Ordered logistic estimates
We use an ordered logistic model in accordance to the ordinal nature of the data on child nutritional status. The dependent variable considers three levels of nourishment: 0 for normal growth, 1 for moderately malnourished and 2 for severely malnourished.
Table
3 reports coefficients for five model variants sequenced from models (1) to (5). Each of the five model variants reports the odds ratio (OR) from five ordered logistic regressions, with the levels of nourishment being the dependent variable. The base model (1) provides the effect of covariates with socio-economic factors without controlling for any of the indices. The subsequent models (2) and (3) build on the base model by adding the child health precaution index, the health seeking precautions index and the medical cost knowledge index, along with their relevant interaction terms as discussed under section 2.6. The difference between models (2) and (3) is the choice of proxy reflecting household wealth; model (2) uses the household asset index whereas model (3) uses per capita income. Finally models (4) and (5) disaggregate the model by gender; model (4) is for male children only and model (5) is for female children only.
Table 3
Odds ratio (OR) estimates of child nutritional status (standard errors)
Child Sex (1 = Female) | 0.729a (0.307) | 0.748a (0.339) | 0.640a (0.348) | 0.1239 | - | - |
Child Age (in months) | −0.021a (0.009) | −0.023a (0.009) | −0.026b (0.009) | 0.0721 | −0.048b (0.015) | −0.002 (0.014) |
Per Capita Income | −0.471b (0.176) | - | −0.519a (0.278) | 0.7717 | −0.699a (0.390) | −1.201a (0.542) |
Household Asset Index | - | −1.643 (2.540) | - | - | - | - |
Maternal Education | −0.425a (0.254) | −0.937a (0.488) | −0.033 (0.565) | 0.2254 | −1.648a (0.908) | 1.317 (1.094) |
Have Older Siblings (1 = Yes) | −0.555a (0.330) | −0.785a (0.359) | −1.046b (0.384) | 0.4068 | −1.551a (0.610) | −1.403a (0.604) |
Child Health Precautions Index | - | 1.893b (0.553) | 1.786b (0.554) | 0.1445 | 1.883a (0.956) | 2.179a (0.885) |
Health Seeking Practices Index | - | 0.649 (0.469) | 0.837a (0.475) | 0.9571 | 0.252 (0.751) | 3.123b (0.823) |
Medical Cost Knowledge Index | - | 1.081b (0.364) | 1.293b (0.375) | 0.6970 | 1.510a (0.646) | 1.221a (0.583) |
Maternal Education × Per Capita Income | - | - | −0.242 (0.245) | 0.2456 | −0.237 (0.410) | 0.893 (0.653) |
Maternal Education × Household Asset Index | - | 1.316 (2.094) | - | - | - | - |
Maternal Education × Child Health Precautions Index | - | 0.161 (0.568) | 0.640 (0.639) | 0.2784 | 1.193 (1.172) | −0.209 (0.995) |
Maternal Education × Health Seeking Practices Index | - | 0.077 (0.533) | −0.268 (0.573) | 0.2233 | −0.254 (1.075) | −2.453a (0.975) |
L0
| −178.4811 | −178.4811 | −178.4811 | - | −90.9896 | −84.4876 |
L | −165.7001 | −147.2901 | −139.4184 | - | −58.6590 | −62.3075 |
LR | 25.56b
| 62.38b
| 78.13b
| - | 64.66b
| 44.36b
|
Observations | 174 | 174 | 174 | - | 86 | 88 |
In order to check the proportional odds assumption of the model, the chi-square test was employed. The chi-square value of the overall model is 12.49 with a prob > chi2 = 0.3281 so the proportional odds assumption is not violated. Since the score test can sometimes be anticonservative, we also conduct single score tests for the explanatory variables. The results of the single score test for the per capita income model (3) are presented in the column next to it and indicate that individually also, none of the variables violate the proportional odds assumption at five percent significance level. Between the household assets model (2) and the per capita income model (3), the likelihood ratio suggests that the goodness of fit is best when we consider per capita income as a proxy for household wealth, that is model (3) is considered the better of the two models.
Note that the covariates can be theoretically differentiated into three categories: the variables of primary interest (proxies of income and maternal education), socio-economic control variables (child sex, child age and whether the child has older siblings) and our three indices of primary interest reflecting health knowledge and health seeking practices.
The coefficient for per capita income under model (3) is negative and significant (OR = −0.519; p-value = 0.061), that is children living in households with more per capita income are less likely to be malnourished. Although the coefficient for the asset index under model (2) is also negative, it is insignificant (OR = −1.64; p-value = 0.518). Since we do not have permanent income or expenditure data, we cannot affirm the role of permanent income on child health. Furthermore, as mentioned earlier, our constructed asset index is somewhat of a rudimentary nature. However, if we do accept it as crude proxy, permanent income is found to be an insignificant factor in determining child health, a result that has also been suggested by Pal (1999). However, given the nature of our asset data, this finding should be taken cautiously, and we will focus on the per capita income model (3) for the remaining of the analysis. It is to be noted however that the remaining explanatory variables maintain the same direction in both models (2) and (3) indicating to some degree the robustness of the result.
Child age is negative and significant for our sample meaning malnourishment is higher for younger children and falls with age. The result is consistent for the all three models (1), (2) and (3). We also find a similar significant relationship with the number of siblings
2 (OR = 0.150; p-value = 0.083). This is in contrast to the finding from Glewwe (1999) [
27] who find a negative relationship. Glewwe (1999) [
27] provided the rationale that with fewer children mothers are able to allocate more time and health inputs per child. In fact such a relationship is what we initially expected.
However, based on the results we reason older siblings help in looking after the young either through increased care or by earning in our study location. It is a common practice in the slums for the elder children to help look after the young and even in household activities. To confirm this, we replaced the number of siblings variable with a binary to indicate whether a child has older siblings and find a statistically significant result, as reported in Table
3 (OR = −1.046; p-value = 0.006 for the per capita income model). Following the same procedure for gender specific regression models (4) and (5), this also resulted in a statistically significant relationship, for both the male (OR = −1.551; p-value = 0.011) and female (OR = −1.403; p-value = 0.020) children. Marginal effects at the mean tell us that having an older sibling means that the child is 22.8 percent less likely to be severely malnourished while 15.5 percent more likely to be normal and 7.4 percent to be moderately malnourished, all statistically significant.
One advantage of retaining the base model (1) in the results table is that it lets us compare how much of the effect of maternal education on child health is soaked by adding our indices. While the odds ratio of maternal education is large and significant (OR = −0.425; p-value = 0.094) for the base model (1), it is small and insignificant (OR = −0.033; p-value = 0.953) for the per capita income model (3) which keeps the indices and interaction terms as covariates. It can be argued that the indices and the interaction terms capture a portion of the 'total' education effect. Thus, after controlling for child health knowledge and health-seeking behavior, the remaining impact of maternal education on child health is no longer significant.
The child health precautions index (OR = 1.786; p-value = 0.001) and the medical cost knowledge index (OR = 1.29; p-value = 0.001) are both positive and highly significant in the per capita income model (3): children are more malnourished in households that are deprived in taking child health precautions and are less knowledgeable regarding medical treatment costs. This is also consistent in the household assets model (2). The health seeking practices index, in contrast, is only significant (OR = 0.837; p-value = 0.078) in the per capita income model (3).
Child sex is positive and significant for all the three models (1), (2) and (3). So, other things remaining the same, female children are more likely to be malnourished than male children. Pal (1999) [
42] suggested that male and female children follow separate nutritional functions and in view of the last result, we run separate gender specific regressions. Given that per capita income model provided a better fit, and for reasons explained earlier, we use per capita income as our proxy for household wealth instead of the household asset index. The results are provided in models (4) and (5) capturing the differences in male and female nutritional functions. Child age, per capita income, maternal education, having older siblings, child health precautions index and medical cost knowledge index, all significantly affect male children. For female children, per capita income, having older siblings, child health precautions index, health seeking practices index and the medical cost knowledge index are significant. The gender unified models (1), (2) and (3) fail to capture this complexity between male and female children nutritional functions.
While older male children face less malnutrition (OR = −0.048; p-value = 0.002) we cannot say so conclusively for the female children due to child age being insignificant (OR = −0.002; p-value = 0.915). Nutritional status for both genders improves with increased per capita income and having older siblings. Even after controlling for child health precautions and health seeking practices, maternal education is significant for male children (OR = −1.648; p-value = 0.070). Although maternal education is insignificant for female children (OR = 1.317; p-value = 0.229), it is interesting that the variable has opposite effects for the two genders: holding everything else the same, educated mothers improve the nutritional status of male children but lowers that of female children. Pal (1999) [
42] also finds similar results in rural India using the World Institute of Development Economics Research (WIDER) dataset.
Finally, while both the child health precautions index and the medical cost knowledge index are significant for both genders, the health seeking practices is highly significant (OR = 3.123; p-value = 0.000) only for female children. In other words, for our study location, female children living in households that are deprived in health-seeking practices are relatively more malnourished, while the same cannot be conclusively said for male children.
Gender specific marginal effects of covariates
The ordered logistic regression allows us to compute the marginal effect of each covariate on child health separately for normal, moderately malnourished and severely malnourished children. In order to further understand the differential nutritional functions by gender, we do this for the male and female regression models (4) and (5) from Table
3. Table
4 thus reports the effect of the covariates on the probability of being normal, moderately malnourished and severely malnourished separately for male and female children.
Table 4
Marginal effects of explanatory variables on child health for male and female children
Child Age | 0.000 | 0.010b
| 0.000 | 0.001 | −0.000 | −0.011b
|
Per Capita Income | 0.122a
| 0.141a
| 0.161a
| 0.013 | −0.283a
| −0.154a
|
Maternal Education | −0.134 | 0.332a
| −0.177 | 0.031 | 0.311 | −0.363a
|
Have Older Siblings (1 = Yes) | 0.143a
| 0.314a
| 0.175a
| 0.009 | −0.318a
| −0.323b
|
Child Health Precautions Index | −0.213a
| −0.320a
| −0.248a
| −0.106 | 0.461b
| 0.426a
|
Health Seeking Practices Index | −0.362b
| −0.050 | −0.272b
| −0.006 | 0.634b
| 0.056 |
Medical Cost Knowledge Index | −0.136a
| −0.331a
| −0.150a
| 0.044 | 0.286a
| 0.287b
|
Maternal Education × Per Capita Income | −0.091 | 0.048 | −0.120 | 0.004 | 0.211 | −0.052 |
Maternal Education × Child Health Precautions Index | 0.021 | −0.024 | 0.028 | −0.022 | −0.049 | 0.263 |
Maternal Education × Health Seeking Practices Index | 0.249a
| 0.051 | 0.330a
| 0.004 | −0.579* | −0.056 |
The marginal effects supplements the inferences made from the parameter estimates. An increase in one unit of per capita income (BDT 1,000) decreases the probability of being severely malnourished by 28.3 percent for female children and 15.4 percent for male children. Thus while per capita income helps all children to step out of malnutrition, it helps female children more. However an increase in one unit of education decreases the probability of being severely malnourished by 36.3 percent for male children while increases the probability by 31.1 percent for female children (the latter is insignificant).
3 This once again reflects the dipartite effect of maternal education on child health in our study location.
While interpreting the estimates on the child health precautions, health seeking practices and medical cost knowledge indices, it is important to remember that an increase in one unit means changing from not deprived status to a deprived status. Thus, for zero maternal education, being deprived in child health precautions increases the probability of being severely malnourished by 42.6 percent for male children and 46.1 percent for female children. Similarly being deprived in medical cost knowledge reduces the probability of having normal health by 33.1 percent for male children and 13.6 percent for female children.
Male children seem to be more sensitive to child health precautions and medical cost knowledge while female children more to health seeking practices. For zero maternal education, being deprived in health seeking practices increases the probability of being severely malnourished by 63.4 percent while decreases the probability of normal growth by 36.2 percent for female children (comparable estimate are 5.6 percent and 5.0 percent for male children, although insignificant). In other words policies looking to improve the nutritional status of female children vis-a-vis male children should focus on health-seeking practices.
The disparity in the results for female children compared to male children is quite noticeable. It is possible that prevailing socio-economic norms in Bangladesh, which tend to place a greater value on male offspring, leads to boys receiving more attention and a better quality of care as infants, which positively affects their health, physical robustness and immunity later on. It is an issue which requires further inquiry.
For mothers with primary education however, being deprived in health seeking practices increases the probability of being severely malnourished by 5.5 percent (for zero maternal education this value was 63.4 percent) while decreases the probability of normal growth by 11.3 percent (for zero maternal education this value was 36.2 percent) for female children. This sharp decline in the probability estimates indicate that education does offset the negative impact caused due to being deprived in health seeking practices for female children.