Data sources

We analyzed demand-side determinants of maternal health services (ANC from a medically-trained provider, facility delivery, and complication management) using data from three rounds (2001, 2010, 2016) of the nationally representative Bangladesh Maternal Mortality and Health Care Survey (BMMS). The BMMS covered large household samples (approximately 100,000 in 2001, 174,000 in 2010, and 298,000 in 2016) and provided comparable data on maternal health indicators, socioeconomic characteristics, and healthcare utilization among ever-married women aged 13–49 who had a live birth in the preceding three years [31–33]. To ensure that the data are nationally representative, BMMS survey weights are not calculated only at the Primary Sampling Unit (PSU) level; they are calculated using a two-stage sample design that accounts for probabilities at both the PSU (first stage) and household (second stage) levels within specific strata. The final household survey datasets were accessed on November 16, 2021, and the authors did not have access to information that could identify individual participants during or after data collection.

Supply-side determinants were assessed using three rounds (2000, 2011, 2017) of the Bangladesh Health Facility Survey (BHFS)/Service Provision Assessment (SPA). These nationally representative surveys employed stratified random sampling (855 facilities in 2000, 1,503 in 2011, and 1,524 in 2017) to assess readiness and provision of essential health services by public, NGO, and private facilities, specifically those registered with the Bangladesh government’s Ministry of Health. BHFS/SPA survey weights are not calculated solely at the PSU level; they are derived using a multi-stage approach that incorporates PSU-level selection probabilities, facility-level sampling, and adjustments for non-response at both the facility and provider/client levels to ensure national representativeness. Details of the sampling methodology and data collection can be found in the BHFS/SPA final reports [34–36].

This study used publicly available secondary, de-identified data from nationally representative household and health facility surveys by the National Institute of Population Research and Training (NIPORT), the DHS Program, and the MEASURE Evaluation project. The surveys were approved by the respective organization’s Institutional Review Board and the National Research Ethics Committee of the Bangladesh Medical Research Council. Informed verbal consent was obtained from survey participants by the original study teams; for the present secondary analysis of anonymized data, additional review was deemed exempt by the Office of Human Research Ethics at the University of North Carolina at Chapel Hill under the regulatory category of Human Subject Research, in accordance with 45 CFR 46.104.

Variables

As outcomes, this study focuses on three dichotomous variables of maternal health services—ANC visits from a medically trained provider during pregnancy, delivery in a health facility, and medical care-seeking for management of pregnancy and/or delivery complications. A summary of the variables used in this analysis is provided in Table 1.

Download:

• PNG
  larger image
• TIFF
  original image

Table 1. Summary of variables for analysis.

https://doi.org/10.1371/journal.pone.0354897.t001

For this analysis, socioeconomic status was measured using a wealth index based on household’s ownership of selected durable assets and dwelling characteristics (e.g., assets like a television or a bicycle, and amenities such as the source of drinking water, toilet facilities, etc.), derived through principal component analysis [37]. Respondents were grouped into five wealth quintiles based on their household’s wealth index [14]. For the multivariate regression analysis to estimate how the poor–nonpoor gap in maternal health services use differed by facility readiness, we categorized poor (comprising the two lowest wealth quintiles) and nonpoor (comprising the remaining three higher wealth quintiles), consistent with definitions used for lower-middle-income countries like Bangladesh [38].

Facility readiness for delivery care was assessed using a composite score based on infrastructure, equipment, supplies, and commodities. Indicators were selected based on BHFS/SPA data and guidance from the World Health Organization (WHO)’s Service Availability and Readiness Assessment (SARA) manual, the Newborn Indicator Technical Working Group, and a recent systematic review [39–41] (see S1 Table in the supporting information). Each indicator was equally weighted, and scores were standardized to range from 0 to 1, representing the percentage of maximum possible readiness [42]. Given this standardization, a facility’s score is interpreted as the percentage of the highest possible readiness that the facility could have. This composite captures structural readiness (i.e., the physical infrastructure, equipment, supplies, and commodities required to provide maternal health services) and therefore measures a facility’s structural capacity to deliver care. It does not directly measure process quality, clinical quality, or patient experience, and we interpret it accordingly throughout this paper.

As the BHFS/SPA and BMMS surveys were independently sampled, we linked facility readiness to households at the district level—the lowest common geographic unit. District-level average readiness scores (the means of the district averages of facility readiness were 0.544 in 2000, 0.340 in 2011, and 0.322 in 2017) were merged into the pooled BMMS dataset (see S2 Table in the supporting information). Each respondent was assigned their district’s average score, and the distribution was divided into terciles (low, medium, high) to examine maternal health service use. For multivariate analysis, we grouped facility readiness into two categories: low (bottom two terciles) and high (top tercile) to assess the differential effects of readiness and socioeconomic status. District readiness classification was time-varying across survey rounds. For each BMMS-BHFS linked year, district-average readiness was calculated from the corresponding facility survey and districts were then classified relative to the distribution in that survey year; therefore, a district could move between readiness categories across waves. Because the household and facility surveys were sampled independently, households were not linked to individual facilities or to the nearest facility; instead, each district containing BMMS respondents was matched to the corresponding district-average readiness score from the same-round BHFS/SPA, so the linkage represents the broader district service environment rather than the specific facility a woman used. All districts with BMMS respondents were successfully linked to a readiness score. The separate proximity measures (nearest public or private facility within one hour) are distinct from this readiness linkage and are based on respondents’ self-reported travel time to the nearest facility.

Statistical analysis

Socioeconomic disparities in maternal health services use in Bangladesh over time were measured by concentration indices, where the index value is bounded between ‒1 to +1, with zero indicating perfect equality. A positive concentration index indicates that service use is concentrated among wealthier (nonpoor) women; a larger positive value reflects greater pro-rich inequality, a decline in the index over time indicates a reduction in pro-rich inequality, and an increase indicates widening pro-rich inequality. As the outcomes were dichotomous, we applied Erreygers’ transformation using Stata’s “conindex” command to correct index limitations, ensuring indices remained within the [–1, 1] interval and satisfied essential analytical requirements [43].

To examine how the poor–nonpoor gap in service use changed over time across districts with differing facility readiness during 2001‒2016, we applied an adaptation of the Difference-in-Differences (DID) model. For this, we used linear probability model (LPM) on a pooled dataset of 2001, 2010, and 2016 BMMSs linked to corresponding BHFS/SPA with the following regression specification:

(1)

where, is the outcome of interest for individual i who lives in district j at time t. takes the value of 1 if the woman i belongs to nonpoor socioeconomic status (i.e., household wealth quintiles 3‒5) and 0 if she belongs to poor socioeconomic status (i.e., household wealth quintiles 1 and 2). is a vector of time dummies to represent the BMMS survey rounds 2010 and 2016. is an indicator variable that takes the value of 1 if the woman i lives in a district with high facility readiness (i.e., facility readiness tercile 3) and 0 if the individual lives in a district with low facility readiness for maternal health services (i.e., readiness terciles 1 and 2). are the relevant socio-demographic covariates influencing the outcomes.

In equation (1), measures the change in socioeconomic equity that occurred during the study period with low facility readiness, and measures the change in socioeconomic equity in districts with high facility readiness (coefficients were combined using “lincom” command in Stata). Finally, indicates the differential change in equity between high- and low-readiness districts, after controlling for relevant covariates, during the study period (see S3 Text in the supporting information for a detailed description of how the LPM coefficients were identified to measure equity). This specification is not intended to evaluate a single policy intervention with a discrete adoption date. Rather, it compares whether poor–nonpoor differences in service use changed differently over time across district-year service environments defined by facility readiness. Accordingly, the triple interaction term should be interpreted as a difference in the change in the poor–nonpoor gap between higher- and lower-readiness districts over time, conditional on observed covariates. In plain terms, the difference-in-differences (DID) estimate captures how much the poor–nonpoor gap in service use changed over time, while the difference-in-difference-in-differences (DIDID) estimate captures whether that change in the gap was larger or smaller in high-readiness districts than in low-readiness districts.

All the analyses were performed using STATA v.16 [44], and observations with missing data on model variables were excluded from regression analyses (complete-case analysis). Missing data on the model variables were limited, with 1.12% of otherwise-eligible observations excluded from the regression models; the extent of missingness did not differ substantially across survey rounds or readiness groups. We used appropriate survey sampling weights, tested for multicollinearity, and controlled for clustering and heteroscedasticity to obtain robust standard errors. Standard errors were clustered at the cluster level. The specification includes survey-round (time) indicators and a district-level readiness indicator but does not include district fixed effects; it is therefore a pooled repeated cross-sectional interaction model rather than a two-way fixed-effects difference-in-differences design, which is what we mean by an “adaptation” of the DID approach. We chose LPM over logistic and probit regression for binary outcomes in this study due to its interpretability and widespread use in economics and social sciences [45]. Additionally, when the binary outcome is common (i.e., prevalence >10%), logistic regression’s odds ratio can misrepresent the true association and LPM is effective for analyzing binary outcomes in such cases [46].

Respondent’s characteristics

Between 2001 and 2016, there were notable increases in urbanization, women’s educational attainment, and availability of health facilities in Bangladesh (see Table 2). In particular, there was a massive growth in the private health sector during this period—only 17% of recent mothers lived within one hour of travel from a private health facility in 2001, which increased to 84% in 2016. Fertility also reduced during this period, as seen in the shift towards lower parity among the respondents.

Download:

• PNG
  larger image
• TIFF
  original image

Table 2. Distribution (%) of demographic, socioeconomic, and structural characteristics of respondents, Bangladesh 2001‒2016.

https://doi.org/10.1371/journal.pone.0354897.t002

Trends in maternal health services use during 2001‒2016

Overall, use of medically trained antenatal care increased from 41% (2001) to 74% (2016), while facility deliveries rose from 9% to 48%. Around half of the women reported pregnancy or delivery complications, with medically trained treatment increasing from 23% to 51%. Despite significant improvements (p < 0.05) in all services, utilization gains were uneven across socioeconomic groups.

Fig 2 illustrates maternal health service use by wealth quintiles over time, highlighting absolute inequities [47]. It shows a modest narrowing in the rich-poor gap for ANC and complication management but a significant widening for facility deliveries. In 2016, service use among the richest quintile was substantially higher compared to the poorest.

Changes in equity in service use by socioeconomic status

By examining the concentration indices for 2001, 2010, and 2016, we can see that inequity in maternal health services use continued to favor the better off in Bangladesh during the study period—i.e., services were used disproportionately more by women belonging to higher wealth quintiles than their counterparts (see Table 3). Because the concentration index is a summary measure that aggregates inequality across the full wealth distribution, it does not show how use changed within individual wealth quintiles; the corresponding wealth-quintile–specific patterns are presented in Fig 2. There were clear indications that inequities in ANC and treatment for delivery/pregnancy complications have reduced significantly (p < 0.001) over the years. However, the concentration index values increased by 135% between 2001 and 2016 for delivery at a health facility, which indicated a statistically significant (p < 0.001) increase in socioeconomic inequity in service use.

Download:

• PNG
  larger image
• TIFF
  original image

Table 3. Concentration indices for maternal health services use by socioeconomic status, Bangladesh 2001‒2016.

https://doi.org/10.1371/journal.pone.0354897.t003

Differential changes in equity by facility readiness in maternal health services use

Fig 3 shows the socioeconomic inequity, measured as the percentage point gap in maternal health services use among the nonpoor and the poor, by facility readiness during the study period. Between 2001 and 2016, socioeconomic inequity for ANC remained similar in low-readiness districts (see points A‒C in Fig 3-a) and increased by 1.9 percentage points in high-readiness districts (see points D‒F in Fig 3-a). The DID model estimated that the gap between poor and nonpoor increased by 1.8 percentage points between low and high facility readiness areas during this period, which is the difference-in-difference-in-difference (DIDID) estimate (see the coefficient of ‘Nonpoor×High×2016’ under the column for ANC in Table 4). Since the DIDID estimate was not statistically significant, we conclude that the differential change in the poor–nonpoor gap for ANC across readiness groups was negligible between 2001 and 2016.

Download:

• PNG
  larger image
• TIFF
  original image

Table 4. Differential changes in socioeconomic equity by facility readiness, Bangladesh 2001‒2016.

https://doi.org/10.1371/journal.pone.0354897.t004

Download:

• PNG
  larger image
• TIFF
  original image

Fig 3. LPM coefficients showing changes in inequity in health services use by facility readiness^a,b.

^a LPM coefficients from Table 4 corresponding to the data points are as follows: A = Nonpoor; B = Nonpoor + Nonpoor×2010; C = Nonpoor + Nonpoor×2016; D = Nonpoor + Nonpoor×High; E = Nonpoor + Nonpoor×2010 + Nonpoor×High + Nonpoor×High×2010; and F = Nonpoor + Nonpoor×2016 + Nonpoor×High + Nonpoor×High×2016. ^b Error bars represent 95% confidence intervals.

https://doi.org/10.1371/journal.pone.0354897.g003

For facility delivery, socioeconomic inequity increased by 15.2 percentage points in low-readiness districts and by 21.7 percentage points in high-readiness districts (see points A‒C and D‒F in Fig 3-b, respectively) between 2001 and 2016. The DID model estimated that the gap between poor and nonpoor increased by 6.5 percentage points between low and high facility readiness areas during this period, which is the DIDID estimate (see the coefficient of ‘Nonpoor×High×2016’ under the column for facility delivery in Table 4). As these DID and DIDID estimates are statistically significant at p < 0.05, socioeconomic inequity increased in both low and high-readiness districts but faster in the high-readiness districts during the study period, resulting in a more inequitable distribution of facility delivery services. The DID and DIDID estimates were also not statistically significant for the treatment of complications, indicating a negligible differential change in equity by facility readiness in the treatment of complications.

Differential changes in equity by access to private facilities in maternal health services use

From the respondent’s characteristics and the factors affecting maternal health services use, we observed that the availability of private health facilities had a statistically significant association with the use of maternal health services in 2016 (analysis not shown). For this reason, we examined how the poor–nonpoor gap in maternal health service use varied with the proximity of private health facilities, after controlling for all other relevant covariates, including facility readiness (see Fig 4). We found that for ANC, socioeconomic inequity increased by 5.1 percentage points in areas more than an hour away from private facilities (see points A‒C in Fig 4-a) but decreased by 1.2 percentage points in the areas closer to private facilities during the study period (see points D‒F in Fig 4-a). The DID model also estimated that socioeconomic inequity decreased by 6.3 percentage points between the areas with >1-hour and <1-hour distance from private facilities during this period, and this DIDID estimate was statistically significant at p < 0.05 (see the coefficient of ‘Nonpoor×<1 hour×2016’ under the column for ANC in S4 Table in the supporting information). We, therefore, conclude that socioeconomic inequity in ANC use decreased between 2001 and 2016 among women living <1 hour from a private facility.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 4. LPM coefficients showing changes in inequity in health services use by distance to private facilities^a,b.

^a LPM coefficients from S4 Table 1 (in the supporting information) corresponding to the data points are as follows: A = Nonpoor; B = Nonpoor + Nonpoor×2010; C = Nonpoor + Nonpoor×2016; D = Nonpoor + Nonpoor×< 1hour; E = Nonpoor + Nonpoor×2010 + Nonpoor×<1 hour + Nonpoor×<1 hour×2010; and F = Nonpoor + Nonpoor×2016 + Nonpoor×< 1 hour + Nonpoor×< 1 hour×2016. ^b Error bars represent 95% confidence intervals.

https://doi.org/10.1371/journal.pone.0354897.g004

For facility delivery, we can see from Fig 4-b that socioeconomic inequity increased similarly in both areas with >1-hour and <1-hour distance to private facilities (see Fig 4-b), and the estimated DIDID was not statistically significant (see the coefficient of ‘Nonpoor×<1 hour×2016’ under the column for facility delivery in S4 Table). For the treatment of delivery complications, socioeconomic inequity slightly decreased in both areas (see Fig 4-c), and the DID and DIDID estimates were also not statistically significant (see S4 Table).

We also examined how changes in socioeconomic equity differed by proximity to private health facilities and by facility readiness for maternal health service use, after controlling for relevant covariates. We found that the difference in changes in socioeconomic inequity between areas with >1-hour and <1-hour distance to private facilities, after accounting for facility readiness, remained not statistically significant for any maternal health service. See S5 Text in the supporting information for a detailed description of how the LPM coefficients were identified to measure equity, and S6 Table for the results of the regression models.

Study limitations

This study’s limitations are largely related to the nature of the data used. First, all analyses relied on women’s recall of details about the ANC and delivery care that they received for a live birth up to three years preceding the survey. Women’s responses were classified into pre-specified survey response options, and there is a possibility of recall bias in their responses despite rigorous training for the field interviewers and several layers of data quality checks in place. Reporting of delivery complications were ‘perceived complications,’ and no medical diagnosis was carried out to ascertain the validity of the responses. Also, data were not collected about the care received by women who experienced non-live birth outcomes such as miscarriages, stillbirths, and induced abortions. Second, linking structural/ programmatic variables (e.g., the readiness of health facilities) to surveyed households at the sampled cluster level was not possible since the BHFS/SPA and BMMS surveys were sampled independently. Instead, we linked facilities and households at the lowest possible geographic unit, which is at the district level. This level of linking provides information on the ‘existing service environment’ for the BMMS respondents rather than the readiness of facilities where delivery care was actually sought (linking women to individual facilities would be problematic conceptually anyway because a lot of women don’t use facilities for delivery care). Third, the data availability of potential structural/ programmatic covariates is not uniform across the BHFS/SPA survey rounds. For example, the 1999‒2000 Bangladesh SPA did not collect information from secondary-level public health facilities such as government district hospitals, for which the bed occupancy rate of district hospitals served as a proxy for service efficiency/readiness and imputed readiness scores for 1999‒2000 based on regression outputs for 2011 and 2017. Fourth, because facility readiness was measured contemporaneously rather than as a discretely timed intervention, the readiness–equity association may be confounded by concurrent secular changes during 2001–2016, including successive national health-sector programs, the rapid expansion of the private sector, and broad improvements in women’s education and household economic conditions. These could independently affect both facility readiness and equitable utilization, and our design cannot fully disentangle these time-varying influences. Fifth, because the analysis pools three independent repeated cross-sections, the composition of respondents (e.g., age, education, parity, and urban–rural residence) shifted across survey rounds and may have shifted differentially across readiness groups. Although these characteristics are adjusted for in the regression models, residual compositional differences could still influence the observed trends, and we did not estimate transition-based models. Sixth, because facility readiness is measured contemporaneously at the district-year level rather than as a single intervention with a common adoption date, and because districts can move between readiness categories across rounds, conventional difference-in-differences diagnostics, such as pre-treatment parallel-trends tests, event-study plots, anticipation or placebo tests, and staggered-adoption estimators, are not directly applicable. Accordingly, we present our findings as differential associations in trends rather than causal effects, and we did not undertake additional sensitivity analyses using alternative readiness thresholds, control groups, or stratified subsamples (e.g., by urban–rural residence, administrative division, baseline poverty, or public/private facility mix), which remain valuable directions for future research. Finally, our readiness measure captures structural capacity only; we were unable to supplement it with process-quality or patient-experience indicators, which were not consistently available across the linked survey rounds, so we cannot empirically establish that structural readiness translated into the quality of care actually experienced by women.

The data sources used for this study did not collect information on ideation and norms in maternal health services use, which remains a gap in this analysis. Further research is needed to understand beliefs and social/cultural norms around pregnancy, delivery, and postpartum, as well as their relation to service utilization and quality of care in health facilities. Despite an ongoing global movement for measuring and improving respectful care, only a few studies have looked into respectful maternity care in Bangladesh [81,82]. Further research is also needed to understand respectful maternity care as an important factor in influencing the use of services.