Methods
Target population
We conducted the analysis on a hypothetical cohort of all LBW newborns born in India in 2022. Data on the total number of live births and proportion born LBW were obtained from the United Nations population estimates and the National Family and Health Survey round 5 (NFHS-5).7 21 Given that 40% of all newborn deaths happen in the first 24 hours of life, our analysis cohort comprised 60% of all stable LBW newborns not requiring emergency care at birth.22
Baseline coverage of interventions
KMC consists of skin-to-skin contact and exclusive breastfeeding. In the absence of direct KMC coverage data, we used the proportion of women receiving postnatal care with counselling on exclusive breastfeeding as a proxy measure, as these services are integral to KMC implementation. We used data from NFHS-5 to generate estimates for each socioeconomic quintile separately. The target coverage was set at 95% in the modelled scenario.
Estimation of OOPE
Data on OOPE associated with care seeking among the enrolled infants were collected during follow-up visits at 28, 90 and 180 days of life. We included provider costs, medicines, diagnostics, transport and other direct medical/non-medical payments. Healthcare spending was collected for all morbidities and not restricted to LBW-attributable diagnoses. In the model, we applied arm-specific mean OOPE per inpatient episode and per outpatient visit, together with probabilities of inpatient-only, outpatient-only and combined care seeking, to estimate expected OOPE per infant by quintile.
Overall costs for care seeking by study arm (inpatient, outpatient and overall) have been presented earlier.19 The current analysis extends the earlier work by reporting the estimates for FRP at national level, overall and by quintiles and across different states in India.
Model structure and outcomes
We used a decision tree model (online supplemental figure 1) to estimate the impact of scaling up KMC nationally to 95% coverage in India. Decision trees represent choices and probabilities leading to various outcomes. A decision node (square) represents a point where a choice between alternatives must be made. A chance node (circle) represents a point where possible outcomes have an associated probability, and the probability of all outcomes adds up to a terminal node (depicted as a triangle) which represents the end of a pathway. The associated outcomes in terms of health effects, costs or both are estimated at this stage, and the decision tree is completed. Further details are available elsewhere.23
Babies would enter the cohort if they were stable and survived the first 24 hours of life. They could either die before or survive until 28 days of life. Those who survived this neonatal period could either die during the postneonatal period or survive infancy. The same model was replicated for all socioeconomic quintiles using quintile-specific estimates.
Socioeconomic quintiles were based on the NFHS-5 household wealth index (asset based), and all NFHS-derived parameters (LBW prevalence, baseline coverage, household size, etc) were generated on asset index-based quintiles. Because CHE is defined relative to household income, we additionally generated quintile-specific household income distributions calibrated to national gross domestic product (GDP) per capita and the Gini coefficient for India. These income distributions were used only for the CHE calculations.
We generated the following estimates across wealth quintiles at national level and at state level in India, that is, geographically:
Health outcomes
Lives saved.
DALYs averted.
Non-health outcomes
OOPE averted.
CHE averted.
Incremental cost-effectiveness ratio (ICER).
Since healthcare decision-making rests primarily with individual states in India, estimating these at state level is essential so that policymakers can determine which regions to focus on when expanding coverage.
Table 1 summarises the input parameters used in the model. Our RCT on KMC initiated in community settings was the primary source of data on effects, care-seeking patterns and costs of inpatient and outpatient care seeking.19 It was conducted among 8402 LBW newborns and showed a substantial reduction in the mortality of infants in the neonatal period and in the first 6 months of life.9 The recent systematic review on the effects of KMC, which included the data from our trial, reports similar estimates.24 All costs were reported in 2022 US$. The OOPE per episode increases with wealth, consistent with differential utilisation/provider choice and ability to pay. The cost estimates were adjusted for inflation between the year of data collection in the trial (2017–2018) and the year for which the model was developed, that is, 2022. Consumer price index available from the World Bank database was used for the adjustment. Cost and care-seeking data were based on RCT and were internally consistent. In addition to adjusting for inflation, we used sensitivity analysis to address the changes in prices or care-seeking patterns which would have occurred after the trial. To the best of our knowledge, updated estimates for healthcare-seeking costs among LBW infants were not available.
Input parameters used in the model*
The income distribution of the participating households represented the lower 50% of the income distribution in India. We used a gamma distribution to generate quintile-specific estimates of per capita income based on GDP per capita and the Gini coefficient for India.16 The GDP estimates were adjusted since only 60% of GDP for India is contributed by expenditure on personal consumption. We used the mean number of household members from NFHS-5 to convert per capita income into household-level estimates. We estimated neonatal and postneonatal mortality rates using the number of live births, reported deaths and age at death from the last 5 years in the NFHS-5.7
Cost of scaling up KMC
The cost of scaling up the intervention in India was based on published literature.25 We used the reported incremental recurrent costs (of health worker time, supplies and operations) and start-up costs for KMC scale-up. The costs were estimated per live newborn eligible for KMC. Recurrent costs comprised staff time (full time equivaentsFTEs) for referral, counselling and discharge preparedness, provision and support of KMC by nurses and physicians, community-based follow-up after discharge, and programme management/supervision and recurrent training. Supplies and operational inputs included KMC wraps/garments, routine recording and educational materials, and food for mothers where provided, as well as transport and communication required for programme delivery. Start-up costs included establishment or renovation of KMC unit space (including toilets), furniture and basic amenities, development of communication materials and initial training. Further detailed information is available elsewhere.25
Estimation of impact on DALYs averted
We estimated DALYs averted by converting deaths averted into years of life lost (YLL) averted. We did not model years lived with disability as mortality is the major driver of DALYs during the first year of life. First, we calculated lives saved among LBW infants as the product of (1) the incremental increase in KMC coverage under scale-up (target coverage minus baseline coverage), (2) the relative mortality effect of KMC from the published meta-analysis, (3) the number of LBW births eligible to benefit (live births by wealth quintile×LBW prevalence×an affected fraction to exclude deaths occurring before the intervention could plausibly act) and (4) quintile-specific mortality risk over the analytical mortality window (day 1 to day 180), expressed per 1000 live births. We then computed DALYs averted as YLL averted by multiplying lives saved by a life expectancy of 70 years.26
Time horizon and discounting
We conducted the analysis for a time horizon of 1 year. Hence, we did not discount lives saved, OOPE averted, cases of CHE averted and cost of scale-up. Since DALYs averted would accrue over the lifetime of the infant, we discounted it at the rate of 3% per year consistent with standard economic evaluation practice.
Statistical analysis
Data on OOPE overall and by quintiles at the national level were generated by multiplying the difference in OOPEs for KMC compared with routine care by the number of LBW children in each quintile based on the trial data. Catastrophic health expenditure (CHE) was modelled because direct household income/consumption data were not available at national scale for the target population. We therefore adopted a distributional modelling approach to estimate the probability that OOPE associated with care seeking for LBW infants would exceed a fixed share of annual household income/consumption. Annual household income was generated from a synthetic national income distribution parameterised using India’s GDP per capita and the national Gini coefficient; mean OOPE for LBW-related inpatient and outpatient care was taken from the community initiated Kangaroo Mother Care (ciKMC) trial, inflated to 2022 US$ and extrapolated from the trial follow-up period to an annual value. CHE cases were then calculated as the proportion of households in each quintile whose healthcare-related OOPE was greater than the CHE threshold of 10%. CHE averted under universal public financing and KMC scale-up was estimated as the difference in expected CHE cases between baseline coverage and the scale-up scenario, stratified by socioeconomic quintile and geography.
ICER was estimated based on the cost and effects for KMC in the scenario of scale-up compared with the existing level of coverage as follows:
We adhered to the Consolidated Health Economic Evaluation Reporting Standards (CHEERS) guidelines.27 All analyses were done in R statistical environment V.4.4 and above.28
Sensitivity analysis
We conducted one-way sensitivity analysis to identify the parameters with the highest impact on the ICER. SEs were applied to each parameter to assess uncertainty; where SEs were missing, we estimated them as ±25% of the mean. This approach helped us pinpoint specific variables that contributed most significantly to the variation in ICER, highlighting which factors most influenced the cost-effectiveness of the intervention including the estimates for cost and care seeking. We also created tornado plots based on the uncertainty ranges of different input parameters.
To account for simultaneous variations in multiple parameters, we performed probabilistic sensitivity analysis (PSA) using a Monte Carlo simulation with 100 000 iterations. For this analysis, we used beta distributions for probabilities, gamma distributions for costs and lognormal distributions for effect sizes. The beta distribution is suitable for probability values as it spans from 0 to 1, capturing a realistic range of outcomes. The gamma distribution is ideal for cost data as it models values starting from 0 and extending upwards without limit, accommodating widely varying costs. The lognormal distribution was used for effect sizes, which are positive and may be right skewed, with most values close to 1 and a few larger values representing stronger effects. We employed a triangular distribution for discounting using only a minimum, maximum and most likely (mode) value, rather than a full set of data points.
These distributions allowed the analysis to simulate real-life variations in both probabilities and costs, providing a more robust and realistic view of ICER under uncertain conditions, overall and by quintiles. The PSA results were illustrated on a cost-effectiveness plane, with median ICER values reported along with 2.5th and 97.5th percentiles.
Ethics approval
The current ECEA was a modelling study based on data from multiple sources including those available in the public domain.
Patient and public involvement
As this was a modelling study, patients or the public were not involved in the design, or conduct, or reporting, or dissemination plans of our research.