Who benefits from the South African Child Support Grant?: The role of gender and birthweight

ABSTRACT

Several studies have suggested that the South African Child Support Grant (CSG) reduces stunting in benefiting children. However, all of these studies have estimated the impact of the CSG on the mean of the height-for-age distribution. This paper investigates how this benefit varies across the quantiles of the height-for-age distribution. The result suggests that the positive effect at the mean is driven by children in the high quantiles and this group of children are more likely to be girls and children that did not experience low birthweight at birth. I argue that the CSG has not been able to address the malnutrition inequality that disadvantage male children and children born with low birthweight.

KEYWORDS:

• Unconditional cash transfer
• child health
• causal inference and nutrition

JEL:

• I38
• H53
• C21
• D13

1. Introduction

The literature has shown that adequate nutrition is important for children. This is because nutritional deprivation and malnutrition early in life have long-term negative consequences on the physical and cognitive development of children (Delany et al., 2008; Walker et al., 2015). One of the manifestations of malnutrition is stunting (low height-for-age) and early childhood stunting is likely to contribute to the intergenerational transmission of poverty (Grantham-McGregor et al., 2007). Walker et al. (2015) found that children born to a stunted parent are likely to have a lower score on the cognitive scale, lower development quotient and are likely to be stunted themselves. These studies suggest that the impact of stunting continues in the next generation of children, highlighting the importance of breaking the cycle of poverty and inequality.

One way to address poverty – a root cause of malnutrition – is through cash transfer programmes like the South Africa Child Support Grants (CSG). In various studies, the CSG has been found to boost child nutrition as measured by children’s height-for-age (Aguero et al., 2006; Coetzee, 2013 and Grinspun, 2016; Oyenubi, 2019, 2020). However, these studies focus on the impact of the CSG on the mean of the height-for-age distribution. From a policy perspective, this may be misleading if the effect of the CSG is heterogeneous across the height-for-age distribution. For example, given that low birthweight increases the risk of stunting in children (Aryastami et al., 2017), and early childhood stunting may not be reversible (Duflo, 2003), these factors can pre-determine the placement of individual children in the height-for-age distribution and implies that the impact of the CSG may be heterogeneous across the distribution. Consequently, the impact of the CSG on children that are born with low birthweight or those that experience early childhood stunting for other reasons and as a result found themselves at the lower quantiles of the height-for-age distribution may be very different from their counterparts at higher quantiles.

Furthermore, most existing studies on the impact of the CSG on height-for-age of children ignore the role of low birthweight and gender (for example, see Aguero et al. (2006) and Coetzee (2013)). These factors may impact on a child’s health differently across the height-for-age distribution. Results from the medical literature suggest that boys show more postnatal complications as a result of low birthweight. This phenomenon is referred to as the male disadvantage hypothesis. According to this hypothesis, boys are the weaker sex and are more sensitive to adverse environmental factors during gestation, infancy and childhood (Kirchengast & Hartmann, 2009). It is, therefore, important to investigate the effect of these variables (gender and low birthweight) on the height-for-age of vulnerable children. This information may assist policymakers to fine-tune the delivery of this policy for enhanced impact.

This study, hence, estimates the quantile effects of the CSG with particular emphasis on the role of gender and low birthweight. If the impact of the CSG varies by placement, then factors that influence this placement can be used to better understand how to enhance the effectiveness of the policy. The entropy balancing method (Hainmueller, 2012; Hainmueller & Xu, 2013) is used to recover the counterfactual distribution of beneficiaries under the ignorability assumption (details in section 2.1). The analysis is disaggregated by gender to investigate the role of gender, and low birthweight.

The results show that children at the top of the height-for-age distribution are driving the positive effect observed at the mean. This result holds even after controlling for low birthweight. The disaggregated analysis suggests that the mean effect of CSG on the height-for-age for boys is weaker than its effect on girls. At the quantiles, there is considerable heterogeneity across gender groups in terms of region of significant positive impact. Furthermore, when low birthweight is used as the treatment variable, the result shows that low birthweight has a negative impact across the height-for-age distribution. It is, hence, argued that ignoring the mitigating effect of low birthweight on the impact of the programme is counterproductive.

The rest of the paper is organised as follows; section 2 discusses the methods and the data used. Sections 2.1 and 2.2 briefly discuss entropy balancing and why the unconditional quantile regression is preferred to the conditional quantile regression in estimating the quantile effects. Section 2.3 discusses the estimation of the caregiver¹ motivation variable while section 2.4 explores the summary of the data used. Section 3 presents the results and section 4 concludes.

2. Methods and data

Existing literature on the impact of CSG (in the South African context) have noted that apart from observed covariates that may confound the impact estimate of the CSG, caregiver motivation is a key factor (Coetzee, 2011, 2013; Oyenubi, 2018). Motivation captures the eagerness of caregivers to take up the grant such that higher caregiver motivation is associated with a lower delay in taking up of the grant. Those with high motivation and consequently high dosage are likely to experience more benefit relative to those with low motivation (Oyenubi, 2018, 2020).

One problem with controlling for caregiver motivation is that motivation is unobserved and, therefore, has to be estimated.² To estimate this variable, existing research uses ‘period of delay before applying for the CSG’ to recover caregiver motivation. This is achieved by modelling the delay as a function of the child’s age and location (rural versus urban) using censored regression. However, Coetzee (2011, 2013) found that when the estimated motivation is included in the set of covariates under Propensity Score Matching (PSM), it becomes difficult to balance the distribution of covariates in the binary treatment case. This is because, by construction, motivation in the treatment group will be higher than motivation in the control group since this variable is a function of delay in applying for the CSG.

More recent methods side-line this problem by focusing directly on balance in the relevant covariates (e.g. Diamond & Sekhon 2013) or estimating propensity scores that incorporate the balancing condition (Imai & Ratkovic, 2014). Another strand of the literature views the balancing problem as a calibration problem (Hainmueller, 2012). Under the calibration approach, weights are calculated for each unit so that the weighted treatment and control groups satisfy prespecified balancing conditions. This approach has also been shown to deliver better performance than the PSM in terms of mean square error and lower model dependency (Hainmueller, 2012).

In this paper, the entropy balancing method is used to balance the distribution of covariates including the estimated motivation variable. This method balances the mean and variance of the distribution of covariates. The balancing weights from entropy scheme can be combined with any standard estimator that one may wish to use to estimate the treatment effect (Hainmueller, 2012: pg. 26). Hence, entropy balancing weights are used in weighted mean and quantile regression to recover the average treatment effect on the treated and the quantile effect of the CSG.

2.1. Entropy weights

Entropy balancing is a pre-processing procedure that allows researchers to create balanced samples for the subsequent estimation of treatment effects (Hainmueller, 2012). This is achieved by reweighting the covariate distributions such that the reweighted data satisfies a set of specified moment conditions. Pre-processing reduces model dependency for the subsequent analysis of the treatment effects using standard methods such as regression (Abadie & Imbens, 2011). Similar to methods in the survey adjustment literature (Deming & Stephan 1940), entropy balancing is based on a maximum entropy-reweighting scheme. It allows exact balance on the first, second and possibly higher moments. The current study uses entropy reweighting to balance the mean and variance of the covariate distributions as implemented in Hainmueller & Xu (2013).³

In plain language, one can think of the regular Logit or Probit propensity score weights as being derived from optimising the likelihood of observing each unit in the treatment group. In the case of entropy balancing, this optimisation includes constraints that specify that the first, second and possibly third moments of the covariates are balanced. In other words, the difference is that entropy balancing includes balance constraints. Details of the numerical implementation can be found in Hainmueller & Xu (2013).

2.2. Conditional versus unconditional quantile effects

The weights recovered from the optimisation discussed in the previous section can be used to recover the unconditional quantile effects since these estimators are based on regression equations.

For the quantile effect, the unconditional quantile estimator proposed by Firpo et al., (2009) is used. The method involves estimating the Regression Influence Function (RIF) of any distributional functional of interest. Since interest is in quantiles, the RIF is given by (1) $\text{RIF}\left(Y_i;q_{\tau }\right)=q_{\tau }+\frac{\tau -1[Y_i<q_{\tau }]}{f_Y\left(q_{\tau }\right)}$(1)

where $Y_i$ is the outcome, $q_{\tau }$ is the ${\tau }^{th}$quantile of $Y_{it}$, $1[Y_i<q_{\tau }]$ is an indicator variable that denotes when $Y_i$ is below $q_{\tau },$ and $f_Y\left(q_{\tau }\right)$ is the estimated density at $q_{\tau }$. The RIF is then regressed on the covariates. Under this method, quantiles are defined pre-regression, therefore, the quantiles do not depend on the value of the covariates (Lass & Wooden, 2019). This implies that the effect at the ${\tau }^{th}$ quantile can be associated with individuals that seat in that quantile in the outcome distribution. Our estimation involves including entropy weights in each (quantile) regression to recover the unconditional quantile effects.

Note that our results are only valid under the conditional independence assumption (i.e. balancing the distribution of the covariates is sufficient to identify the counterfactual distribution and consequently the treatment effect). While this approach does not control for unobserved variables, it allows this analysis to qualify existing results about the impact of CSG in South Africa (i.e. qualify existing result that are based on similar assumptions).

2.3. Estimation of caregiver motivation

The estimation of the unobserved caregiver motivation follows the work of Oyenubi (2018, 2020). Caregiver delay for treated observations is estimated as the number of years between the birthday of the child and the day the CSG was first received for the child. For control observations, the delay is approximated by the number of years between the birthday of the child and the date the caregiver applied for CSG on behalf of the child (where the data is available). If the date of application is not available, or CSG has never been applied for on behalf of the child, the delay is equal to the child’s age. To accommodate the control observations without data on application date, censored regression is employed to estimate the expected delay equation. These observations are regarded as being right censored, with a variable censoring point that is equal to the age of the child (since an application has not been made for these children but may be made sometime in the future). The expected delay is calculated as the predicted delay from the censored regression of the delay on child’s age, relationship with primary caregiver, and location (rural or urban). The difference between actual and expected delay is then standardized⁴ to arrive at a variable that represents the unobserved variation in caregiver’s motivation.⁵ Existing studies (Aguero, et al. (2006) and Coetzee (2011)) include only the child’s age and location in the delay equation. Relationship with primary caregiver is included in this analysis because it may influence the eagerness of the caregiver to apply for the grant (Oyenubi, 2020).

2.4. Data

Wave 1 data of the National Income Dynamic Study (NIDS) is used for the analysis. The choice of this dataset is informed by the fact that a good number of existing research that focuses on the mean, and for which the comparison of the quantile results is based on, uses wave 1 of NIDS (see Coetzee (2011), and Oyenubi (2018, 2019, 2020)). NIDS is a nationally representative panel dataset, with the first wave undertaken in 2008. Eligibility for CSG is determined by age and a means test⁶ as it was applied in 2008. The age and means test conditions are used to identify children who should benefit from the CSG. The treatment group in this study consists of children for whom CSG is currently being received, while the control group consists of children in eligible households that are not receiving CSG. Although as at 2008, the CSG was well known, accessing the grant remained problematic for caregivers for a number of reasons (SASSA & UNICEF, 2013). These included the high level of confusion about the means test, lack of prescribed documentation such as birth certificates, identity documents and death certificates, etc. These created a pool of children in eligible households who were not yet benefiting from the programme. For analysis purpose, this pool is assumed to represent the counterfactual, and under the relevant assumption, the entropy weighting allows us to control for selection due to observable characteristics. The sample is restricted to black children since they are the majority in the population and genetic factors may affect the height-for-age score of children differently across race groups.

To derive the height-for-age score, NIDS took two height measures, and then a third if the first two measures are more than one centimetre apart. The average of the first two measures is used to calculate the Z-scores. When the first two measures differ by one centimetre or more, the third measurement is used. NIDS, however, noted that third measures are seldomly taken (Brophy et al., 2018). The WHO international child growth standards and WHO growth standard for school aged children and adolescents (WHO, 2006; de Onis et al., 2007) were used as reference in calculating the height-for-age scores. Following WHO guidelines, NIDS set all biologically implausible values (i.e. values below -6 and above 6) to missing. Consequently, 20% of the data is missing, overall. Following existing studies, the missing data on other variables were ignored.⁷

Table 1 presents the summary statistics for the sample. Caregivers of benefiting children delay for about 2 years on average before accessing the grant while caregivers of eligible non-benefiting children delay for around 7 years on average. This shows that motivation is significantly higher in the treatment group. However, even in the treated group, the average delay is relatively high because the literature suggests that stunting may be irreversible in children older than two years.

Table 1. Summary statistics.

	Treated		Comparison
Variable	Mean	SD	Mean	SD	Mean difference	t-statistic
Delay (in years)	2.05	2.07	7.67	5.39	5.623***	−43.9
Motivation	0.57	0.79	−0.87	0.56	−1.439***	(−59.27)
Employed	0.24	0.43	0.24	0.43	−0.00449	(−0.31)
Married	0.35	0.48	0.41	0.49	0.0541**	−3.28
Primary caregivers’ education	7.36	4.38	6.1	4.79	−1.269***	(−8.17)
Primary caregivers age	38.16	13.65	43.87	17.68	5.706***	−10.87
HH head gender	0.36	0.48	0.35	0.48	−0.00914	(−0.56)
Child’s gender	0.5	0.5	0.5	0.5	−0.00038	(−0.02)
Electricity	0.64	0.48	0.63	0.48	−0.00616	(−0.37)
Water	0.17	0.37	0.19	0.39	0.0245	−1.89
Telephone	0.03	0.16	0.04	0.19	0.00902	−1.52
Toilet	0.2	0.4	0.24	0.43	0.0482***	−3.44
Food expenditure (AES)	183.92	120.82	195.55	151.87	11.63*	−2.54
Child’s age	6.64	3.74	7.86	5.39	1.215***	−7.97
Rural	0.74	0.44	0.72	0.45	−0.0278	(−1.84)
Child’s relationship to grant recipient	0.74	0.44	0.51	0.5	−0.236***	(−14.91)
Informal dwelling	0.08	0.27	0.08	0.26	−0.00334	(−0.37)
Hospital visits (No=1)	0.45	0.5	0.49	0.5	−0.0109	(−1.91)
Observations	2170		1418		3588

t-statistics in parentheses.

*p < 0.05, **p < 0.01, *** p < 0.00.

Another strand of the literature, however, suggests the possibility of catch-up growth (e.g. Zhang et al, 2016). Catch-up growth is defined as rapid linear growth that allows the child to accelerate forward, and in favourable circumstances, resume his/her pre-illness growth curve (Boersma & Wit, 1997). This suggests that even if some children are malnourished in the first two years of life, income from CSG can help them recover some of the lost grounds relative to eligible non-benefiting children. The NIDS dataset captures hight-for-age score for respondents that are 15 years or younger, hence, based on the possibility of catch-up growth, this study’s sample includes children up to age 15 years. It is, however noted that Desmond & Casale (2017) provide evidence that while the catch-up growth is possible, stunted children at age 2 may not fully recover in terms of cognitive ability.

On average, there are significant differences at the mean for most of the covariates. Variables used for estimation includes caregiver characteristics (i.e. marital status, employment status, caregiver years of education, age, health-seeking behaviour in form of the number of times the child has visited the hospital in the last 12 months⁸ and a variable that describes the relationship between the child and the primary caregiver (the variable is coded 1 if the primary caregiver is the father or mother of the child and zero otherwise)); household characteristics (i.e. household location (rural or urban), type of dwelling (informal or otherwise), the gender of head of household, availability of electricity, water, telephone, toilet, and household food expenditure per adult equivalent⁹); and child characteristics (i.e. gender and age of the child).

3. Results

Table 2 shows the censored regression that is used to estimate caregiver motivation. The result justifies the inclusion of the caregiver relationship (as a robustness check, a set of results that excludes caregiver relationship is presented in Appendix A). Caregiver relationship is significant in explaining variation in delay. Children who live with their father or mother are likely to apply 2 years earlier than those who live with other relatives. The sign of the other variables is as expected.

Table 2. Delay equation.

VARIABLES	Years of delay
Child age	0.787***
	(0.017)
Rural location	−0.309*
	(0.163)
Care giver relationship	−2.053***
	(0.156)
Constant	1.187***
	(0.222)
Observations	3,588

Standard errors in parentheses.

*** p < 0.01, ** p < 0.05, * p < 0.1.

3.1. Effect of CSG at the mean

Table 3 presents the summary statistics after applying entropy weights. Note that birthweight is excluded from this analysis.¹⁰ The result shows that the entropy method balances the mean and variance of all variables.¹¹ The result of the weighted mean regression is shown in columns 1 and 2 of Table 4. The result shows that at the mean, the treatment effects are 15% and 14% of a standard deviation¹² with and without control for low birthweight. These effects are not significant at conventional levels. Note that other covariates are included (i.e. covariates listed in Table 2) in the weighted regression for efficiency but the results are not shown in the table since focus is on the effect of the CSG. These results indicate that the CSG has a positive, but insignificant, effect on height-for-age at the mean. Column 2 shows that low birthweight¹³ has a negative effect on height-for-age, as one would expect, and this effect is large (44% of a standard deviation) and significant. Since birthweight was not included in the processing, this can be interpreted as the correlation between low birthweight and height-for-age score at the mean.

Table 3. First three moments after entropy balancing.

	Treat			Control
	Mean	Variance	Skewness	Mean	Variance	Skewness
Motivation	0.569	0.628	0.660	0.568	0.629	0.795
Employed dummy	0.243	0.184	1.199	0.243	0.184	1.199
Married dummy	0.352	0.228	0.619	0.352	0.228	0.619
*Primary caregivers’ education (<5years)	0.313	0.215	0.805	0.313	0.215	0.805
Primary caregivers’ education (between 5 and 10 years)	0.396	0.239	0.426	0.396	0.239	0.426
*Primary caregivers age (<40years)	0.636	0.232	−0.567	0.637	0.232	−0.568
Primary caregivers age (between 40 and 60 years)	0.277	0.200	0.999	0.276	0.200	1.000
Household head’s gender	0.362	0.231	0.575	0.362	0.231	0.575
Child’s gender	0.505	0.250	−0.018	0.505	0.250	−0.019
Electricity	0.637	0.231	−0.571	0.637	0.231	−0.571
Water	0.166	0.138	1.796	0.166	0.139	1.796
Telephone	0.028	0.027	5.762	0.028	0.027	5.759
Toilet	0.196	0.158	1.533	0.196	0.158	1.532
Food expenditure (AES)	183.9	14598.	3.710	183.9	14601	3.496
Informal dwelling dummy	0.079	0.073	3.127	0.079	0.073	3.126
Hospital visits	0.454	0.248	0.187	0.454	0.248	0.187

Note that by definition any difference between the moments are not significant

*Primary caregivers age and education are feed to the algorithm as dummy variables to ease convergence

Table 4. Effect of CSG with and without low birthweight.

Dependent variable height-for-age score	(1)	(2)
Treatment	0.14	0.15
	(0.18)	(0.18)
Low birthweight		−0.44**
		(0.21)
Constant	−0.21	−0.18
	(0.43)	(0.43)
Other covariates	(included)	(included)
Observations	2,837	2,837
R-squared	0.08	0.08

Standard errors in parentheses *** p < 0.01, ** p < 0.05, * p < 0.1.

By other covariates I mean the covariates listed in Table 2.

The estimate of 0.14 or 0.15 is around the mid range of other estimates reported in the literature, especially in the binary treatment case. Coetzee (2011) and DSD, SASSA and UNICEF (2012). reported 0.07 and 0.072 respectively. Aguero et al. (2006) did not present the estimate (in the binary treatment case) but noted that their estimate is not statistically significant. Oyenubi (2018, 2019, 2020), however, report significant estimates of 0.16, 0.44 and 0.39¹⁴ respectively. These authors use, arguably, more efficient matching algorithims (i.e. Covariate Balancing Propensity Scores and Genetic Matching). Lastly, in the continuous treatment case, Aguero et al. (2006) and Coetzee (2013) report maximum effects of 0.25 and 0.04 respectively.¹⁵

Next, the analysis is disaggregated by gender – the result is presented in Table 5. Columns 3 and 4 show the result for girls where the treatment effect is 27% and 29% of a standard deviation with and without a control for low birthweight. These estimates are significant at the 1% level showing that CSG has a significant effect on the height-for-age score of benefiting girls. The result also shows that the correlation between low birthweight and the outcome for girls is negative (as one will expect) but not significant. Perhaps, this is because girls recover faster than boys when affected by low birthweight. For boys, the estimate of the CSG’s impact is 3% of a standard deviation with or without controlling for birthweight. These estimates are not significant in both cases. This means that the CSG has no significant effect on the height-for-age of boys at the mean. The estimate for low birthweight shows that this factor is significantly correlated with height-for-age for boys. This set of results is consistent with the male disadvantage hypothesis (Kirchengast & Hartmann, 2009) – that is, if boys are more sensitive to adverse factors during gestation, infancy, and childhood, then it is plausible that the relationship between low birthweight and height-for-age will remain significant while this relationship may dissipate for girls because of catch-up growth. Furthermore, this may result in girls being more responsive to the impact of CSG relative to boys. Qualitatively this result is similar to the result reported in DSD, SASSA and UNICEF (2012) in that significant effect is found only for girls.¹⁶

Table 5. Results disaggregated by gender.

	Boys		Girls
Dependent variable height-for-age score	(1)	(2)	(3)	(4)
CSG	−0.03	0.03	0.27***	0.29***
	(0.27)	(0.25)	(0.11)	(0.11)
		−0.44**		−0.38
Low birthweight		(0.21)		(0.35)
Constant	−0.67	−0.84	−1.43***	−1.42***
	(0.57)	(0.56)	(0.18)	(0.31)
Other covariates	(included)	(included)	(included)	(included)
Observations	1,436	1,436	1,401	1,401

3.2. Effect of CSG at the quantiles

One implication of the results in Tables 4 and 5 is that the effect of CSG could be heterogeneous across the height-for-age distribution. This is because, if the effect of an adverse factor like low birthweight persists longer for boys, then boys should have a relatively lower placement in the height-for-age distribution. One can get an indication of how gender is distributed across the outcome distribution by checking the proportion of the gender variable in the quintiles of the height-for-age distribution. The proportion of boys in the quantiles of the outcome distribution in the treatment group are 55%, 48%, 53%, 50% and 45% while the figures for the control are 60%, 52%,49%, 49% and 42%. The common trend here is that the proportion of boys reduces as one approaches the upper tail of the outcome distribution. This suggests that under the RIF unconditional quantile regression¹⁷, the effect of CSG should be stronger at higher quantiles (where there are more girls) as suggested by results in Tables 4 and 5. Figure 1 shows the unconditional quantile treatment effects using RIF regressions.

Figure 1. Unconditional quantile effect of CSG.

It is evident that children at the high quantiles (i.e. >80th quantile)¹⁸ of the height-for-age distribution drive the positive effect observed at the mean. A significant positive effect of about 1 standard deviation exists at the highest quantiles, while there is no significant positive effect for children below the 75th quantile of the height-for-age distribution. Given that the proportion of boys tend to decrease towards higher quantiles, the quantile result (in Figure 1) suggests that the positive effect is being driven by the girls at higher quantiles. This confirms the possibility of gender disparity in the effect of the programme.

The next set of analyses disaggregates the quantile results by gender. The results are presented in Figures 2 and 3 and suggest that the effect is heterogeneous across gender groups.

Figure 2. Unconditional quantile regression for boys.

Figure 3. Unconditional quantile regression for girls.

For boys, there is a significant effect only at the top quantile (i.e. 95^th quantile), and a negative significant effect for boys between the 60^th and the 80th quantile, while for boys below the 60^th quantile, there is no significant effect. This explains the small positive effect found at the mean for boys (in Table 5). For girls, there is a significant positive effect above the 65^th quantile (with the exception of the effect at the 80th quantile which is not significant), between the 20^th quantile and the 65^th quantile, the effect is not significant (with the exception of the 50^th and 55^th quantile where the effect is significant). Lastly, there is a significant negative effect below the 20^th quantile for girls. It is clear that the region of a positive and significant effect is larger for girls relative to boys. More important is the fact that across gender groups only children at the top quantiles have positive and significant benefit from receiving the CSG.

One can argue that policymakers will be more interested in making a difference for children at the lower quantiles since they are more likely to be stunted. Attention is drawn to the fact that children at the 25^th quantile of the unconditional outcome distribution have a height-for-age score of about -2 standard deviations.¹⁹ This means that children below the 25^th quantile of the unconditional outcome distribution are actually stunted and receipt of CSG does not appear to improve this condition. The inequality in the effect of CSG on height-for-age, hence, tends to favour girls. This result is not strange since gender inequality in height-for-age has been documented by Zere & McIntyre (2003) for South Africa and Wamani et al. (2007) for sub-Saharan Africa. The point here is that one would have expected CSG to address this inequality.

3.3. Relationship between birthweight and height-for-age

In this section, the correlation between low birthweight and height-for-age score is further investigated. Following WHO’s standard definition of low birthweight (i.e. a child is born with low birthweight if the weight of the child at birth is 2.5 kg or less (Wardlaw, 2004)). To do this, low birthweight is used as the treatment indicator and the distribution of the other covariates (including CSG treatment) is balanced using entropy weights. The balance statistic after entropy balancing is reported in the appendix. The results of the weighted mean regression is presented in Table 6 and shows that on average, being born with birthweight below 2.5 kg is associated with a reduction in height-for-age of about 50% of a standard deviation.

Table 6. Effect of low birthweight.

Dependent variable height-for-age score	(1)
Low birthweight	−0.49***
	(0.19)
Constant	1.50***
	(0.37)
Other covariates	(included)
Observations	2,837
R-squared	0.08

Standard errors in parentheses *** p < 0.01, ** p < 0.05, * p < 0.1.

This is substantial since the highest impact observed for the CSG in this study is only about 1 standard deviation and this is for children at the top quantiles (see Figures 1–3). Figure 4 shows the quantile results for low birthweight. It is clear from the result that low birthweight is associated with lower height-for-age. The only exception is at the top of the height-for-age distribution where the relationship is not significant.

Figure 4. Quantile effect of low birthweight.

This, again, supports earlier results in that it shows that the negative effect of low birthweight is weakest at the upper quantiles which means that, all things being equal, the effect of CSG should be strongest in this region, and this is exactly what the quantile effect shows in Figures 1–3.

These results imply that placement of individuals in the height-for-age distribution matters when it comes to the effect of the CSG. The male disadvantage hypothesis suggests that boys are more vulnerable to environmental and health factors that impact height-for-age, and implies that boys are more likely to be born with low birthweight or affected by adverse conditions in general. The dataset used suggests that even in the treatment group, it takes 2 years, on average, before a child starts receiving the grant. This misses the most important period of the child’s life where significant gains can be made in terms of catch-up growth. The implication is that these factors will pre-determine the placement of children in the height-for-age distribution. This, then, explains the heterogeneity in the effect of CSG since there are differences in the characteristics of children at different parts of the outcome distribution. This helps qualify existing results – it is not the case that CSG is ineffective when it comes to stunting, it is rather the fact that it is more effective for children with better health as measured by height-for-age.

For example, the current study found that the effect of low birthweight is stronger for boys. This means that placement in the height-for-age distribution will be such that boys who experience low birthweight will find themselves in the lower part of the distribution. This scenario clearly predicts that the effect of CSG on height-for-age must be heterogeneous because benefiting children are not starting at the same base. Children at higher quantiles are likely to be more responsive to the positive impact of the programme than those at the lower quantiles. This means a significant or insignificant mean effect misses significant details/information about who is actually benefiting from the programme. This is the main contribution of this paper.

3.4. Robustness of results

In Appendix 2 & 3, some robustness checks for the results are presented. Appendix 2 check the sensitivity of the result to the model used to estimate caregiver motivation. To do this, the caregiver relationship dummy is excluded from the censored regression equation. Note that this means the new motivation variable is similar to the one used by Coetzee (2011, 2013) and Aguero et al. (2006). The results show that while the point estimates are slightly different for the mean analysis, the estimates are qualitatively similar. Furthermore, the quantile results show a similar pattern to the one presented in the text. Specifically, children at the upper deciles experience significant impact, and the region of significant impact is larger for girls relative to boys. Lastly, low birthweight is associated with lower height-for-age score.

Appendix 3 presents the results when the cut-off for low birthweight is changed to 2.3 kg. Again, the result is very similar to the one presented so far, leading to the conclusion that the results presented in this study are fairly robust to the specification.

4. Conclusion

This paper explores the impact of the CSG beyond the mean, to uncover the dynamics of its effect on the quantiles of the height-for-age distribution. To do this, the unconditional quantile effect of the CSG is estimated after balancing the first two moments of the relevant covariates. Result shows that in the combined population of boys and girls, children above the 80th quantile of the height-for-age distribution drive the positive effect observed at the mean. For children below this threshold, CSG has no significant effect. The results also show that the positive impact is stronger for girls relative to boys. These results support the male disadvantage hypothesis.

Lastly, the study demonstrates that low birthweight is a significant correlate of height-for-age in the NIDS wave 1 data, and the effect of this factor differs at different quantiles of the outcome distribution. This indicates that the male disadvantage hypothesis, coupled with the delay in receipt of the CSG, means that the heterogeneous effect observed should not be a surprise.

The implication of these results is that, while it is known that the effect of low birthweight continues to manifest in the growth of children, it seems this effect is stronger for male children. Therefore, reducing the prevalence of low birthweight (in general) for the beneficiaries of the CSG may unlock more benefit for the grant in terms of the height-for-age score.

Policy proposals that can mitigate the effect of low birthweight include providing support for women during pregnancy (Chersich et al., 2016) or making an effort to reduce the delay in receipt. While the current practice of providing support only when the child is born is commendable, the data and results suggests that this limits the effectiveness of the support through delay. Especially since the damage done by maternal deprivation during pregnancy and early childhood, may be hard to reverse. Hence, extending CSG to cover pregnancy period may, to a large extent, help address the problem of delay. While this may be costly, the gains may justify such a policy move. As it stands the government has increased the age limit for children that can benefit from 6 to 18 years over the last two decades. While this is justified because CSG affects other areas of a child’s life, it is arguably more costly than providing support for pregnant women.

However, apart from the cost, a pregnancy grant may be subject to perverse behaviours.²⁰ For example, confirming pregnancy and dealing with cases of abortion can make such policy move complicated and ineffective. Perhaps targeting vulnerable children at the point of birth (i.e. at hospitals) is the best approach to mitigate the effect of delay.

One limitation of this research is that the results are based on the conditional independence assumption which means that the result might be sensitive to unobserved attributes. However, the paper still makes an important contribution for two reasons: First, the estimation of caregiver motivation does control for any unobserved attribute that is correlated with delay. Second, these results help qualify existing results in the literature that are based on the same assumption and highlight why it is important to look beyond the effect of the programme on the average beneficiary.

Acknowledgement

The author is grateful for the comments received from the anonymous referee at Economic Research Southern Africa (ERSA) and for their useful suggestions. This paper has a working paper version that was published in ERSA working paper series.

Disclosure statement

No potential conflict of interest was reported by the author(s).

Additional information

Funding

The author is grateful for the comments received from the anonymous referee at Economic Research Southern Africa (ERSA) and for their useful suggestions.

Notes

1 The primary caregiver is the person that takes care of the needs of the child without payment.

2 Note that one could argue that motivation is time invariant and make an argument for a fixed effect model. First, this will put considerable difference between the methodology adopted in this study and the existing results it wishes to shed light on. Second, this will mean that the current study cannot make sense of the gender and birth weight coefficients since they are also time invariant. Lastly this will shrink the sample size significantly in latter waves of the data because year on year the number of children that benefit from the grant increases, meaning that control units become more and more scares.

3 Note that there is no need to check for balance after running the entropy weight algorithm because it is designed to automatically achieve balance.

4 Calculated as delay minus mean of delay over standard deviation of delay.

5 Note that while unobserved factors are not directly controlled for, unobserved characteristics related to delay (in receiving the CSG) may be captured by the motivation variable.

6 At the time of the NIDS 2008 survey a caregiver (who does not have to be a family member of the child) must have a monthly income below R800 in urban areas or R1 100 in rural areas (see Coetzee 2011).

7 An exception here is the income and expenditure variables. These variables were imputed by NIDS.

8 A dummy variable is used for this. The variable is equal to 1 if the child never visited the hospital in the last 12 months and 0 otherwise.

9 Food expenditure per adult equivalent scale (AES) = (Food expenditure/(Adult + 0.5 * children)^0.9).

10 An indicator for low birthweight is added in subsequent analysis to isolate its relationship with height-for-age after accounting for the effect of the treatment and other covariates. This provides an initial idea of how this variable is correlated with height-for-age.

11 The algorithm did not converge when an attempt was made to balance the skewness. This is, however, not a significant source of bias because, even without balancing the skewness, only two variables show differences in skewness (food expenditure and motivation). Since this variable will be included in the subsequent weighted regressions, this will help to mitigate the effect of problematic imbalance in skewness (if any).

12 Sometimes height-for-age result is reported in terms of the percentage of a standard deviation it represents since height-for-age is measured by z-scores see (Coetzee, 2011) for example.

13 A child is born with low birthweight if the weight of the child at birth is below 2.5 kg according to the World Health Organization’s definition.

14 These papers report a range of estimates under different tweaks to their method, the one reported here is the maximum of the reported values.

15 Note that Aguero et al. (2006) use the KwaZulu-Natal Income Dynamics Study while Coetzee (2011, 2013) use the National Income Dynamic Study.

16 The size of the estimate in DSD, SASSA and UNICEF (2012) is 19% of a standard deviation (0.19). The difference in size of the effect compared with the 27% of a standard deviation reported for girls in Table 5 can be attributed to difference in the samples and methodology used. The effect reported for boys in in DSD, SASSA and UNICEF (2012) is 0.007 (i.e. similar to the result in the current study the disparity in the estimates is large).

17 i.e. the quantiles are defined before the regression so that the quantile effects do not depend on the covariates and represent the difference between the marginal distribution of outcomes across treatment arms.

18 The treatment effect is also estimated using conditional quantile estimator. The result is presented in Figure A1 of the Appendix. It shows that the effect is not significant anywhere on the outcome distribution, although the effect is stronger at the higher quantiles. This is attributable to the difference between the conditional and unconditional quantile regression.

19 Stunting is defined as height-for-age score less than −2 standard deviations of World Health Organization’s reference standard.

20 The author is grateful to one of the reviewers for raising this point.

Appendices

Appendix 1

Table A1. Summary statistics when low birthweight is the treatment variable.

Entropy weighting
	Treat			Control
	Mean	Variance	Skewness	Mean	Variance	Skewness
Treatment	0.70	0.21	−0.85	0.69	0.21	−0.84
Motivation	0.11	0.65	0.76	0.11	0.65	0.76
Employed	0.24	0.18	1.25	0.24	0.18	1.23
Married	0.40	0.24	0.40	0.40	0.24	0.39
Primary caregivers’ education (<5years)	0.27	0.20	1.01	0.28	0.20	1.00
Primary caregivers’ education (between 5 and 10 years)	0.39	0.24	0.44	0.39	0.24	0.43
Primary caregivers age (<40years)	0.70	0.21	−0.85	0.69	0.21	−0.84
Primary caregivers age (between 40 and 60 years)	0.21	0.17	1.46	0.21	0.16	1.44
HH head gender	0.43	0.25	0.28	0.43	0.25	0.27
Child’s gender	0.43	0.25	0.28	0.43	0.25	0.27
Electricity	0.70	0.21	−0.85	0.69	0.21	−0.84
Water	0.27	0.20	1.01	0.28	0.20	1.00
Telephone	0.04	0.04	4.75	0.04	0.04	4.73
Toilet	0.21	0.17	1.46	0.21	0.16	1.44
Food expenditure (AES)	187	15990	1.81	187	15967	1.82
Informal dwelling	0.11	0.10	2.53	0.11	0.10	2.51
Hospital visits	0.35	0.23	0.62	0.36	0.23	0.60

Figure A1. Quantile effect (Conditional quantile regression).

Appendix 2. Analysis excluding caregiver relationship from the estimation of motivation variable

As a robustness check, the analysis below excludes caregiver relationship from the estimation of the motivation variable. This led to a situation where the entropy balancing algorithm did not converge and converged only when the health seeking dummy was excluded from the model. Health seeking behaviour is however controlled for in the weighted regressions.

Table A2. Effect of CSG with and without low birthweight when the estimation of motivation excludes the caregiver relationship.

Dependent variable height-for-age score	(1)	(2)
Treatment	0.22	0.21
	(0.18)	(0.18)
Low birthweight		−0.42**
		(0.21)
Constant	−1.20***	−1.23***
	(0.37)	(0.37)
Other covariates	(included)	(included)
Observations	2,837	2,837
R-squared	0.08	0.08

Standard errors in parentheses *** p < 0.01, ** p < 0.05, * p < 0.1.

By other covariates I mean the covariates listed in Table 2.

Figure A2. Unconditional quantile effect of CSG when the estimation of motivation excludes the caregiver relationship.

Table A3. Results disaggregated by gender when the estimation of motivation excludes the caregiver relationship.

	Boys		Girls
Dependent variable height-for-age score	(1)	(2)	(3)	(4)
CSG	−0.08	−0.14	0.68***	0.72***
	(0.24)	(0.23)	(0.19)	(0.20)
		−0.73***		−0.41
Low birthweight		(0.25)		(0.36)
Constant	−0.42	−0.31	−1.36***	−1.45***
	(0.58)	(0.53)	(0.27)	(0.25)
Other covariates	(included)	(included)	(included)	(included)
Observations	1,436	1,436	1,401	1,401

Figure A3. Unconditional quantile regression for boys when the estimation of motivation excludes the caregiver relationship.

Figure A4. Unconditional quantile regression for girls when the estimation of motivation excludes the caregiver relationship.

Table A4. Effect of low birthweight when the estimation of motivation excludes the caregiver relationship.

Dependent variable height-for-age score	(1)
Low birthweight	−0.49**
	(0.20)
Constant	−1.57***
	(0.38)
Other covariates	(included)
Observations	2,837
R-squared	0.08

Standard errors in parentheses *** p < 0.01, ** p < 0.05, * p < 0.1.

Figure A5. Quantile effect of low birthweight when the estimation of motivation excludes the caregiver relationship

Appendix 3. Analysis for different low birthweight cut-off is 2.3 kg

Table A5. Effect of CSG with and without low birthweight (birthweight<2.3 kg).

Dependent variable height-for-age score	(1)	(2)
Treatment	0.14	0.22
	(0.18)	(0.19)
Low birthweight		−0.58***
		(0.22)
Constant	−0.21	−0.39
	(0.43)	(0.36)
Other covariates	(included)	(included)
Observations	2,837	2,837
R-squared	0.08	0.11

Standard errors in parentheses *** p < 0.01, ** p < 0.05, * p < 0.1.

By other covariates I mean the covariates listed in Table 2.

Figure A6. Unconditional quantile effect of CSG (birthweight<2.3 kg).

Table A6. Results disaggregated by gender (birthweight<2.3 kg).

	Boys		Girls
Dependent variable height-for-age score	(1)	(2)	(3)	(4)
CSG	−0.04	0.03	0.27**	0.29***
	(0.27)	(0.25)	(0.11)	(0.11)
		−0.44*		−0.61
Low birthweight		(0.24)		(0.46)
Constant	−0.65	−0.84	−1.43***	−1.43***
	(0.56)	(0.55)	(0.22)	(0.23)
Other covariates	(included)	(included)	(included)	(included)
Observations	1,436	1,436	1,401	1,401

Standard errors in parentheses *** p < 0.01, ** p < 0.05, * p < 0.1.

Figure A7. Unconditional quantile regression for boys (birthweight<2.3 kg).

Figure A8. Unconditional quantile regression for girls (birthweight<2.3 kg).

Table A7. Effect of low birthweight (birthweight<2.3 kg)

Dependent variable height-for-age score	(1)
Low birthweight	−0.41*
	(0.24)
Constant	−1.34***
	(0.46)
Other covariates	(included)
Observations	2,837
R-squared	0.08

Standard errors in parentheses *** p < 0.01, ** p < 0.05, * p < 0.1.

Figure A9. Quantile effect of low birthweight when the estimation of motivation excludes the caregiver relationship.