Introduction

Drowning is among the leading causes of unintentional-injury death in Indian children, yet India is one of the few high-burden countries without a dedicated national prevention strategy.(1,2) The World Health Organization's 2024 global status report asks every member state to halve drowning deaths by 2030 and sets out ten recommended interventions spanning supervised day care, swimming and water-safety skills, barriers around water, life jackets, bystander rescue, flood warning, resilient infrastructure, boating standards, public awareness and school curricula.(1)

Translating that target into Indian policy requires answers to four questions the published literature has not addressed. How completely do routine systems record child drowning? Where does burden concentrate once measurement error is accounted for? Which states combine high burden with weak capacity to deliver a programme? And what does the evidence actually support about which interventions to fund first?

Each question is obstructed by a data limitation. The National Crime Records Bureau publication Accidental Deaths and Suicides in India reports drowning deaths by state and, separately, by age, but publishes no state-by-age cross-tabulation, and collapses the highest-risk under-five band into a single below-14 category.(3) The Sample Registration System reports an aggregate injuries category and does not disaggregate drowning.(4) Global Burden of Disease estimates provide age-specific state figures but are themselves modelled outputs carrying wide uncertainty.(2)

A further limitation shapes what any prioritisation can claim. No Indian state publishes an administrative indicator of water-safety programme activity: there is no series for swim-training coverage, no register of fenced water bodies and no published rescue-service density. That absence is not a gap in this analysis but a finding about the Indian policy environment.

This study therefore matches its methods to the evidence available. Rather than treating either data source as correct, it combines them in a measurement-error model that estimates how completely deaths are recorded. Rather than scoring state readiness on judgement, it uses measured indicators and reports classification as a probability. Rather than ranking interventions on effect sizes the literature does not support, it propagates the uncertainty in that literature and reports what can be concluded.

Material and Methods

Study design and reporting

This was a cross-sectional secondary analysis of nationally representative aggregate data for 2014-2023, combining Bayesian hierarchical modelling with multi-criteria decision analysis. Reporting follows the Strengthening the Reporting of Observational Studies in Epidemiology statement for cross-sectional studies, and a completed checklist is provided as Supplementary Table S1.(5)

Ethical considerations

This study used only publicly available aggregated secondary data with no individual-level or identifiable information. No individual was enrolled, contacted or intervened upon, and no identifiable personal record or biological specimen was accessed. Institutional review board approval and informed consent were not applicable.

Data sources

State-level drowning deaths with 95% uncertainty intervals for the three paediatric age bands, both sexes, 2014-2023, were taken from the Global Burden of Disease 2023 release.(2) Drowning deaths by state and by age for 2023 were extracted from the National Crime Records Bureau.(3) State-level capacity and contextual indicators were computed from the National Family Health Survey-5 district factsheet, aggregating 707 district records to state level.(6) The national 0-14 denominator came from World Population Prospects 2024.(7) The intervention set came from the World Health Organization 2024 global status report.(1)

Analyses covered the 30 states for which the Global Burden of Disease reported state-level paediatric drowning and the National Family Health Survey-5 reported the capacity indicators. The Global Burden of Disease aggregate for other union territories has no counterpart in any state-level covariate source and was excluded by name. The National Crime Records Bureau state table for 2023 sums to 37,518 of 37,738 national drowning deaths, a shortfall of 220 (0.6%) attributable to Delhi, which the table omits; Delhi therefore contributes through the Global Burden of Disease likelihood alone.

Child population denominators

State child populations for 2023 were obtained by applying the National Family Health Survey-5 measured under-15 share to a 2023 state population and normalising to the World Population Prospects national 0-14 total. This was preferred to apportioning the national total by Census 2011 state shares, because the published share table resolves shares to three decimal places and therefore assigns an identical share, and an identical child population, to several small states, distorting their rates. The Census-share denominator was retained as a sensitivity analysis.

Two-source burden model

For each state s, the paediatric drowning rate per 100 000 was modelled hierarchically on the log scale, giving true deaths D(s) for child population N(s). Two likelihoods were specified: the Global Burden of Disease estimate as a normal measurement of D(s) with standard deviation implied by its published uncertainty interval, and the recorded all-age drowning count as a negative-binomial observation with mean D(s) multiplied by a recording ratio k(s). The logarithm of k(s) was regressed on standardised death-registration completeness with a state-level random effect.

The quantity k(s) is the number of recorded all-age drowning deaths per true child drowning death in state s. Because no state-by-age cross-tabulation is published, k(s) deliberately absorbs the paediatric share together with recording completeness rather than fixing the paediatric share at its national value. Only k(s) is identified by the recorded count alone; partial pooling across states, and the regression of k(s) on measured registration completeness, are what allow the second source to carry information about D(s). The prior on the intercept was centred on the national ratio of recorded all-age to estimated paediatric deaths.

To establish whether the second source materially changes the answer, the identical model was refitted with the second likelihood removed and posterior rates compared state by state.

Trajectory and projection

The national 0-14 rate was modelled as a Bayesian segmented regression with a break at 2020. The level-change parameter identifies a discontinuity in the series; it does not identify its cause, and no causal interpretation is placed on it.

The national rate was projected to 2030 with a smooth-trend local linear model fitted on the log scale. Beyond the observed series the trend was damped by a factor. Ten annual observations do not identify that factor, so rather than fixing it, a Beta(6, 2) prior was placed over it and integrated out, concentrating mass on the range used in the damped-trend forecasting literature while retaining mass on weak persistence.(8) Because the data do not update it, its posterior equals its prior, and the reported probability expresses uncertainty about trend persistence rather than evidence about it. Probabilities conditional on fixed damping, under alternative damping priors, and under alternative variance priors are reported alongside.

State priority index

A priority index was computed by the Technique for Order of Preference by Similarity to Ideal Solution over six criteria: posterior burden, where higher burden indicates higher priority, and five capacity indicators, where lower capacity indicates higher priority.(9) The capacity indicators were birth registration, death registration, institutional births, four or more antenatal visits and household health-insurance coverage, all measured in the National Family Health Survey-5 and all covering every state at 96% district completeness or better. Indicators with poorer coverage were excluded from the primary index and used in a specification sensitivity analysis.

Two sources of uncertainty were propagated jointly: posterior uncertainty in state burden, by drawing from the two-source model, and sampling uncertainty in each state mean, by resampling districts within states. Two thousand replicates were computed, and results are reported as a score with a credible interval, a rank with an interval, and the probability that a state falls in the highest-priority third. Sensitivity to weighting was assessed under burden-favouring and capacity-favouring schemes.

Intervention appraisal

Each of the ten recommended interventions was matched to its strongest published evidence for an effect on drowning mortality, and effect estimates were re-derived from the primary sources. Where an interval was published, risk ratios were drawn lognormally to match it. Where evidence established direction but not magnitude, the draw was deliberately broad. A single transferability discount, drawn uniformly and applied identically to every intervention in each draw, widened the uncertainty arising from importing effects estimated elsewhere without allowing a judgement about which intervention suits India to drive the ordering. Interventions were ranked over effect and certainty of evidence across 20 000 draws, and results are reported as rank-acceptability probabilities.

Contextual layers

State-level contextual indicators were correlated with posterior burden by Spearman rank correlation, and repeated as partial correlations adjusting for a composite development index defined as the first principal component of six socioeconomic and health-system indicators. These analyses are ecological and exploratory and are reported as associations only.

Statistical analysis

All inference was Bayesian except the contextual correlations. Convergence was assessed by the Gelman-Rubin statistic, bulk effective sample size and divergent-transition count, all of which are reported for every model in Supplementary Table S2. Analyses used Python 3.14 with PyMC 6.0, ArviZ, pandas, NumPy, SciPy and Matplotlib.(10) Analytic code and processed data are available from the author.

Results

National trajectory

Paediatric drowning deaths fell from 24,562 in 2014 to 14,913 in 2023, a reduction of 39.3%, with the rate falling from 6.73 to 4.25 per 100 000 (Figure 1). The segmented model fitted without divergent transitions and with a maximum Gelman-Rubin statistic of 1.00. The pre-2020 slope was negative, indicating continuous decline. The level change at 2020 was 1.08 per 100 000 (95% credible interval 0.17 to 1.92), with posterior probability 0.986 of being positive, and the post-2020 slope change was negative, indicating that the decline resumed.

Figure 1: Paediatric drowning rate, India, 2014-2023. Points show the estimated annual rate per 100 000 children aged 0-14. The dashed line marks 2020, the year at which a level change was estimated. The level change identifies a discontinuity in the series and does not identify its cause.

This identifies a discontinuity, not its cause. School closures, monsoon flooding and pandemic-period changes in death recording all coincided with 2020, and a ten-point annual series cannot distinguish among them or exclude ordinary year-to-year variation.

Recorded against modelled counts

The National Crime Records Bureau recorded 4,829 below-14 drowning deaths in 2023, 12.8% of the 37,738 drowning deaths it reported at all ages.(3) The corresponding Global Burden of Disease estimate for ages 0-14 was 14,913 deaths, 3.1-fold the recorded count.

Both figures estimate different things and neither is a census of events. Global Burden of Disease figures are model outputs produced by redistributing ill-defined causes and correcting for incomplete registration, and they carry wide intervals. The ratio indicates the direction and approximate scale of under-recording rather than an exact multiplier. It is nonetheless large enough that routine police records cannot serve as a stand-alone monitoring source for paediatric drowning.

State burden

Admitting the second source changed state estimates materially. Posterior rates differed from a single-source fit by a median of 6.0%, by up to 22.3%, and by more than 10% in 11 of 30 states; they differed from the raw modelled input rates by a median of 13.4% (Figure 2). Death-registration completeness predicted the recording ratio strongly (coefficient 0.65, 95% credible interval 0.42 to 0.88, posterior probability of a positive coefficient 1.00), confirming that states registering deaths more completely also record proportionately more drowning deaths.

Figure 2: State paediatric drowning burden, 2023. Left panel: posterior rate with 95% credible interval from the two-source model (circles) against the single-source input rate (crosses), for the twenty highest-burden states. Right panel: percentage shift of the two-source posterior from a fit using the modelled source alone.

Posterior rates ranged from 1.52 per 100 000 in Delhi to 5.82 in Madhya Pradesh, a 3.8-fold spread. The five highest-burden states were Madhya Pradesh (5.82, 95% CrI 3.10-9.29); Arunachal Pradesh (5.32, 95% CrI 2.78-8.61); Jharkhand (4.91, 95% CrI 2.78-7.42); Karnataka (4.55, 95% CrI 2.46-7.27); Bihar (4.41, 95% CrI 2.66-6.39) (Table 1). The two-source model showed 21 divergent transitions in 8,000 draws and a maximum Gelman-Rubin statistic of 1.01.

Under the Census-share denominator, state ranks correlated with the primary ranking at a Spearman coefficient of 0.90, with a median rank shift of 1 place but a maximum shift of 12 places. Denominator choice therefore leaves the broad ordering intact while moving individual states substantially, and the states it moves most are those whose Census shares are least precisely resolved.

State priority

Combining burden with measured capacity produced an ordering in which 7 states fell in the highest-priority third with probability at least 0.90: Nagaland, Bihar, Arunachal Pradesh, Uttar Pradesh, Jharkhand, Madhya Pradesh and Manipur (Figure 3, Table 2). A further 8 states could not be classified with confidence, their probability of belonging to the highest-priority third lying between 0.10 and 0.90, and 15 states were confidently outside it. The median width of the rank credible interval was 9 places.

Figure 3: State priority index. Left panel: priority score with 95% credible interval, propagating posterior uncertainty in burden and district sampling uncertainty in the capacity indicators. Right panel: probability that each state falls in the highest-priority third.

The ordering was robust to capacity-favouring weights (Spearman coefficient 0.97) but only moderately robust to burden-favouring weights (0.60), and moderately robust to adding two further capacity indicators for the states reporting them (0.73). The identity of the highest-priority group is therefore reasonably secure; the ordering within it is not, and the boundary of the group depends on how burden and capacity are weighted against each other.

Projection to 2030

The projected 2030 rate was 3.74 per 100 000 (95% credible interval 1.22 to 6.68) against a halving target of 2.13 (Figure 4). Integrating over the damping prior, the probability of missing the target was 88%. The model showed 14 divergent transitions in 8,000 draws.

Figure 4: Projection of the national paediatric drowning rate to 2030. The shaded band is the 95% credible interval with the trend damping factor integrated out. The dashed line is the trajectory implied by halving the 2023 rate by 2030. Probabilities conditional on fixed damping values are shown because the damping factor is not identified by the observed series.

That figure is sensitive to assumptions the data cannot inform. Conditional on fixed damping, the probability is 65% if the fitted decline continues undamped and above 99% if the trend does not persist at all. Across alternative damping priors it ranges from 88% to 96%, and across the full set of damping and variance specifications examined from 65% to above 99% (Supplementary Table S3).

The variance specifications warrant a further caution. Ten annual observations cannot separate observation noise from trend innovation, and the two variances trade off along a ridge. Under deliberately loose variance priors the sampler does not explore that geometry reliably, returning 923 divergent transitions in 4,000 draws, so the probability it reports should not be relied upon. The primary specification uses tighter priors justified by the scale of the observed series, and the sensitivity analysis is reported chiefly to show that the choice matters. The defensible conclusion is that attainment of the 2030 target is unlikely but highly uncertain under plausible extrapolation assumptions, and that no single probability should be treated as a planning input.

Intervention evidence

Of the ten recommended interventions, 2 had a published quantitative effect estimate on drowning mortality with an interval, 2 had moderate-certainty evidence of benefit without a pooled magnitude, and 6 had no identified controlled evaluation reporting a mortality effect (Table 3; provenance for each estimate is given in Supplementary Table S4). A 2025 systematic review of swimming and water-safety skills training reported that heterogeneity across populations, interventions and outcomes prevented any meta-analysis.(11)

Propagating that evidence, no intervention exceeded a probability of 0.46 of ranking first (Supplementary Figure S1). Day care and basic swimming and water-safety training were indistinguishable from one another at the top of the ordering, with probabilities of ranking in the top three of 0.84 and 0.85 respectively, and both separated clearly from the six interventions with no identified mortality evidence, whose probability of ranking in the top three did not exceed 0.001. Public-awareness media campaigns, which have low unit cost and wide reach but no published estimate of an effect on drowning mortality in any setting, ranked 10 of 10.

These results do not establish a ranking. They establish that only two interventions can currently be distinguished from the field on evidential grounds, that those two cannot be distinguished from each other, and that the remainder cannot be ordered at all.

Contextual associations

Of 6 contextual layers tested, 4 were associated with posterior state burden before adjustment. After adjusting for the composite development index, which accounted for 43% of the variance across its six constituent indicators, 2 remained: female literacy (coefficient -0.68 unadjusted, -0.62 adjusted) and child marriage prevalence (0.60 unadjusted, 0.52 adjusted) (Supplementary Table S5, Supplementary Figure S2). Associations with clean cooking fuel and electricity access did not survive adjustment, attenuating from -0.48 to -0.20 and from -0.41 to -0.07 respectively.

Most of the apparent contextual signal is therefore attributable to general state development rather than to specific pathways. These are ecological associations across 30 state-level observations and support no inference about individual risk or causal mechanism.

Discussion

Principal findings

Four findings emerge. First, routine police records capture roughly a third of the modelled paediatric drowning burden, and how completely a state registers deaths predicts how completely it records drowning deaths. Second, combining two sources rather than propagating one changes state burden estimates by a median of 6.0% and by more than 10% in 11 states, so the choice of source is consequential for targeting. Third, a small group of states is robustly identified as highest priority while a comparable number cannot be classified at all. Fourth, the evidence base does not support ranking the recommended interventions, and reporting that is more useful than producing a ranking the evidence cannot bear.

What the data can and cannot show

The gap between recorded and modelled counts plausibly reflects incomplete registration together with age categories that obscure the under-five band. Two administrative changes would materially improve monitoring: disaggregating the below-14 category into under-5, 5-9 and 10-14 bands, and publishing a state-by-age cross-tabulation for drowning. The second would remove the principal identification constraint in the model presented here, which absorbs the paediatric share into the recording ratio precisely because that cross-tabulation does not exist.

The relationship between registration completeness and the recording ratio has a practical implication that cuts against intuition. States that register deaths well record proportionately more drowning deaths relative to their estimated child burden. Under-recording is therefore concentrated where registration is weakest, which is also where burden is highest, so uncorrected routine data will systematically understate burden in precisely the states that most need investment.

The projection illustrates a general limitation of short series. Seven years of extrapolation from ten annual observations produces intervals wide enough to accommodate very different futures, and the headline probability moves substantially with an assumption about trend persistence that the data cannot inform. Reporting the conditional range rather than a single number is the honest presentation, and the policy-relevant conclusion, that the target is unlikely to be met without deliberate action, does not depend on resolving the assumption.

Interpreting the intervention findings

The finding that eight of ten recommended interventions have no identified controlled evaluation reporting an effect on drowning mortality should not be read as evidence that they do not work. Absence of evaluation is not evidence of absence of effect. It does mean that comparative-effectiveness claims cannot presently be supported, and that a national strategy adopting them would be doing so on grounds of plausibility, cost and deliverability rather than demonstrated mortality benefit.

Day care and basic swimming and water-safety skills training are the two interventions with moderate-certainty evidence, and both were evaluated principally in rural Bangladesh.(12,13) Their transferability to India is a matter of judgement rather than measurement. Day care has the stronger practical case in India because delivery machinery already exists in the national child-development network, whereas swim training requires safe training water and trained instructors that are scarce in most states. That is an argument about feasibility, not about relative effect size, and it is presented as such.

The two interventions with quantified estimates illustrate the transferability problem directly. The pooled estimate for barriers around water comes from case-control studies of domestic swimming pools in high-income countries, whereas Indian child drowning occurs predominantly in open wells, ponds, tanks and rivers.(14) The estimate for mandatory life jackets comes from an uncontrolled before-and-after study of recreational boating in a single high-income jurisdiction with strong enforcement capacity.(15) Neither transfers numerically to the Indian setting, and treating either as a common effect measure applicable to Indian children would not be justified.

For the remaining interventions the position is weaker still. Bystander rescue and resuscitation is described in a descriptive study of rescue reports without a controlled estimate of its effect on drowning mortality, and the rescues it documents frequently placed the rescuer in danger.(16) An earlier systematic review of interventions to prevent drowning in children and adolescents reached the same conclusion about the field as a whole, finding few evaluations reporting mortality outcomes.(17) The Bangladeshi work that anchors the two moderate-certainty interventions was, moreover, conducted where drowning accounts for a far larger share of child deaths than it does in India, which is part of why effect sizes estimated there should not be transported as though they were a common effect measure.(18)

Implications for a national strategy

The results support a strategy built around measurement first. A costed national drowning prevention strategy would benefit from age-disaggregated targets, reform of the reporting categories described above, and adoption of a small, explicitly uncertain priority list rather than a long one presented with false precision. The states identified here with high probability are a defensible starting point; the states that could not be classified are an argument for better data rather than for exclusion.

Because no Indian state publishes a water-safety-specific administrative indicator, no state can currently demonstrate readiness to deliver such a strategy, and no national body can monitor whether readiness is improving. Establishing even a minimal set of programme indicators would be a low-cost precondition for everything else.

Strengths and limitations

The analysis uses only published measurements, combines two independent sources rather than restating one, propagates uncertainty from both model and sampling into every reported classification, and reports what the evidence base cannot support as well as what it can.

Its limitations are substantial. The two-source model is identified through partial pooling and a single covariate rather than through direct observation of the paediatric share, because the required cross-tabulation is not published; the recording ratio consequently confounds age composition with recording completeness. Delhi contributes through one source only. The segmented model cannot attribute the 2020 discontinuity to any cause. The projection depends on an unidentified damping assumption and its interval is wide; residual divergent transitions (14 of 8,000 draws) warrant caution in reading precise posterior probabilities from it. The priority index measures general child-health-system capacity, not drowning-specific readiness, because no drowning-specific indicator exists; it is a proxy, and its boundary is sensitive to weighting. Intervention effect estimates are drawn from other settings, and the transferability discount is an assumption applied uniformly rather than a measurement. The contextual analysis is ecological and exploratory. Estimates for small states remain sensitive to denominator construction. Flood exposure was not examined, because the state-wise flood-prone-area figures in general circulation trace to a national commission estimate made in 1980 that could not be verified against a currently retrievable primary source; given that monsoon flooding is a plausible driver of paediatric drowning, the absence of a maintained, citable state flood-exposure series is itself a data gap worth closing.

Conclusion

India's paediatric drowning burden fell substantially over 2014-2023, but routine police records capture only about a third of it, and under-recording concentrates in the states with the weakest civil registration and the highest burden. Combining sources rather than relying on one changes state estimates materially and should be standard for injury burden estimation in India. A small group of states is robustly identified as highest priority, and a comparable group cannot be classified at all, which is an argument for better data rather than for confident targeting. Attainment of the 2030 halving target is unlikely but highly uncertain under any plausible extrapolation. The evidence base cannot presently rank the recommended interventions; only supervised day care and basic swimming and water-safety skills training separate from the field, and they do not separate from each other. A national strategy should therefore begin with measurement, because without age-disaggregated reporting, a state-by-age cross-tabulation and a minimal set of programme indicators, neither targeting nor evaluation is possible.

Acknowledgements

The author thanks the Institute for Health Metrics and Evaluation, the World Health Organization, the National Crime Records Bureau, the International Institute for Population Sciences and the United Nations Population Division for maintaining the public data resources on which this analysis depends.

Declarations

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 board approval and informed consent were therefore not required.

Data availability: All source datasets are publicly available: the Global Burden of Disease Results Tool (Institute for Health Metrics and Evaluation), Accidental Deaths and Suicides in India 2023 (National Crime Records Bureau), the National Family Health Survey-5 district factsheet (International Institute for Population Sciences) and World Population Prospects 2024 (United Nations Population Division). Processed data files and the analytic code that generated every result are available from the author on request.

Use of artificial intelligence: Generative artificial-intelligence tools were used solely for editorial and language refinement. No scientific content, analysis, interpretation or conclusion was generated by such tools. The author takes full responsibility for the integrity and accuracy of the work.

References

1. World Health Organization. Global status report on drowning prevention 2024. Geneva: World Health Organization; 2024. Available from: https://www.who.int/publications/i/item/9789240103962 [cited 2026 Sep 1].
2. Institute for Health Metrics and Evaluation. GBD Results: Global Burden of Disease Study 2023 [database]. Seattle (WA): Institute for Health Metrics and Evaluation, University of Washington. Available from: https://vizhub.healthdata.org/gbd-results/ [cited 2026 Sep 1].
3. National Crime Records Bureau. Accidental deaths and suicides in India 2023. New Delhi: National Crime Records Bureau, Ministry of Home Affairs, Government of India; 2025. Available from: https://ncrb.gov.in/accidental-deaths-suicides-in-india-adsi.html [cited 2026 Sep 1].
4. Office of the Registrar General and Census Commissioner, India. Sample registration system (SRS)-cause of death in India 2022-2024. New Delhi: Office of the Registrar General and Census Commissioner, India; 2026. Available from: https://censusindia.gov.in/census.website/data/SRSCOD [cited 2026 Sep 1].
5. 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. Lancet. 2007;370(9596):1453-7. doi:10.1016/S0140-6736(07)61602-X
6. International Institute for Population Sciences, ICF. National family health survey (NFHS-5), 2019-21: India. Mumbai: International Institute for Population Sciences; 2021. Available from: https://dhsprogram.com/publications/publication-FR375-DHS-Final-Reports.cfm [cited 2026 Sep 1].
7. United Nations, Department of Economic and Social Affairs, Population Division. World population prospects 2024. New York: United Nations; 2024. Available from: https://population.un.org/wpp/ [cited 2026 Sep 1].
8. Gardner ES Jr, McKenzie E. Forecasting trends in time series. Manage Sci. 1985;31(10):1237-46. doi:10.1287/mnsc.31.10.1237
9. Hwang CL, Yoon K. Multiple attribute decision making: methods and applications - a state-of-the-art survey. Berlin: Springer-Verlag; 1981. doi:10.1007/978-3-642-48318-9
10. Abril-Pla O, Andreani V, Carroll C, et al. PyMC: a modern, and comprehensive probabilistic programming framework in Python. PeerJ Comput Sci. 2023;9:e1516. doi:10.7717/peerj-cs.1516
11. Criel Y, Severijns P, Scheers H, et al. Basic swimming or water safety skills training for drowning prevention in children: an updated systematic review. Front Public Health. 2025;13:1698353. doi:10.3389/fpubh.2025.1698353
12. Rahman F, Bose S, Linnan M, et al. Cost-effectiveness of an injury and drowning prevention program in Bangladesh. Pediatrics. 2012;130(6):e1621-8. doi:10.1542/peds.2012-0757
13. Alfonso YN, Hyder AA, Alonge O, et al. Cost-effectiveness analysis of a large-scale creche intervention to prevent child drowning in rural Bangladesh. Inj Epidemiol. 2021;8(1):61. doi:10.1186/s40621-021-00351-9
14. Thompson DC, Rivara FP. Pool fencing for preventing drowning of children. Cochrane Database Syst Rev. 1998;(1):CD001047. doi:10.1002/14651858.CD001047
15. Bugeja L, Cassell E, Brodie LR, et al. Effectiveness of the 2005 compulsory personal flotation device (PFD) wearing regulations in reducing drowning deaths among recreational boaters in Victoria, Australia. Inj Prev. 2014;20(6):387-92. doi:10.1136/injuryprev-2014-041169
16. Venema AM, Groothoff JW, Bierens JJ. The role of bystanders during rescue and resuscitation of drowning victims. Resuscitation. 2010;81(4):434-9. doi:10.1016/j.resuscitation.2010.01.005
17. Wallis BA, Watt K, Franklin RC, et al. Interventions associated with drowning prevention in children and adolescents: systematic literature review. Inj Prev. 2015;21(3):195-204. doi:10.1136/injuryprev-2014-041216
18. Rahman A, Mashreky SR, Chowdhury SM, et al. Analysis of the childhood fatal drowning situation in Bangladesh: exploring prevention measures for low-income countries. Inj Prev. 2009;15(2):75-9. doi:10.1136/ip.2008.020123

Supplementary Material

STROBE Checklist

Capacity indicator selection and uncertainty propagation

Five National Family Health Survey-5 indicators were selected on conceptual relevance to delivering a child drowning prevention programme and on district-level completeness. The factsheet suppresses estimates built on too few unweighted cases, and three otherwise attractive indicators fail on coverage: care-seeking for childhood acute respiratory infection is 31% complete, full immunisation 65% and pre-primary attendance 78%. Retaining them would require imputing the states they omit. All five retained indicators are reported for every state in the survey, at 96% district completeness or better. Pre-primary attendance and full immunisation were used in a specification sensitivity analysis restricted to the states reporting them.

State means carry sampling uncertainty because they aggregate a finite number of districts. That uncertainty was obtained by resampling districts within each state, and paired with independent draws from the burden posterior, so that the reported credible intervals and tier probabilities carry both sources jointly.

Convergence diagnostics

Sensitivity of the 2030 projection

Evidence provenance for each intervention

Contextual associations

Supplementary figures

Supplementary Figure S1: Rank acceptability of ten recommended drowning prevention interventions. Left panel: median rank with 95% uncertainty interval. Right panel: probability of ranking in the top three. Colour denotes certainty of evidence.

Supplementary Figure S2: Contextual associations with state paediatric drowning burden, before and after adjustment for a composite state development index.