Extent of socio-economic inequity in childhood stunting
As mentioned earlier, the successive waves of NFHS in India indicates a declining trend in the prevalence of child malnutrition among children aged below three years (Figure ).
Trend in Malnutrition in India among Children (0-35 months)
Except for wasting, across the two different established anthropometric measures of malnutrition; stunting and underweight, a consistent decline is evident during 1992-2005 period (Figure ). Overall, NFHS-3 reveals a differential scenario of child malnutrition across the fifteen major states of India (Table ).
Prevalence of Malnutrition among children (0-59 months) across fifteen major states of India (NFHS-3)
To describe further, the state of Kerala showed the lowest prevalence of stunting among children (25 percent) across all the major states, where the rural-urban differential is virtually nonexistent. Whereas the opposite side of the spectrum, more than half the children below five years were stunted in Uttar Pradesh (57 percent), Bihar (56 percent), Gujarat (52 percent) and Madhya Pradesh (50 percent) (Table ). The rural-urban differentials are also considerably high in these states, along with West Bengal; which showed the highest (19 percent) differential between rural-urban prevalence of child malnutrition which is unfavorable for rural areas, during NFHS-3.
Overall, all the three indicators of malnourishment are found highly correlated with each other and hence it was worthwhile to explore their association with the incidence of poverty in the states, following the established line of argument. It can be said that the optimal growth of the children (having standard height for their age and weight for their height) have been strongly associated with economic status of the population.
The CI values for chronic malnutrition in respect to fifteen major states and at the country level consistently return negative values, reflecting a heavy burden of malnutrition among the poor in India (Table ).
Concentration Index Values for Stunting across States and Urban-Rural Locations, India, NFHS-3
The above table (Table ) confirms the fact that across all fifteen major states and the rural-urban locations, children from poorer households share the higher burden of sub-optimal growth due to undernourishment. It needs special mention that chronic malnutrition among children is more concentrated among urban poor comparing their counterpart living in rural areas. This trend is consistent across all thirteen states, except for Bihar and Kerala; where concentration of stunting is observed higher among poor children from rural areas.
It is also seen in a similar vein that aggregate economic status of a population is associated with child nutritional status. CI values for stunting and Net State Domestic Product (NSDP; considered as the indicator for economic development for the aggregate level of the state) share an inverse association (Figure ), common for most of the states.
Scatter plot showing relationship between NSDP and CI for Stunting, across the states.
Overall, the negative correlation established between CI values for stunting and NSDP per capita stands at r = -0.603 (p < .001). The scatter plot of NSDP per capita and CI values for stunted children across the states (Figure ) emerges few specific patterns. The states like, Bihar, Uttar Pradesh, Madhya Pradesh and Rajasthan exhibit a typical situation where per capita NSDP is lower than the India average and the burden of malnutrition among the poor is also lesser (indicated by lower negative values of CI). The second group consisting of Gujarat, Tamil Nadu, Karnataka and Andhra located at the north-east quadrant of the scatter diagram, shows a situation where NSDP per capita either equals or surpasses the national average; however a relatively higher burden of chronic malnourishment is found concentrated among the poor. This characteristic is further intensified in the case of Punjab, along with somewhat closer scenario for Kerala and West Bengal and Haryana; where a considerably higher share of stunting can be found among the poor. However, the worst case is noticed for Orissa, where with much lower per-capita NSDP as compared to the national average, the state exhibits a noticeably higher burden of chronic malnourishment among the poor.
Role of household socio-economic conditions determining long-term nutritional status among children
The results shows (Table ), significant association between household asset quintiles and nutritional status of children.
Table 3 Association (βs) from Ordinary Least Squares and Multilevel Linear Regression Models (Main Effects) between Child Stunting (Height for Age) and Household Socio-Economic Status, controlling for various other covariates; Fifteen Major States, India, (more ...)
Given the form of the dependent variable in the subsequent models a higher coefficient indicates better nutritional status among children from better off socio-economic status quintiles (SES). It shows (Table ) nearly 50 percent better nutritional status (0.31 - (-0.18)) among children from richest SES quintiles, compared to ones those from the poorest quintile.
The variance component models (i.e., Model_kids, Model_moms and Model_full) and the random slope model (Table ) also support such finding. By introducing covariates at each level, the variance component models allow to examine the extent to which observed differences in the HAZ scores are attributed to the factors operating at each level. With the introduction of child's individual characteristics in the Model_kids along with the state level fixed effect, the impact of richest & poorest SES quintiles become much stronger. The result shows over 80 percent (0.54 - (-0.29)) higher incidence of worse nutritional status among children in the poorest quintile, than the ones hailing from richest SES group. However, such richest-poorest gap decreases with the phased introduction of covariates related to mother's characteristics, household ethnicity and place of residence in the models (Table ). Finally, similar to the initial estimate by OLS, the variance component models and random slope model indicates that the children with the most favorable SES background enjoy almost 50 percent better nutritional status than their counterpart from the poorest SES groups.
The calculated ICC coefficient values presented in Table differ from zero. This indicates that child nutrition is indeed correlated with households and communities (PSUs). The ICC for household level shows much higher correlation than the case of PSUs.
Random Coefficients, Intra-class correlation and Variance Decomposition estimates from comparative models
The lower panel of the Table shows how the residual variance is distributed across PSUs and households. Estimates from model 1(null model), which contains no observed covariates, indicate that the variation in height-for-age has substantial group level components. The total variance 0.548 (combined for PSU and households estimates), of which 63 percent is attributed to household level variation in anthropometric scores. Consistent with this observation on null model variance decomposition, other model specifications show similar variance distribution pattern across state, PSU and household levels.
Estimation of household random effects (Table ) indicates that household heterogeneity is accounted for only partially by the covariates in our model (Model_Full & Model_random_slope). In other words the significantly different values of σ2c and σ2h indicates that the homogeneity within cluster observations is not explained by the observed covariates specified in the model. The intra-household correlation remains as large as 0.242 suggesting that the height outcomes of two children belonging to the same family are more homogenous than those of two children chosen at random, even after adjusting for other observed covariates (Model_Full).
Overall, the higher value of σ2h ; i.e., variance at household level denotes existence of higher homogeneity at the household level. These results further imply that choice of one-level model with the similar data set might yield underestimation of parameters.