Moreover, given the enormous variation in the geographic, socioeconomic, and health profile within India, it is likely that different parts of the country will experience differential rates of improvement in child survival. Hence, monitoring the status of mortality at the national or state level may conceal important heterogeneity occurring at local levels [11-14]. Instead, evaluating the progress (or its lack thereof) in districts – the lowest administrative unit at which infrastructural, developmental, and other services are planned and where demographic data are consistently provided [15,16] – may allow interventions and programmes to be targeted with greater precision. There are 640 districts in India as per the 2011 Census. In order to produce robust district-wise estimations based on survey data, reliability and sampling variability resulting from the survey must be incorporated. While such precision-weighted estimations based on pooling data and borrowing strength is well-established for small area estimation [17,18], they have not been readily applied in global health research and certainly not in prior studies focused on district level mortality rates in India [11,19-23].[REMOVED HYPERLINK FIELD]

In this study, we derive precision-weighted estimates of NM, PNM, and CM for 640 districts in India using data from the latest nationally representative survey. While national and state-specific NM, PNM, and CM estimates have been reported previously [24], these estimates were not precision-weighted nor broken down by districts. The geographic distribution of mortality across these three disaggregated and mutually exclusive life periods are further visualized in maps to aid identification of spatial clustering of districts with high burden.

The overall mean probability of NM, PNM, and CM in India, based on our random effects models, were 24.4 deaths per 1000 live births (95% CI = 20.9, 28.4), 10.4 (95% CI = 9.0, 11.9), and 3.8 (95% CI = 3.1, 4.7), respectively (Figure 1). 73.6% of the districts (N = 471) had a NM probability that was significantly different from the nationwide mean (P < 0.05). Similarly, the precision-weighted PNM and CM for 74.1% (N = 474) and 64.1% (N = 410) of the districts significantly differed from the nationwide means (P < 0.05). The predicted NM, PNM, and CM for each district are presented in the Table S1 in the Online Supplementary Document. Based on the sum of our precision-weighted NM, PNM, and CM, so far only 75 districts have realized the SDG3 goal of reducing the total U5M to less than 25 deaths per 1000 births (Table S1 in the Online Supplementary Document). Estimates stratified by child’s sex (boys vs girls) and place of residence (urban vs rural) are presented in Table S2 in the Online Supplementary Document.

Figure 1.  Distribution of precision-weighted probability of neonatal, post-neonatal, and child mortality across 640 districts in India, NFHS 4. The precision-weighted estimates for each district are reported in Table S1 of the Online Supplementary Document.

Mapping the geographical distribution of mortality

Figure 2 illustrates that while some districts have consistently high (or low) probability of mortality across all age groups, other districts show inconsistency. In Chhattisgarh, all districts had the highest quintile of NM and the majority (67%) fell in the fourth quintile of CM, but 50% were doing fairly well in terms of PNM (second quintile). In Tamil Nadu and Arunachal Pradesh, almost all of the districts had the lowest level of probability for NM, but had higher probabilities for PNM and CM. The overall correlation between district-wide probabilities of NM, PNM, and CM were moderate (Pearson correlation = 0.47-0.58, Spearman’s rank correlation = 0.58-0.64) (Table 3). In addition, a relatively strong spatial clustering was detected across districts that transcended state boundaries (Moran’s I = 0.61-0.76), with large patches of spatially contiguous districts from the states of Uttar Pradesh, Bihar, Madhya Pradesh, and Chhattisgarh with high burden of mortality (Figure S2 in the Online Supplementary Document).

Figure 2.  Maps illustrating the precision-weighted probability of neonatal, post-neonatal, and child mortality (in quintiles) across 640 districts in India, NFHS 4. The quintile cutpoints were 17.7, 22.7, 29.1, and 36.1 deaths per 1000 live births for neonatal mortality; 6.7, 8.3, 10.1, and 12.9 deaths per 1000 live births for post-neonatal mortality; and 2.3, 3.3, 4.0, and 5.2 deaths per 1000 live births for child mortality. State index map provided in Figure S1 of the Online Supplementary Document.

Table 3.  Correlation statistics for the precision-weighted probability of neonatal, post-neonatal, and child mortality across 640 districts in India, NFHS 4

	Neonatal mortality	Post-neonatal mortality	Child mortality
Pearson correlation:
Neonatal mortality	1
Post-neonatal mortality	0.475*	1
Child mortality	0.581*	0.514*	1
Spearman's rank correlation:
Neonatal mortality	1
Post-neonatal mortality	0.595*	1
Child mortality	0.581*	0.638*	1
Moran' s I bi-variate spatial correlation:
Neonatal mortality	0.764*
Post-neonatal mortality	0.414*	0.675*
Child mortality	0.478*	0.421*	0.614*

*P < 0.0001.

We present four salient findings from the precision-weighted estimates of NM, PNM, and CM for all 640 districts in India. First, we found the majority of U5M was concentrated in the neonatal period followed by the post-neonatal age group and a substantially lower probability of CM. Aggregating deaths from these three distinct age groups into the conventional indicators of IM or U5M masks the greater investment needed in health care quality to improve child survival for distinct age groups. Second, we revealed substantial variability in the probability of mortality across districts that are overlooked in summaries based on national and state-wide averages. The precision-weighted estimations accounting for sampling variability could thus be used to monitor and identify performances at smaller geographic and administrative units to motivate accountability and targeted interventions. Third, the overall correlation between district-wide probabilities of NM, PNM, and CM was moderate, meaning that districts that had high probability of NM did not necessarily have high PNM or CM, and vice versa. Fourth, strong regional spatial patterning in mortality estimates was detected suggesting opportunities for districts from different states to collaborate to achieve the common goals of reducing NM, PNM, and CM. Taken together, we found that many districts have distinct age-specific mortality challenges, indicating that state-level mortality patterns are too non-specific to make effective targeted policy changes.

It is important to note that while the NFHS applied the synthetic cohort life table approach to estimate mortality rates based on the number of deaths over the number of births in a given time period [24], we have estimated the probability of death per 1000 live births. Compared to the mortality rates reported by the NFHS (ie, 29.5 for NM, 11.3 for PNM, and 9.4 for CM) [24], our probability estimates were more conservative due to precision-weighting, such that districts with smaller sample sizes and unreliable estimates were down-weighted more towards the overall mean [18,26], and the largest difference was seen for CM. The required sample size needed to achieve unbiased estimates from the synthetic cohort approach ranges from at least 500 [31] to thousands of observations [32]. Similarly, we present more reliable and appropriately conservative estimates compared to a recent study [23] that has also used the synthetic cohort probability approach to estimate district-wide NM. However, we acknowledge that a future validation study should be performed comparing the estimates of different estimation methods by, for instance, quantifying their coefficients of variance. Nevertheless, many of the districts that were identified to have the highest NM in their study [23] aligned with our findings (Gonda, Sitapur, Budaun, Dakshin Bastar Dantewada, Kaushambi, Shahjahanpur), which were also consistent with the priority districts identified in the Ministry of Health and Family Welfare’s annual report on health [25].

While we provide additional stratifications by child’s sex and place of residence (urban vs rural) (Table S2 in the Online Supplementary Document) as resources for health planning, our study still lacks stratified estimates by socioeconomic status and smaller geographic areas within districts. In India, NM was shown to be greater for boys while PNM and CM were higher for girls, and the difference has increased over time for the former while the opposite was found for the latter [33]. A negative wealth gradient in IM and U5M has been well documented elsewhere [13,14,33]. While substantial variation in mortality is expected within districts [34], it is not possible to obtain village level estimates in the absence of complete registration of births and deaths in India. Lastly, information on children’s birth date, survival status, and age of death if a child died were self-reported by mothers, so there exists the potential for underreporting bias in particular. However, the DHS is known for its high response rates, national coverage, high quality interviewer training, standardized data collection procedures across countries that are consistent over time [35].

Despite these potential limitations, our findings have important implications to renew the policy discussion in India for more efficient prioritization and resource allocation to reduce U5M. Historically, decentralized planning in India through the panchayat system allowed to create health action plans, train health professionals, deliver resources, and monitor facilities either with or without state involvement. While different districts had varying levels of success, a case study showed that district-led health planning with state-level oversight was able to achieve the majority of its stated goals, such as providing antenatal care kits to midwives and establishing Mother and Child Protection antenatal service camps with proper staffing and facility maintenance [36]. For instance, different components of Integrated Child Development Services (ICDS) may be emphasized more depending on the share of NM, PNM, and CM within each district. In India, 78% of neonatal deaths were attributed to prematurity, neonatal infections, and birth asphyxia [37]. Previously, birth asphyxia and neonatal infection rates were found to be the highest in central India, in states such as Madhya Pradesh and Chhattisgarh, and prematurity and low birthweight were found to be the highest in West India, in states such as Rajasthan and Gujarat. Given our district-level estimates, areas with the highest probability of NM may be targeted for improvement in delivery care to ensure that all births occur in health centres with comprehensive facilities, such as Special New Born Care Units (SNCUs), Newborn Stabilization Units (NBSUs), and Newborn Care Corners (NBCCs), to provide sterile and proper neonatal care [25]. Investments in improved diet and nutrition of mothers, such as with iron tablet supplementation, may also be further emphasized in these areas [38].

Districts that do not suffer from high probability of NM may suffer from high PNM or CM. Since pneumonia and diarrhea are common causes of deaths after the first month but before the first year [37], districts with higher burden of PNM may prioritize introduction of vaccines as well as education on sanitation practices (eg, hand-washing and safe drinking water) for prevention of diarrhea and access to oral rehydration solution (ORS)-zinc tablet treatments [25,39]. To reduce CM, our estimates suggest that nutrition interventions [40] and UIP should be intensified in not only districts in Central India, but also select districts in Tamil Nadu, Arunachal Pradesh, Assam, Meghalaya, and Nagaland, which we have identified as among the highest mortality quintile in CM. The accredited social health activities (ASHA) program should continue to grow in the worst-performing districts via their bi-annual vitamin A supplements and bi-weekly iron supplements program to combat poor nutrition [25].

By disaggregating U5M into NM, PNM, and CM, we highlight distinct distributions of age-specific mortality across 640 districts in India. Our precision-weighted estimates may serve as evidence to assist district administration develop appropriate interventions and programmes to prioritize the most precarious age groups within their district. To reduce U5M as a whole, focusing more attention on neonatal period may be most effective and efficient. Efforts to improve nutritional status of the population should adopt a life cycle perspective since interventions targeting adolescent population can go a long way in improving the nutritional status of the reproductive age group as well as birth outcomes. Further etiological studies on the causative factors of mortality in these three periods of early childhood are needed to develop more specific treatments and prevention programmes.