.67 [2.20, 6.06]	-	-	2.87 [1.71, 4.74]
ANC visits
No visit (Reference)	-	1.00 (Ref)	-	-	1.00 (Ref)
1–4 visits	-	5.18 [3.10, 6.78]	-	-	4.77 [3.67, 6.28]
4 + visits	-	9.30 [6.53, 13.32]	-	-	7.83 [5.51, 11.26]
Birth order
Low (Reference)	-	1.00 (Ref)	-	-	1.00 (Ref)
High	-	0.55 [0.36, 0.85]	-	-	0.57 [0.37, 0.86]
Number of living children
Low (Reference)	-	1.00 (Ref)	-	-	1.00 (Ref)
High	-	0.89 [0.57, 1.37]	-	-	0.92 [0.59, 1.42]
Reading newspaper
Not at all (Reference)	-	1.00 (Ref)	-	-	1.00 (Ref)
Sometimes	-	1.27 [0.86, 1.90]	-	-	1.16 [0.79, 1.71]
Listening to radio
Not at all (Reference)	-	1.00 (Ref)	-	-	1.00 (Ref)
Sometimes	-	0.97 [0.60, 1.57]	-	-	0.87 [0.54, 1.42]
Watching television
Not at all (Reference)	-	1.00 (Ref)	-	-	1.00 (Ref)
Sometimes	-	1.60 [1.37, 1.89]	-	-	1.10 [0.92, 1.42]
Use of the internet
Never (Reference)	-	1.00 (Ref)	-	-	1.00 (Ref)
Often use	-	1.35 [1.11, 1.65]	-	-	1.19 [0.89, 1.48]
Literacy
Illiterate (Reference)	-	1.00 (Ref)	-	-	1.00 (Ref)
Literate	-	1.12 [0.81, 1.53]	-	-	1.10 [0.80, 1.49]
Household electricity
No (Reference)	-	-	1.00 (Ref)	-	1.00 (Ref)
Yes	-	-	2.53 [1.10, 6.24]	-	2.21 [0.94, 5.16]
Household television
No (Reference)	-	-	1.00 (Ref)	-	1.00 (Ref)
Yes	-	-	1.42 [1.20, 1.67]	-	1.22 [0.97, 1.53]
Distance to health facility
Not a big problem (Reference)	-	-	1.00 (Ref)	-	1.00 (Ref)
Big problem	-	-	0.92 [0.79, 1.07]	-	0.98 [0.83, 1.15]
Sex of household head
Male (Reference)	-	-	1.00 (Ref)	-	1.00 (Ref)
Female	-	-	0.90 [0.72, 1.14]	-	0.89 [0.69, 1.17]
Religion
Islam (Reference)	-	-	1.00 (Ref)	-	1.00 (Ref)
Hinduism	-	-	3.58 [2.43, 5.29]	-	3.16 [2.13, 4.72]
Buddhist	-	-	0.96 [0.28, 3.27]	-	1.86 [0.54, 6.83]
Christianity	-	-	8.15 [1.21, 81.13]	-	4.86 [0.63, 53.45]
Wealth index
Poorest (Reference)	-	-	1.00 (Ref)	-	1.00 (Ref)
Poorer	-	-	1.58 [1.26, 1.95]	-	1.22 [0.96, 1.54]
Middle	-	-	2.21 [1.76, 2.82]	-	1.51 [1.17, 1.96]
Richer	-	-	3.64 [2.83, 4.67]	-	1.98 [1.50, 2.63]
Richest	-	-	7.10 [5.26, 9.71]	-	2.76 [1.94, 3.89]
Division
Barishal (Reference)	-	-	-	1.00 (Ref)	1.00 (Ref)
Chattogram	-	-	-	1.03 [0.67, 1.57]	1.14 [0.76, 1.69]
Dhaka	-	-	-	1.28 [0.81, 1.98]	1.26 [0.84, 1.90]
Khulna	-	-	-	4.42 [2.70, 7.26]	4.14 [2.68, 6.48]
Mymensingh	-	-	-	0.91 [0.57, 1.44]	1.15 [0.77, 1.74]
Rajshahi	-	-	-	2.28 [1.44, 3.65]	2.03 [1.32, 3.15]
Rangpur	-	-	-	1.48 [0.93, 2.34]	1.62 [1.06, 2.46]
Sylhet	-	-	-	0.88 [0.56, 1.40]	1.06 [0.70, 1.60]
Residence type
Urban (Reference)	-	-	-	1.00 (Ref)	1.00 (Ref)
Rural	-	-	-	0.44 [0.34, 0.57]	0.73 [0.59, 0.92]
Random effects
Cluster Variation	[1.09, 1.36]	[0.82, 1.08]	[0.81, 1.05]	[0.94, 1.19]	[0.64, 0.9]
Between-Cluster Variance	1.49	0.90	0.86	1.12	0.60
Within-Cluster Variance	3.29	3.29	3.29	3.29	3.29
ICC	0.31	0.22	0.21	0.25	0.15
MOR	3.20	2.47	2.43	2.74	2.09
PCV (%)	0.00	39.6	42	24.8	59.9
LOOIC	5045	4565	4822	5002	4468
WAIC	5034	4559	4816	4994	4464

Convergence diagnostics were satisfactory, with R-hat = 1 and bulk-ESS and tail-ESS > 1000 for all parameters (Table S1). Moreover, model diagnostics, including trace plots for the most influential covariates (Figure S2), confirm good convergence with well-mixed chains. Additionally, the posterior predictive fit (Figure S3) visually confirms that the Bayesian model effectively captures the observed data patterns and uncertainty.

ANC visits were strongly associated with institutional delivery. The odds of institutional delivery for women with 1–4 ANC visits were 4.77 times higher compared to those with no visits (AOR = 4.77; 95% CrI: 3.67–6.28). Furthermore, women who received 4 + visits had 7.83 times greater odds (AOR = 7.83; 95% CI: 5.51–11.26), suggesting that higher ANC frequency is associated with a greater likelihood of institutional delivery.

The odds ratio of women with higher education was 2.87 times greater than that of women with no education (AOR = 2.87; 95% CrI: 1.71–4.74), whereas there was no significant association with primary education (AOR = 0.98; 95% CrI: 0.65–1.47). Compared with the poorest women, those in the wealthiest households had 2.76 times higher odds of institutional delivery (AOR = 2.76; 95% CrI: 1.94–3.89). Women in the richer quintile had significantly higher odds (AOR = 1.98; 95% CrI: 1.50–2.63), and women in the middle quintile had 51% greater odds of having an AOR of 1.51 (95% CrI: 1.17–1.96).

While the odds for women aged 40–44 were elevated (AOR = 1.65; 95% CrI: 0.83–3.38), the credible interval included 1.0, indicating statistical uncertainty, and the finding should not be considered statistically significant. Women aged 30–34 had significantly higher odds of institutional delivery (AOR = 1.43; 95% CrI: 1.04–1.96), while the association for women aged 25–29 was borderline (AOR = 1.30; 95% CrI: 1.00–1.72). In contrast, the odds ratios for women aged 20–24 years (AOR = 1.10; 95% CrI: 0.85–1.40) and 35–39 years (AOR = 1.15; 95% CrI: 0.77–1.72) were not statistically significant. Furthermore, for the oldest age group (45–49), there was a trend toward decreasing odds (AOR = 0.63; 95% CrI: 0.08–3.68). However, the CrI includes the null value (AOR = 1.0), and the wide credible interval reflects the small number of observations in this category.

Women with higher birth orders had 43% lower odds of experiencing the outcome than women with lower birth orders did (AOR = 0.57; 95% CrI: 0.37–0.86). The point estimate suggested lower odds for currently working women (AOR = 0.83), but the 95% CrI (0.68–1.00) included the null value, indicating high uncertainty and no statistically clear effect. Rural women had 27% lower odds of delivering in a health facility compared with urban residents (AOR = 0.73; 95% CrI: 0.59–0.92), suggesting an association with lower institutional delivery in rural areas. Residents in Khulna (AOR = 4.14; 95% CrI: 2.68–6.48) had 4.14× higher odds than did those in Barishal, whereas there was no significant difference in the odds for Sylhet (AOR = 1.06; 95% CrI: 0.70–1.60). On the other hand, the odds ratio of Hindu women was 3.16 times greater than that of Muslims (AOR = 3.16; 95% CrI: 2.13–4.72). Christian women had higher but imprecise odds of delivering in a health facility (AOR = 4.86; 95% CrI: 0.63–53.45), reflecting uncertainty due to the small sample size.

The Bayesian full model revealed that factors such as media exposure, the sex of the household head, household assets, distance to health facilities, and literacy had no meaningful effects on institutional delivery. For example, the frequency of watching television (AOR = 1.10; 95% CrI: 0.92–1.42), use of internet (AOR = 1.19; 0.89–1.42), sex of the household head (AOR = 0.89; 95% CrI: 0.69–1.17), having electricity (AOR = 2.21; 0.94–5.16), having television (AOR = 1.22; 0.97–1.53), distance to health facility (AOR = 0.98; 0.83–1.15), and literacy (AOR = 1.10; 0.80–1.69) all had credible intervals containing 1, indicating no strong evidence of effect (posterior probability < 90%).

In addition, the effects of the strongest predictors were illustrated via a forest plot of adjusted odds ratios with 95% CrI, as shown in Fig. 2. ANC visits, religion, geographic division, education level, and the wealth index were identified as having the strongest significant associations. The plot confirmed that socioeconomic and healthcare access factors dominate over media exposure or other proxy variables.

Fig. 2.

Bayesian forest plot: strongest predictors of institutional delivery (maximum category effect for each variable with 95% credible intervals)

Discussion

Maternal healthcare in Bangladesh has improved substantially over the past 15 years [15]. According to the BDHS 2007, only 15% of births occurred in health facilities [12]. This rate increased to 29% in 2011 and 49% in 2017–18 [47, 48]. Our analysis of the BDHS 2022 data reveals a further increase to 63.5%, representing a 48.5% increase over the past 15 years. While this progress is substantial, our analysis highlights the persistence of socioeconomic and structural inequalities that shape access to institutional delivery.

Using a Bayesian hierarchical framework, this study examined individual-, household-, and community-level determinants of institutional delivery. ANC visits (4 + visits) and household wealth (richest) emerged as the strongest predictors. Similarly, previous studies reported that ANC attendance substantially increases the likelihood of facility delivery [21] and that wealth and education enhance maternal healthcare use [9]. Use of health facilities for childbirth was higher among women who had contact with antenatal services, as these visits help build understanding of risks around childbirth and support advanced planning for delivery [8]. During prenatal checkups, many women learn about government programs that offer payments or vouchers for giving birth at a health facility, information that often leads them to choose a hospital or clinic for delivery [9]. Evidence from South Asian countries also confirms that highly educated mothers are more likely to deliver in health facilities [38]. While a mother’s education can empower her personal health choices, in Bangladesh’s male-centric societal framework, it critically enables the household support and joint decision-making required for her to seek institutional delivery [37]. The present findings align with this pattern, suggesting that women with better socioeconomic status are typically more informed and better able to access and afford modern healthcare services [10, 22]. Women from wealthier households are more likely to deliver in hospitals, while financial constraints continue to limit access to facility-based childbirth among poorer women [38]. It may be that underprivileged households spend almost all their financial resources on food, leaving little for health and education [7]. Bangladesh faces a widening gap in institutional delivery access, requiring urgent, targeted interventions to address rising disparities across different socioeconomic segments [49].

A key insight from our study is that educational and economic inequalities, rather than media exposure, were more influential in determining delivery choices. While some studies reported greater institutional delivery among women without media exposure [7], others reported the opposite trend [9, 37]. Our Bayesian results suggest that media exposure primarily serves as a proxy: women with strong educational and economic backgrounds tend to use institutional delivery regardless of their media exposure, whereas women who are media-exposed but less educated or poorer still have lower facility delivery rates.

In contrast to patterns in India, Nepal, and Ethiopia, where younger women tend to have higher facility delivery rates [18, 23, 50], our results from Bangladesh suggest a different, non-linear pattern. This nonlinear pattern is similar to evidence from 34 sub-Saharan African countries [51]. Notably, the odds ratio point estimate peaked among women aged 40–44 years. However, the substantial statistical uncertainty precludes a definitive conclusion and highlights this as a key demographic relationship that requires further investigation. These findings suggest that country-specific cultural norms, socioeconomic conditions, and decision-making structures have a significant association with healthcare utilization. Older women in Bangladesh may possess greater autonomy, resources, and awareness of pregnancy risks, enabling them to utilize delivery services more frequently, even though globally, the opposite trend is more common [52].

Our results also revealed that employed women were less likely to use health facilities for childbirth than unemployed women, although the credible interval included the null value. This pattern is consistent with earlier research [7, 24]. One explanation may be time constraints, job pressures, or reduced access to ANC services, all of which may indirectly reduce facility-based delivery utilization [53]. Although employed women showed a trend toward lower facility delivery rates, the high statistical uncertainty suggests that work alone may not be a barrier; rather, the type of employment and the lack of supportive policies are likely more influential.

Birth order was another important determinant. Women with first or second births were more likely to choose institutional delivery, corroborating previous findings [7]. First-time mothers are typically more cautious and prefer a professionally assisted delivery [9], whereas those with multiple prior births may rely on past experiences and be more inclined toward home delivery [8].

Consistent with previous studies, Muslim women were less likely to deliver in health facilities than non-Muslims [8, 37]. This association may reflect broader social and contextual factors that often overlap with religious affiliation, such as socioeconomic conditions, gender norms, and differences in access to healthcare, rather than religious beliefs alone [7, 17, 24, 54]. These interpretations should be viewed cautiously, as the present study did not directly measure religious practices or women’s decision-making autonomy.

Persistent geographic disparities in access to institutional delivery remain a key challenge for maternal health policy in Bangladesh [55]. Previous studies have reported that the Khulna division [14, 48] has been in a relatively privileged position regarding the use of institutional deliveries, whereas divisions such as Sylhet and Barisha [14, 48, 55] have shown limited access to such care. Although a recent study identified that the lowest level of institutional delivery occurred in Barishal [11], our model uses it as the reference group for regional comparison. Relative to Barishal, the model indicates significantly higher odds of institutional delivery in Khulna, Rajshahi, and Rangpur, whereas the odds in Sylhet, Chattogram, Dhaka, and Mymensingh were not statistically distinguishable from the reference group. These findings highlight the need for targeted, region-specific strategies to strengthen maternal, newborn, and child health services in underperforming divisions [55].

Despite national progress, regional inequalities remain substantial. Such differences may arise from the relative advantages often found in urban settings, including higher maternal education levels, greater health awareness, and easier healthcare accessibility compared to rural areas [18]. To address the urban-rural disparity in institutional delivery, policies must prioritize equitable investment in rural healthcare infrastructure, reliable transportation, and improved hospital resources to ensure all mothers have access to safe delivery facilities [9]. Our findings reinforce the ‘inverse care law’, which states that those most in need of healthcare often have the least access, particularly evident in many LMICs [56]. These findings indicate that national-level summaries often hide deep inequalities arising from inefficiencies in the distribution of health care and resources [57]. Therefore, instead of general policies, targeted actions are needed to address specific geographic and socioeconomic challenges [58].

Key predictors such as antenatal care visits and household wealth may be influenced by unobserved factors, like health-seeking behaviour or access to information, which could bias the estimated associations [59, 60]. Although our multilevel model captures unobserved differences between clusters, it cannot fully account for potential endogeneity; therefore, the findings should be interpreted as associations rather than definitive causal effects. Future research using longitudinal data [60, 61] or instrumental variable [61], propensity score [59, 60] methods would help clarify causal pathways.

Conclusions

Bayesian modelling revealed that factors at multiple levels are strongly linked to health facility delivery in Bangladesh. ANC visits, maternal education, household wealth, maternal age, birth order, and geographic region are significantly associated with the likelihood of institutional delivery, indicating pronounced socioeconomic and regional disparities. To make these findings actionable, policies should focus on targeted interventions in specific regions rather than general equity goals.

Among the key determinants, ANC visits and household wealth are strongly associated with institutional delivery. Mothers from wealthier and more educated households are more likely to complete adequate ANC, which is linked to improved monitoring during pregnancy and preparation for facility-based childbirth. At the community level, policymakers should facilitate at least 4 + visits to support the utilization of all essential ANC services. Health facilities in Barishal, Sylhet, and Mymensingh, where maternal service coverage remains relatively low, may require targeted strengthening to better support emergency childbirth care, ensuring consistent, high-quality services. Additionally, increasing the frequency of ANC visits in low-performing divisions and reducing disparities between public and private facilities may help build trust in health facilities and encourage women to choose institutional delivery. Increasing the national education budget and implementing regular community programs may contribute to women’s health awareness, decision-making power, and timely use of maternal health services. Keeping girls in school, especially those at risk of dropping out due to financial hardship, has a long-term impact on increasing antenatal care use and institutional delivery. Targeted policies are essential for vulnerable groups, including working mothers, adolescents, and Muslim women, who often face social, cultural, and economic barriers to facility-based childbirth. Strategies such as reducing early pregnancies, creating maternity-friendly workplace policies, and increasing access to trained female healthcare providers may help enhance confidence in health facilities and be associated with increased institutional delivery among these groups. Maternity leave policies across all sectors may further support the timely use of maternal healthcare services.

Equitable access to healthcare remains a critical policy goal nationwide, with special attention paid to under-resourced regions and disadvantaged groups. Mothers in low-income, low-education, and remote communities require priority attention throughout pregnancy to reduce preventable risks. Persisting disparities may pose long-term challenges to the nation’s health and sustainable development.

Our Bayesian analysis indicates substantial cluster-level inequality. An MOR of 2.09 reveals that women in higher-performing clusters have approximately twice the odds of those in lower-performing clusters, indicating a need for targeted investment in clusters. Furthermore, PCV = 59.9% suggests that over half of this between-cluster variation is statistically linked to disparities in household wealth and education. Therefore, comprehensive strategies must address both immediate geographic inequities in healthcare access and the underlying socioeconomic determinants through combined efforts in poverty reduction, education, and health promotion.

Strengths and limitations

This study utilized recent nationally representative BDHS 2022 data by employing a survey-weighted Bayesian multilevel framework that accounted for the complex survey design and hierarchical data structure. This approach enables robust estimation across individual, household, and community levels, along with intuitive probabilistic summaries (e.g., credible intervals, MOR, PCV) that effectively quantify uncertainty and cluster-level heterogeneity.

However, the cross-sectional design limits causal inference, and self-reported data may introduce recall bias. Despite adjusting for a comprehensive set of covariates, residual confounding may persist due to the unavailability of contextual and behavioural variables, and uncertainty remains for some associations, as indicated by credible intervals that cross unity. Relatedly, potential endogeneity cannot be ruled out, as unobserved factors may influence both the predictor and the outcome. Finally, the study examined institutional delivery as a binary outcome and did not address qualitative aspects such as timeliness of care, patient satisfaction, or clinical outcomes.

Authors’ contributions

Mahmudul Hasan: Conceptualization, Data Curation, Methodology, Software, Formal Analyses, Writing- Draft. Suman Biswas: Conceptualization, Data Curation, Methodology, Software, Formal Analyses, Writing-Draft, Review & Editing, Supervision.

Funding

The author (s) received no specific funding for this work.

Data availability

The dataset was obtained from the DHS Program website (https://dhsprogram.com/data/available-datasets.cfm) in response to a formal request outlining the study objectives. For the analysis, the 2022 Bangladesh individual recode file was used.

Declarations

Ethics approval and consent to participate

Approval and participant consent were not required for this study, as it utilized secondary data that was publicly available through the DHS Program. Additionally, ethical clearance for the BDHS was initially granted by the ICF Institutional Review Board and the NIPORT Review Committee. The highest ethical standards were employed throughout the study, which was conducted in accordance with the Declaration of Helsinki.

Consent for publication

Not applicable. This manuscript does not contain any person’s data in any form (including individual details, images, or videos) that would require consent for publication.

Competing interests

The authors declare no competing interests.

Abbreviations

SDG Sustainable Development Goal.

MDG Millennium Development Goal.

LMIC Low- and middle-income countries.

BDHS Bangladesh Demographic and Health Survey.

AOR Adjusted odds ratio.

CrI Credible interval.

WAIC Widely applicable information criterion.

LOOIC Leave-one-out information criterion.

ANC Antenatal care.

MMR Maternal mortality ratio.

ICC Intraclass correlation.

MOR Median odds ratio.

PCV Proportional change in variance.

Bulk ESS Bulk Effective Sample Size.

Tail ESS Tail Effective Sample Size.