Introduction

Civil registration and vital statistics systems are the only continuous source of individual-level birth and death data, and their completeness determines whether a country can measure its own mortality.(1,2) Where registration is incomplete, mortality estimates depend on sample surveys and modelling rather than on counted events, and subnational planning suffers most.(3)

India operates a statutory Civil Registration System under the Registration of Births and Deaths Act.(4) Its birth registration has improved markedly, but death registration has lagged, and completeness varies widely between states.(5,6) Several assessments conclude that the system is approaching usability for subnational mortality estimation in some states while remaining far from it in others.(7-9) The COVID-19 pandemic exposed the consequences directly, when incomplete routine death surveillance made excess mortality difficult to quantify.(10,11)

Attention has concentrated on births. Determinants of birth registration in India are well described, and the expansion of institutional delivery is widely credited with driving improvement, because a facility birth creates a documentary event that can be notified.(12,13) International work similarly treats the facility birth as the natural point of capture for registration.(14-16)

Whether that same health-system contact improves death registration has not been tested at district level. The question matters because the policy inference differs sharply. If facility contact raises both, then expanding service coverage will close the gap in time. If it raises only birth registration, then death certification requires a separate intervention, and continued investment in delivery capacity will not help.

Two features of the Indian evidence base make a district-level analysis necessary. First, state averages conceal large internal variation in health-system performance.(17) Second, an association observed across all districts may reflect state context rather than any district-level mechanism, and only a within-state comparison can separate them.(18)

This analysis describes the birth-minus-death registration gap across 705 Indian districts and identifies its correlates. It then tests whether those correlates survive a comparison of districts inside the same state, and locates the districts where the two registration systems are most discordant.

Material and Methods

Study design and data source

We conducted an ecological cross-sectional analysis of publicly available aggregated district-level data from the National Family Health Survey-5, conducted in India during 2019-21.(19) District factsheet indicators were compiled into a numeric analytical file. Reporting follows the Strengthening the Reporting of Observational Studies in Epidemiology statement for cross-sectional studies.(20)

Analytical units

The analytical panel comprised 705 districts across 36 states and union territories with complete data on all study indicators. A district listed as "Chandel" under Mizoram in the source file is a known listing error, Chandel being a district of Manipur, and was excluded; the exclusion was re-verified during analysis. Three state names carrying source misspellings or non-standard forms were corrected for display only, with no value altered and no unit merged.

Measures and their denominators

The two registration indicators do not share a denominator or a reference period, and this is carried explicitly through the analysis rather than assumed away. Birth registration is the percentage of children under five years whose birth was registered with the civil authority, measured among surviving children over a five-year window. Death registration is the percentage of deaths in the household during the last three years that were registered, measured over a three-year window among decedents who are predominantly older adults. Their arithmetic difference, the registration gap, is therefore a descriptive contrast between two completeness measures rather than a single coherent estimand, and is interpreted as such throughout.

Five district correlates were used: women aged 15-49 with ten or more years of schooling; population living in households with improved sanitation; households using clean cooking fuel; households with any member covered by health insurance or a financing scheme; and institutional births in the five years before the survey. Each is a percentage. All were standardised, so coefficients express the difference in an outcome per one standard deviation higher indicator.

Statistical analysis

Models were weighted by the number of households surveyed in each district. This is a precision weight reflecting the sampling effort behind each district estimate; because district samples are of broadly similar size, weighting does not make the estimates population-representative, and an unweighted sensitivity analysis is reported. Variance was clustered at the state or union territory.

Two model families were fitted for each of the three outcomes. The first regressed the outcome on all five correlates across all districts, so that comparisons span state boundaries. The second added state fixed effects, restricting comparison to districts within the same state. A correlate that holds in the first but disappears in the second is a marker of state context rather than a district-level lever. Because 30 correlate-outcome tests were performed across the two model families, P values were corrected together using the Benjamini-Hochberg false discovery rate, and q values below 0.05 were treated as surviving correction.(21)

Inference was strengthened in four ways, because 36 clusters is few enough that the cluster-robust sandwich is known to reject too often. First, p values for the principal contrast were recomputed by wild cluster bootstrap with 4,999 Rademacher draws and the null imposed. Second, because both registration outcomes are bounded percentages, models were refitted as fractional (quasi-binomial) logits and reported as average marginal effects on the percentage-point scale. Third, each model was refitted 36 times, omitting one state at a time. Fourth, Huber M-estimation was used to check that influential districts do not drive the linear estimates.

To distinguish a genuine loss of association from ordinary regression dilution, the change between the two specifications was tested for every predictor, treating the estimates as independent, which is conservative because they share data. Generic dilution would shrink all predictors; a predictor-specific pattern would not.

We decomposed the variance in each outcome into a between-state and a within-state component using the coefficient of determination of a model containing state indicators alone. Districts were classified into four groups by birth registration of 90% or more and death registration of 70% or more, and the concentration of the death-registration shortfall across districts and states was quantified. Because the published district estimates carry no standard errors, the sensitivity of that classification was assessed by moving both thresholds and by counting districts lying within five percentage points of either boundary. Analyses used Python 3.14 with pandas, statsmodels and SciPy; all result files are generated directly by the analysis script without manual editing.

Ethical approval

Not applicable. The study used only publicly available aggregated secondary data with no individual-level or identifiable information, so institutional review and informed consent were not required.

Results

Registration levels and the gap

Across 705 districts, median birth registration was 94.9% and median death registration 76.2%, giving a median registration gap of 17.1 percentage points and a mean of 19.8 (Table 1). Death registration was far more variable than birth registration, ranging from 8.3% to 100% with a standard deviation of 21.0 points, against 51.6% to 100% and a standard deviation of 9.4 points for births. Death registration fell below 70% in 279 districts and below 50% in 137 districts. The gap was negative, meaning deaths were better registered than births, in 23 districts, 3.3% of the total (Figure 1).

The shortfall was concentrated. The worst decile of districts accounted for 24.4% of the total national death-registration deficit and the worst quartile for 51.7%. Five states, Uttar Pradesh, Bihar, Jharkhand, Arunachal Pradesh and Madhya Pradesh, together accounted for 50.5% of the deficit.

Figure 1: Birth registration against death registration, 705 Indian districts, National Family Health Survey-5. Each point is one district. Dashed lines mark birth registration of 90% and death registration of 70%. The shaded band identifies districts registering fewer than seven in ten deaths. Colours denote the four registration groups defined in Table 4.

Correlates across districts

Across all districts, institutional births showed the strongest association with both outcomes (Table 2). Each standard deviation higher coverage went with 5.05 percentage points higher birth registration (95% confidence interval 3.61 to 6.50) and 9.13 points higher death registration (6.77 to 11.49). Both survived correction for multiple testing. Improved sanitation was associated with higher birth registration (2.51; 0.88 to 4.13) and higher death registration (4.54; 0.94 to 8.13). Women's schooling, clean cooking fuel and health insurance showed no association that survived correction. Four of the 30 tests across both model families survived the false discovery rate procedure at this stage, with three more surviving in the within-state models reported below.

Because institutional births were more strongly associated with death registration than with birth registration, they were associated with a narrower gap across districts (-4.08 points per standard deviation; -5.96 to -2.20).

The association reverses within states

Adding state fixed effects changed the picture (Table 3, Figure 2). The association between institutional births and death registration disappeared entirely, from 9.13 points per standard deviation across districts to -0.06 (-2.13 to 2.02; P=0.959) within states, an attenuation of essentially the whole effect. The association with birth registration attenuated but persisted (5.05 to 2.87; 0.39 to 5.35; P=0.023).

Consequently the gap association reversed sign. Across districts, higher institutional births went with a narrower gap; within a state, districts with more institutional births had a wider gap (2.93 points per standard deviation; 0.87 to 4.98; P=0.005). Health insurance and clean cooking fuel, by contrast, retained or strengthened their association with death registration within states.

Figure 2: Adjusted associations of five district indicators with birth registration, death registration and the registration gap, comparing estimates across all districts with estimates within states. Circles are models spanning all 705 districts; squares are models with state fixed effects. Bars are 95% confidence intervals. Coefficients are percentage points per one standard deviation higher indicator. The association between institutional births and death registration disappears within states.

Three features rule out ordinary regression dilution as the explanation. First, the attenuation was specific rather than general: institutional births lost 100.6% of their association with death registration, whereas the other four predictors of the same outcome changed by a median of -5.9%, that is, on balance they did not shrink, and no individual change was significant (all P≥0.55; Table S6). Second, the difference between the between-district and within-state estimates for institutional births was itself significant (P=1.0 x 10^-8) on a conservative test. Third, every alternative method reproduced the reversal. A wild cluster bootstrap with 4,999 draws gave P=0.96 within states against P=0.0002 across districts. A fractional logit, which respects the bounds of a percentage outcome, gave an average marginal effect of -0.55 points within states against 7.47 across districts. Huber M-estimation gave 0.18 against 9.77. Omitting each state in turn produced within-state estimates from -0.55 to 0.69, crossing zero, against a stable 8.56 to 9.52 across districts.

Variance decomposition explained why. Death registration was 82.3% a between-state phenomenon and only 17.7% within-state, whereas birth registration was 69.6% between-state and the gap 67.5%. Death registration is therefore predominantly determined at the level of the state system, though about a third of the variation in the gap still lies between districts of the same state.

Districts where registration systems are most discordant

Classifying districts on both indicators produced four groups (Table 4). Four hundred districts registered both well. One hundred and ninety-one registered both poorly, with mean birth registration 78.3% and mean death registration 44.8%. Twenty-six combined birth registration below 90% with death registration of 70% or more.

The remaining 88 districts, spread across 17 states, combined birth registration of 90% or more with death registration below 70% (Figure 3). Their mean birth registration was 94.3% and mean death registration 58.6%, a mean gap of 35.8 points. Critically, their mean institutional birth coverage was 87.7%, close to the national picture. Among the 417 districts where institutional births reached 90% or more, mean birth registration was 95.2% but mean death registration only 81.5%. Of the 88 districts, Assam contributed 21 such districts, Arunachal Pradesh 11, and Madhya Pradesh and Uttar Pradesh nine each.

Figure 3: States containing districts with birth registration of 90% or more and death registration below 70%. Bars show the number of such districts per state; annotations give the mean death registration in those districts. Eighty-eight districts across 17 states meet both criteria.

This grouping is a planning device rather than a precise list. Moving the thresholds by five percentage points changed the group size from 41 to 142 districts, and 264 districts lie within five points of one boundary or the other. Individual district membership should therefore be treated as indicative, although the existence of a substantial group registering most births and few deaths is not sensitive to the cut-points.

Sensitivity analyses

Unweighted models produced closely similar estimates, indicating that the precision weighting did not drive the findings. The wild cluster bootstrap, fractional logit, leave-one-state-out and Huber estimates are reported in full in the supplementary file, and all reproduce the pattern described above.

Discussion

Across 705 Indian districts, birth registration exceeded death registration by a median 17 percentage points, and the shortfall was concentrated in a minority of districts and a handful of states. The central finding, however, is not the size of the gap but its mechanism. Institutional birth coverage, the strongest apparent correlate of death registration across the country, carried no association once districts were compared within their own state. Facility contact tracks birth registration; it does not track death registration. Because the design is cross-sectional and ecological, this is an association rather than a demonstrated mechanism, but the pattern is consistent across every estimator applied.

This distinction is easy to miss and consequential. A cross-sectional analysis that pools districts across states recovers a strong positive association between institutional births and death registration, which invites the conclusion that strengthening delivery services will improve mortality statistics. The within-state comparison shows that this association is carried by state context: states that have built delivery capacity have also, separately, built registration administration. Inside any one state, a district with more facility births registers no more deaths than its neighbour, and in fact shows a wider birth-death gap because facility contact lifts one side of the ledger only.

Alternative explanations deserve consideration. Districts with more institutional births are, within a state, likely to be more urban, and urban death registration faces its own barriers, including deaths among migrants and in private facilities that may not be notified locally. Differential measurement error is a second possibility, since district death-registration estimates rest on relatively few household deaths and are noisier than birth estimates. Neither explanation accounts for the specificity of the finding. Ordinary noise would attenuate every predictor. Yet the other four correlates of death registration showed no significant attenuation, and the change for institutional births was large and highly significant.

The most economical explanation is structural. A facility birth generates an institutional record, a notification pathway and an administrative actor with an incentive to report.(13,14) Most deaths in India occur at home, outside any facility record, and the medical certification of cause of death applies chiefly to institutional deaths. Even where certification is required, its completeness and quality are known to be poor, and studies from tertiary hospitals repeatedly find substantial error rates that respond to targeted training.(22-25) Certification difficulties were amplified during the pandemic.(26) The consequence for national statistics is that a large share of deaths reaching the system carries uninformative cause codes.(27) Structured district-level programmes to strengthen cause-of-death statistics are being trialled precisely because the routine pathway does not deliver.(28)

Our findings align with, and extend, the established literature. Assessments of the Civil Registration System have documented improving but uneven death registration and persistent inter-state inequality.(5,6) Analyses of birth registration determinants have emphasised institutional delivery.(12) Recent work has shown that female deaths are particularly under-registered, indicating that the shortfall is not merely geographic but also social.(29) Surveys themselves undercount some vital events, as demonstrated for stillbirths in both the Sample Registration System and the National Family Health Survey, which is a reminder that no single source is complete.(30) The concentration of the deficit in a few large states echoes the geography of other district-level health inequalities and of programme coverage such as the Janani Suraksha Yojana.(17,31) That the Aspirational Districts Programme has been evaluated using the same survey underlines that district-level targeting is an established policy instrument in India.(32)

Implications for policy and practice

Four implications follow. First, death registration should be monitored and targeted as a distinct indicator, not assumed to follow health-service expansion. The variance decomposition indicates that state-level administrative reform is the dominant lever, but a third of the variation in the gap lies between districts of the same state, so district action is not futile.

Second, the 88 districts identified here are an immediately actionable list. They already register nine in ten births and have near-average institutional delivery coverage, meaning the administrative apparatus and the health-system contact both exist. What is missing is the death-side workflow. These are the districts where an intervention has the best prior odds of success.

Third, because most deaths occur at home, the intervention cannot be facility-centred. It requires community-level notification by frontline workers, simplified reporting for non-institutional deaths, and linkage between the health system and the registrar.

Fourth, for clinicians, medical certification of cause of death is a professional duty rather than a clerical formality, and the evidence indicates that its quality improves with brief, targeted training.(24,25) Incorporating certification into undergraduate and internship teaching is a low-cost contribution to national mortality statistics.

Strengths and limitations

Strengths include national district-level coverage, explicit handling of the differing denominators of the two registration indicators, separation of between-state from within-state variation, correction for multiple testing, and a fully scripted and reproducible analysis.

The limitations are substantial. The design is ecological and cross-sectional; associations between district aggregates need not hold for individuals, and no causal claim is made.(18) Both registration indicators are self-reported by households rather than drawn from the registration system itself, so they measure reported registration and may misclassify. The two indicators differ in denominator, reference period and age structure, so the gap is a contrast and not a difference in a common quantity. Household deaths in the preceding three years are relatively rare events, so district death-registration estimates carry more sampling error than birth estimates, which may attenuate within-state associations. Weighting by households surveyed is a precision weight and does not confer population representativeness. Only five district correlates were available from the factsheet, and administrative variables such as registrar density or distance to a registration office, which are plausibly the operative factors, were not observable. Finally, the survey was fielded in 2019-21, overlapping the onset of the COVID-19 pandemic.

Conclusion

Birth registration exceeded death registration by a median 17 percentage points across 705 Indian districts, with death registration below 70% in 279 districts. Institutional birth coverage predicted death registration across the country but not within states, showing that health-facility contact converts into birth registration alone. Eighty-eight districts already register most births yet register fewer than seven in ten deaths, and constitute a defined target for a death-side registration intervention.

Acknowledgements

The author acknowledges the International Institute for Population Sciences and ICF for conducting the National Family Health Survey and making district factsheet data publicly available.

Declarations

Author contributions (CRediT): S H S - conceptualisation, methodology, software, formal analysis, data curation, investigation, visualisation, writing - original draft, writing - review and editing.

Conflicts of interest: The author declares no competing financial or non-financial interests.

Funding: No specific funding was received for this work from any agency in the public, commercial or not-for-profit sectors.

Ethical approval and informed consent: Not applicable. The study used only publicly available aggregated secondary data with no individual-level or identifiable information; institutional review and informed consent were not required.

Reporting guideline: The study is reported in accordance with the STROBE statement for cross-sectional observational studies. The completed STROBE checklist is provided as a supplementary file.

Data availability: All source data are publicly available from the National Family Health Survey-5 district factsheets. The derived analytical file, the complete analysis code and all result tables are available from the corresponding author on reasonable request.

Use of artificial intelligence: Generative artificial-intelligence tools were used to assist with literature retrieval, analysis code drafting and language editing. The author verified all data, analyses, results and interpretations, and accepts full responsibility for the content. No artificial-intelligence system is listed as an author.

References

1. Mills S, Lee JK, Rassekh BM. An introduction to the civil registration and vital statistics systems with applications in low- and middle-income countries. J Health Popul Nutr 2019;38(Suppl 1):23. doi: 10.1186/s41043-019-0177-1.
2. World Health Organization. Civil registration and vital statistics (CRVS) [Internet]. Geneva: World Health Organization. Available from: https://www.who.int/data/data-collection-tools/civil-registration-and-vital-statistics-(crvs) [Accessed 19 July 2026].
3. Rao C. Elements of a strategic approach for strengthening national mortality statistics programmes. BMJ Glob Health 2019;4(5):e001810. doi: 10.1136/bmjgh-2019-001810.
4. Office of the Registrar General, India. Civil Registration System, India [Internet]. New Delhi: Ministry of Home Affairs, Government of India. Available from: https://crsorgi.gov.in/ [Accessed 19 July 2026].
5. Kumar GA, Dandona L, Dandona R. Completeness of death registration in the Civil Registration System, India (2005 to 2015). Indian J Med Res 2019;149(6):740-747. doi: 10.4103/ijmr.IJMR_1620_17.
6. Basu JK, Adair T. Have inequalities in completeness of death registration between states in India narrowed during two decades of civil registration system strengthening? Int J Equity Health 2021;20(1):195. doi: 10.1186/s12939-021-01534-y.
7. Rao C, Gupta M. The civil registration system is a potentially viable data source for reliable subnational mortality measurement in India. BMJ Glob Health 2020;5(8):e002586. doi: 10.1136/bmjgh-2020-002586.
8. Gupta M, Rao C, Lakshmi PVM, et al. Estimating mortality using data from civil registration: a cross-sectional study in India. Bull World Health Organ 2016;94(1):10-21. doi: 10.2471/BLT.15.153585.
9. Gupta A, Asadullah M, Kumar R, et al. Exploratory analysis of the suitability of data from the civil registration system for estimating excess mortality due to COVID-19 in Faridabad district of India. Indian J Med Res 2022;156(3):421-428. doi: 10.4103/ijmr.ijmr_1421_21.
10. Shewade HD, Parameswaran GG, Mazumder A, et al. Adjusting Reported COVID-19 Deaths for the Prevailing Routine Death Surveillance in India. Front Public Health 2021;9:641991. doi: 10.3389/fpubh.2021.641991.
11. Natarajan S, Subramanian P. Systematic Review of Excess Mortality in India during the Covid-19 Pandemic with Differentiation between Model-Based and Data-Based Mortality Estimates. Indian J Community Med 2022;47(4):491-494. doi: 10.4103/ijcm.ijcm_11_22.
12. Kumar K, Saikia N. Determinants of birth registration in India: Evidence from NFHS 2015-16. PLoS One 2021;16(9):e0257014. doi: 10.1371/journal.pone.0257014.
13. Verma S, Dwivedi LK, Kamal R. Impact of Institutionalisation of Births on Health Policies and Birth Registration in India. Ann Glob Health 2024;90(1):53. doi: 10.5334/aogh.4474.
14. Paleker M, Boggs D, Jackson D, et al. Closing the birth registration gap for Every newborn facility birth: literature review and qualitative research. Glob Health Action 2023;16(1):2286073. doi: 10.1080/16549716.2023.2286073.
15. Bhatia A, Krieger N, Beckfield J, et al. Are inequities decreasing? Birth registration for children under five in low-income and middle-income countries, 1999-2016. BMJ Glob Health 2019;4(6):e001926. doi: 10.1136/bmjgh-2019-001926.
16. Bhatia A, Ferreira LZ, Barros AJD, et al. Who and where are the uncounted children? Inequalities in birth certificate coverage among children under five years in 94 countries using nationally representative household surveys. Int J Equity Health 2017;16(1):148. doi: 10.1186/s12939-017-0635-6.
17. Dandona R, Majumder M, Kumar GA. Population-level trends over a decade in geographical inequality for opportunity in access to maternal care services: a cross-sectional analysis from the National Family Health Surveys in India. BMJ Open 2024;14(11):e083922. doi: 10.1136/bmjopen-2024-083922.
18. Morgenstern H. Ecologic Studies in Epidemiology: Concepts, Principles, and Methods. Annu. Rev. Public Health 1995;16:61-81. doi: 10.1146/annurev.pu.16.050195.000425.
19. International Institute for Population Sciences (IIPS), ICF. National Family Health Survey (NFHS-5), 2019-21: India: Volume I. Mumbai: IIPS; 2021. Available from: https://dhsprogram.com/pubs/pdf/FR375/FR375.pdf [Accessed 19 July 2026].
20. von Elm E, Altman DG, Egger M, et al. The Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) statement: guidelines for reporting observational studies. PLoS Med 2007;4(10):e296. doi: 10.1371/journal.pmed.0040296.
21. Benjamini Y, Hochberg Y. Controlling the False Discovery Rate: A Practical and Powerful Approach to Multiple Testing. J R Stat Soc Series B: Methodol 1995;57(1):289-300. doi: 10.1111/j.2517-6161.1995.tb02031.x.
22. Anand A, Khanna D, Singh P, et al. Evaluation of Medical Certification of Cause of Death in Tertiary Cancer Hospitals in Northern India. Health Inf Manag 2024;53(2):121-128. doi: 10.1177/18333583221144665.
23. Dash SK, Behera BK, Patro S. Accuracy in certification of cause of death in a tertiary care hospital--a retrospective analysis. J Forensic Leg Med 2014;24:33-36. doi: 10.1016/j.jflm.2014.03.006.
24. Chaudhary A, Singh G, Hooda A, et al. Impact of an Educational Measure on the Quality of Medical Certificate of Cause of Death at a Tertiary Care Teaching Hospital: A Pre-post Interventional Study. Cureus 2024;16(2):e54721. doi: 10.7759/cureus.54721.
25. Thanigasalam A, Ramu R, Rajarethinam K. Improving completeness and reducing errors in medical certification of cause of death: The impact of electronic mortality software in a tertiary care centre in South India. Indian J Med Res 2025;161(4):420-424. doi: 10.25259/IJMR_1394_2024.
26. Juyal D, Kumar A, Pal S, et al. Medical certification of cause of death during COVID-19 pandemic - a challenging scenario. J Family Med Prim Care 2020;9(12):5896-5898. doi: 10.4103/jfmpc.jfmpc_1435_20.
27. Naghavi M, Richards N, Chowdhury H, et al. Improving the quality of cause of death data for public health policy: are all 'garbage' codes equally problematic? BMC Med 2020;18(1):55. doi: 10.1186/s12916-020-01525-w.
28. Grover A, Nair S, Sharma S, et al. Strengthening Cause of Death Statistics in Selected Districts of 3 States in India: Protocol for an Uncontrolled, Before-After, Mixed Method Study. JMIR Res Protoc 2024;13:e51493. doi: 10.2196/51493.
29. Verma S, Kamal R, Dwivedi LK, et al. Who gets counted? Understanding low female death registration in India. PLoS One 2025;20(11):e0337224. doi: 10.1371/journal.pone.0337224.
30. Dandona R, George S, Majumder M, et al. Stillbirth undercount in the sample registration system and national family health survey, India. Bull World Health Organ 2023;101(3):191-201. doi: 10.2471/BLT.22.288906.
31. Mishra PS, Kumar P, Srivastava S. Regional inequality in the Janani Suraksha Yojana coverage in India: a geo-spatial analysis. Int J Equity Health 2021;20(1):24. doi: 10.1186/s12939-020-01366-2.
32. Agarwal SK, Mishra S. Health impact evaluation of Aspirational Districts Program in India: Evidence from National Family Health Survey. Econ Hum Biol 2024;54:101411. doi: 10.1016/j.ehb.2024.101411.

Supplementary Material

Birth and death registration gap across 705 Indian districts, NFHS-5

Contents

S1. STROBE checklist for cross-sectional studies (mapped to manuscript locations)

S2. Variance decomposition of registration outcomes between and within states

S3. Sensitivity analysis without precision weights

S4. Districts with birth registration of 90% or more and death registration below 70%, by state

S5. Health-facility contact and registration

S6. Change between between-district and within-state estimates

S7. Wild cluster bootstrap inference

S8. Fractional logit models with average marginal effects

S9. Influence and robust estimation

S10. Sensitivity of the priority-district classification

S1: STROBE checklist for cross-sectional studies

This study is reported in accordance with the Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) statement for cross-sectional observational studies.

S2: Variance decomposition of registration outcomes

The proportion of variation in each outcome attributable to differences between states, obtained as the coefficient of determination of a model containing state indicators alone. A high between-state share indicates that the outcome is determined predominantly at the level of the state system rather than the district, and it is also the reason a reader might expect within-state estimates to attenuate through regression dilution alone. Table S6 addresses that objection directly.

S3: Sensitivity analysis without precision weights

Models identical to Table 2 of the manuscript but fitted without weighting by the number of households surveyed. Closely similar estimates indicate that the precision weighting does not drive the reported findings.

S4: Priority districts by state

All districts with birth registration of 90% or more and death registration below 70%, aggregated by state. These districts already register the large majority of births and have near-average institutional birth coverage, so the registration apparatus and the health-system contact both exist; what is absent is the death-side workflow. Table S10 shows how sensitive this grouping is to its thresholds.

Table S5: Health-facility contact and registration

Institutional birth coverage is a marker of health-system contact with the household. This table contrasts its association with birth and with death registration in models containing institutional births as the only district indicator, estimated first across all districts and then within states. The note gives mean registration levels among the 417 districts where institutional births reached 90% or more, showing that near-universal facility delivery coexists with substantially incomplete death registration.

Table S6: Change between between-district and within-state estimates

Attenuation is the percentage reduction in a coefficient when state fixed effects are added. Ordinary regression dilution would attenuate every predictor, because the fixed effects absorb most of the variance in death registration and leave a noisier residual. Here only institutional births loses its association; none of the other four correlates of death registration changes significantly. The difference between the two estimates is tested treating them as independent, which is conservative because they share data and are positively correlated, so the true standard error of the difference is smaller than the one used.

Table S7: Wild cluster bootstrap inference

With 36 state clusters the cluster-robust sandwich is known to reject too often, because asymptotic cluster inference assumes many clusters. P values for the principal contrast were therefore recomputed by wild cluster bootstrap with 4,999 Rademacher draws and the null imposed on the restricted fit. Conclusions are unchanged.

Table S8: Fractional logit models with average marginal effects

Birth and death registration are bounded percentages, so a linear model can predict outside the admissible range and imposes constant marginal effects across it. Models were refitted as fractional (quasi-binomial) logits with a logit link and cluster-robust variance, and are reported as average marginal effects in percentage points so that they remain comparable with the linear estimates. The pattern matches the linear models.

Table S9: Influence and robust estimation

Each model was refitted 36 times, omitting one state in turn, and additionally by Huber M-estimation to limit the leverage of influential districts. The between-district estimate is stable across omissions; the within-state estimate crosses zero, as expected for an estimate centred on the null.

Table S9a: Leave-one-state-out estimates for institutional births and death registration

Table S9b: Huber M-estimation

S10. Sensitivity of the priority-district classification

The published district estimates carry no standard errors, so classification uncertainty is assessed by moving both thresholds and by counting districts near a boundary. Group size varies from 41 to 142 districts across the threshold combinations examined, and 264 districts lie within five percentage points of one boundary or the other. Individual district membership is therefore indicative rather than definitive, although the existence of a substantial group registering most births and few deaths is not sensitive to the cut-points.

Reproducibility

Analyses used Python 3.14 with pandas, NumPy, statsmodels, SciPy and Matplotlib, with a fixed random seed. All tables and figures are generated directly by the analysis and figure scripts without manual editing. Every reference in the manuscript was retrieved programmatically from PubMed or Crossref, or recorded with a URL confirmed reachable; none was transcribed by hand, and the reference list was independently re-verified against freshly retrieved PubMed records before submission. The derived analytical file and the complete analysis code accompany this submission.