Review Papers Claim by Claim

Visual summary — Key outcome distribution (from paper)

Top-line assessment

The paper provides a focused, policy-relevant descriptive and multivariable analysis of delivery attendants in Bangladesh's rural population using the nationally representative BDHS-2014 dataset; principal findings are low skilled attendance (35.9%), dominance of TBAs (52.1%), and strong socioeconomic/regional correlates (education, wealth, antenatal care, division) — results reported with survey weighting and multivariable adjustment but constrained by omitted supply-side variables and cross-sectional design Korim et al., Birth attendance (preprint)

Strengths (visual + bullet)

• Nationally representative DHS sampling frame, appropriate use of survey weights and 'svy' commands for variance estimation, and restriction to births in the preceding 3 years to limit recall bias Korim et al., methods
• Clear operationalization of 'skilled' vs 'unskilled' attendants aligned with WHO definitions as used by BDHS (doctors, nurses, midwives, FWVs, CSBAs = skilled) which aids comparability across studies Korim et al., skill definitions

Key limitations and blindspots (critical)

1. Omitted supply-side/care-access variables: Distance to facility, local availability of skilled staff, facility capacity, transportation cost and user fees are not included in BDHS variables used and therefore not controlled for; these are major confounders for delivery attendant choice (authors note this) Korim et al., limitations
2. Cross-sectional & survival bias: DHS cross-sectional retrospective design prevents causal claims and the analytic sample excludes women who died during childbirth, potentially biasing associations if mortality correlates with unskilled attendance (authors acknowledge survival bias) Korim et al., study limits
3. CSBA program measured as exposure but with tiny observed coverage (0.1%): the near-zero CSBA coverage limits inference on program impact; reporting (0.1%) should prompt verification and disaggregation by year/area because program roll-out might vary locally Korim et al., CSBA finding
4. Residual confounding & model selection: Covariate selection used p<0.20 in bivariate screening and hierarchical two-stage multilevel approach; while defensible, this approach risks omitted confounders and over/under-adjustment; VIFs were checked but full multilevel random effects structure (cluster, EA, union) is not fully described in text for reproducibility Korim et al., modelling
5. Limited reporting of effect sizes for all covariates and missing data handling: The paper reports selected AORs (e.g., education, wealth, Khulna division) but does not fully tabulate model diagnostics, missing data counts or sensitivity analyses—these reduce reproducibility and confidence in effect magnitude robustness Korim et al., reporting gaps

Interpretation & implications (evidence-weighted)

The associations (higher maternal/husband education, antenatal care by skilled provider, and wealth linked with greater SBA use) are consistent with prior DHS-based literature and are biologically/plausibly mediated via knowledge, care-seeking behaviour, and affordability; however, because supply-side variables were unobserved, we cannot quantify how much of the observed socioeconomic gradient is mediated by physical access or local facility supply rather than household demand alone Korim et al., results interpretation

Immediate methodological improvements I would request / reproduce steps

• Provide full model specification: list fixed effects, random effects (levels: cluster/EA/household), survey-weight incorporation method in multilevel modelling, and full VIF/multicollinearity diagnostics.
• Report missing-data patterns, how records were excluded, and sensitivity analyses using multiple imputation for covariates with missingness.
• Disaggregate CSBA by administrative unit and year (if possible) to confirm the 0.1% result is not a coding or rare-sample artefact.
• If feasible, add proxy supply-side variables available in BDHS (e.g., urban/rural division, reported facility delivery rates, travel time proxies) or merge with facility-census datasets to assess supply constraints.

What would change my conclusions (falsifiability)

• If addition of supply-side variables (local facility density, distance/time, skilled-staff availability) and transportation/cost eliminated the wealth/education associations, the original interpretation that socioeconomic status primarily drives SBA uptake would need revision.
• If longitudinal data showed SBA uptake rises strongly with proximate facility improvements but not with household wealth rises, that would also revise causal inference toward supply-side drivers.

Recommended next empirical steps

1. Merge BDHS cluster-level coordinates (or administrative unit) with facility-census data (health facility registry) to model how local supply (distance/time, facility type, midwife density) modifies individual determinants using multilevel models.
2. Conduct sensitivity analysis using multiple imputation and alternative variable selection strategies (e.g., directed acyclic graph (DAG)-guided adjustment) rather than p-value screening.
3. Qualitative sampling in divisions with low SBA (e.g., Sylhet) and high SBA (Khulna) to identify contextual supply and cultural drivers explaining the observed division-level differences.

Author follow-up request (for reproducibility)

Please share: (1) Stata do-file(s) for analytic steps including survey weight application and model commands; (2) full Table 3 model output and VIFs; (3) counts of missing data per covariate; (4) any cluster identifiers used for multilevel steps so independent reanalysis can be performed.

Run deeper science agent to (a) re-run multilevel models with DHS public files, (b) merge facility data, (c) produce sensitivity analyses and richer figures.