Abstract
The Millennium Development Goals prompted renewed international efforts to reduce under-five mortality and measure progress at the national level. However, scant evidence exists about the distribution of child mortality at low sub-national levels, which in diverse and decentralised countries like India are required to inform policy-making. This study estimates changes in under-five and neonatal mortality across a range of markers of inequities in Orissa and Madhya Pradesh, two of India’s largest, poorest, and most disadvantaged states. Estimates were computed using seven datasets and inequalities were gauged by comparison of mortality rates within four sub-state populations defined by the following characteristics: rural-urban location, ethnicity, wealth, and district. Trend estimates suggest that progress has been made in neonatal and under-five mortality rates at the state levels. However, reduction rates have been modest, particularly for neonatal mortality. Different mortality rates are observed across all the equity markers, although there is a pattern of convergence between rural and urban areas, largely due to inadequate progress in urban settings. Inter-district disparities and differences between socioeconomic groups are also evident. Polices targeting rural populations and scheduled caste and tribe groups appear to have achieved some success in reducing mortality differentials, but less progress has been made in reducing under-five and neonatal mortality rates for the urban poor and near-poor. The results of this study thus add weight to recent government initiatives targeting these groups. Equitable progress, particularly for neonatal mortality, requires continuing efforts to strengthen health systems and overcome barriers to identify and reach vulnerable groups. Keywords: MDG 4, under-five mortality, neonatal mortality, health inequalities, IndiaEquation Chapter 1 Section 1
Introduction
In recent times the topic of inequalities in health services and outcomes has drawn increasing attention by scholars and policy makers alike (Barros et al., 2012; Boerma et al., 2008). While interest has historically centred on socio-economic status (Hosseinpoor et al., 2011), there is increasing recognition of the importance of other dimensions, such as ethnicity and geography, in identifying disadvantage (Bauze et al., 2012; Mulholland et al., 2008). Yet, much is still unknown about the causes of health inequalities, a major research and policy priority in developing countries seeking to attain the Millennium Development Goals (MDG) and general societal advancement (Östlin et al., 2011). India contributes to almost a quarter of under-five deaths and a third of neonatal deaths worldwide (Lozano et al., 2011; United Nations Children’s Fund, 2011a, 2011b). While inequity in child health outcomes has been documented between the states of India and amongst broad equity markers within India, such as rural-urban location and wealth (Balarajan et al., 2011; Subramanian et al., 2006), less is known about equity trends within states and whether any progress in reducing inequities has occurred of late. Amongst India’s twenty-eight states, Orissa and Madhya Pradesh (MP) rank within the top three highest states in terms of under-five mortality rates (United Nations Children’s Fund, 2010), with levels similar to Zimbabwe and Kenya (Kumar et al., 2012). These rates have remained persistently high over the last decade (Office of the Registrar General and Census Commissioner, 2009). Both states have weak infrastructure compared to other states, low socio-economic status, and high levels of fertility. Subsequently, they have been included along with six other states – Bihar, Chhattisgarh, Jharkhand, Rajasthan, Uttarakhand, and Uttar Pradesh – in the Empowered Action Group (EAG) requiring high focus by the Government of India, and are also two of eighteen states targeted under the Indian Government’s flagship program for health sector reform, the National Rural Health Mission (NRHM). With a collective population of over 114 million, the two states provide fertile ground to assess the degree of health inequalities across sub-populations defined by wealth and non-wealth equity markers. In this paper, we estimated the levels and trends in under-five and neonatal mortality rates between rural-urban populations, caste groups, wealth quintiles, and sub-state districts within Orissa and MP. State-level inequalities are of particular importance, since, at finer levels of disaggregation, such information provides policy makers information on the early impacts of recent initiatives, such as the NRHM, and guidance for future programming and policies to meet the MDG 4. Moreover, this study provides crucial evidence on India’s progression towards the MDG, since persistent within-country inequities reduce the effectiveness of the national objective of improving livelihoods.
Methods
2. 1 DataThe data were taken from seven datasets after a total of eight surveys were reviewed and validated. We utilised microdata derived from a collection of surveys, supplemented with crude death rates stemming from sample registration systems. An overview of the datasets and their usage in our study is presented in Table 1. The first main data source was the Demographic Health Surveys (DHS) series – known as the Indian National Family Health Surveys – conducted in India in 1992-93, 1998-99, and 2005-2006. Like other DHS, these surveys provide consistent and reliable estimates of mortality and fertility, family planning, the utilization of maternal and child health care services, other related health indicators, and socio-economic measures. The sampling design was a systematic, stratified random sample of households, with two stages in rural areas and three stages in urban areas. Unequal sample sizes were collected per state depending on the size of the state, resources available for the survey, and the desire for sub-population estimates, for example states, urban/rural, or metropolitan cities (International Institute for Population Sciences & Macro International, 1994, 2000, 2007). Consequently, the national sample size, by survey standards, was unusually large. In 1992-93, a near (99%) nationally representative sample of 89, 777 ever-married women aged 13-49 were interviewed from 88, 562 households, with 6, 254 (5, 857) and 4, 257 (4, 602) women (households) from MP and Orissa, respectively. In 1998-99, the nationally representative sample of 89, 199 ever-married women aged 15-49 from 91, 196 households was collected, including 6, 941 (6, 749) and 4, 425 (4, 689) women (households) from MP and Orissa, respectively. Lastly, in 2005-06, a national total of 124, 385 women aged 15-49 from 109, 041 households were sampled, with 6, 427 (5, 488) and 4, 540 (3, 910) women (households) from MP and Orissa, respectively. The second main data source used was the District Level Household and Facility Surveys (DLHS) series undertaken in 1998-99, 2002-04, and 2007-08. The DLHS is a collection of nationally representative household surveys, primarily conducted to monitor and assess the implementation and operation of the Reproductive and Child Health program across the districts of India. Similar to the DHS, the DLHS were undertaken using a systematic, multi-stage stratified sampling design and the national sample sizes are large (International Institute for Population Sciences & Ministry of Health and Family Welfare, 2001, 2006, 2010). For the 1998-99 wave, 474, 463 currently married women age 15-44 from 529, 817 households were interviewed, with 41, 250 (46, 355) and 28, 757 (32, 214) women (household) from MP and Orissa, respectively. In 2002-04, a nationally representative sample of 507, 622 currently married women age 15-44 years from 620, 107 households were questioned. The number of women (households) from MP and Orissa were 38, 024 (46, 413) and 24, 972 (31, 909), respectively. Lastly, the 2007-08 survey wave covered 643, 944 ever-married women aged 15-49 and 166, 260 never-married women aged 15-24 from 720, 320 households, with 56, 574 (51, 419) and 35, 105 (33, 172) women (households) from MP and Orissa, respectively. The final dataset utilised by the study was the Sample Registration System (SRS) data. The SRS is a sample of birth and death registrations under the Office of the Registrar General of India, and it provides annual estimates of the population, birth rates, fertility, mortality, live births, maternal mortality, life expectancy, death rate, and other indicators at the national and state level and separately for rural and urban place of residence. Generally, the sample design adopted for the SRS is a single-stage stratified random sample (Office of the Registrar General and Census Commissioner, 2009). The sampling frame is revised every ten years based on the latest census frame and the sample size has increased over time to approximately 1. 5 households and 7. 27 million people in 2010. We also attempted to utilise the World Health Survey (WHS). However, it is representative at the national level only, and thus was not used in the final analysis. In 2000, the state of Chhattisgarh was formed via the partitioning of 16 south-eastern districts of MP. As a result, the 1992-93 and 1998-99 DHS were not usable for MP since the DHS is only representative at the state level. Fortunately, given that the DLHS were representative at the district-level, we were able to map the data to fit into the structure of the newly formed states in 2000. Similarly, the SRS data were available on a yearly basis, and thus, we were able to account for the changes in the state boundaries. The datasets were cleaned by deleting duplicates and dropping children that had unreasonable birthdays and death ages (e. g. child reported to die after the interview date). The collection of datasets resulted in a sample period from 1990 to 2007. Estimates were produced at the state level and across four equity markers: urban-rural location, ethnicity, wealth, and districts. These dimensions were selected based on the previous literature and the data that were available in order to best represent the diversity within the two states (Gang et al., 2006; International Institute for Population Sciences & Macro International, 2007; International Institute for Population Sciences & Ministry of Health and Family Welfare, 2010; Pande & Yazbeck, 2003; Pradhan & Arokiasamy, 2010). Data on the equity markers are available in all datasets, with the exception of SRS which includes measures at the state level and for rural/urban location only. As is common with health surveys, data on income and expenditure is not collected. Previous studies have utilised questions on household assets and housing characteristics to construct a wealth index using principal components analysis (Filmer & Pritchett, 2001). We follow this practice. While it is understood that there is a certain degree of correlation between these equity markers – for instance, the richest households are likely to reside in urban areas – the relative variations in mortality identified for different equity markers can reveal the main driving forces behind disparities in mortality and child health service utilisation. 2. 2 Mortality EstimatesSurvey measures of under-five and neonatal mortality were derived from complete birth histories (CBH) and summary birth histories (SBH) using the methods developed by Rajaratnam, Tran, et al. (2010). Mortality from CBH were computed by pooling data from all available surveys and restructuring the datasets such that the life of each child is quantified into months of observation, where a binary variable indicates if the child is alive or died each month of the first five years of the child’s life. The pooling approach aims to migrate some of the biases associated with CBH, such as recall bias, especially where the surveys overlap. We also imposed a cut-off such that estimates are not produced for periods with less than 10, 000 person-months of observation, which has the effect of limiting the duration of the recall period. While single-year estimates are possible, due to the relative rarity of observed deaths, precise estimates are only available over two-year periods. Using the person-month structure, survival rates, accounting for sampling weights, were computed for the age groups: 0-1 month, 1-11 months, 1-2 years, 2-3 years, 3-4 years, and 4 years-59 months. Under-five mortality rates (U5MR) are derived by combining the survival rates from all the age groups and subtracting from one, while the neonatal mortality rates (NMR) are similarly computed by subtracting the survival rate for the first age group (0-1 month) from one. In the cases where there is no other type of data available, a continuous series is created from the biennial period estimates using Loess regression, using a smoothing parameter (bandwidth) of 0. 75 (Cleveland & Loader, 1996). In such cases, confidence intervals are generated by running for each time-period/age-category 1, 000 simulations of the survival probability by assuming a binomial distribution where the probability of success (p) is equal to the mean survival probability and the sequence n is equated to the number of person-months observed in the time-period/age-category. Mortality rates are then calculated for each time-period in each simulation and the 2. 5th and 97. 5th percentiles from the lower and upper confidence bounds for each time-period. In the absence of CBH, under-five mortality rates were indirectly estimated from SBH using the combined method developed by Rajaratnam, Tran, et al. (2010) that incorporates the cohort-derived and period-derived techniques into a single measure. We utilised the full set of four methods available: the time since first birth cohort-derived method (TFBC), the maternal age cohort-derived method (MAC), the time since first birth period-derived method (TFBP), and the maternal age period-derived method (MAP). A combined measure was created by again applying Loess regression, and following the suggested practice, we choose an alpha value of 0. 5 to smooth the data. Confidence intervals on the summary measure are computed by accounting for the parameter uncertainty in the SBH. This is achieved by performing 1, 000 simulations from the multivariate normal distribution described by the point estimates and variance-covariance matrices for the coefficients in each model. From these simulated sets of coefficients we estimate under-five mortality for each data point. Loess regression is applied for each draw to combine the four methods. We calculate the standard deviation of the 1, 000 loess series and then multiply this by 4 to reflect the fact that the standard error is artificially deflated by a factor of one-quarter since we have used all the available data a total of four times. Finally, we use this corrected standard error to estimate a 95% confidence interval. This procedure was detailed by Rajaratnam, Tran, et al. (2010), and we introduced one modification to the process. We applied multipliers to the standard error of the summary measures to ensure that the lower bound of the uncertainty coverage was at least 95%. This more conservative approach places a larger likelihood of estimating high under-five mortality rate, resulting in wide asymmetrical confidence intervals. To convert indirect estimates of U5MR into NMR, the relationships between U5MR and NMR rates were explored using state and sub-state direct estimates of the mortality rates from the other datasets with CBH (Murray et al. 2007). This was achieved using three steps. First, U5MR and NMR were transformed into logit space. Second, in this space a hierarchical model with random intercept and random slope was fitted at the state and sub-state levels to relate NMR to U5MR. The model was then used to predict NMR in logit space from the indirectly estimated U5MR computed using the SBH. Finally, the NMR were transformed back into their proper values. Again this approach results in rather sizeable uncertainty bounds since a number of sources of uncertainty are taken into account (i. e. uncertainty from the CBHs, the model itself, and the SBHs estimates). The final type of estimated U5MR were derived from the SRS. The SRS only provided under-five crude death rates aggregated across both sexes by state. Accordingly, we applied the commonly used technique outlined by Preston et al. (2001) to convert the crude death rates to mortality rates. Having produced various estimates of under-five and neonatal mortality from the different sources and data types using either direct estimation, indirect estimation, or through conversions from crude death rates, we produced a single summary measure following the technique developed by Murray et al. (2007), which averages all the estimators into one estimator across time, and which has been updated in recent years (Lozano et al., 2011; Rajaratnam, Marcus, et al., 2010). In brief, the summary measure is computed via a modified version of Loess regression that subsumes the choice of the smoothing parameter into the uncertainty, thus eliminating the need to choose a smoothing parameter. We followed the methods of Murray and co-authors with a few modifications as outlined below. We adapted the basic model:* MERGEFORMAT ()by estimating the following:* MERGEFORMAT ()where 5q0 is under-five mortality, T is the calendar year, VR is a dummy variable for vital registration system, and S is a dummy variable for the survey type (e. g. all DHSs are considered to be one survey type). The intercept for the final estimator is taken to be the inverse variance weighted mean of the coefficients on each dummy variable of S, which means that the overall level of child mortality is more heavily influenced by the more precise survey types. We also slightly modified the standard tricubic weighting used in Loess regression procedures to ensure that useful information in the trends of child mortality is not eliminated. Specifically, let i denote the index of the point of interest, j denote the index of the point with the lowest weight, W denote the weight, and Dj denote the distance from the point of interest. Then,* MERGEFORMAT ()Lastly, while the original model captured numerous sources of uncertainty – for example, the uncertainty associated with the model parameters – it does not take into account data uncertainty originating from uncertainty associated with each survey measurement, which is a measure of the accuracy of the estimator. For example, an estimator with large uncertainty will have less weight, and an estimator with small uncertainty will have more weight. In our methods, we did capture this data uncertainty. Since the final estimator’s uncertainty is already large mostly due to the smoothing parameter uncertainty, adding data uncertainty would cause the final estimator’s uncertainty to be even larger. Thus, the uncertainty coming from the choice of the smoothing parameter was eliminated – a smoothing parameter of 0. 5 was chosen for equity markers with only one source of data, and a smoothing parameter of 0. 25 was chosen for equity markers with multiple sources. This is not an uncommon practice as nonlinear local regression was historically developed as a tool where the user chooses the smoothing parameter and checks the stability of the results via a sensitivity analysis. Finally, there are three issues of note. First, since in India it is known that the type of assets owned by the rural populations different from that of the urban populations, the asset-based wealth index is derived for both rural and urban areas separately, as are the mortality trends. Partially as a consequence, the modes for the fourth and the highest quintiles are very close to each other. This indicates that the wealth index might have low discriminatory power between the two groups. Furthermore, the number of child deaths within the highest quintile is fairly small, leading to less reliable estimation. Thus, these two wealthiest groups are combined into one wealth group for the analysis. Second, since neonatal rates are converted from indirect estimates of U5MR, in the case of wealth quintiles, data restrictions imply that such rates are computed with an excessive degree of uncertainty. Consequently, we only estimated direct estimates across wealth quintiles, which are associated with a lower but still high degree of uncertainty. Accordingly, these estimates should be interpreted with some caution. Finally, only the DLHS datasets are representative at the district level and the district estimates are produced using the most recent wave. Given the relative size of the sample, in this instance some caution is required when interpreting the results. All statistical analyses described were carried out using two statistical packages, Stata and R.
Results
In both states, across all the equity markers inequalities are evident. Tables 2 and 3 provide estimates of under-five and neonatal mortality rates for selected years, with corresponding confidence intervals, at the state level, and for urban-rural areas, three ethnic groupings and wealth quintiles. The state-level results, reported in Figure 1, demonstrate that there has been a reduction in the U5MR in both MP and Orissa from 156 deaths per 1, 000 live births (95% CI: 140-173) in 1990 to 95 (95% CI: 75-124) in 2007 and from 138 per 1, 000 live births (95% CI: 124-153) to 87 (95% CI: 65-114), respectively. In comparison to recent comparable national estimates (Lozano et al., 2011; Rajaratnam, Marcus, et al., 2010), the U5MR exceeded the national figure in both states, and the average per annum reduction over the sample period of 2. 91 percent in MP and 2. 48 percent in Orissa is slightly below the national figure of approximately 3 percent. The reductions in neonatal mortality have been even more modest. In MP, the NMR has declined from 72 per 1, 000 live births (95% CI: 61-84) in 1990 to 54 (95% CI: 39-80) in 2007, while in Orissa a corresponding decline from 69 (95% CI: 59-82) to 47 (95% CI: 27-84) was observed. However, despite this overall decline, in MP, NMR has remained almost completely stagnated since 2001. If these trends continue, in Orissa by 2015 neonatal deaths will constitute more than half of all under-five deaths and over three-quarters of all infant deaths, with corresponding percentage contributions of approximately 82 and 85 in MP, respectively. The general pattern of decline in under-five and neonatal mortality is observed in both urban and rural areas (see Figure 2). Children residing in urban areas remain better off than their rural counterparts in terms of the level of mortality. Yet, the rates of decline are observed to be higher in rural areas, where, for example, in MP the average annual decline in U5MR in urban areas is 2. 13 percent and 3. 25 percent in rural areas. This might point to some early outcomes of programs such as the NRHM to scale up maternal and child health coverage in rural areas. Nonetheless, at the end of the sample period the differences in mortality outcomes remained high, with the U5MR in urban areas of Orissa estimated at approximately 67 compared to 92 in rural areas. Moreover, while the gaps between rural and urban areas in both states are predicted to reduce substantially by 2015, the pattern of convergence, which is most obvious from 2003 onwards, is largely due to the inadequate progress in urban areas. In fact, the urban NMR in both states was fairly constant in the latter years of the sample, with some slight increases since 2004. Estimates of under-five and neonatal mortality displayed in Figure 3 confirmed higher rates among the Scheduled Castes (SC) and the Scheduled Tribes (ST) caste groups, compared to the remainder of the states’ populations (denoted as Other), although these differences are smaller for NMR as compared to U5MR. In both states, convergence between SC and ST is clearly observed. However, in MP the U5MR among the SC and the ST has fallen at a faster rate compared to that of the Other caste grouping, with average rates of reduction of 3. 3 and 3. 9 per cent per annum compared to approximately 2. 8 per cent, respectively. In Orissa, on the other hand, the average rate of reduction has been greatest for the Other grouping, with the SC and ST experiencing an average annual decline below 2 percent. The differences between the ethnic groups are smaller in terms of NMR in both states. In Orissa, the estimates of NMR suggest that both the ST and the SC experienced steady reductions in mortality rates since 2000. At the same time, the NMR of the Other has stagnated at approximately 48 deaths per 1, 000 live births. In MP, the performance of all ethnic groups has stagnated since 2001, again highlighting the difficulties with reducing the number of neonatal deaths in the latter years of the sample. The trends in U5MR and NMR across wealth quintiles for rural and urban areas are presented in Figures 4 and 5, respectively. In MP, the estimates suggest that all socioeconomic groups experienced mortality reduction over the past two decades. However, wealth-related inequality in mortality persists in both areas for both under-five and neonatal mortality, and is unlikely to diminish by 2015. In Orissa, the results are more variable, with the mortality profiles for the four wealth groups in urban areas very different from those of the rural populations. Consequently, wealth-based inequality among the rural population is found to be far less than that of the urban population. Both rural and urban areas, however, do share one common characteristic: the divergence between wealth groups appears to begin around 1995-96. In both states across urban and rural regions, we observe that one wealth quintile experiences markedly worse NMR compared to the other wealth-groups. In MP, the second quintile suffers from a lack of progress in rural areas; while in urban areas, this lack of improved NMR is incurred by the lowest quintile group. In Orissa, a large gap in NMR between the second quintile and the remaining wealth groups is observed in urban areas. In rural areas, children from middle quintile households experience greater NMR. Finally, we reassess the spatial distribution of under-five mortality by estimating district-level estimates within both states. The results are presented in Figure 6. Two patterns are observed from these results. First, substantial variations in the levels of under-five mortality are detected in both states. In MP, the worst performing district in 1990 is found to have an U5MR of approximately 239 (95% CI: 164-304), while the best performing district achieved a rate of 84 (95% CI: 68-100). Similarly, in Orissa, the rates of under-five mortality in the worst and best performing districts are 231 (95% CI: 167-290) and 90 (95% CI: 65-117), respectively. Second, no uniform pattern of reduction is observed over time. While some of the worst performing districts in 1990 have been able to reduce their mortality rates, others continued to experience high relative levels of under-five mortality. Likewise, some of the average performing districts have been able to reduce mortality substantially while others have stagnated. Accordingly, the results suggest that geography-related inequalities persist within both states.
Discussion
While the number of child deaths in India continues to decline at a national level (Lozano et al., 2011), the disaggregation of mortality trends for two of the country’s poorest states – MP and Orissa – reveals that various sub-populations suffer disproportionately. Inequalities in under-five and neonatal mortality rates are observed between rural and urban households, social/ethnic groups, wealth strata, and sub-state governmental administrative districts. Both states have made progress in reducing high mortality rates of certain equity groups (e. g. Scheduled tribes and rural populations). However, as a consequence, other groups have been left behind and stagnating or rising mortality levels amongst these groups has led to a lack of progress in overall U5MR and NMR. The prevalence of the within-state heterogeneity suggests that caution should be heeded when interpreting the success of nations at the aggregate level and indicates the importance of assessing the sub-national performance of countries seeking to further reduce the levels of child and maternal mortality. Both states have shown signs of rural and urban convergence, driven primarily by the decline in rural U5MR. However further efforts are required in urban areas to see a faster rate of U5MR decline. A number of factors could be at play here. Other inequities between rural and urban populations have been documented, with those living in rural areas experiencing greater barriers to accessing health services and a higher incidence and severity of poverty (International Institute for Population Sciences & Macro International, 2007; International Institute for Population Sciences & Ministry of Health and Family Welfare, 2010). The faster rate of decline in U5MR observed in the rural areas might be partly attributed to the success of the National Rural Health Mission and innovative policies, such as the Navjyoti scheme in Orissa, which focus on rural areas. Another possibility is that children living in urban areas, especially those who are poor, face distinct difficulties that prevent further mortality reduction, which has led to the relatively lower rate of U5MR reduction observed. It has been suggested that crowding, indoor air pollution, and poor access to health services could be partially responsible for the disappointing progress in urban under-five mortality outcomes in many countries (World Health Organisation & United Nations Human Settlements Programme, 2010). The situation is exacerbated if we also take into account the influx of the rural poor into urban slums. Figures from census data indicate almost 10 per cent of the urban population in the MP migrated from rural areas between 1991 and 2001, with livelihood opportunities a major driver of migration (Mitra & Murayama, 2008). Increasing rural-urban migration may place upward pressure on the mortality rates of the urban segments, if poorer, low health status rural households are more likely to migrate to urban areas than better-off households. Unfortunately, the DLHS series does not allow one to formally test for this hypothesis because they contain different waves of cross-sectional data (different households/women were surveyed for each wave; and hence, it is not possible to trace the regional movement of families). Such population pressures may have seriously affected the capacity of the health system to deliver good quality services and to achieve further improvements in mortality reduction in urban areas. Additionally, the increasing privatisation of health services in urban centres may have also posed financial barriers for the poor to good quality and affordable MNCH services. Consequently, some caution is required when assessing the effectiveness of the rural programmes, such as the NRHM, on under-five and neonatal mortality reduction. In Orissa, the impact of current programs aimed at reducing high under-five mortality amongst scheduled castes and tribes is obvious, but the analysis also shows that other groups are being left behind. The role of government policies could be critical in this case. For example, the ST grouping has made more progress towards the end of the sample period than the SC in reducing neonatal mortality. This might be the result of the Navajyoti scheme – launched in 2005 in 14 districts with higher-than-state-average infant mortality rates – to target tribal groups in Orissa with community-based home safe delivery and newborn care, and referrals to health centres (Government of Orissa, 2004). Similarly, in MP, the converging pattern observed for U5MR may be partly attributed to state government programs aimed at improving service delivery for disadvantaged groups, such as the Deendayal Mobile Hospital Scheme, which provides outreach services to tribal areas. Other state-wide schemes, such as Bal Shakti Yojana, target malnourished children based on physical (anthropometric) measurements. The higher prevalence of under-nutrition amongst Scheduled Tribe and Scheduled Caste children could prompt greater targeting of these disadvantaged groups and could also explain some of the observed gains. Unfortunately, neither DLHS-II nor DLHS-III collects data on nutrition to enable us to examine this hypothesis. The trends we present suggest that all socio-economic groups experienced mortality reduction over the past two decades. However, wealth-related inequality in mortality endures in rural and urban areas. While the lowest wealth quintile, especially in rural areas, has experienced rapid reduction in under-five mortality, since the mid-1990s the trends for the second and middle wealth quintiles, especially in urban areas, have stagnated or risen in both states, especially in Orissa. Orissa has done well in targeting the lowest wealth quintile in rural and urban areas, but the second and middle quintiles appear to have been neglected, while in urban areas of MP, the lowest wealth quintile continues to lag behind the other groups. It should be noted, however, that many pro-poor programmes and policies were introduced in India, and in MP in particular, around 2005. Consequently, their effects on maternal and child health service utilisations and outcomes might not be fully captured by the DLHS-III. As suggested earlier, children of the poorest urban households might face some distinct difficulties, which have not yet been addressed by pro-poor policies in MP. For example, children and women in slums experience much higher rates of under-nutrition and anaemia than those in non-slum areas of India (Athreya et al., 2010). Such difficulties are usually compounded by the influx of rural migrants into urban slums. Apart from the poor living conditions in slums, poor immigrants lack social networks and are rarely aware of their entitlements and the availability of free or affordable health services in urban centres. The consequence is delays in seeking health care or, in some cases, no health care at all. These findings support the renewed policy focus on the urban poor. A National Urban Health Mission (NUHM) targeting the urban poor, particularly slum dwellers, has been proposed to cover towns with populations over 100, 000 and the duration of the mission would be the remaining period of the 11th five-year plan (2008-2012) (Agarwal et al., 2006). Other initiatives have also started to address the problems facing slum dwellers. For example, a submission under the Jawaharlal Nehru National Urban Renewal Mission (JNNURM) covers the provision of affordable housing and public services, such as water supply, sewage and community toilets, in selected urban slums. The results of this study add weight to the importance of such initiatives and suggest that continual efforts to identify and target marginalised sub-populations are required to reduce mortality rates. The importance of geography-related inequalities is confirmed by the district-level results. While a number of factors previously mentioned may explain the large variations between districts, we submit one additional explanation. It is reasonable to expect that the problems associated with systematic poverty and disadvantage would have stronger effects in districts where local capacity is already the weakest. The substantial demands on local governments and health departments to improve general health and living conditions are likely to weight heavier in areas with a limited pool of resources and expertise. It is possible that the district estimates are detecting the consequences of such realities. A noteworthy pattern observed at the state-level as well as across most of the equity markers was the stagnation of neonatal mortality rates despite higher institutional births. In fact, NMR in urban areas has increased in both states, while such rates either remained stagnant or declined very slowly in rural areas. Such findings suggest that current policies have achieved a greater impact on the level of mortality in older children (those between one and four years of age), as well as on the degree of associated inequity. This is to be expected since reducing mortality in older children, unlike neonates, is more amenable to interventions that can be easily scaled up through vertical programs, such as immunisation. Interventions addressing neonatal mortality, such as basic and comprehensive emergency obstetric, and neonatal care (e. g. BEmONC and CEmONC) are more complex, and therefore, depend on broad strengthening of the health systems. In this study, we have tried to systematically collate the evidence on levels and trends in child mortality at a number of sub-national levels for two of India’s poorest states over the period 1990–2007. Nevertheless, this study has several important limitations. First, direct estimation of child mortality rates may be subject to recall bias and/or the under-reporting of deaths by mothers (Murray et al., 2007). However, we seek to reduce the extent of recall bias by pooling data from multiple surveys and for the periods with fewer than 10, 000 person-months of observation we did not generate estimates. Second, in addition to the well-known measurement errors created by survey-based data, the limitations of indirect estimation methods have previously been documented by Rajaratnam, Tran, et al. (2010). As they note, the main limitation of indirect methods is the need to infer information on statistics, such as the location in time of births and deaths, from observed patterns in surveys with CBH. This leads to the reliance on generalised patterns across states and across time. The impact of these generalisations is minimised by the application of local regression methods. Third, the large sampling errors associated with some of the trends where only a limited number of observations are currently available implies some caution is required when interpreting those results. Finally, our forecasts are based on recent time trends, and consequently do not indicate the possible impact achieved via intensified efforts to reduce child mortality in certain areas or by targeting specific sub-populations. In conclusion, in spite of the many development challenges facing the extremely poor Indian states of MP and Orissa, both have taken great strides in reducing child mortality, both in terms of absolute levels and in reducing the comparative disadvantage experienced by its most vulnerable groups. Nevertheless, based on the most recent data available, by 2015, MP and Orissa are unlikely to meet the national targets set by Millennium Development Goal 4. Of importance are the considerable gaps between different sub-populations, as defined by location, ethnicity, or socio-economic status. Future improvements will increasingly rely on the more difficult task of strengthening health systems and overcoming the barriers facing disadvantaged segments of the population.
