Do Cash Transfers Promote Food Security? The Case of the South African Child Support Grant

Do Cash Transfers Promote Food Security? The Case of the South African Child Support Grant Abstract This paper evaluates the causal effect of the Child Support Grant (CSG) implemented in South Africa on household food consumption and dietary diversity. The analysis uses the National Income Dynamics Study (NIDS) covering 2008, 2010–2011, and 2012, and carries out a regression-discontinuity design exploiting the increase in the age limit requirement for eligibility for the programme. Our results show that the CSG has proved to be effective in increasing total food expenditure per adult equivalent but has not significantly changed the dietary habits of the beneficiary households, nor has the programme resulted in any stronger effect for the most vulnerable subgroups of the beneficiary population. To analyse the external and internal validities of the results, a comparison between non-parametric, semi-parametric and parametric estimates is presented. 1. Introduction A number of studies have recently evaluated the impact of cash transfers (CTs) targeted to the poorest and most vulnerable people in sub-Saharan African countries, and a positive role in promoting human capital, health and productive activities has been documented (World Bank 2015). CTs may also affect food security because, by addressing the lack of purchasing power of poor households, they contribute to improving food purchases and access to more good-quality food (Bassett, 2008, Alderman, 2014, Burchi et al., 2016). However, the evidence for the impact of CTs on food security in sub-Saharan Africa is not conclusive since, even if most of the literature shows a positive impact on the consumption level of food, it is still not clear whether the consumption gains are translated into improved nutritional status for household members (Manley et al., 2013, Alderman, 2014, Slater et al., 2014, Burchi et al., 2016, Davis and Handa, 2016). In this paper, we contribute to this literature by evaluating the impact on food security of the Child Support Grant (CSG), a CT programme introduced in South Africa in 1998 with the aim of supporting vulnerable children and poor households. At the demise of the apartheid regime in 1994, South Africa underwent significant political and social advances and rapid economic growth, both of which have led to steady progress in reducing poverty (Agüero et al., 2007, Leibbrandt et al., 2011, Leibbrandt and Levinsohn, 2011). However, food insecurity remains widespread: 35% of the population is vulnerable (Kirsten, 2012) and about 25% of children under age of 6 are classified as stunted by malnutrition (Manyamba et al., 2012). Food insecurity in South Africa is not due to a shortage of food, but rather to insufficient access as a result of structural poverty and inequality dynamics with a strong racial footprint (Aliber, 2001, Du Toit, 2011, Manyamba et al., 2012). South Africa has no CT directly aimed at reducing food insecurity. In this context, CT programmes aimed at eradicating extreme poverty1 play a central role in addressing food insecurity and represent one of the pillars of the Integrated Food Security Strategy established in 1996 (South African Department of Agriculture, 2002). Although the CSG is the major CT programme in South Africa, its impact on food security has only been analysed in a few studies. Using data from the General Household Survey and the Labour Force Survey, Williams (2007) found that CSG beneficiaries aged between 7 and 17 years more probably attended school and suffered less hunger. Samson et al. (2008) showed a reduction in hunger among children under 7 years of age, using a household panel extracted by the Economic Policy Research Institute (EPRI) from repeated cross-sections of the National General Household Surveys. Agüero et al. (2010), using data from the KwaZulu-Natal Income Dynamic Study, evaluated the impact of the CSG on anthropometric indicators which are widely used in nutritional assessment. Focusing on children in the first 36 months of life, the study found that the CSG improved childhood nutrition as measured by child height-for-age. Heinrich et al. (2012) evaluated the impact of the CSG on the well-being of children in early life. Considering aspects related to food security, the study showed that the CSG significantly improved nutrition, as measured by the height for age z-scores, in the cases of children whose mothers have more than 8 grades of schooling. Using data extracted from the first wave of a longitudinal survey provided by the National Income Dynamics Study (NIDS), Coetzee (2013) evaluated the impact of the CSG on health, height-for-age and education of beneficiary children under 14 and she found a few positive effects on their well-being. Similarly, she found a positive, albeit small, impact of the programme on household food security measured by the per capita expenditure on food items. Unlike previous studies, Delany et al. (2008) considered the impact of the programme at the households level. Analysing an original survey of households in low-income areas and focus group discussions, Delany et al. (2008) found that beneficiary households allocated a larger proportion of their expenditure to essential goods, such as food, and that more than three-quarters of the CSG recipients stated that food was the main expenditure covered by the grant. Overall, these studies indicate that the CSG has positive effects on outcome variables connected to the food security of beneficiary children and that the programme also slightly increased households’ food expenditure. With respect to the previous literature, our study provides an evaluation of the impact of the CSG which is focused on food expenditure at the household level and extends the analysis by considering different drivers of food security. We explore whether participation in the CSG is effective in improving the total food expenditure per adult equivalent as well as dietary diversity for the beneficiary families who have a child that was eligible up to the age of 18.2 To analyse dietary diversity, which gives us information about the quality of the food consumed, we use the share in food expenditure of the primary food groups (e.g. carbohydrates, dairy products, proteins and vitamins) and a synthetic index defined as the number of different food groups consumed in that period (Hoddinott and Yohannes, 2002, Ruel, 2003, Hoddinott and Bassett, 2008). By considering these measures jointly, we provide a more comprehensive investigation of the impact of the CSG on food security in South Africa and, at least partially, we fill a gap in the literature. Within this framework, we exploit the discontinuous variation induced by the expansion in eligibility in terms of the child’s age. We take into consideration the fact that from 1 January 2010, eligibility for the CSG was extended for children born on or after 1 January 1994 until their 18th birthday; those born before 1994 had lost their eligibility at age 14. As a result of this policy change, a discontinuous increase occurred in the probability of being a CSG beneficiary during the age interval 14–17 for children born after 1 January 1994. This discontinuity provides a natural experiment for examining the causal effects of the programme across birth cohorts, suitable for running a Regression-Discontinuity Design (RDD). Using this identification structure, we estimate the effect of the CSG by using the local-polynomial (LP) estimator (Calonico et al., 2014a) and then compare the results with the parametric estimates obtained with an instrumental variable (IV) approach. To overcome problems related to potential selection bias and threats to internal validity, we apply, following van der Klaauw (2002), a two-step propensity score (PS) procedure and placebo experiments which use the panel structure of the data to replicate the analysis only for the 2008 wave.3 We also employ a formal test strategy implemented by Cerulli et al. (2017) to address other important threats to the stability of the compliers (i.e., the eligible people who have received the treatment after the policy change) in the neighbourhood of the cutoff and to the external validity of the analysis. Finally, as a further robustness check, we ascertain that the causal estimates are robust within population subgroups and use elasticity measures based on the two-step propensity score to compare the potential outcome of the subgroups with that of the entire population and to account for the differences between them (Angrist, 2004). Our results show that the monetary transfers provided by the CSG led to an increase in total food expenditure per adult equivalent, with an impact which ranges between R54 and R55 (corresponding to a mean increase of 10%). This result was confirmed by the comparison between non-parametric, semi-parametric, and parametric estimations, and thus is very robust. Our analysis also shows that the estimated parameters are constant away from the cutoff and that the results are homogeneous across various population subgroups. We may thus interpret the causal effects of the CSG unconditionally to the birth cohorts which are not close to the cutoff. In addition, the analysis shows that the CSG did not have a greater effect on the population subgroups which were more vulnerable to food insecurity than it did on the others. When dietary diversity is analysed, we find robust positive results only for carbohydrates, which represent the major food group in the total food expenditure of the treated households (with a share of 23%). However, we do not find any significant result when the synthetic index of dietary diversity is considered. This analysis indicates that the CSG did not lead to significant changes in the dietary habits of the beneficiary households and did not improve the dietary diversity of the poor households. While there are caveats pertaining to our outcome variables and estimates, as usual in a non-experimental approach, we think that our work provides additional and valuable information about the impact of the CSG on several dimensions of the food security of the beneficiary households. This paper proceeds as follows: Section 2 describes the dataset and the identification strategy; Section 3 provides an overview of the empirical framework used to estimate the causal impact of the CSG on food expenditure and dietary diversity; Section 4 discusses the results, and Section 5 draws some conclusions and outlines avenues for future research 2. Identification strategy and dataset 2.1 Identification strategy CSG benefits are provided each month to eligible beneficiaries4 and are paid to the primary caregiver.5 As mentioned in Section 1, the eligible population is determined according to a means test and the child’s age, and these criteria have changed over time. Figure 1 shows the timeline of policy changes in the eligible population and in the amount of the grant, from 1998, when the programme was introduced, to the last year of our evaluation (2012). Figure 1: View largeDownload slide Timeline of CSG Implementation. Figure 1: View largeDownload slide Timeline of CSG Implementation. In the case of the income criteria, the means test was initially based on household income, and the ceiling was fixed at the nominal level of R800 in urban areas and R1,100 in rural areas for 10 years. However, in 1999, to increase take-up rates, the government altered this rule to one which considered only the income of the primary caregiver plus her/his spouse (Agüero et al., 2010, Woolard et al., 2011). Eligibility was expanded again in 2008: the Department of Social Development defined the income ceiling as 10 times the value of the grant paid to the single primary caregiver of the child (double for married caregivers), so that the means test would automatically keep pace with inflation (Agüero et al., 2010, Woolard et al., 2011). Figure 1 also shows changes in age limit criteria. When the programme was introduced in 1998, age eligibility was limited to children under 7 years old, but it was later gradually raised: in 2003, it was extended to children up to their 9th birthday, in 2004 up to their 11th, and in 2005 up to their 14th. From 1 January 2010, eligibility was further extended, so that children born after 1 January 1994 were eligible until their 18th birthday, whereas those born before that date lost eligibility at 14 (van der Berg et al., 2010, Woolard et al., 2011, Bor 2013). A number of problems in the implementation of the CSG and of policy changes have been well documented (Heinrich et al., 2012, Heinrich and Brill 2015, Samson et al., 2016, Heinrich et al., 2017) and we are aware that, for this reason, serious threats to the internal and external validity arise in our identification strategy. In particular, programme rules and administrative capacity affected the access and the duration of the reception of the benefits (Heinrich and Brill 2015, Heinrich et al., 2017), influencing the effectiveness of the programme. In more detail, Heinrich et al. (2012) stressed the barriers to application due to misinformation about the eligibility criteria and difficulties in ensuring that awareness of the changes in the targeting rules were disseminated to populations that are often disconnected from reliable sources of information.6 Problems with getting the correct documentation were another frequent cause of delay in applications or deterrence from would-be recipients. These and related factors associated with barriers to access for eligible families have been reduced over time but they are still present, and challenge the assumption that all eligible children are enroled in the programme once the policy change is implemented. In addition, administrative burdens may vary geographically or politically, and thus the influence on the take-up rates may be uneven. Following this outline, three major concerns arise in our analysis. First of all, not all households with a child in the eligible age range receive the CSG, since they must also meet an eligibility requirement based on an income means test, and this criterion has changed over time. A selection bias may therefore emerge if we do not take into account how the participation in the programme varies with changes in the income requirement. Secondly, administrative costs, bureaucratic requirements, and implementation problems have limited the participation in the CSG (Heinrich et al., 2012, Heinrich and Brill, 2015, Samson et al., 2016, Heinrich et al., 2017). These issues may violate both the the internal and external validity assumptions in the RD design and may imply instability in the population of compliers in the neighbourhood of the cutoff. Lastly, significant differences may emerge in this analysis because the CSG was designed to focus on poor households which had been excluded from social assistance programmes during apartheid (Pauw and Mncube, 2007), particularly Africans, and those living in marginalised rural areas (Lund et al., 2008). Note that these groups are particularly vulnerable to the problems related to the design and implementations of the CSG. Thus, we expected that the potential outcome would not be homogeneous in these subgroups with respect to the entire population (Angrist, 2004). 2.2 Dataset The NIDS is a dataset implemented by the South African Labour and Development Research Unit (SALDRU) of the University of Cape Town, and is the first nation-wide panel study in South Africa.7 The NIDS is available for the waves 2008, 2010–2011, and 2012, and allows a face-to-face longitudinal survey of households residing in South Africa. Its aim is to follow a sample of household members and register the changes in household compositions, migration, and several dimensions of well-being (e.g., income, expenditure, assets, access to social services, education, health, employment). From the entire dataset, we extracted a sample of households that responded to the three waves of the survey and whose composition remained unchanged during this period, to ensure the absence of migration between different provinces and areas. In the second step, households with no children or with children who were not born in 1990–1998 were dropped. This procedure was followed because the evaluation required a sample of households with children in the age range 14–17 who experienced the policy change. Since our panel data covers the period 2008–2012, we needed to include children born after 1990, and to remove those born from 1998 onwards, who were always eligible for the CSG. Using this framework, we obtained a sample of households with one child in the age range from 10 (in 2008) to 22 (in 2012). This sample consists of 1,336 households for a total of 4,011 observations.8 The dataset provides expenditure data at the household level and includes both the total food expenditure and the food expenditure disaggregated by food group. We used these variables according to our interest in analysing aspects of food security related to economic access to a sufficient quantity of food (Coetzee, 2013, Burchi et al., 2016) and to the the quality of the food consumed (Hoddinott and Yohannes, 2002, Ruel, 2003, Hoddinott and Bassett, 2008). In more detail, the first variable of interest is monthly food expenditure at the household level, adjusted by an adult equivalent scale.9 The variables related to food groups were obtained by disaggregating the total food expenditure per adult equivalent into four groups: (i) carbohydrates, given by the sum of the expenditure on cereals such as samp, flour and bread, mealie meal, rice and pasta; (ii) dairy products; (iii) proteins, given by the sum of expenditure on meat and fish; (iv) vitamins, comprising expenditure on fruits and vegetables. The analysis of the food groups was carried out for 2008–2011, since the NIDS stopped reporting such data after 2011.10 Taking into account that there is no consensus on which food group and classification system is the most adequate (Ruel, 2003), we have also considered a synthetic indicator of dietary diversity, obtained by following the recommendations of the FAO’s ‘Guidelines for measuring household and individual dietary diversity’ (Kennedy et al., 2011, Vorster et al., 2013). First of all, we defined 12 food groups: (i) cereals; (ii) roots and white tubers; (iii) vegetables; (iv) fruits; (v) meat; (vi) eggs; (vii) fish and other seafood; (viii) legumes, nuts and seeds; (ix) milk and milk products; (x) fats and oils; (xi) sweets; (xii) spices, condiments and beverages. Secondly, we generated 12 dummy variables that register when a household consumed, in the last 30 days, at least one item in the corresponding food group. Lastly, we summed these dummy variables to obtain a score that ranges from 1 to 12, where 1 indicates households consuming only one food group, and 12 households consuming all the available food groups.11 Table 1 presents the means and standard deviations (s.d.) for the outcome variables in question, separated into control and treated groups. As an initial outcome, the first row of the table shows that considerable differences arise when the total food expenditure per adult equivalent is considered: the mean in the treated group is about R394, compared with about R556 in the control group. The standard deviation is also much higher in the control group. Table 1: Total food expenditure per adult equivalent and shares of food groups Full-sample Control Treatment Mean s.d. Mean s.d. Mean s.d. Total food expenditure per adult equivalent 500.460 497.196 556.419 532.096 394.815 403.085 Share of food groups in total food expenditure:  Carbohydrates 19.401 18.532 17.476 17.366 23.035 20.065  Dairy products 4.203 5.123 4.257 5.119 4.101 5.130  Proteins 15.939 15.911 16.384 16.480 15.099 14.747  Vitamins 5.857 6.386 5.828 6.342 5.912 6.471  Dietary diversity 9.777 2.244 9.954 2.223 9.264 2.228 Full-sample Control Treatment Mean s.d. Mean s.d. Mean s.d. Total food expenditure per adult equivalent 500.460 497.196 556.419 532.096 394.815 403.085 Share of food groups in total food expenditure:  Carbohydrates 19.401 18.532 17.476 17.366 23.035 20.065  Dairy products 4.203 5.123 4.257 5.119 4.101 5.130  Proteins 15.939 15.911 16.384 16.480 15.099 14.747  Vitamins 5.857 6.386 5.828 6.342 5.912 6.471  Dietary diversity 9.777 2.244 9.954 2.223 9.264 2.228 Notes: Total food expenditure per adult equivalent is monthly expenditure at household level, adjusted by an adult equivalent scale. Analysis of food groups and of synthetic index of dietary diversity is valid for period 2008–2011, as NIDS stopped reporting such data after 2011. View Large Table 1: Total food expenditure per adult equivalent and shares of food groups Full-sample Control Treatment Mean s.d. Mean s.d. Mean s.d. Total food expenditure per adult equivalent 500.460 497.196 556.419 532.096 394.815 403.085 Share of food groups in total food expenditure:  Carbohydrates 19.401 18.532 17.476 17.366 23.035 20.065  Dairy products 4.203 5.123 4.257 5.119 4.101 5.130  Proteins 15.939 15.911 16.384 16.480 15.099 14.747  Vitamins 5.857 6.386 5.828 6.342 5.912 6.471  Dietary diversity 9.777 2.244 9.954 2.223 9.264 2.228 Full-sample Control Treatment Mean s.d. Mean s.d. Mean s.d. Total food expenditure per adult equivalent 500.460 497.196 556.419 532.096 394.815 403.085 Share of food groups in total food expenditure:  Carbohydrates 19.401 18.532 17.476 17.366 23.035 20.065  Dairy products 4.203 5.123 4.257 5.119 4.101 5.130  Proteins 15.939 15.911 16.384 16.480 15.099 14.747  Vitamins 5.857 6.386 5.828 6.342 5.912 6.471  Dietary diversity 9.777 2.244 9.954 2.223 9.264 2.228 Notes: Total food expenditure per adult equivalent is monthly expenditure at household level, adjusted by an adult equivalent scale. Analysis of food groups and of synthetic index of dietary diversity is valid for period 2008–2011, as NIDS stopped reporting such data after 2011. View Large In addition, when the four shares of food expenses in the total food expenditure per adult equivalent are compared, Table 1 shows that the largest difference between treated and control groups is in the mean value of the share of carbohydrates, which is about 6% higher in the treated group. As has been shown in the literature, the poorest people in South Africa have little dietary variety, and consume more cereals and fewer items in the other food groups, than does the rest of the population.12 However, the last row of the table shows that there are small differences in terms of dietary diversity, when the treated and control groups are analysed. Table 2 lists household characteristics in the entire sample (for both 2008–2012 and 2008–2011) and both control and treated groups. In the sample for 2008–2012, the first remarkable difference between the groups is due to the share of households living under the food poverty line,13 defined as the threshold under which individuals cannot purchase sufficient food to provide them with an adequate diet. This share is 44% of treated units and only 32% of controls. Other significant differences are found when geographical location, race, gender and employment status are considered. In more detail, we find that a higher percentage of households in the treated group live in rural areas (57%), and have household heads who are Africans (86%), women (69%), or receiving no education (26.5%), compared with the control group. Table 2: Descriptive statistics Sample 2008–2012 Sample 2008–2011 Sample Control Treated Difference Sample Control Treated Difference Province   Western Cape 0.146 0.175 0.088 0.087* 0.144 0.173 0.088 0.085*   Eastern Cape 0.134 0.124 0.154 −0.030 0.135 0.126 0.154 −0.028   Northern Cape 0.064 0.070 0.053 0.017 0.063 0.069 0.053 0.016   Free State 0.084 0.093 0.067 0.026 0.085 0.094 0.067 0.027   KwaZulu-Natal 0.235 0.197 0.310 −0.113* 0.237 0.198 0.311 −0.113*   North West 0.051 0.043 0.066 −0.023* 0.051 0.044 0.066 −0.022*   Gauteng 0.117 0.140 0.074 0.066* 0.114 0.135 0.075 0.060*   Mpumalanga 0.070 0.065 0.081 −0.016 0.070 0.065 0.078 −0.013   Limpopo 0.098 0.094 0.107 −0.013 0.100 0.096 0.108 −0.012 Poverty   Food poverty line 0.362 0.321 0.442 −0.121* 0.293 0.255 0.368 −0.113*   Poverty line 0.329 0.309 0.369 −0.060* 0.338 0.307 0.399 −0.092* Geographical location   Rural 0.442 0.374 0.573 −0.199* 0.450 0.383 0.580 −0.197*   Urban 0.558 0.626 0.427 0.199* 0.550 0.617 0.420 0.197* Number of household members   1 or 2 0.135 0.147 0.113 0.034* 0.124 0.133 0.105 0.028*   3 0.227 0.211 0.258 −0.047* 0.235 0.215 0.274 −0.059*   4 0.212 0.213 0.211 0.002 0.216 0.215 0.218 −0.003   5 0.174 0.179 0.165 0.014 0.173 0.179 0.162 0.017   ≥6 0.252 0.250 0.254 −0.004 0.252 0.258 0.240 0.018 Head of household by race   African 0.768 0.719 0.862 −0.143* 0.768 0.719 0.862 −0.143*   Coloured 0.171 0.195 0.124 0.071* 0.171 0.195 0.124 0.071*   Asian/Indian/White 0.061 0.086 0.014 0.072* 0.061 0.085 0.014 0.071* Head of household by gender   Female 0.617 0.578 0.692 −0.114* 0.588 0.550 0.661 −0.111*   Male 0.383 0.422 0.308 0.114* 0.412 0.450 0.339 0.111* Head of household by education   No schooling 0.182 0.138 0.265 −0.127* 0.201 0.153 0.291 −0.138*   Primary 0.185 0.154 0.246 −0.092* 0.196 0.165 0.255 −0.090*   Secondary 0.515 0.542 0.464 0.078* 0.499 0.532 0.437 0.095*   Tertiary 0.118 0.167 0.025 0.142* 0.105 0.151 0.018 0.133* Head of household by age 49.085 48.166 50.869 −2.703 49.967 49.181 51.491 −2.310 Sample 2008–2012 Sample 2008–2011 Sample Control Treated Difference Sample Control Treated Difference Province   Western Cape 0.146 0.175 0.088 0.087* 0.144 0.173 0.088 0.085*   Eastern Cape 0.134 0.124 0.154 −0.030 0.135 0.126 0.154 −0.028   Northern Cape 0.064 0.070 0.053 0.017 0.063 0.069 0.053 0.016   Free State 0.084 0.093 0.067 0.026 0.085 0.094 0.067 0.027   KwaZulu-Natal 0.235 0.197 0.310 −0.113* 0.237 0.198 0.311 −0.113*   North West 0.051 0.043 0.066 −0.023* 0.051 0.044 0.066 −0.022*   Gauteng 0.117 0.140 0.074 0.066* 0.114 0.135 0.075 0.060*   Mpumalanga 0.070 0.065 0.081 −0.016 0.070 0.065 0.078 −0.013   Limpopo 0.098 0.094 0.107 −0.013 0.100 0.096 0.108 −0.012 Poverty   Food poverty line 0.362 0.321 0.442 −0.121* 0.293 0.255 0.368 −0.113*   Poverty line 0.329 0.309 0.369 −0.060* 0.338 0.307 0.399 −0.092* Geographical location   Rural 0.442 0.374 0.573 −0.199* 0.450 0.383 0.580 −0.197*   Urban 0.558 0.626 0.427 0.199* 0.550 0.617 0.420 0.197* Number of household members   1 or 2 0.135 0.147 0.113 0.034* 0.124 0.133 0.105 0.028*   3 0.227 0.211 0.258 −0.047* 0.235 0.215 0.274 −0.059*   4 0.212 0.213 0.211 0.002 0.216 0.215 0.218 −0.003   5 0.174 0.179 0.165 0.014 0.173 0.179 0.162 0.017   ≥6 0.252 0.250 0.254 −0.004 0.252 0.258 0.240 0.018 Head of household by race   African 0.768 0.719 0.862 −0.143* 0.768 0.719 0.862 −0.143*   Coloured 0.171 0.195 0.124 0.071* 0.171 0.195 0.124 0.071*   Asian/Indian/White 0.061 0.086 0.014 0.072* 0.061 0.085 0.014 0.071* Head of household by gender   Female 0.617 0.578 0.692 −0.114* 0.588 0.550 0.661 −0.111*   Male 0.383 0.422 0.308 0.114* 0.412 0.450 0.339 0.111* Head of household by education   No schooling 0.182 0.138 0.265 −0.127* 0.201 0.153 0.291 −0.138*   Primary 0.185 0.154 0.246 −0.092* 0.196 0.165 0.255 −0.090*   Secondary 0.515 0.542 0.464 0.078* 0.499 0.532 0.437 0.095*   Tertiary 0.118 0.167 0.025 0.142* 0.105 0.151 0.018 0.133* Head of household by age 49.085 48.166 50.869 −2.703 49.967 49.181 51.491 −2.310 Notes: Descriptive statistics of district of residence of household and interview’s month are omitted. For definitions of food poverty line and poverty line, see Appendix A. Asterisk mark indicates when the difference between the treated and control samples is statistically significant at 5% level. View Large Table 2: Descriptive statistics Sample 2008–2012 Sample 2008–2011 Sample Control Treated Difference Sample Control Treated Difference Province   Western Cape 0.146 0.175 0.088 0.087* 0.144 0.173 0.088 0.085*   Eastern Cape 0.134 0.124 0.154 −0.030 0.135 0.126 0.154 −0.028   Northern Cape 0.064 0.070 0.053 0.017 0.063 0.069 0.053 0.016   Free State 0.084 0.093 0.067 0.026 0.085 0.094 0.067 0.027   KwaZulu-Natal 0.235 0.197 0.310 −0.113* 0.237 0.198 0.311 −0.113*   North West 0.051 0.043 0.066 −0.023* 0.051 0.044 0.066 −0.022*   Gauteng 0.117 0.140 0.074 0.066* 0.114 0.135 0.075 0.060*   Mpumalanga 0.070 0.065 0.081 −0.016 0.070 0.065 0.078 −0.013   Limpopo 0.098 0.094 0.107 −0.013 0.100 0.096 0.108 −0.012 Poverty   Food poverty line 0.362 0.321 0.442 −0.121* 0.293 0.255 0.368 −0.113*   Poverty line 0.329 0.309 0.369 −0.060* 0.338 0.307 0.399 −0.092* Geographical location   Rural 0.442 0.374 0.573 −0.199* 0.450 0.383 0.580 −0.197*   Urban 0.558 0.626 0.427 0.199* 0.550 0.617 0.420 0.197* Number of household members   1 or 2 0.135 0.147 0.113 0.034* 0.124 0.133 0.105 0.028*   3 0.227 0.211 0.258 −0.047* 0.235 0.215 0.274 −0.059*   4 0.212 0.213 0.211 0.002 0.216 0.215 0.218 −0.003   5 0.174 0.179 0.165 0.014 0.173 0.179 0.162 0.017   ≥6 0.252 0.250 0.254 −0.004 0.252 0.258 0.240 0.018 Head of household by race   African 0.768 0.719 0.862 −0.143* 0.768 0.719 0.862 −0.143*   Coloured 0.171 0.195 0.124 0.071* 0.171 0.195 0.124 0.071*   Asian/Indian/White 0.061 0.086 0.014 0.072* 0.061 0.085 0.014 0.071* Head of household by gender   Female 0.617 0.578 0.692 −0.114* 0.588 0.550 0.661 −0.111*   Male 0.383 0.422 0.308 0.114* 0.412 0.450 0.339 0.111* Head of household by education   No schooling 0.182 0.138 0.265 −0.127* 0.201 0.153 0.291 −0.138*   Primary 0.185 0.154 0.246 −0.092* 0.196 0.165 0.255 −0.090*   Secondary 0.515 0.542 0.464 0.078* 0.499 0.532 0.437 0.095*   Tertiary 0.118 0.167 0.025 0.142* 0.105 0.151 0.018 0.133* Head of household by age 49.085 48.166 50.869 −2.703 49.967 49.181 51.491 −2.310 Sample 2008–2012 Sample 2008–2011 Sample Control Treated Difference Sample Control Treated Difference Province   Western Cape 0.146 0.175 0.088 0.087* 0.144 0.173 0.088 0.085*   Eastern Cape 0.134 0.124 0.154 −0.030 0.135 0.126 0.154 −0.028   Northern Cape 0.064 0.070 0.053 0.017 0.063 0.069 0.053 0.016   Free State 0.084 0.093 0.067 0.026 0.085 0.094 0.067 0.027   KwaZulu-Natal 0.235 0.197 0.310 −0.113* 0.237 0.198 0.311 −0.113*   North West 0.051 0.043 0.066 −0.023* 0.051 0.044 0.066 −0.022*   Gauteng 0.117 0.140 0.074 0.066* 0.114 0.135 0.075 0.060*   Mpumalanga 0.070 0.065 0.081 −0.016 0.070 0.065 0.078 −0.013   Limpopo 0.098 0.094 0.107 −0.013 0.100 0.096 0.108 −0.012 Poverty   Food poverty line 0.362 0.321 0.442 −0.121* 0.293 0.255 0.368 −0.113*   Poverty line 0.329 0.309 0.369 −0.060* 0.338 0.307 0.399 −0.092* Geographical location   Rural 0.442 0.374 0.573 −0.199* 0.450 0.383 0.580 −0.197*   Urban 0.558 0.626 0.427 0.199* 0.550 0.617 0.420 0.197* Number of household members   1 or 2 0.135 0.147 0.113 0.034* 0.124 0.133 0.105 0.028*   3 0.227 0.211 0.258 −0.047* 0.235 0.215 0.274 −0.059*   4 0.212 0.213 0.211 0.002 0.216 0.215 0.218 −0.003   5 0.174 0.179 0.165 0.014 0.173 0.179 0.162 0.017   ≥6 0.252 0.250 0.254 −0.004 0.252 0.258 0.240 0.018 Head of household by race   African 0.768 0.719 0.862 −0.143* 0.768 0.719 0.862 −0.143*   Coloured 0.171 0.195 0.124 0.071* 0.171 0.195 0.124 0.071*   Asian/Indian/White 0.061 0.086 0.014 0.072* 0.061 0.085 0.014 0.071* Head of household by gender   Female 0.617 0.578 0.692 −0.114* 0.588 0.550 0.661 −0.111*   Male 0.383 0.422 0.308 0.114* 0.412 0.450 0.339 0.111* Head of household by education   No schooling 0.182 0.138 0.265 −0.127* 0.201 0.153 0.291 −0.138*   Primary 0.185 0.154 0.246 −0.092* 0.196 0.165 0.255 −0.090*   Secondary 0.515 0.542 0.464 0.078* 0.499 0.532 0.437 0.095*   Tertiary 0.118 0.167 0.025 0.142* 0.105 0.151 0.018 0.133* Head of household by age 49.085 48.166 50.869 −2.703 49.967 49.181 51.491 −2.310 Notes: Descriptive statistics of district of residence of household and interview’s month are omitted. For definitions of food poverty line and poverty line, see Appendix A. Asterisk mark indicates when the difference between the treated and control samples is statistically significant at 5% level. View Large 3. Empirical framework We will consider a standard RDD model, where T is a binary treatment indicator (treated by CSG), X is an assignment variable (in our analysis, the birth cohort of the child) and c is the threshold for X at which the probability of treatment changes discontinuously (1 January 1994). In the fuzzy RD design, the treatment effect is obtained by dividing the increase in the relation between Y (the outcome variable) and X at c by the difference of the probability that an applicant household will be treated before or after the cutoff. The treatment effect (denoted by τF) n the fuzzy RD design is τF=limϵ↓0E(Y∣X=c+ϵ)−limϵ↑0E(Y∣X=c+ϵ)limϵ↓0E(T∣X=c+ϵ)−limϵ↑0E(T∣X=c+ϵ). (1) where ϵ defines the neighbourhood in which local random assignment is satisfied. When local random assignment holds, equation (1) provides an unbiased estimate of a weighted version of the LATE, which evaluates the impact of the programme on individuals who were assigned to the treatment group and actually participated in it (compliers) compared with those who were assigned to the treatment group but did not participate in the programme (non-compliers) (Jacob et al., 2012). Equation (1) can be estimated using either parametric or non-parametric methods. The first approach followed here is to estimate equation (1) through a local-polynomial (LP) estimator (Calonico et al., 2014b), which approximates the regression functions above and below the cutoff by means of weighted polynomial regressions, with weights computed by applying a kernel function to the distance of each observation from the cutoff.14 The LP does not impose any functional form or introduce any relevant characteristic of the household and, hence, represents a benchmark for parametric and semi-parametric estimations. In addition, when the results estimated through the LP are not statistically different from the ones estimated with parametric methods, described below, we can indirectly show that the relevant covariates are balanced at the cutoff.15 Further, since equation (1) shows a close analogy between the fuzzy RD design with the Wald formulation of the treatment effect, an IV setting can also be applied. The only further requirement is the monotonicity and excludability related to the assignment variable when crossing the cutoff (Hahn et al., 2001). To estimate equation (1) in an IV framework, when the monotonicity assumption holds, we impose a linear form on the two-sided relation between the assignment variable, the treatment dummy, and the outcome variable.16 The IV equation is Y=δ0+δ1T+W1′δ2+P′δ3+Ω. (2) where W1 is a matrix of the relevant characteristics of the households and heads of households (described in Section 2.2) and Ω is an error term. The matrix P includes a second-order polynomial time trend ( trend and trend2) to take into account non-linear patterns of food expenditure,17 a set of regional Dr and district Dm dummies and their interactions with the second-order polynomial time trend. To take into account non-linearities, we also present estimation results obtained by introducing a third-degree polynomial of the assignment variable into equation (3).18 As shown in Section 2.1, the first shortcoming of the analysis concerns the reliability of the internal validity of the RDD design, since we can only identify the impact of the CSG on the outcome variable caused by the variation of the age limit, but we do not have complete control over the effects of changes in the income eligibility rule. In addition, households are able to manipulate their income threshold. Thus, the internal validity assumption may be not valid. Similarly, a threat to internal validity emerges considering implementation concerns, as outlined in Section 2.1. To overcome these problems, we propose the following robustness analyses. First of all, following van der Klaauw (2002), the selection bias may be overcome by replacing T with its propensity score E[T∣X] in equation (3). In this case, a two-step procedure is required and, thus, in the first stage we specify the PS function in the fuzzy RD design (van der Klaauw 2002, You 2013). Hence, the propensity score in the RD framework is E[T∣X]=f(X)+μ1(S≥c) (3) where f(X) is a continuous function in c which may be estimated parametrically or semi-parametrically. In the present context, it is defined as a third-degree polynomial of the assignment variable. The estimated propensity score can then replace the treatment variable T in equation (3) to estimate δ1. The second-step equation is Y=δ0+δ1[T∣X]+W′δ2+P′δ3+Ψ (4) so that we can compare the parameters estimated with the IV method with those obtained with the two-step PS procedure. When we find a statistically significant difference between them, we can consider whether the internal validity assumption may be violated.19 A second related issue concerns the administrative costs, bureaucratic requirements, and implementation problems, in terms of limiting or delaying participation in the CSG (Heinrich et al., 2012, Heinrich and Brill 2015, Samson et al., 2016, Heinrich et al., 2017). In the present case, we could have two concerns. First of all, the population of compliers might be unstable. Statistically, this implies that the population of compliers may change dramatically with small changes in the birth cohort of the children. Secondly, given that the estimates of the RD treatment effect only apply to people such that X=c, it is important to investigate the stability of the RD estimates, that is, to examine whether people with other values of X near c would have expected treatment effects of similar sign and magnitude. If not, i.e., if ceteris paribus a small change in X away from c would greatly change the average effect of treatment, then one would have serious doubts about the general usefulness and external validity of the estimates, since other contexts are likely to differ from the given one in even more substantial ways than a marginal change in X. With this in mind, we apply the method proposed by Cerulli et al. (2017), based on the Complier Probability Derivative (CPD) and on the Treatment Effect Derivative (TED).20 Estimates of TED that are statistically significant and large in magnitude are evidence of instability and, hence, a potential lack of external validity. In contrast, having TED near zero, or not statistically significant, provides evidence supporting the stability of the RD estimates since if the threshold had been somewhat lower or higher, the estimated LATE would probably have still been close to zero. On the other hand, significant and high values of the CPD show that there is instability in the complier population. However, as pointed out by Cerulli et al. (2017), when the TED is near zero, or not statistically significant, then even a certain instability in the complier population (i.e., CPD≠0) does not affect the estimated results and does not violate the assumption of external validity. Lastly, we recall that the potential outcome in a specific population subgroup may not be homogeneous with respect to the entire population (Angrist 2004). Hence, we calculate elasticity measures based on the PS estimates obtained in the two-step procedure, and carry out a further robustness check by comparing the results for the subgroups with that of the entire population, to verify whether significant differences emerge. In more detail, we extend equation (4) by replacing W1 with W2, which includes the interaction terms between the PS estimates and the dummy variables describing, one by one, five subgroups of the population which may present substantial differences in the potential outcomes. To allow for the correct identification of equation (4), we must assume that the interaction variables are continuous at the cutoff and uncorrelated with the error term, conditional on W2 (Becker et al., 2013). With the estimated parameters, we construct elasticity measures using the interaction terms and PS estimates. The elasticity measures allow us to compare the effects of policy changes on varying population subgroups and to interpret the results in terms of the percentage variation in the outcome variable caused by a 1% increase in the treated population. 4. Results and robustness analysis This section presents the estimates of the effects of the CSG on total food expenditure per adult equivalent, the shares of the food groups and the synthetic index of dietary diversity. The results for the full-sample 2008–2012 are given, together with those excluding 2012 for the shares of food groups and the synthetic index of dietary diversity. Figure 2 shows the results of the estimation of equation (1), which links the assignment variable X (the birth cohort of the child) to treatment status T (the probability of participating in the CSG), with the LP estimator. Children in birth cohorts outside the age eligibility for the CSG lie on the left of the cutoff, whereas children in both the treated and control groups lie on the right. The figure confirms the fuzzy nature of the RDD and, by using the corresponding robust estimator,21 we find that about 5.9% (s.e. 0.02, p-value 0.003) of the households in the first sample and about 10.5% (s.e. 0.034, p-value 0.002) in the second sample, participated in the CSG. Figure 2: View largeDownload slide LP Estimates: Birth Cohort and Treatment Status. (a) Sample 2008–2012 and (b) sample 2008–2011. Notes: Construction of evenly spaced bins follows Calonico et al. (2014a). Figure 2: View largeDownload slide LP Estimates: Birth Cohort and Treatment Status. (a) Sample 2008–2012 and (b) sample 2008–2011. Notes: Construction of evenly spaced bins follows Calonico et al. (2014a). Figure 3 panel (a) shows the estimates of the effects of the CSG on the first outcome variable, total food expenditure per adult equivalent, obtained through the LP estimator. Panels (b)–(e) replicate the same analysis for the outcome variables of the shares in total food expenditure of (i) carbohydrates, (ii) dairy products, (iii) proteins, and (iv) vitamins. Panel (f) considers the synthetic dietary diversity variable. Figure 3: View largeDownload slide LP Estimates: Impact of CSG on Total Food Expenditure Per Adult Equivalent and Shares of Food Groups. (a) Total food expenditure per adult equivalent, (b) share of carbohydrates *, (c) share of dairy products *, (d) share of proteins *, (e) share of vitamins *, (f) sietary diversity *. Notes: (*) Analysis of food groups and of synthetic index of dietary diversity is valid for 2008–2011. Construction of evenly spaced bins follows Calonico et al. (2014a). Figure 3: View largeDownload slide LP Estimates: Impact of CSG on Total Food Expenditure Per Adult Equivalent and Shares of Food Groups. (a) Total food expenditure per adult equivalent, (b) share of carbohydrates *, (c) share of dairy products *, (d) share of proteins *, (e) share of vitamins *, (f) sietary diversity *. Notes: (*) Analysis of food groups and of synthetic index of dietary diversity is valid for 2008–2011. Construction of evenly spaced bins follows Calonico et al. (2014a). Panel (a) of Figure 3 clearly shows a discontinuity in the food expenditure per adult equivalent around the 1 January 1994 cutoff: before that date, the expenditure pattern remained quite constant along the birth cohorts, but afterwards, participant households showed an increase in food expenditure per adult equivalent of about R55. This variation is significant, given that the mean value of food expenditure for adult equivalent is only R500 in our sample. A similar pattern is also found in the shares of carbohydrates and dairy products in total food expenditure, in panels (b) and (c), respectively. In both cases, there is a positive variation in expenditure shares after the cutoff due to the expanded CSG age eligibility rule. Conversely, a poorly defined result appears for the protein share (panel d), and no discontinuity in the vitamin share (panel e) or in the index of dietary diversity (panel f). As the last five plots were obtained with data up to 2011, i.e., only 1 year after the policy change, and as the LP estimator requires an extensive number of observations, these results should be viewed with caution. Lastly, the six plots show that a third-degree polynomial is used by the LP estimator to approximate the food expenditure patterns although, except for protein, a linear functional form approximates the behaviour of the variables in the neighbourhood of the cutoff quite well. To complete the preliminary analysis, we perform the continuity test for each variable listed in Table 2.22 The smoothness graphs are reported in Appendix B and show that there is no evidence of discontinuity at the cutoff.23 Table 3 lists the estimates of food expenditure per adult equivalent, comparing the results for the IV and two-stage PS estimators.24 In each case, we consider a linear functional form (1) and then a non-linear one (2) that uses a third-degree polynomial, as suggested by the LP estimator. When the results obtained from the two-step PS estimator are analysed, the non-linear form is introduced in f(X), as in equation (4). Table 3: Impact of CSG on total food expenditure per adult equivalent Instrumental variable Propensity score (1) (2) (1) (2) Child Support Grant (CSG) 55.337*** 54.202*** 52.357*** 53.964*** (20.040) (19.589) (19.547) (19.176) Constant term 1046.445*** 1232.174*** 1221.254*** 1432.057*** (184.109) (78.246) (77.864) (187.934) Fixed effects Yes Yes Yes Yes Linear and quadratic trends Yes Yes Yes Yes Selected covariates Yes Yes Yes Yes Cragg–Donald F statistic 1359.645 499.282 (0.000) (0.000) Kleibergen–Paap F test statistic † 1433.471 510.756 (0.000) (0.000) R2 0.676 0.676 0.681 0.682 Adjusted R2 0.658 0.658 0.663 0.664 No. of observations 3399 3399 3399 3399 Instrumental variable Propensity score (1) (2) (1) (2) Child Support Grant (CSG) 55.337*** 54.202*** 52.357*** 53.964*** (20.040) (19.589) (19.547) (19.176) Constant term 1046.445*** 1232.174*** 1221.254*** 1432.057*** (184.109) (78.246) (77.864) (187.934) Fixed effects Yes Yes Yes Yes Linear and quadratic trends Yes Yes Yes Yes Selected covariates Yes Yes Yes Yes Cragg–Donald F statistic 1359.645 499.282 (0.000) (0.000) Kleibergen–Paap F test statistic † 1433.471 510.756 (0.000) (0.000) R2 0.676 0.676 0.681 0.682 Adjusted R2 0.658 0.658 0.663 0.664 No. of observations 3399 3399 3399 3399 Notes: In all specifications, dependent variable is total food expenditure per adult equivalent (see the list of the selected covariates in Section 2). Robust standard errors, clustered at household level, are shown in brackets. Asterisks: p-value levels ( *p<0.1;**p<0.05;***p<0.01). For each model, linear (1) and non-linear specification (2) is shown. When two-step PS procedure is applied, non-linearities are introduced only in first-stage regression. Tests for weak instrument hypothesis and first-stage F statistics and Wald statistics based on Cragg and Donald (1993) and Kleibergen and Paap (2006). †Confidence intervals for Kleibergen–Paap F test statistic follow Bazzi and Clemens (2013). View Large Table 3: Impact of CSG on total food expenditure per adult equivalent Instrumental variable Propensity score (1) (2) (1) (2) Child Support Grant (CSG) 55.337*** 54.202*** 52.357*** 53.964*** (20.040) (19.589) (19.547) (19.176) Constant term 1046.445*** 1232.174*** 1221.254*** 1432.057*** (184.109) (78.246) (77.864) (187.934) Fixed effects Yes Yes Yes Yes Linear and quadratic trends Yes Yes Yes Yes Selected covariates Yes Yes Yes Yes Cragg–Donald F statistic 1359.645 499.282 (0.000) (0.000) Kleibergen–Paap F test statistic † 1433.471 510.756 (0.000) (0.000) R2 0.676 0.676 0.681 0.682 Adjusted R2 0.658 0.658 0.663 0.664 No. of observations 3399 3399 3399 3399 Instrumental variable Propensity score (1) (2) (1) (2) Child Support Grant (CSG) 55.337*** 54.202*** 52.357*** 53.964*** (20.040) (19.589) (19.547) (19.176) Constant term 1046.445*** 1232.174*** 1221.254*** 1432.057*** (184.109) (78.246) (77.864) (187.934) Fixed effects Yes Yes Yes Yes Linear and quadratic trends Yes Yes Yes Yes Selected covariates Yes Yes Yes Yes Cragg–Donald F statistic 1359.645 499.282 (0.000) (0.000) Kleibergen–Paap F test statistic † 1433.471 510.756 (0.000) (0.000) R2 0.676 0.676 0.681 0.682 Adjusted R2 0.658 0.658 0.663 0.664 No. of observations 3399 3399 3399 3399 Notes: In all specifications, dependent variable is total food expenditure per adult equivalent (see the list of the selected covariates in Section 2). Robust standard errors, clustered at household level, are shown in brackets. Asterisks: p-value levels ( *p<0.1;**p<0.05;***p<0.01). For each model, linear (1) and non-linear specification (2) is shown. When two-step PS procedure is applied, non-linearities are introduced only in first-stage regression. Tests for weak instrument hypothesis and first-stage F statistics and Wald statistics based on Cragg and Donald (1993) and Kleibergen and Paap (2006). †Confidence intervals for Kleibergen–Paap F test statistic follow Bazzi and Clemens (2013). View Large All specifications include province and district dummies, and also linear and quadratic trends to take into account the non-linear patterns of food expenditure (Pieroni et al., 2013, Pieroni and Salmasi 2015), together with household and head of household characteristics (see Section 2). Robust standard errors, clustered at household level, are shown in brackets.25 For each specification, we test for weak instruments. In more detail, we run a test for weak instruments and present first-stage F statistics and Wald statistics based on the Cragg and Donald (1993) and Kleibergen and Paap (2006) generalisation to non-independently and non-identically distributed errors, together with the p-values (Bazzi and Clemens 2013). The under-identification and weak instrument tests show that the assignment variable accounts for the entire endogeneity appearing in the IV estimations. We find that the non-parametric and parametric estimations show very similar variations due to the CSG policy change. That is, from the linear (1) and non-linear (2) specifications of the IV estimator, we find a variation in food expenditure per adult equivalent that ranges between R54 and R55, and this variation is significant at 1%. Considering the full sample (see Table 1), this variation corresponds to a mean increase of 10% in total food expenditure per adult equivalent. This result is stronger than the findings of Coetzee (2013) for the CSG and almost in line with other evaluations of the impact of social grants on several indicators of food security in other countries of sub-Saharan Africa (Pellerano et al., 2014, Burchi et al., 2016, Davis and Handa 2016, Pellerano et al., 2016).26 Moving to columns 3 and 4, we use the two-step PS estimator to check for the robustness of the IV results. It should be noted that the two-step PS is used to analyse the robustness of the results when the identification mechanism is not completely known. In the present case, it shows that the multiple sources of variations in participation in the CSG, caused by the change in the income cutoff and by problems in access related to administrative burdens, do not affect the IV results, since there are no differences in the standard deviations between the IV and the two-step PS estimates. Table 4 extends the previous analysis to the shares of food groups in the total food expenditure per adult equivalent and to the synthetic index of dietary diversity. In depth, the table compares the results from a linear specification of the IV and two-step PS estimations.27 Table 4: Impact of CSG on food expenditure by food group and on dietary diversity index Carbohydrates Dairy products Proteins Vitamins Dietary diversity IV PS IV PS IV PS IV PS IV PS Child Support Grant (CSG) 2.267** 2.022** 0.808** 0.753 0.196 −0.233 1.230** 1.019 −0.074 −0.043 (1.150) (1.030) (0.405) (0.483) (1.216) (1.176) (0.557) (0.588) (0.176) (0.171) Constant term 32.921*** 32.716*** 5.768*** 5.685*** 9.164 9.270 14.986*** 14.898*** 9.506*** 9.506*** (7.440) (7.359) (2.156) (2.185) (5.772) (5.774) (4.379) (4.391) (0.897) (0.895) Fixed effects Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Linear and quadratic trends Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Selected covariates Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Cragg–Donald F statistic 866.207 866.207 866.207 866.207 848.632 (0.000) (0.000) (0.000) (0.000) (0.000) Kleibergen–Paap F test statistic † 930.615 930.615 930.615 930.615 913.688 (0.000) (0.000) (0.000) (0.000) (0.000) R2 0.361 0.361 0.090 0.094 0.222 0.222 0.097 0.106 0.384 0.385 Adjusted R2 0.321 0.321 0.034 0.038 0.174 0.174 0.042 0.051 0.346 0.346 No. of observations 2193 2193 2193 2193 2193 2193 2193 2193 2175 2175 Carbohydrates Dairy products Proteins Vitamins Dietary diversity IV PS IV PS IV PS IV PS IV PS Child Support Grant (CSG) 2.267** 2.022** 0.808** 0.753 0.196 −0.233 1.230** 1.019 −0.074 −0.043 (1.150) (1.030) (0.405) (0.483) (1.216) (1.176) (0.557) (0.588) (0.176) (0.171) Constant term 32.921*** 32.716*** 5.768*** 5.685*** 9.164 9.270 14.986*** 14.898*** 9.506*** 9.506*** (7.440) (7.359) (2.156) (2.185) (5.772) (5.774) (4.379) (4.391) (0.897) (0.895) Fixed effects Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Linear and quadratic trends Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Selected covariates Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Cragg–Donald F statistic 866.207 866.207 866.207 866.207 848.632 (0.000) (0.000) (0.000) (0.000) (0.000) Kleibergen–Paap F test statistic † 930.615 930.615 930.615 930.615 913.688 (0.000) (0.000) (0.000) (0.000) (0.000) R2 0.361 0.361 0.090 0.094 0.222 0.222 0.097 0.106 0.384 0.385 Adjusted R2 0.321 0.321 0.034 0.038 0.174 0.174 0.042 0.051 0.346 0.346 No. of observations 2193 2193 2193 2193 2193 2193 2193 2193 2175 2175 Notes: The dependent variables are shares of food groups in total food expenditure per adult equivalent. In all specifications, dependent variable is total food expenditure per adult equivalent (see the list of the selected covariates in Section 2). Robust standard errors, clustered at household level, are shown in brackets. Asterisks: p-value levels ( *p<0.1;**p<0.05;***p<0.01). For each model, linear (1) and non-linear specification (2) is shown. When two-step PS procedure is applied, non-linearities are introduced only in first-stage regression. Tests for weak instrument hypothesis and first-stage F statistics and Wald statistics based on Cragg and Donald (1993) and Kleibergen and Paap (2006). †Confidence intervals for Kleibergen–Paap F test statistic follow Bazzi and Clemens (2013). View Large Table 4: Impact of CSG on food expenditure by food group and on dietary diversity index Carbohydrates Dairy products Proteins Vitamins Dietary diversity IV PS IV PS IV PS IV PS IV PS Child Support Grant (CSG) 2.267** 2.022** 0.808** 0.753 0.196 −0.233 1.230** 1.019 −0.074 −0.043 (1.150) (1.030) (0.405) (0.483) (1.216) (1.176) (0.557) (0.588) (0.176) (0.171) Constant term 32.921*** 32.716*** 5.768*** 5.685*** 9.164 9.270 14.986*** 14.898*** 9.506*** 9.506*** (7.440) (7.359) (2.156) (2.185) (5.772) (5.774) (4.379) (4.391) (0.897) (0.895) Fixed effects Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Linear and quadratic trends Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Selected covariates Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Cragg–Donald F statistic 866.207 866.207 866.207 866.207 848.632 (0.000) (0.000) (0.000) (0.000) (0.000) Kleibergen–Paap F test statistic † 930.615 930.615 930.615 930.615 913.688 (0.000) (0.000) (0.000) (0.000) (0.000) R2 0.361 0.361 0.090 0.094 0.222 0.222 0.097 0.106 0.384 0.385 Adjusted R2 0.321 0.321 0.034 0.038 0.174 0.174 0.042 0.051 0.346 0.346 No. of observations 2193 2193 2193 2193 2193 2193 2193 2193 2175 2175 Carbohydrates Dairy products Proteins Vitamins Dietary diversity IV PS IV PS IV PS IV PS IV PS Child Support Grant (CSG) 2.267** 2.022** 0.808** 0.753 0.196 −0.233 1.230** 1.019 −0.074 −0.043 (1.150) (1.030) (0.405) (0.483) (1.216) (1.176) (0.557) (0.588) (0.176) (0.171) Constant term 32.921*** 32.716*** 5.768*** 5.685*** 9.164 9.270 14.986*** 14.898*** 9.506*** 9.506*** (7.440) (7.359) (2.156) (2.185) (5.772) (5.774) (4.379) (4.391) (0.897) (0.895) Fixed effects Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Linear and quadratic trends Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Selected covariates Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Cragg–Donald F statistic 866.207 866.207 866.207 866.207 848.632 (0.000) (0.000) (0.000) (0.000) (0.000) Kleibergen–Paap F test statistic † 930.615 930.615 930.615 930.615 913.688 (0.000) (0.000) (0.000) (0.000) (0.000) R2 0.361 0.361 0.090 0.094 0.222 0.222 0.097 0.106 0.384 0.385 Adjusted R2 0.321 0.321 0.034 0.038 0.174 0.174 0.042 0.051 0.346 0.346 No. of observations 2193 2193 2193 2193 2193 2193 2193 2193 2175 2175 Notes: The dependent variables are shares of food groups in total food expenditure per adult equivalent. In all specifications, dependent variable is total food expenditure per adult equivalent (see the list of the selected covariates in Section 2). Robust standard errors, clustered at household level, are shown in brackets. Asterisks: p-value levels ( *p<0.1;**p<0.05;***p<0.01). For each model, linear (1) and non-linear specification (2) is shown. When two-step PS procedure is applied, non-linearities are introduced only in first-stage regression. Tests for weak instrument hypothesis and first-stage F statistics and Wald statistics based on Cragg and Donald (1993) and Kleibergen and Paap (2006). †Confidence intervals for Kleibergen–Paap F test statistic follow Bazzi and Clemens (2013). View Large When the share of carbohydrates is analysed, we find a variation of about 2% in expenditure produced by the policy change. Again, this result is similar to that found by the LP estimator and does not change in the two-step PS estimation. Less robust but still significant results are found when the shares of dairy products and vitamins are accounted for in the IV specification: the beneficiary families increase the share of expenditure on diary products by 0.8% and that on vitamins by 1.2%. However, these results become non-significant at the 10% level when the two-step PS estimators are used. Instead, no significant variations are found when the share of proteins in total expenditure per adult equivalent or the synthetic index of dietary diversity are examined. It should be emphasised that carbohydrates are the major food item consumed by the treated households (on average, 23% of total food expenditure per adult equivalent), and these results thus indicate that the programme grants were not sufficient to allow the beneficiary households to make significant changes in their dietary habits. However, we must use caution because we are using the restricted sample (2008–2011) when considering food shares and dietary diversity whereas, in the case of total food expenditure, we use the full sample (2008–2012). We now consider the robustness tests of our estimates. Appendix C presents the results for a restricted sample including only 2008 of the same estimations as were run for Tables 3 and 4. Indeed, since the policy change took place in 2010, we can use these estimates to run placebo tests showing that the estimated variations are due to the extension in the age eligibility of the CSG and not to the self-selection of individuals across different age cohorts. These tests show that the estimated average treatment effect is not only valid for the population of compliers but can be generalised to the entire population since, for each outcome variable, we do not find any statistically significant effect due to the extension of the CSG. Following the outline described in Section 3, Table 5 presents the two formal tests based on the CPD and TED to assess the validity of the RD design. The CPD estimates are presented for the two sample 2008–2012 and 2008–2011, at the bottom of the table. Comparing the two estimated parameters, we find statistically significant evidence of a certain instability in the population of compliers. As shown in Figure 2, the probability of being enroled in the CSG at the cutoff is about 2% (7% when we consider the sample 2008–2011), whereas we find that it is about 4.4% (9.2% in the second sample) when we consider a birth cohort that is a month older. As shown in the upper part of the table, we do not find any statistically significant estimate of the TED for each outcome variable. Together, these results indicate that despite the fact that the complier population is unstable, this result does not affect the constancy of the average treatment effect. Thus, the external validity assumption still holds for each specification. Table 5: Impact of CSG on food expenditure by food group and on dietary diversity index Food expenditure Carbohydrates Diary products Proteins Vitamins Dietary diversity TED −380.548 −2.415 0.053 −3.333 −0.014 0.329 (387.730) (2.647) (0.773) (3.217) (0.905) (0.368) No. of observation 3399 2183 2183 2183 2183 2183 Sample Sample (2008–2012) (2008–2011) CPD 0.022*** 0.019*** (0.003) (0.003) No. of observation 3399 2183 Food expenditure Carbohydrates Diary products Proteins Vitamins Dietary diversity TED −380.548 −2.415 0.053 −3.333 −0.014 0.329 (387.730) (2.647) (0.773) (3.217) (0.905) (0.368) No. of observation 3399 2183 2183 2183 2183 2183 Sample Sample (2008–2012) (2008–2011) CPD 0.022*** 0.019*** (0.003) (0.003) No. of observation 3399 2183 Notes: Bootstrapped Standard errors based on 500 simulations are reported in brackets. Asterisks: p-value levels ( *p<0.1;**p<0.05;***p<0.01). View Large Table 5: Impact of CSG on food expenditure by food group and on dietary diversity index Food expenditure Carbohydrates Diary products Proteins Vitamins Dietary diversity TED −380.548 −2.415 0.053 −3.333 −0.014 0.329 (387.730) (2.647) (0.773) (3.217) (0.905) (0.368) No. of observation 3399 2183 2183 2183 2183 2183 Sample Sample (2008–2012) (2008–2011) CPD 0.022*** 0.019*** (0.003) (0.003) No. of observation 3399 2183 Food expenditure Carbohydrates Diary products Proteins Vitamins Dietary diversity TED −380.548 −2.415 0.053 −3.333 −0.014 0.329 (387.730) (2.647) (0.773) (3.217) (0.905) (0.368) No. of observation 3399 2183 2183 2183 2183 2183 Sample Sample (2008–2012) (2008–2011) CPD 0.022*** 0.019*** (0.003) (0.003) No. of observation 3399 2183 Notes: Bootstrapped Standard errors based on 500 simulations are reported in brackets. Asterisks: p-value levels ( *p<0.1;**p<0.05;***p<0.01). View Large To complete the analysis, Figure 4 shows the elasticity measures for total food expenditure per adult equivalent and the share of carbohydrates. The elasticity measures are based on equation (4) (Section 3) and are obtained by jointly considering the interaction term between the estimated PS and the dummy variables, describing, one by one, five subgroups of the population which may present substantial differences in the potential outcomes. We do not present the other panels (for dairy products, proteins and vitamins and for the index of dietary diversity) since the estimated parameters are not robust across differing estimation methods (see Table 4). Figure 4 shows the results for the total population and some population subgroups, which are the households under both the food poverty line and the poverty line, households in which the head is an African and households living in rural and urban areas. It should be recalled that the main targets of the CSG programme are the subgroups most vulnerable to poverty and food insecurity, i.e. households living under the food poverty line, Africans, and those living in rural areas. The elasticity measure allows us to compare the various subgroups with the entire population in terms of the percentage effect of the variation in their participation in the programme after the policy change and to check its homogeneity. Figure 4: View largeDownload slide Elasticity Measures: Total Food Expenditure Per Adult Equivalent and Share of Carbohydrates. (a) Total food expenditure per adult equivalent and (b) Share of carbohydrates. Notes: The elasticity measures are based on the the two-step PS estimated parameters (Table 3 specification 2 and in Table 4 specification 2). Standard errors are estimated by Delta-Method. Asterisks: p-value levels ( *p<0.1;**p<0.05;***p<0.01). Figure 4: View largeDownload slide Elasticity Measures: Total Food Expenditure Per Adult Equivalent and Share of Carbohydrates. (a) Total food expenditure per adult equivalent and (b) Share of carbohydrates. Notes: The elasticity measures are based on the the two-step PS estimated parameters (Table 3 specification 2 and in Table 4 specification 2). Standard errors are estimated by Delta-Method. Asterisks: p-value levels ( *p<0.1;**p<0.05;***p<0.01). Figure 4 panel (a) shows that although some differences between the subgroups and full-sample estimates are found, all the subgroups elasticities are plotted in the same confidence interval as the full-sample elasticity. In this case, the analysis shows that a 1% variation in participation in the programme produces a 3% increase in total food expenditure per adult equivalent, at a significance level which is less than 5%. We find a higher variability when the rural/urban subgroups are analysed. Figure 4 shows that a 1% increase in participation in the programme for the households living in urban areas produces a positive variation in the food expenditure per adult equivalent, which is double that estimated for rural areas. This result may be explained by the fact that rural areas in South Africa are geographically isolated and marginalised, so that food availability is a serious problem (Kirsten 2012). Several studies have found a higher percentage of food insecurity in rural areas in South Africa (Kirkland et al., 2013). However, we must also recall that, as Aliber (2009) stresses, rural households spend less than their urban counterparts on food purchases because they have their own production. In principle, measures of food expenditure are designed to capture this information, but the imputed value is probably smaller than the true one. In this case, the effect of the CSG on food security in rural areas in our analysis would be underestimated. When the share of carbohydrates is analysed, we find more variability between the full-sample elasticity measure and those for the population subgroups, but the results are significant at the 5% level only for households under the food poverty line or ones in which the head is African. In more detail, the elasticity measure for beneficiaries under the food poverty line is about double that of the full sample, even though it falls in the same confidence interval. This result is in line with the existing literature, which states that poor dietary variety mainly concerns the poorest people, who consume far more cereals and less fruit, vegetables, dairy products, and meat, than other population subgroups (Labadarios et al., 2011, South African Department of Health 2013). In the case of the subgroup of households with an African head, the elasticity is very close to that of the full sample. Overall, these results indicate that the CSG has proved to be effective in supporting the food expenditure of its beneficiary households. However, the strategy of providing a basic grant to a large share of the poor population has not been effective in producing a significant change in the dietary habits of the very poor beneficiaries or at guaranteeing a nutritionally varied food basket. This result is not unexpected since the grant money is used to meet the broader needs of the families, including education, in low-income families. However, this outcome should not be overlooked since increasing food consumption without modifying the dietary habits may shift poor households from undernutrition to malnutrition, worsening their health condition (Tzioumis and Adair 2014). This suggests that ancillary social services, starting with the improvement of mother’s and children’s nutritional education, should complement the CSG in strengthening food security impacts (Heinrich et al., 2012, Burchi et al., 2016). Our findings also show that the CSG has failed to improve the food security of the most disadvantaged groups to a greater extent than for the others. This result, that the CSG is not effectively targeting those most in need, may be partially explained by several administrative constraints in the implementation of the programme, serving as an operational barrier to reaching the most vulnerable caregivers and children (Delany et al., 2008, Heinrich et al., 2012, Heinrich and Brill 2015, Davis and Handa 2016, Samson et al., 2016). This study suggests that an integrated approach and nutrition-sensitive social protection programme, specifically targeting additional resources and complementary social services to the most food-insecure households, would contribute more effectively to increasing the food and nutrition status of the very poor households in South Africa. 5. Conclusions This paper estimates the impact of the South African Child Support Grant (CSG) on food security. We used the dataset provided by the National Income Dynamics Study covering 2008, 2010–2011, and 2012, and carried out a Regression Discontinuity design to estimate the effect of the programme on total food expenditure per adult equivalent, as well as on dietary diversity. Our results show that the transfers provided by the CSG significantly increased the total food expenditure per adult equivalent. Even if our findings must be treated with caution, given that they are based on a non-experimental study, we show that the results are very robust, being confirmed by a comparison between non-parametric, semi-parametric, and parametric estimations. Our analysis also shows that the estimated parameters are constant when the cutoff diverges, which means that it provides plausible casual effects of the programme. When dietary diversity is analysed, we find robust positive results only when the share of carbohydrates in the total food expenditure per adult equivalent is included. Since this is the largest food group consumed by the treated households (23% of total food expenditure), this result indicates that the CSG has not been effective in making significant improvements in the dietary habits of the beneficiary households. The estimates of the elasticity measures for the total food expenditure and the share of carbohydrates do not show any significant difference between various population subgroups. This means that the statistical results are homogeneous. It also suggests that the CSG, at variance with its stated objective, has not been effectively targeted to the poorest population subgroups who are the most vulnerable to food insecurity. Overall, the policy implication of our study is that the current design of the CSG, which only provides a small grant for each beneficiary child, and its strategy of gradually widening the eligible population, has not been sufficient for guaranteeing a significant reduction of deprivation for the most vulnerable households. A more effective approach would be to integrate this policy with a specific, comprehensive strategy to reduce food insecurity and deliver additional grants to those most in need. In addition, ancillary social services, including investments in mother’s and children’s nutritional education, should be introduced to improve the food habits of poor households. Lastly, income-generating programmes and the enhancement of small-scale agricultural activities remain crucial in increasing household access to food in the most poverty-stricken areas, which are the rural ones. Acknowledgements We would like to thank Ingrid Woolard, J. Paul Dunne and the participants at the SITES/IDEAs 2nd Annual Conference for their insightful comments. We thank SALDRU (Southern Africa Labour & Development Research Unit), University of Cape Town, for its support. The authors wish also to acknowledge the Editor and two anonymous reviewers of this journal for their helpful suggestions. The usual disclaimers apply. References Agüero J. , Carter M. R. , May J. ( 2007 ) ‘ Poverty and iNequality in the First Decade of South Africa’s Democracy: What Can Be Learnt From Panel Data From KwaZulu-Natal? ’, Journal of African Economies , 16 ( 5 ): 782 – 812 . Google Scholar CrossRef Search ADS Agüero J. , Carter M. , Woolard I. ( 2010 ) The impact of unconditional cash transfers on nutrition: The South African Child Support Grant. International Poverty Centre Discussion Papers No. 39, United Nations Development Programme (UNDP), New York. Alderman H. ( 2014 ) Can transfer programs be made more nutrition-sensitive? International Food Policy Research Institute (IFPRI) Discussion Paper No. 1342, IFPRI, Washington D.C. Aliber M. A. ( 2001 ) Study of the incidence and nature of chronic poverty and development policy in South Africa: An overview. Chronic Poverty Research Centre Working Papers No.1, University of Manchester, Manchester. Aliber M. ( 2009 ) ‘ Exploring Statistics South Africa’s National Household Surveys as Sources of Information About Household-level Food Security ’, Agrekon , 48 ( 4 ): 384 – 409 . Google Scholar CrossRef Search ADS Angrist J. D. ( 2004 ) ‘ Treatment Effect Heterogeneity in Theory and Practice ’, The Economic Journal , 114 ( 494 ): 52 – 83 . Google Scholar CrossRef Search ADS Bassett L. ( 2008 ) Can conditional cash transfer programs play a greater role in reducing child undernutrition? World Bank Social Policy Discussion Papers No. 835, The World Bank, Washington D.C. Bazzi S. , Clemens M. A. ( 2013 ) ‘ Blunt Instruments: Avoiding Common Pitfalls in Identifying the Causes of Economic Growth ’, American Economic Journal: Macroeconomics , 5 ( 2 ): 152 – 86 . Google Scholar CrossRef Search ADS Becker S. O. , Egger P. H. , von Ehrlich M. ( 2013 ) ‘ Absorptive Capacity and the Growth and Investment Effects Of Regional Transfers: A Regression Discontinuity Design With Heterogeneous Treatment Effects ’, American Economic Journal: Economic Policy , 5 ( 4 ): 29 – 77 . Google Scholar CrossRef Search ADS Bor J. ( 2013 ) Cash transfers and teen pregnancy in South Africa: Evidence from a natural experiment. Northeast Universities Development Consortium Conference, Harvard University, Cambridge, MA. Burchi F. , Scarlato M. , d’Agostino G. ( 2016 ) Addressing food insecurity in sub-Saharan Africa: The role of cash transfers. German Development Institute Discussion Papers No. 17, German Development Institute Discussion, Bonn. Calonico S. , Cattaneo M. D. , Titiunik R. ( 2014 a) ‘ Robust Data-driven Inference in the Regression-Discontinuity Design ’, Stata Journal , 14 ( 4 ): 909 – 46 . Calonico S. , Cattaneo M. D. , Titiunik R. ( 2014 b) ‘ Robust Non-parametric Confidence Intervals for Regression-discontinuity Designs ’, Econometrica: Journal of the Econometric Society , 82 ( 6 ): 2295 – 2326 . Google Scholar CrossRef Search ADS Cameron A. C. , Miller D. L. ( 2015 ) ‘ A Practitioner’s Guide to Cluster-Robust Inference ’, Journal of Human Resources , 50 ( 2 ): 317 – 72 . Google Scholar CrossRef Search ADS Cerulli G. , Dong Y. , Lewbel A. , Poulsen A. ( 2017 ) ‘Testing Stability of Regression Discontinuity Models’, in Cattaneo M. D. , Escanciano J. C. (eds) , Regression discontinuity designs, Advances in econometrics , Vol. 38 . Bingley (UK) : Emerald Publishing Limited , pp. 317 – 39 . Google Scholar CrossRef Search ADS Coetzee M. ( 2013 ) ‘ Finding the Benefits: Estimating the Impact of the South African Child Support Grant ’, South African Journal of Economics , 81 ( 3 ): 427 – 50 . Google Scholar CrossRef Search ADS Cragg J. G. , Donald S. G. ( 1993 ) ‘ Testing Identifiability and Specification in Instrumental Variable Models ’, Econometric Theory , 9 ( 2 ): 222 – 40 . Google Scholar CrossRef Search ADS Davis B. , Handa S. ( 2016 ) ‘Implementing rigorous evaluations in the real world: the quantitative approach to evaluation design in the transfer project’, in Davis B. , Handa S. , Hypher N. , Winder Rossi N. , Winters P. , Yablonski J. (eds) , From Evidence to Action. The Story of Cash Transfers and Impact Evaluation in Sub-Saharan Africa . Oxford : Oxford University Press . Google Scholar CrossRef Search ADS Deaton A. , Grosh M. ( 2000 ) ‘Consumption’, in Grosh M. , Glewwe P. (eds) , Designing Household Survey Questionnaires for Developing Countries: Lessons from 15 Years of the Living Standards Measurement Study , vol.1 . Washington, DC : The World Bank , pp. 91 – 133 . Delany A. , Ismail Z. , Graham L. , Ramkissoon Y. ( 2008 ) Review of the Child Support Grant: Uses, Implementation and Obstacles . Johannesburg : United Nations Children’s Fund . del Ninno C. , Mills B. ( 2015 ) ‘Introduction’, in Del Ninno C. , Mills B. (eds) , Safety Nets in Africa . Washington D.C : The World Bank . Department of Social Development ( 2010 ). An Overview of South Africa’s Social Security System . Republic of South Africa , Pretoria . Dong Y. , Lewbel A. ( 2015 ) ‘ Identifying the Effect of Changing The Policy Threshold In Regression Discontinuity Models ’, Review of Economics and Statistics , 97 ( 5 ): 1081 – 92 . Google Scholar CrossRef Search ADS Du Toit D. ( 2011 ) Food Security. Department of Agriculture, Forestry and Fisheries . Pretoria : Republic of South Africa . Eyal K. , Woolard I. ( 2013 ). School Enrolment and the Child Support Grant: Evidence from South Africa. Southern African Labour and Development Research Unit (SALDRU) Working Papers No. 12, University of Cape Town, Cape Town. Hahn J. , Todd P. , Van der Klaauw W. ( 2001 ) ‘ Identification and Estimation of Treatment Effects With a Regression-discontinuity Design ’, Econometrica: Journal of the Econometric Society , 69 ( 1 ): 201 – 9 . Google Scholar CrossRef Search ADS Heinrich C. J. , Brill R. ( 2015 ) ‘ Stopped in the Name of the Law: Administrative Burden and its Implications For Cash Transfer Program Effectiveness ’, World Development , 72 ( C ): 277 – 95 . Google Scholar CrossRef Search ADS Heinrich C. J. , Hoddinott J. , Samson M. , Mac Quene K. , van Niekerk I. , Bryant R. ( 2012 ) The South African Child Support Grant impact assessment. Prepared for the Department of Social Development (DSD), the South African Social Security Agency (SASSA) and the United Nations Children’s Fund (UNICEF) . Pretoria : Republic of South Africa . Heinrich C. J. , Hoddinott J. , Samson M. ( 2017 ) ‘ Reducing Adolescent Risky Behaviors In A High-risk Context: The Effects Of Unconditional Cash Transfers in South Africa ’, Economic Development and Cultural Change , 65 ( 4 ): 619 – 52 . Google Scholar CrossRef Search ADS Hoddinott J. , Bassett L. ( 2008 ) Conditional Cash Transfer Programs And Nutrition In Latin America: Assessment Of Impacts And Strategies For Improvement . Washington DC/Rome : United Nations (UN) and Food and Agriculture Organization (FAO) . Hoddinott J. , Yohannes Y. ( 2002 ) Dietary Diversity as a Food Security Indicator. Food and Nutrition Technical Assistance Project . Washington DC : Academy for Educational Development . Jacob R. T. , Zhu P. , Somers M.-A. , Bloom H. S. ( 2012 ) A Practical Guide to Regression Discontinuity . New York : MDRC . Kennedy G. , Ballard T. , Dop M. ( 2011 ) Guidelines for Measuring Household and Individual Dietary Diversity . Rome : Food and Agriculture Organization of the United Nations (FAO) . Kirkland T. M. , Kemp R. J. , Hunter L. M. , Twine W. M. ( 2013 ) ‘ Toward imProved Understanding Of Food Security: A Methodological Examination Based in Rural South Africa ’, Food, Culture & Society , 16 ( 1 ): 65 – 84 . Google Scholar CrossRef Search ADS Kirsten J. F. ( 2012 ) The political economy of food price policy in South Africa. United Nations University, World Institute for Development Economics Research (UNU-WIDER) Working Papers No.102, UNU-WIDER, Helsinki. Kleibergen F. , Paap R. ( 2006 ) ‘ Generalized Reduced Rank Tests Using The Singular Value Decomposition ’, Journal of Econometrics , 133 ( 1 ): 97 – 126 . Google Scholar CrossRef Search ADS Labadarios D. , Steyn N. P. , Nel J. ( 2011 ) ‘ How Diverse is the Diet of Adult South Africans? ’, Nutrition Journal , 10 ( 33 ): 1 – 11 . Google Scholar PubMed Lee D. S. ( 2008 ) ‘ Randomized Experiments From Non-random Selection in U.S. House Elections ’, Journal of Econometrics , 142 ( 2 ): 675 – 97 . Google Scholar CrossRef Search ADS Lee D. S. , Lemieux T. ( 2010 ) ‘ Regression Discontinuity Designs in Economics ’, Journal of Economic Literature , 48 ( 2 ): 281 – 355 . Google Scholar CrossRef Search ADS Leibbrandt M. , Levinsohn J. ( 2011 ) Fifteen years on: Household incomes in South Africa. NBER Working Papers No. 16661, NBER, Cambridge Mass. Leibbrandt M. , Wegner E. , Finn A. ( 2011 ) The policies for reducing income inequality and poverty in South Africa. Southern African Labour and Development Research Unit (SALDRU) Working Papers No. 64, University of Cape Town, Cape Town. Lund F. , Noble M. , Barnes H. , Wright G. ( 2008 ) ‘ Is There a Rationale For Conditional Cash Transfers For Children in South Africa? ’, Transformation: Critical Perspectives on Southern Africa , 70 : 70 – 91 . Google Scholar CrossRef Search ADS Manley J. , Gitter S. , Slavchevska V. ( 2013 ) ‘ How Effective Are Cash Transfers At Improving Nutritional Status? ’, World Development , 48 ( issue C ): 133 – 55 . Google Scholar CrossRef Search ADS Manyamba C. , Hendriks S. , Chilonda P. , Musaba E. ( 2012 ) Factors contributing to inequalities in food security in South Africa: implications for agricultural policy. Paper presented at conference ‘Strategies to overcome poverty & inequality’, University of Cape Town, Cape Town. McEwen H. , Kannemeyer C. , Woolard I. ( 2009 ) Social assistance grants: Analysis of the NIDS wave 1 dataset. NIDS Discussion papers No.10, National Income Dynamics Study (NIDS). Southern African Labour and Development Research Unit (SALDRU) and University of Cape Town, Cape Town. Pauw K. , Mncube L. ( 2007 ) Expanding the social security net in South Africa: opportunities, challenges and constraints. Development Policy Research Unit (DPRU) Working Paper No. 127, University of Cape Town, Cape Town. Pellerano L. , Daidone S. , Davis B. , Farooq M. , Homayoun M. , Kardan A. , Masasa M. , Niang O. , Ramirez B. , Safi N. ( 2016 ) ‘Does Evidence Matter? Role of the Evaluation of the Child Grants Programme in the Consolidation of the Social Protection sector in Lesotho’, in Davis B. , Handa S. , Hypher N. , Winder Rossi N. , Winters P. , Yablonski J. (eds) , From Evidence to Action. The Story of Cash Transfers and Impact Evaluation in sub-Saharan Africa . Oxford : Oxford University Press . Pellerano L. , Moratti M. , Jakobsen M. , Bajgar M. , Barca V. ( 2014 ) Child Grants Programme Impact Evaluation (Follow-up Report) . Oxford : Oxford Policy Management . Pieroni L. , Lanari D. , Salmasi L. ( 2013 ) ‘ Food Prices and Overweight Patterns in Italy ’, The European Journal of Health Economics , 14 ( 1 ): 133 – 51 . Google Scholar CrossRef Search ADS PubMed Pieroni L. , Salmasi L. ( 2015 ) ‘ Does Cigarette Smoking Affect Body Weight? Causal Estimates From the Clean Indoor Air Law Discontinuity ’, Economica , 82 ( 328 ): 671 – 704 . Google Scholar CrossRef Search ADS Ravallion M. ( 1998 ) Poverty Lines in Theory and Practice . Washington DC : The World Bank . Google Scholar CrossRef Search ADS Ruel M. T. ( 2003 ) ‘ Operationalizing Dietary Diversity: A Review of Measurement Issues and Research Priorities ’, The Journal of Nutrition , 133 ( 11 ): 3911 – 26 . Google Scholar CrossRef Search ADS Samson M. , Heinrich C. J. , Hoddinott J. , Laryea-Adjei G. , Buthelezi T. , Cluver L. , Jehoma S. , Mogotsi M. , Stevens T. , van Nieckerk I. , Nyokangi E. ( 2016 ) ‘The impact of a promise realized’, in Davis B. , Handa S. , Hypher N. , Winder Rossi N. , Winters P. , Yablonski J. (eds) , From Evidence to Action. The Story of Cash Transfers and Impact Evaluation in sub-Saharan Africa . Oxford : Oxford University Press . Samson M. , Heinrich C. , Williams M. , Kaniki S. , Muzondo T. , Mac Quene K. , Van Niekerk I. ( 2008 ) Quantitative Analysis of the Impact of the Child Support Grant. Report to the Department of Social Development, South Africa, Statistics South Africa and United Nations Children’s Fund (UNICEF) . Pretoria : Republic of South Africa . Slater R. , Holmes R. , Mathers N. ( 2014 ) Food and nutrition (in)security and social protection. OECD Development Co-operation Working Papers No. 15, OECD, Paris. South African Department of Agriculture ( 2002 ) The Integrated Food Security Strategy for South Africa . Pretoria : Republic of South Africa . South African Department of Health ( 2013 ) South African National Health and Nutrition Examination Survey (SANHANES1) . Pretoria : Republic of South Africa . South African Social Security Agency, SASSA ( 2015 ) A statistical summary of social grants in South Africa. Fact sheet: Issue No. 2, Pretoria. Statistics South Africa, StatSA ( 2007 a) A Discussion Note: Constructing Comparable Household Survey Data for the Analysis Of Poverty in South Africa (1995–2000) . Pretoria : Statistics South Africa . Statistics South Africa, StatSA ( 2007 b) A National Poverty Line for South Africa . Pretoria : Statistics South Africa . Statistics South Africa, StatSA ( 2014 ) Poverty Trends in South Africa. An Examination of Absolute Poverty between 2006 and 2011 . Pretoria : Statistics South Africa . Tzioumis E. , Adair L. S. ( 2014 ) ‘ Childhood Dual Burden of Under- And Over-nutrition In Low- And Middle-income Countries: A Critical Review ’, Food and Nutrition Bulletin , 35 ( 2 ): 230 – 43 . Google Scholar CrossRef Search ADS PubMed van der Berg S. , Siebrits F. K. , Lekezwa B. ( 2010 ) Efficiency and equity effects of social grants in South Africa. Stellenbosch Economic Working Papers No.15, Stellenbosch University, Cape Town. van der Klaauw W. ( 2002 ) ‘ Estimating the Effect of Financial Aid Offers on College Enrolment: A Regression-discontinuity Approach ’, International Economic Review , 43 ( 4 ): 1249 – 87 . Google Scholar CrossRef Search ADS Vorster H. H. , Badham J. B. , Venter C. S. ( 2002 ) ‘ An Introduction to the Revised Food-based Dietary Guidelines for South Africa ’, South African Journal of Clinical Nutrition , 26 ( 3 ): 1 – 164 . Williams M. J. 2007 . The social and economic impacts of South Africa’s Child Support Grant. Economic Policy Research Institute (EPRI) Working Paper No. 39, EPRI, Cape Town. Woolard I. , Harttgen K. , Klasen S. ( 2011 ) ‘ The History and Impact of Social Security in South Africa: Experiences and Lessons ’, Canadian Journal of Development Studies , 32 ( 4 ): 357 – 80 . Google Scholar CrossRef Search ADS Woolard I. , Klasen S. ( 2005 ) ‘ Determinants of Income Mobility and Household Poverty Dynamics in South Africa ’, Journal of Development Studies , 41 ( 5 ): 865 – 97 . Google Scholar CrossRef Search ADS World Bank ( 2015 ) The State of Social Safety Nets . Washington DC : The World Bank . You J. ( 2013 ) ‘ The Role of Microcredit in Older Children’s Nutrition: Quasi-experimental Evidence from Rural China ’, Food Policy , 43 ( C ): 167 – 79 . Google Scholar CrossRef Search ADS Footnotes 1 South Africa has several types of social grants targeted to children, older persons and people with disabilities, amounting to 3.5%–4% of GDP (Department Social Development, 2010). By March 2015, these programmes had reached almost 16.5 million people, representing more than 25% of the population (South African Social Security Agency, SASSA, 2015). 2 Although food expenditure is only a proxy for food security, in the case of low-income households it estimates information closely related to access to food, which is an important dimension of food security (Burchi et al., 2016). See also Coetzee (2013) for a discussion of the shortcomings related to this index. 3 The 2008 wave can be used as baseline for running placebo experiments because the policy change in the age eligibility was introduced in 2010. 4 The grant is given for each beneficiary child up to a maximum of six children per caregiver. The transfer was fixed at a level of R100 per month in 1998, but this has increased over the years, reaching R280 in 2012 (and R320 in 2014). As from 2008, the amount of the grant is now adjusted every year for inflation. 5 The primary caregiver is defined as the person who takes primary responsibility for meeting the daily care needs of the child, without payment. In 98% of the cases, the caregiver is a woman of the household in which the child lives (Agüero et al., 2010). 6 Similarly, Heinrich and Brill (2015) showed that some children who should have been enroled in the CSG following the increase in the eligible age were disconnected from the programme because of lack of knowledge of the social welfare offices regarding the policy changes, or because of misinformation among households and high burdens placed on families in the re-application process which discouraged participation of the poorer households. 7 The NIDS is the only South African dataset that provides some detailed information (e.g., CSG receipts, reliable date of birth of beneficiary children) which are crucial for evaluating the impact of the CSG on food expenditure at the national level. In particular, the dates of birth of children younger than 15 years are written down by the interviewers using birth certificates, and a computer check allows comparing the answers of the three waves. See also Eyal and Woolard (2013) and McEwen et al. (2009). 8 Note that the attrition rate in NIDS is very low or negative. Indeed, 7,296 households successfully completed the interview during the NIDS wave 1, 6,787 households were re-interviewed during wave 2, and more households than in wave 2 (8,040) were interviewed again during wave 3. 9 The variable is the result of the aggregation of 4 separate sources of food expenditure: (i) expenditure for food items; (ii) value of food items received as gifts; (iv) value of food items received as payment; (v) value of self-produced food items. To obtain the total food expenditure, the survey considers 32 food items in 10 major categories: (i) cereals; (ii) meat; (iii) fish; (iv) dairy products; (v) fats; (vi) fruits; (vii) vegetables; (viii) sweets; (ix) beverages; (x) other food expenses. In order to yield expenditure in terms of constant prices, it was adjusted according to the monthly and provincial Consumer Price Index (CPI) (December 2012=100). The household food expenditure was also adjusted according to a per adult equivalent scale, to take into account economies of scale at the household level. Following Woolard and Klasen (2005), we applied the formula commonly used for poverty and welfare analyses in South Africa, thus obtaining the total food expenditure per adult equivalent at constant prices (AdultEquivalentScale=HouseholdIncome(Adult+0.5*Children)0.9 ). 10 The corresponding sample is composed of 1,331 households with one beneficiary child, for a total of 2,673 observations. 11 Some caution is necessary when interpreting this index. The recall period of the NIDS was fixed at 30 days to reduce the bias related to over-reporting of low frequency food purchases (Deaton and Grosh 2000). Nonetheless, recall errors in the surveys are still possible for those households that have not been purchasing certain goods. 12 For an extensive discussion of dietary diversity in South Africa, see Labadarios et al. (2011) and South African Department of Health (2013). 13 For the definition of the two poverty lines, see del Ninno and Mills (2015) and Appendix A. 14 The main shortcoming of the LP approach concerns the choice of the most appropriate bandwidth. We follow Calonico et al. (2014a) for the best choice. 15 One of the assumptions behind the RDD is that the only observed discontinuity should be on the access to CSG. 16 As stressed by Jacob et al. (2012), more sophisticated functional forms may also be used as robustness checks of the linear formulation. 17 We apply to food expenditure the procedure used by Pieroni et al. (2013) and Pieroni and Salmasi (2015) to capture non-linear patterns in food consumption. 18 The degree of the polynomial in equation (3) is derived from the polynomial degree used by the LP estimator. 19 A second robustness check is run by using placebo experiments to assess the reliability of the internal validity assumption. Since the policy change did not occur before 2010, we can use the 2008 wave to determine if and to what extent our identification strategy is able to remove the selection bias in our estimates. 20 See also Dong and Lewbel (2015). The CPD uses a linear approximation to estimate the relation between the treatment and running variable, whereas the TED uses a linear approximation to estimate the relation between the outcome and the running variable. The corresponding treatment effect is given by the ratio of the parameters estimated in the two equations. Furthermore, we take the derivative of the estimated ratio and obtain a formal test for the stability of the complier population and of the treatment effect. 21 We use a local-polynomial regression-discontinuity point estimator with robust confidence intervals proposed by Calonico et al. (2014a). 22 We do not perform the continuity test on the province and district dummy variables. 23 As showed in Table 2, the treatment and control samples are not well balanced with respect to several covariates. However, when there is no evidence of discontinuity at the cutoff, the RDD provides unbiased estimated results. 24 The results using food expenditure in per capita terms, available upon request from the authors, do not show any statistically significant difference from those presented in Table 3. 25 Using the clustering method at household level may involve a large number of clusters and, in turn, this could produce wrong standard errors. In this case, we obtain robust results also when clustering the standard errors using the provincial and district variables. Comparing different clustering methods, allows us to be confident with the estimated standard errors. For a discussion of the different clustering procedures, see Cameron and Miller (2015). 26 Note that Coetzee (2013), as most of the literature, uses food expenditure in per capita terms, whereas we are using food expenditure in terms of adult equivalents. However, we do not find any statistical difference when assessing the impact of the CSG using per capita food expenditure. 27 Omitted non-linear specifications are available from the authors. Appendix A. Poverty line boundaries We examine the Statistics South Africa (StatsSA) money metric measure of poverty in the country (Statistics South Africa, StatSA 2007a, 2007b) following a ‘cost of basic needs’ approach, as reported by Ravallion (1998). This approach determines a consumption bundle considered adequate for basic consumption needs and its estimated cost. Following del Ninno and Mills (2015), we examine two poverty lines: food poverty line (FPL): the level of consumption below which individuals are unable to purchase sufficient food to provide them with an adequate diet; poverty line (PL): the level of consumption which allows individuals to purchase both adequate food and non-food items. The FPL is calculated as the cost of satisfying the daily energy requirement for an average person for one month, which, according to the South African Medical Research Council, is 2,261 calories per person (Statistics South Africa, StatSA, 2007b). The food basket is composed of the food items commonly consumed by all expenditure-ranked household groups and usually recommended for a balanced diet. Median quantities of the reference food basket, as purchased by reference households, were then derived from data on food expenditure at the household level, with the CPI for food in September 2000. The PL includes non-food expenditure, such as accommodation, electricity, clothing, schooling for children, transport, and medical services, among other things. The two poverty lines are measured in per adult equivalent terms, assuming that resources are equally shared in the households without any differences based on age, gender, or spousal status. The two poverty lines are also expressed in rand per month and are annually adjusted according to CPI data, which track the rate of change in the price of goods and services purchased by consumers (see Table A1). Table A1: Poverty line boundaries Year Food poverty line Poverty line 2008 259 507 2010 307 594 2011 321 620 2012 339 655 Year Food poverty line Poverty line 2008 259 507 2010 307 594 2011 321 620 2012 339 655 Notes: Poverty line boundaries extracted from Statistics South Africa (StatsSA) (Statistics South Africa, StatSA 2014). View Large Table A1: Poverty line boundaries Year Food poverty line Poverty line 2008 259 507 2010 307 594 2011 321 620 2012 339 655 Year Food poverty line Poverty line 2008 259 507 2010 307 594 2011 321 620 2012 339 655 Notes: Poverty line boundaries extracted from Statistics South Africa (StatsSA) (Statistics South Africa, StatSA 2014). View Large Appendix B. Falsification tests: balance of selected covariates Figure B1. Figure B1: View largeDownload slide Figure B1: View largeDownload slide Appendix C. Falsification tests: impact of CSG on selected outcome variable, year 2008 Food expenditure Carbohydrates Dairy products IV PS IV PS IV PS Child Support Grant (CSG) 19.274 21.748 0.941 1.243 −0.032 0.044 (34.971) (33.668) (1.692) (1.512) (0.764) (0.672) Constant term 1103.221*** 1018.377*** 21.772*** 21.135*** 2.961* 2.955* (88.104) (106.131) (6.099) (6.095) (1.641) (1.608) Cragg–Donald Wald F statistic 413.206 413.206 413.206 Kleibergen–Paap Wald rk F statistic 460.191 460.191 460.191 R2 0.652 0.653 0.453 0.452 0.095 0.095 Adjusted R2 0.628 0.630 0.416 0.415 0.034 0.034 No. of observations 1139 1139 1139 1139 1139 1139 Food expenditure Carbohydrates Dairy products IV PS IV PS IV PS Child Support Grant (CSG) 19.274 21.748 0.941 1.243 −0.032 0.044 (34.971) (33.668) (1.692) (1.512) (0.764) (0.672) Constant term 1103.221*** 1018.377*** 21.772*** 21.135*** 2.961* 2.955* (88.104) (106.131) (6.099) (6.095) (1.641) (1.608) Cragg–Donald Wald F statistic 413.206 413.206 413.206 Kleibergen–Paap Wald rk F statistic 460.191 460.191 460.191 R2 0.652 0.653 0.453 0.452 0.095 0.095 Adjusted R2 0.628 0.630 0.416 0.415 0.034 0.034 No. of observations 1139 1139 1139 1139 1139 1139 Notes: The dependent variables are shares of food groups in total food expenditure per adult equivalent. In all specifications, dependent variable is total food expenditure per adult equivalent (see the list of the selected covariates in Section 2). Robust standard errors, clustered at household level, are shown in brackets. Asterisks: p-value levels ( *p<0.1;**p<0.05;***p<0.01). For each model, linear (1) and non-linear specification (2) is shown. When two-step PS procedure is applied, non-linearities are introduced only in first-stage regression. Tests for weak instrument hypothesis and first-stage F statistics and Wald statistics based on Cragg and Donald (1993) and Kleibergen and Paap (2006). †Confidence intervals for Kleibergen–Paap F test statistic follow Bazzi and Clemens (2013). Food expenditure Carbohydrates Dairy products IV PS IV PS IV PS Child Support Grant (CSG) 19.274 21.748 0.941 1.243 −0.032 0.044 (34.971) (33.668) (1.692) (1.512) (0.764) (0.672) Constant term 1103.221*** 1018.377*** 21.772*** 21.135*** 2.961* 2.955* (88.104) (106.131) (6.099) (6.095) (1.641) (1.608) Cragg–Donald Wald F statistic 413.206 413.206 413.206 Kleibergen–Paap Wald rk F statistic 460.191 460.191 460.191 R2 0.652 0.653 0.453 0.452 0.095 0.095 Adjusted R2 0.628 0.630 0.416 0.415 0.034 0.034 No. of observations 1139 1139 1139 1139 1139 1139 Food expenditure Carbohydrates Dairy products IV PS IV PS IV PS Child Support Grant (CSG) 19.274 21.748 0.941 1.243 −0.032 0.044 (34.971) (33.668) (1.692) (1.512) (0.764) (0.672) Constant term 1103.221*** 1018.377*** 21.772*** 21.135*** 2.961* 2.955* (88.104) (106.131) (6.099) (6.095) (1.641) (1.608) Cragg–Donald Wald F statistic 413.206 413.206 413.206 Kleibergen–Paap Wald rk F statistic 460.191 460.191 460.191 R2 0.652 0.653 0.453 0.452 0.095 0.095 Adjusted R2 0.628 0.630 0.416 0.415 0.034 0.034 No. of observations 1139 1139 1139 1139 1139 1139 Notes: The dependent variables are shares of food groups in total food expenditure per adult equivalent. In all specifications, dependent variable is total food expenditure per adult equivalent (see the list of the selected covariates in Section 2). Robust standard errors, clustered at household level, are shown in brackets. Asterisks: p-value levels ( *p<0.1;**p<0.05;***p<0.01). For each model, linear (1) and non-linear specification (2) is shown. When two-step PS procedure is applied, non-linearities are introduced only in first-stage regression. Tests for weak instrument hypothesis and first-stage F statistics and Wald statistics based on Cragg and Donald (1993) and Kleibergen and Paap (2006). †Confidence intervals for Kleibergen–Paap F test statistic follow Bazzi and Clemens (2013). Proteins Vitamins Dietary diversity IV PS IV PS IV PS Child Support Grant (CSG) 0.900 0.717 1.249 1.092 0.008 −0.074 (1.679) (1.451) (0.887) (0.807) (0.233) (0.206) Constant term 19.652*** 19.194*** 8.747*** 8.080** 9.502*** 9.525*** (4.160) (4.166) (3.268) (3.322) (0.684) (0.682) Cragg–Donald Wald F statistic 413.206 413.206 407.654 Kleibergen–Paap Wald rk F statistic 460.191 460.191 452.837 R2 0.215 0.215 0.109 0.117 0.394 0.394 Adjusted R2 0.162 0.162 0.049 0.058 0.353 0.353 No. of observations 1139 1139 1139 1139 1133 1133 Proteins Vitamins Dietary diversity IV PS IV PS IV PS Child Support Grant (CSG) 0.900 0.717 1.249 1.092 0.008 −0.074 (1.679) (1.451) (0.887) (0.807) (0.233) (0.206) Constant term 19.652*** 19.194*** 8.747*** 8.080** 9.502*** 9.525*** (4.160) (4.166) (3.268) (3.322) (0.684) (0.682) Cragg–Donald Wald F statistic 413.206 413.206 407.654 Kleibergen–Paap Wald rk F statistic 460.191 460.191 452.837 R2 0.215 0.215 0.109 0.117 0.394 0.394 Adjusted R2 0.162 0.162 0.049 0.058 0.353 0.353 No. of observations 1139 1139 1139 1139 1133 1133 Notes: The dependent variables are shares of food groups in total food expenditure per adult equivalent. In all specifications, dependent variable is total food expenditure per adult equivalent (see the list of the selected covariates in Section 2). Robust standard errors, clustered at household level, are shown in brackets. Asterisks: p-value levels ( *p<0.1;**p<0.05;***p<0.01). For each model, linear (1) and non-linear specification (2) is shown. When two-step PS procedure is applied, non-linearities are introduced only in first-stage regression. Tests for weak instrument hypothesis and first-stage F statistics and Wald statistics based on Cragg and Donald (1993) and Kleibergen and Paap (2006). †Confidence intervals for Kleibergen–Paap F test statistic follow Bazzi and Clemens (2013). Proteins Vitamins Dietary diversity IV PS IV PS IV PS Child Support Grant (CSG) 0.900 0.717 1.249 1.092 0.008 −0.074 (1.679) (1.451) (0.887) (0.807) (0.233) (0.206) Constant term 19.652*** 19.194*** 8.747*** 8.080** 9.502*** 9.525*** (4.160) (4.166) (3.268) (3.322) (0.684) (0.682) Cragg–Donald Wald F statistic 413.206 413.206 407.654 Kleibergen–Paap Wald rk F statistic 460.191 460.191 452.837 R2 0.215 0.215 0.109 0.117 0.394 0.394 Adjusted R2 0.162 0.162 0.049 0.058 0.353 0.353 No. of observations 1139 1139 1139 1139 1133 1133 Proteins Vitamins Dietary diversity IV PS IV PS IV PS Child Support Grant (CSG) 0.900 0.717 1.249 1.092 0.008 −0.074 (1.679) (1.451) (0.887) (0.807) (0.233) (0.206) Constant term 19.652*** 19.194*** 8.747*** 8.080** 9.502*** 9.525*** (4.160) (4.166) (3.268) (3.322) (0.684) (0.682) Cragg–Donald Wald F statistic 413.206 413.206 407.654 Kleibergen–Paap Wald rk F statistic 460.191 460.191 452.837 R2 0.215 0.215 0.109 0.117 0.394 0.394 Adjusted R2 0.162 0.162 0.049 0.058 0.353 0.353 No. of observations 1139 1139 1139 1139 1133 1133 Notes: The dependent variables are shares of food groups in total food expenditure per adult equivalent. In all specifications, dependent variable is total food expenditure per adult equivalent (see the list of the selected covariates in Section 2). Robust standard errors, clustered at household level, are shown in brackets. Asterisks: p-value levels ( *p<0.1;**p<0.05;***p<0.01). For each model, linear (1) and non-linear specification (2) is shown. When two-step PS procedure is applied, non-linearities are introduced only in first-stage regression. Tests for weak instrument hypothesis and first-stage F statistics and Wald statistics based on Cragg and Donald (1993) and Kleibergen and Paap (2006). †Confidence intervals for Kleibergen–Paap F test statistic follow Bazzi and Clemens (2013). © The Author(s) 2017. Published by Oxford University Press on behalf of the Centre for the Study of African Economies, all rights reserved. For Permissions, please email: journals.permissions@oup.com This article is published and distributed under the terms of the Oxford University Press, Standard Journals Publication Model (https://academic.oup.com/journals/pages/about_us/legal/notices) http://www.deepdyve.com/assets/images/DeepDyve-Logo-lg.png Journal of African Economies Oxford University Press

Do Cash Transfers Promote Food Security? The Case of the South African Child Support Grant

Loading next page...
 
/lp/ou_press/do-cash-transfers-promote-food-security-the-case-of-the-south-african-W9HMkCkRzU
Publisher
Oxford University Press
Copyright
© The Author(s) 2017. Published by Oxford University Press on behalf of the Centre for the Study of African Economies, all rights reserved. For Permissions, please email: journals.permissions@oup.com
ISSN
0963-8024
eISSN
1464-3723
D.O.I.
10.1093/jae/ejx041
Publisher site
See Article on Publisher Site

Abstract

Abstract This paper evaluates the causal effect of the Child Support Grant (CSG) implemented in South Africa on household food consumption and dietary diversity. The analysis uses the National Income Dynamics Study (NIDS) covering 2008, 2010–2011, and 2012, and carries out a regression-discontinuity design exploiting the increase in the age limit requirement for eligibility for the programme. Our results show that the CSG has proved to be effective in increasing total food expenditure per adult equivalent but has not significantly changed the dietary habits of the beneficiary households, nor has the programme resulted in any stronger effect for the most vulnerable subgroups of the beneficiary population. To analyse the external and internal validities of the results, a comparison between non-parametric, semi-parametric and parametric estimates is presented. 1. Introduction A number of studies have recently evaluated the impact of cash transfers (CTs) targeted to the poorest and most vulnerable people in sub-Saharan African countries, and a positive role in promoting human capital, health and productive activities has been documented (World Bank 2015). CTs may also affect food security because, by addressing the lack of purchasing power of poor households, they contribute to improving food purchases and access to more good-quality food (Bassett, 2008, Alderman, 2014, Burchi et al., 2016). However, the evidence for the impact of CTs on food security in sub-Saharan Africa is not conclusive since, even if most of the literature shows a positive impact on the consumption level of food, it is still not clear whether the consumption gains are translated into improved nutritional status for household members (Manley et al., 2013, Alderman, 2014, Slater et al., 2014, Burchi et al., 2016, Davis and Handa, 2016). In this paper, we contribute to this literature by evaluating the impact on food security of the Child Support Grant (CSG), a CT programme introduced in South Africa in 1998 with the aim of supporting vulnerable children and poor households. At the demise of the apartheid regime in 1994, South Africa underwent significant political and social advances and rapid economic growth, both of which have led to steady progress in reducing poverty (Agüero et al., 2007, Leibbrandt et al., 2011, Leibbrandt and Levinsohn, 2011). However, food insecurity remains widespread: 35% of the population is vulnerable (Kirsten, 2012) and about 25% of children under age of 6 are classified as stunted by malnutrition (Manyamba et al., 2012). Food insecurity in South Africa is not due to a shortage of food, but rather to insufficient access as a result of structural poverty and inequality dynamics with a strong racial footprint (Aliber, 2001, Du Toit, 2011, Manyamba et al., 2012). South Africa has no CT directly aimed at reducing food insecurity. In this context, CT programmes aimed at eradicating extreme poverty1 play a central role in addressing food insecurity and represent one of the pillars of the Integrated Food Security Strategy established in 1996 (South African Department of Agriculture, 2002). Although the CSG is the major CT programme in South Africa, its impact on food security has only been analysed in a few studies. Using data from the General Household Survey and the Labour Force Survey, Williams (2007) found that CSG beneficiaries aged between 7 and 17 years more probably attended school and suffered less hunger. Samson et al. (2008) showed a reduction in hunger among children under 7 years of age, using a household panel extracted by the Economic Policy Research Institute (EPRI) from repeated cross-sections of the National General Household Surveys. Agüero et al. (2010), using data from the KwaZulu-Natal Income Dynamic Study, evaluated the impact of the CSG on anthropometric indicators which are widely used in nutritional assessment. Focusing on children in the first 36 months of life, the study found that the CSG improved childhood nutrition as measured by child height-for-age. Heinrich et al. (2012) evaluated the impact of the CSG on the well-being of children in early life. Considering aspects related to food security, the study showed that the CSG significantly improved nutrition, as measured by the height for age z-scores, in the cases of children whose mothers have more than 8 grades of schooling. Using data extracted from the first wave of a longitudinal survey provided by the National Income Dynamics Study (NIDS), Coetzee (2013) evaluated the impact of the CSG on health, height-for-age and education of beneficiary children under 14 and she found a few positive effects on their well-being. Similarly, she found a positive, albeit small, impact of the programme on household food security measured by the per capita expenditure on food items. Unlike previous studies, Delany et al. (2008) considered the impact of the programme at the households level. Analysing an original survey of households in low-income areas and focus group discussions, Delany et al. (2008) found that beneficiary households allocated a larger proportion of their expenditure to essential goods, such as food, and that more than three-quarters of the CSG recipients stated that food was the main expenditure covered by the grant. Overall, these studies indicate that the CSG has positive effects on outcome variables connected to the food security of beneficiary children and that the programme also slightly increased households’ food expenditure. With respect to the previous literature, our study provides an evaluation of the impact of the CSG which is focused on food expenditure at the household level and extends the analysis by considering different drivers of food security. We explore whether participation in the CSG is effective in improving the total food expenditure per adult equivalent as well as dietary diversity for the beneficiary families who have a child that was eligible up to the age of 18.2 To analyse dietary diversity, which gives us information about the quality of the food consumed, we use the share in food expenditure of the primary food groups (e.g. carbohydrates, dairy products, proteins and vitamins) and a synthetic index defined as the number of different food groups consumed in that period (Hoddinott and Yohannes, 2002, Ruel, 2003, Hoddinott and Bassett, 2008). By considering these measures jointly, we provide a more comprehensive investigation of the impact of the CSG on food security in South Africa and, at least partially, we fill a gap in the literature. Within this framework, we exploit the discontinuous variation induced by the expansion in eligibility in terms of the child’s age. We take into consideration the fact that from 1 January 2010, eligibility for the CSG was extended for children born on or after 1 January 1994 until their 18th birthday; those born before 1994 had lost their eligibility at age 14. As a result of this policy change, a discontinuous increase occurred in the probability of being a CSG beneficiary during the age interval 14–17 for children born after 1 January 1994. This discontinuity provides a natural experiment for examining the causal effects of the programme across birth cohorts, suitable for running a Regression-Discontinuity Design (RDD). Using this identification structure, we estimate the effect of the CSG by using the local-polynomial (LP) estimator (Calonico et al., 2014a) and then compare the results with the parametric estimates obtained with an instrumental variable (IV) approach. To overcome problems related to potential selection bias and threats to internal validity, we apply, following van der Klaauw (2002), a two-step propensity score (PS) procedure and placebo experiments which use the panel structure of the data to replicate the analysis only for the 2008 wave.3 We also employ a formal test strategy implemented by Cerulli et al. (2017) to address other important threats to the stability of the compliers (i.e., the eligible people who have received the treatment after the policy change) in the neighbourhood of the cutoff and to the external validity of the analysis. Finally, as a further robustness check, we ascertain that the causal estimates are robust within population subgroups and use elasticity measures based on the two-step propensity score to compare the potential outcome of the subgroups with that of the entire population and to account for the differences between them (Angrist, 2004). Our results show that the monetary transfers provided by the CSG led to an increase in total food expenditure per adult equivalent, with an impact which ranges between R54 and R55 (corresponding to a mean increase of 10%). This result was confirmed by the comparison between non-parametric, semi-parametric, and parametric estimations, and thus is very robust. Our analysis also shows that the estimated parameters are constant away from the cutoff and that the results are homogeneous across various population subgroups. We may thus interpret the causal effects of the CSG unconditionally to the birth cohorts which are not close to the cutoff. In addition, the analysis shows that the CSG did not have a greater effect on the population subgroups which were more vulnerable to food insecurity than it did on the others. When dietary diversity is analysed, we find robust positive results only for carbohydrates, which represent the major food group in the total food expenditure of the treated households (with a share of 23%). However, we do not find any significant result when the synthetic index of dietary diversity is considered. This analysis indicates that the CSG did not lead to significant changes in the dietary habits of the beneficiary households and did not improve the dietary diversity of the poor households. While there are caveats pertaining to our outcome variables and estimates, as usual in a non-experimental approach, we think that our work provides additional and valuable information about the impact of the CSG on several dimensions of the food security of the beneficiary households. This paper proceeds as follows: Section 2 describes the dataset and the identification strategy; Section 3 provides an overview of the empirical framework used to estimate the causal impact of the CSG on food expenditure and dietary diversity; Section 4 discusses the results, and Section 5 draws some conclusions and outlines avenues for future research 2. Identification strategy and dataset 2.1 Identification strategy CSG benefits are provided each month to eligible beneficiaries4 and are paid to the primary caregiver.5 As mentioned in Section 1, the eligible population is determined according to a means test and the child’s age, and these criteria have changed over time. Figure 1 shows the timeline of policy changes in the eligible population and in the amount of the grant, from 1998, when the programme was introduced, to the last year of our evaluation (2012). Figure 1: View largeDownload slide Timeline of CSG Implementation. Figure 1: View largeDownload slide Timeline of CSG Implementation. In the case of the income criteria, the means test was initially based on household income, and the ceiling was fixed at the nominal level of R800 in urban areas and R1,100 in rural areas for 10 years. However, in 1999, to increase take-up rates, the government altered this rule to one which considered only the income of the primary caregiver plus her/his spouse (Agüero et al., 2010, Woolard et al., 2011). Eligibility was expanded again in 2008: the Department of Social Development defined the income ceiling as 10 times the value of the grant paid to the single primary caregiver of the child (double for married caregivers), so that the means test would automatically keep pace with inflation (Agüero et al., 2010, Woolard et al., 2011). Figure 1 also shows changes in age limit criteria. When the programme was introduced in 1998, age eligibility was limited to children under 7 years old, but it was later gradually raised: in 2003, it was extended to children up to their 9th birthday, in 2004 up to their 11th, and in 2005 up to their 14th. From 1 January 2010, eligibility was further extended, so that children born after 1 January 1994 were eligible until their 18th birthday, whereas those born before that date lost eligibility at 14 (van der Berg et al., 2010, Woolard et al., 2011, Bor 2013). A number of problems in the implementation of the CSG and of policy changes have been well documented (Heinrich et al., 2012, Heinrich and Brill 2015, Samson et al., 2016, Heinrich et al., 2017) and we are aware that, for this reason, serious threats to the internal and external validity arise in our identification strategy. In particular, programme rules and administrative capacity affected the access and the duration of the reception of the benefits (Heinrich and Brill 2015, Heinrich et al., 2017), influencing the effectiveness of the programme. In more detail, Heinrich et al. (2012) stressed the barriers to application due to misinformation about the eligibility criteria and difficulties in ensuring that awareness of the changes in the targeting rules were disseminated to populations that are often disconnected from reliable sources of information.6 Problems with getting the correct documentation were another frequent cause of delay in applications or deterrence from would-be recipients. These and related factors associated with barriers to access for eligible families have been reduced over time but they are still present, and challenge the assumption that all eligible children are enroled in the programme once the policy change is implemented. In addition, administrative burdens may vary geographically or politically, and thus the influence on the take-up rates may be uneven. Following this outline, three major concerns arise in our analysis. First of all, not all households with a child in the eligible age range receive the CSG, since they must also meet an eligibility requirement based on an income means test, and this criterion has changed over time. A selection bias may therefore emerge if we do not take into account how the participation in the programme varies with changes in the income requirement. Secondly, administrative costs, bureaucratic requirements, and implementation problems have limited the participation in the CSG (Heinrich et al., 2012, Heinrich and Brill, 2015, Samson et al., 2016, Heinrich et al., 2017). These issues may violate both the the internal and external validity assumptions in the RD design and may imply instability in the population of compliers in the neighbourhood of the cutoff. Lastly, significant differences may emerge in this analysis because the CSG was designed to focus on poor households which had been excluded from social assistance programmes during apartheid (Pauw and Mncube, 2007), particularly Africans, and those living in marginalised rural areas (Lund et al., 2008). Note that these groups are particularly vulnerable to the problems related to the design and implementations of the CSG. Thus, we expected that the potential outcome would not be homogeneous in these subgroups with respect to the entire population (Angrist, 2004). 2.2 Dataset The NIDS is a dataset implemented by the South African Labour and Development Research Unit (SALDRU) of the University of Cape Town, and is the first nation-wide panel study in South Africa.7 The NIDS is available for the waves 2008, 2010–2011, and 2012, and allows a face-to-face longitudinal survey of households residing in South Africa. Its aim is to follow a sample of household members and register the changes in household compositions, migration, and several dimensions of well-being (e.g., income, expenditure, assets, access to social services, education, health, employment). From the entire dataset, we extracted a sample of households that responded to the three waves of the survey and whose composition remained unchanged during this period, to ensure the absence of migration between different provinces and areas. In the second step, households with no children or with children who were not born in 1990–1998 were dropped. This procedure was followed because the evaluation required a sample of households with children in the age range 14–17 who experienced the policy change. Since our panel data covers the period 2008–2012, we needed to include children born after 1990, and to remove those born from 1998 onwards, who were always eligible for the CSG. Using this framework, we obtained a sample of households with one child in the age range from 10 (in 2008) to 22 (in 2012). This sample consists of 1,336 households for a total of 4,011 observations.8 The dataset provides expenditure data at the household level and includes both the total food expenditure and the food expenditure disaggregated by food group. We used these variables according to our interest in analysing aspects of food security related to economic access to a sufficient quantity of food (Coetzee, 2013, Burchi et al., 2016) and to the the quality of the food consumed (Hoddinott and Yohannes, 2002, Ruel, 2003, Hoddinott and Bassett, 2008). In more detail, the first variable of interest is monthly food expenditure at the household level, adjusted by an adult equivalent scale.9 The variables related to food groups were obtained by disaggregating the total food expenditure per adult equivalent into four groups: (i) carbohydrates, given by the sum of the expenditure on cereals such as samp, flour and bread, mealie meal, rice and pasta; (ii) dairy products; (iii) proteins, given by the sum of expenditure on meat and fish; (iv) vitamins, comprising expenditure on fruits and vegetables. The analysis of the food groups was carried out for 2008–2011, since the NIDS stopped reporting such data after 2011.10 Taking into account that there is no consensus on which food group and classification system is the most adequate (Ruel, 2003), we have also considered a synthetic indicator of dietary diversity, obtained by following the recommendations of the FAO’s ‘Guidelines for measuring household and individual dietary diversity’ (Kennedy et al., 2011, Vorster et al., 2013). First of all, we defined 12 food groups: (i) cereals; (ii) roots and white tubers; (iii) vegetables; (iv) fruits; (v) meat; (vi) eggs; (vii) fish and other seafood; (viii) legumes, nuts and seeds; (ix) milk and milk products; (x) fats and oils; (xi) sweets; (xii) spices, condiments and beverages. Secondly, we generated 12 dummy variables that register when a household consumed, in the last 30 days, at least one item in the corresponding food group. Lastly, we summed these dummy variables to obtain a score that ranges from 1 to 12, where 1 indicates households consuming only one food group, and 12 households consuming all the available food groups.11 Table 1 presents the means and standard deviations (s.d.) for the outcome variables in question, separated into control and treated groups. As an initial outcome, the first row of the table shows that considerable differences arise when the total food expenditure per adult equivalent is considered: the mean in the treated group is about R394, compared with about R556 in the control group. The standard deviation is also much higher in the control group. Table 1: Total food expenditure per adult equivalent and shares of food groups Full-sample Control Treatment Mean s.d. Mean s.d. Mean s.d. Total food expenditure per adult equivalent 500.460 497.196 556.419 532.096 394.815 403.085 Share of food groups in total food expenditure:  Carbohydrates 19.401 18.532 17.476 17.366 23.035 20.065  Dairy products 4.203 5.123 4.257 5.119 4.101 5.130  Proteins 15.939 15.911 16.384 16.480 15.099 14.747  Vitamins 5.857 6.386 5.828 6.342 5.912 6.471  Dietary diversity 9.777 2.244 9.954 2.223 9.264 2.228 Full-sample Control Treatment Mean s.d. Mean s.d. Mean s.d. Total food expenditure per adult equivalent 500.460 497.196 556.419 532.096 394.815 403.085 Share of food groups in total food expenditure:  Carbohydrates 19.401 18.532 17.476 17.366 23.035 20.065  Dairy products 4.203 5.123 4.257 5.119 4.101 5.130  Proteins 15.939 15.911 16.384 16.480 15.099 14.747  Vitamins 5.857 6.386 5.828 6.342 5.912 6.471  Dietary diversity 9.777 2.244 9.954 2.223 9.264 2.228 Notes: Total food expenditure per adult equivalent is monthly expenditure at household level, adjusted by an adult equivalent scale. Analysis of food groups and of synthetic index of dietary diversity is valid for period 2008–2011, as NIDS stopped reporting such data after 2011. View Large Table 1: Total food expenditure per adult equivalent and shares of food groups Full-sample Control Treatment Mean s.d. Mean s.d. Mean s.d. Total food expenditure per adult equivalent 500.460 497.196 556.419 532.096 394.815 403.085 Share of food groups in total food expenditure:  Carbohydrates 19.401 18.532 17.476 17.366 23.035 20.065  Dairy products 4.203 5.123 4.257 5.119 4.101 5.130  Proteins 15.939 15.911 16.384 16.480 15.099 14.747  Vitamins 5.857 6.386 5.828 6.342 5.912 6.471  Dietary diversity 9.777 2.244 9.954 2.223 9.264 2.228 Full-sample Control Treatment Mean s.d. Mean s.d. Mean s.d. Total food expenditure per adult equivalent 500.460 497.196 556.419 532.096 394.815 403.085 Share of food groups in total food expenditure:  Carbohydrates 19.401 18.532 17.476 17.366 23.035 20.065  Dairy products 4.203 5.123 4.257 5.119 4.101 5.130  Proteins 15.939 15.911 16.384 16.480 15.099 14.747  Vitamins 5.857 6.386 5.828 6.342 5.912 6.471  Dietary diversity 9.777 2.244 9.954 2.223 9.264 2.228 Notes: Total food expenditure per adult equivalent is monthly expenditure at household level, adjusted by an adult equivalent scale. Analysis of food groups and of synthetic index of dietary diversity is valid for period 2008–2011, as NIDS stopped reporting such data after 2011. View Large In addition, when the four shares of food expenses in the total food expenditure per adult equivalent are compared, Table 1 shows that the largest difference between treated and control groups is in the mean value of the share of carbohydrates, which is about 6% higher in the treated group. As has been shown in the literature, the poorest people in South Africa have little dietary variety, and consume more cereals and fewer items in the other food groups, than does the rest of the population.12 However, the last row of the table shows that there are small differences in terms of dietary diversity, when the treated and control groups are analysed. Table 2 lists household characteristics in the entire sample (for both 2008–2012 and 2008–2011) and both control and treated groups. In the sample for 2008–2012, the first remarkable difference between the groups is due to the share of households living under the food poverty line,13 defined as the threshold under which individuals cannot purchase sufficient food to provide them with an adequate diet. This share is 44% of treated units and only 32% of controls. Other significant differences are found when geographical location, race, gender and employment status are considered. In more detail, we find that a higher percentage of households in the treated group live in rural areas (57%), and have household heads who are Africans (86%), women (69%), or receiving no education (26.5%), compared with the control group. Table 2: Descriptive statistics Sample 2008–2012 Sample 2008–2011 Sample Control Treated Difference Sample Control Treated Difference Province   Western Cape 0.146 0.175 0.088 0.087* 0.144 0.173 0.088 0.085*   Eastern Cape 0.134 0.124 0.154 −0.030 0.135 0.126 0.154 −0.028   Northern Cape 0.064 0.070 0.053 0.017 0.063 0.069 0.053 0.016   Free State 0.084 0.093 0.067 0.026 0.085 0.094 0.067 0.027   KwaZulu-Natal 0.235 0.197 0.310 −0.113* 0.237 0.198 0.311 −0.113*   North West 0.051 0.043 0.066 −0.023* 0.051 0.044 0.066 −0.022*   Gauteng 0.117 0.140 0.074 0.066* 0.114 0.135 0.075 0.060*   Mpumalanga 0.070 0.065 0.081 −0.016 0.070 0.065 0.078 −0.013   Limpopo 0.098 0.094 0.107 −0.013 0.100 0.096 0.108 −0.012 Poverty   Food poverty line 0.362 0.321 0.442 −0.121* 0.293 0.255 0.368 −0.113*   Poverty line 0.329 0.309 0.369 −0.060* 0.338 0.307 0.399 −0.092* Geographical location   Rural 0.442 0.374 0.573 −0.199* 0.450 0.383 0.580 −0.197*   Urban 0.558 0.626 0.427 0.199* 0.550 0.617 0.420 0.197* Number of household members   1 or 2 0.135 0.147 0.113 0.034* 0.124 0.133 0.105 0.028*   3 0.227 0.211 0.258 −0.047* 0.235 0.215 0.274 −0.059*   4 0.212 0.213 0.211 0.002 0.216 0.215 0.218 −0.003   5 0.174 0.179 0.165 0.014 0.173 0.179 0.162 0.017   ≥6 0.252 0.250 0.254 −0.004 0.252 0.258 0.240 0.018 Head of household by race   African 0.768 0.719 0.862 −0.143* 0.768 0.719 0.862 −0.143*   Coloured 0.171 0.195 0.124 0.071* 0.171 0.195 0.124 0.071*   Asian/Indian/White 0.061 0.086 0.014 0.072* 0.061 0.085 0.014 0.071* Head of household by gender   Female 0.617 0.578 0.692 −0.114* 0.588 0.550 0.661 −0.111*   Male 0.383 0.422 0.308 0.114* 0.412 0.450 0.339 0.111* Head of household by education   No schooling 0.182 0.138 0.265 −0.127* 0.201 0.153 0.291 −0.138*   Primary 0.185 0.154 0.246 −0.092* 0.196 0.165 0.255 −0.090*   Secondary 0.515 0.542 0.464 0.078* 0.499 0.532 0.437 0.095*   Tertiary 0.118 0.167 0.025 0.142* 0.105 0.151 0.018 0.133* Head of household by age 49.085 48.166 50.869 −2.703 49.967 49.181 51.491 −2.310 Sample 2008–2012 Sample 2008–2011 Sample Control Treated Difference Sample Control Treated Difference Province   Western Cape 0.146 0.175 0.088 0.087* 0.144 0.173 0.088 0.085*   Eastern Cape 0.134 0.124 0.154 −0.030 0.135 0.126 0.154 −0.028   Northern Cape 0.064 0.070 0.053 0.017 0.063 0.069 0.053 0.016   Free State 0.084 0.093 0.067 0.026 0.085 0.094 0.067 0.027   KwaZulu-Natal 0.235 0.197 0.310 −0.113* 0.237 0.198 0.311 −0.113*   North West 0.051 0.043 0.066 −0.023* 0.051 0.044 0.066 −0.022*   Gauteng 0.117 0.140 0.074 0.066* 0.114 0.135 0.075 0.060*   Mpumalanga 0.070 0.065 0.081 −0.016 0.070 0.065 0.078 −0.013   Limpopo 0.098 0.094 0.107 −0.013 0.100 0.096 0.108 −0.012 Poverty   Food poverty line 0.362 0.321 0.442 −0.121* 0.293 0.255 0.368 −0.113*   Poverty line 0.329 0.309 0.369 −0.060* 0.338 0.307 0.399 −0.092* Geographical location   Rural 0.442 0.374 0.573 −0.199* 0.450 0.383 0.580 −0.197*   Urban 0.558 0.626 0.427 0.199* 0.550 0.617 0.420 0.197* Number of household members   1 or 2 0.135 0.147 0.113 0.034* 0.124 0.133 0.105 0.028*   3 0.227 0.211 0.258 −0.047* 0.235 0.215 0.274 −0.059*   4 0.212 0.213 0.211 0.002 0.216 0.215 0.218 −0.003   5 0.174 0.179 0.165 0.014 0.173 0.179 0.162 0.017   ≥6 0.252 0.250 0.254 −0.004 0.252 0.258 0.240 0.018 Head of household by race   African 0.768 0.719 0.862 −0.143* 0.768 0.719 0.862 −0.143*   Coloured 0.171 0.195 0.124 0.071* 0.171 0.195 0.124 0.071*   Asian/Indian/White 0.061 0.086 0.014 0.072* 0.061 0.085 0.014 0.071* Head of household by gender   Female 0.617 0.578 0.692 −0.114* 0.588 0.550 0.661 −0.111*   Male 0.383 0.422 0.308 0.114* 0.412 0.450 0.339 0.111* Head of household by education   No schooling 0.182 0.138 0.265 −0.127* 0.201 0.153 0.291 −0.138*   Primary 0.185 0.154 0.246 −0.092* 0.196 0.165 0.255 −0.090*   Secondary 0.515 0.542 0.464 0.078* 0.499 0.532 0.437 0.095*   Tertiary 0.118 0.167 0.025 0.142* 0.105 0.151 0.018 0.133* Head of household by age 49.085 48.166 50.869 −2.703 49.967 49.181 51.491 −2.310 Notes: Descriptive statistics of district of residence of household and interview’s month are omitted. For definitions of food poverty line and poverty line, see Appendix A. Asterisk mark indicates when the difference between the treated and control samples is statistically significant at 5% level. View Large Table 2: Descriptive statistics Sample 2008–2012 Sample 2008–2011 Sample Control Treated Difference Sample Control Treated Difference Province   Western Cape 0.146 0.175 0.088 0.087* 0.144 0.173 0.088 0.085*   Eastern Cape 0.134 0.124 0.154 −0.030 0.135 0.126 0.154 −0.028   Northern Cape 0.064 0.070 0.053 0.017 0.063 0.069 0.053 0.016   Free State 0.084 0.093 0.067 0.026 0.085 0.094 0.067 0.027   KwaZulu-Natal 0.235 0.197 0.310 −0.113* 0.237 0.198 0.311 −0.113*   North West 0.051 0.043 0.066 −0.023* 0.051 0.044 0.066 −0.022*   Gauteng 0.117 0.140 0.074 0.066* 0.114 0.135 0.075 0.060*   Mpumalanga 0.070 0.065 0.081 −0.016 0.070 0.065 0.078 −0.013   Limpopo 0.098 0.094 0.107 −0.013 0.100 0.096 0.108 −0.012 Poverty   Food poverty line 0.362 0.321 0.442 −0.121* 0.293 0.255 0.368 −0.113*   Poverty line 0.329 0.309 0.369 −0.060* 0.338 0.307 0.399 −0.092* Geographical location   Rural 0.442 0.374 0.573 −0.199* 0.450 0.383 0.580 −0.197*   Urban 0.558 0.626 0.427 0.199* 0.550 0.617 0.420 0.197* Number of household members   1 or 2 0.135 0.147 0.113 0.034* 0.124 0.133 0.105 0.028*   3 0.227 0.211 0.258 −0.047* 0.235 0.215 0.274 −0.059*   4 0.212 0.213 0.211 0.002 0.216 0.215 0.218 −0.003   5 0.174 0.179 0.165 0.014 0.173 0.179 0.162 0.017   ≥6 0.252 0.250 0.254 −0.004 0.252 0.258 0.240 0.018 Head of household by race   African 0.768 0.719 0.862 −0.143* 0.768 0.719 0.862 −0.143*   Coloured 0.171 0.195 0.124 0.071* 0.171 0.195 0.124 0.071*   Asian/Indian/White 0.061 0.086 0.014 0.072* 0.061 0.085 0.014 0.071* Head of household by gender   Female 0.617 0.578 0.692 −0.114* 0.588 0.550 0.661 −0.111*   Male 0.383 0.422 0.308 0.114* 0.412 0.450 0.339 0.111* Head of household by education   No schooling 0.182 0.138 0.265 −0.127* 0.201 0.153 0.291 −0.138*   Primary 0.185 0.154 0.246 −0.092* 0.196 0.165 0.255 −0.090*   Secondary 0.515 0.542 0.464 0.078* 0.499 0.532 0.437 0.095*   Tertiary 0.118 0.167 0.025 0.142* 0.105 0.151 0.018 0.133* Head of household by age 49.085 48.166 50.869 −2.703 49.967 49.181 51.491 −2.310 Sample 2008–2012 Sample 2008–2011 Sample Control Treated Difference Sample Control Treated Difference Province   Western Cape 0.146 0.175 0.088 0.087* 0.144 0.173 0.088 0.085*   Eastern Cape 0.134 0.124 0.154 −0.030 0.135 0.126 0.154 −0.028   Northern Cape 0.064 0.070 0.053 0.017 0.063 0.069 0.053 0.016   Free State 0.084 0.093 0.067 0.026 0.085 0.094 0.067 0.027   KwaZulu-Natal 0.235 0.197 0.310 −0.113* 0.237 0.198 0.311 −0.113*   North West 0.051 0.043 0.066 −0.023* 0.051 0.044 0.066 −0.022*   Gauteng 0.117 0.140 0.074 0.066* 0.114 0.135 0.075 0.060*   Mpumalanga 0.070 0.065 0.081 −0.016 0.070 0.065 0.078 −0.013   Limpopo 0.098 0.094 0.107 −0.013 0.100 0.096 0.108 −0.012 Poverty   Food poverty line 0.362 0.321 0.442 −0.121* 0.293 0.255 0.368 −0.113*   Poverty line 0.329 0.309 0.369 −0.060* 0.338 0.307 0.399 −0.092* Geographical location   Rural 0.442 0.374 0.573 −0.199* 0.450 0.383 0.580 −0.197*   Urban 0.558 0.626 0.427 0.199* 0.550 0.617 0.420 0.197* Number of household members   1 or 2 0.135 0.147 0.113 0.034* 0.124 0.133 0.105 0.028*   3 0.227 0.211 0.258 −0.047* 0.235 0.215 0.274 −0.059*   4 0.212 0.213 0.211 0.002 0.216 0.215 0.218 −0.003   5 0.174 0.179 0.165 0.014 0.173 0.179 0.162 0.017   ≥6 0.252 0.250 0.254 −0.004 0.252 0.258 0.240 0.018 Head of household by race   African 0.768 0.719 0.862 −0.143* 0.768 0.719 0.862 −0.143*   Coloured 0.171 0.195 0.124 0.071* 0.171 0.195 0.124 0.071*   Asian/Indian/White 0.061 0.086 0.014 0.072* 0.061 0.085 0.014 0.071* Head of household by gender   Female 0.617 0.578 0.692 −0.114* 0.588 0.550 0.661 −0.111*   Male 0.383 0.422 0.308 0.114* 0.412 0.450 0.339 0.111* Head of household by education   No schooling 0.182 0.138 0.265 −0.127* 0.201 0.153 0.291 −0.138*   Primary 0.185 0.154 0.246 −0.092* 0.196 0.165 0.255 −0.090*   Secondary 0.515 0.542 0.464 0.078* 0.499 0.532 0.437 0.095*   Tertiary 0.118 0.167 0.025 0.142* 0.105 0.151 0.018 0.133* Head of household by age 49.085 48.166 50.869 −2.703 49.967 49.181 51.491 −2.310 Notes: Descriptive statistics of district of residence of household and interview’s month are omitted. For definitions of food poverty line and poverty line, see Appendix A. Asterisk mark indicates when the difference between the treated and control samples is statistically significant at 5% level. View Large 3. Empirical framework We will consider a standard RDD model, where T is a binary treatment indicator (treated by CSG), X is an assignment variable (in our analysis, the birth cohort of the child) and c is the threshold for X at which the probability of treatment changes discontinuously (1 January 1994). In the fuzzy RD design, the treatment effect is obtained by dividing the increase in the relation between Y (the outcome variable) and X at c by the difference of the probability that an applicant household will be treated before or after the cutoff. The treatment effect (denoted by τF) n the fuzzy RD design is τF=limϵ↓0E(Y∣X=c+ϵ)−limϵ↑0E(Y∣X=c+ϵ)limϵ↓0E(T∣X=c+ϵ)−limϵ↑0E(T∣X=c+ϵ). (1) where ϵ defines the neighbourhood in which local random assignment is satisfied. When local random assignment holds, equation (1) provides an unbiased estimate of a weighted version of the LATE, which evaluates the impact of the programme on individuals who were assigned to the treatment group and actually participated in it (compliers) compared with those who were assigned to the treatment group but did not participate in the programme (non-compliers) (Jacob et al., 2012). Equation (1) can be estimated using either parametric or non-parametric methods. The first approach followed here is to estimate equation (1) through a local-polynomial (LP) estimator (Calonico et al., 2014b), which approximates the regression functions above and below the cutoff by means of weighted polynomial regressions, with weights computed by applying a kernel function to the distance of each observation from the cutoff.14 The LP does not impose any functional form or introduce any relevant characteristic of the household and, hence, represents a benchmark for parametric and semi-parametric estimations. In addition, when the results estimated through the LP are not statistically different from the ones estimated with parametric methods, described below, we can indirectly show that the relevant covariates are balanced at the cutoff.15 Further, since equation (1) shows a close analogy between the fuzzy RD design with the Wald formulation of the treatment effect, an IV setting can also be applied. The only further requirement is the monotonicity and excludability related to the assignment variable when crossing the cutoff (Hahn et al., 2001). To estimate equation (1) in an IV framework, when the monotonicity assumption holds, we impose a linear form on the two-sided relation between the assignment variable, the treatment dummy, and the outcome variable.16 The IV equation is Y=δ0+δ1T+W1′δ2+P′δ3+Ω. (2) where W1 is a matrix of the relevant characteristics of the households and heads of households (described in Section 2.2) and Ω is an error term. The matrix P includes a second-order polynomial time trend ( trend and trend2) to take into account non-linear patterns of food expenditure,17 a set of regional Dr and district Dm dummies and their interactions with the second-order polynomial time trend. To take into account non-linearities, we also present estimation results obtained by introducing a third-degree polynomial of the assignment variable into equation (3).18 As shown in Section 2.1, the first shortcoming of the analysis concerns the reliability of the internal validity of the RDD design, since we can only identify the impact of the CSG on the outcome variable caused by the variation of the age limit, but we do not have complete control over the effects of changes in the income eligibility rule. In addition, households are able to manipulate their income threshold. Thus, the internal validity assumption may be not valid. Similarly, a threat to internal validity emerges considering implementation concerns, as outlined in Section 2.1. To overcome these problems, we propose the following robustness analyses. First of all, following van der Klaauw (2002), the selection bias may be overcome by replacing T with its propensity score E[T∣X] in equation (3). In this case, a two-step procedure is required and, thus, in the first stage we specify the PS function in the fuzzy RD design (van der Klaauw 2002, You 2013). Hence, the propensity score in the RD framework is E[T∣X]=f(X)+μ1(S≥c) (3) where f(X) is a continuous function in c which may be estimated parametrically or semi-parametrically. In the present context, it is defined as a third-degree polynomial of the assignment variable. The estimated propensity score can then replace the treatment variable T in equation (3) to estimate δ1. The second-step equation is Y=δ0+δ1[T∣X]+W′δ2+P′δ3+Ψ (4) so that we can compare the parameters estimated with the IV method with those obtained with the two-step PS procedure. When we find a statistically significant difference between them, we can consider whether the internal validity assumption may be violated.19 A second related issue concerns the administrative costs, bureaucratic requirements, and implementation problems, in terms of limiting or delaying participation in the CSG (Heinrich et al., 2012, Heinrich and Brill 2015, Samson et al., 2016, Heinrich et al., 2017). In the present case, we could have two concerns. First of all, the population of compliers might be unstable. Statistically, this implies that the population of compliers may change dramatically with small changes in the birth cohort of the children. Secondly, given that the estimates of the RD treatment effect only apply to people such that X=c, it is important to investigate the stability of the RD estimates, that is, to examine whether people with other values of X near c would have expected treatment effects of similar sign and magnitude. If not, i.e., if ceteris paribus a small change in X away from c would greatly change the average effect of treatment, then one would have serious doubts about the general usefulness and external validity of the estimates, since other contexts are likely to differ from the given one in even more substantial ways than a marginal change in X. With this in mind, we apply the method proposed by Cerulli et al. (2017), based on the Complier Probability Derivative (CPD) and on the Treatment Effect Derivative (TED).20 Estimates of TED that are statistically significant and large in magnitude are evidence of instability and, hence, a potential lack of external validity. In contrast, having TED near zero, or not statistically significant, provides evidence supporting the stability of the RD estimates since if the threshold had been somewhat lower or higher, the estimated LATE would probably have still been close to zero. On the other hand, significant and high values of the CPD show that there is instability in the complier population. However, as pointed out by Cerulli et al. (2017), when the TED is near zero, or not statistically significant, then even a certain instability in the complier population (i.e., CPD≠0) does not affect the estimated results and does not violate the assumption of external validity. Lastly, we recall that the potential outcome in a specific population subgroup may not be homogeneous with respect to the entire population (Angrist 2004). Hence, we calculate elasticity measures based on the PS estimates obtained in the two-step procedure, and carry out a further robustness check by comparing the results for the subgroups with that of the entire population, to verify whether significant differences emerge. In more detail, we extend equation (4) by replacing W1 with W2, which includes the interaction terms between the PS estimates and the dummy variables describing, one by one, five subgroups of the population which may present substantial differences in the potential outcomes. To allow for the correct identification of equation (4), we must assume that the interaction variables are continuous at the cutoff and uncorrelated with the error term, conditional on W2 (Becker et al., 2013). With the estimated parameters, we construct elasticity measures using the interaction terms and PS estimates. The elasticity measures allow us to compare the effects of policy changes on varying population subgroups and to interpret the results in terms of the percentage variation in the outcome variable caused by a 1% increase in the treated population. 4. Results and robustness analysis This section presents the estimates of the effects of the CSG on total food expenditure per adult equivalent, the shares of the food groups and the synthetic index of dietary diversity. The results for the full-sample 2008–2012 are given, together with those excluding 2012 for the shares of food groups and the synthetic index of dietary diversity. Figure 2 shows the results of the estimation of equation (1), which links the assignment variable X (the birth cohort of the child) to treatment status T (the probability of participating in the CSG), with the LP estimator. Children in birth cohorts outside the age eligibility for the CSG lie on the left of the cutoff, whereas children in both the treated and control groups lie on the right. The figure confirms the fuzzy nature of the RDD and, by using the corresponding robust estimator,21 we find that about 5.9% (s.e. 0.02, p-value 0.003) of the households in the first sample and about 10.5% (s.e. 0.034, p-value 0.002) in the second sample, participated in the CSG. Figure 2: View largeDownload slide LP Estimates: Birth Cohort and Treatment Status. (a) Sample 2008–2012 and (b) sample 2008–2011. Notes: Construction of evenly spaced bins follows Calonico et al. (2014a). Figure 2: View largeDownload slide LP Estimates: Birth Cohort and Treatment Status. (a) Sample 2008–2012 and (b) sample 2008–2011. Notes: Construction of evenly spaced bins follows Calonico et al. (2014a). Figure 3 panel (a) shows the estimates of the effects of the CSG on the first outcome variable, total food expenditure per adult equivalent, obtained through the LP estimator. Panels (b)–(e) replicate the same analysis for the outcome variables of the shares in total food expenditure of (i) carbohydrates, (ii) dairy products, (iii) proteins, and (iv) vitamins. Panel (f) considers the synthetic dietary diversity variable. Figure 3: View largeDownload slide LP Estimates: Impact of CSG on Total Food Expenditure Per Adult Equivalent and Shares of Food Groups. (a) Total food expenditure per adult equivalent, (b) share of carbohydrates *, (c) share of dairy products *, (d) share of proteins *, (e) share of vitamins *, (f) sietary diversity *. Notes: (*) Analysis of food groups and of synthetic index of dietary diversity is valid for 2008–2011. Construction of evenly spaced bins follows Calonico et al. (2014a). Figure 3: View largeDownload slide LP Estimates: Impact of CSG on Total Food Expenditure Per Adult Equivalent and Shares of Food Groups. (a) Total food expenditure per adult equivalent, (b) share of carbohydrates *, (c) share of dairy products *, (d) share of proteins *, (e) share of vitamins *, (f) sietary diversity *. Notes: (*) Analysis of food groups and of synthetic index of dietary diversity is valid for 2008–2011. Construction of evenly spaced bins follows Calonico et al. (2014a). Panel (a) of Figure 3 clearly shows a discontinuity in the food expenditure per adult equivalent around the 1 January 1994 cutoff: before that date, the expenditure pattern remained quite constant along the birth cohorts, but afterwards, participant households showed an increase in food expenditure per adult equivalent of about R55. This variation is significant, given that the mean value of food expenditure for adult equivalent is only R500 in our sample. A similar pattern is also found in the shares of carbohydrates and dairy products in total food expenditure, in panels (b) and (c), respectively. In both cases, there is a positive variation in expenditure shares after the cutoff due to the expanded CSG age eligibility rule. Conversely, a poorly defined result appears for the protein share (panel d), and no discontinuity in the vitamin share (panel e) or in the index of dietary diversity (panel f). As the last five plots were obtained with data up to 2011, i.e., only 1 year after the policy change, and as the LP estimator requires an extensive number of observations, these results should be viewed with caution. Lastly, the six plots show that a third-degree polynomial is used by the LP estimator to approximate the food expenditure patterns although, except for protein, a linear functional form approximates the behaviour of the variables in the neighbourhood of the cutoff quite well. To complete the preliminary analysis, we perform the continuity test for each variable listed in Table 2.22 The smoothness graphs are reported in Appendix B and show that there is no evidence of discontinuity at the cutoff.23 Table 3 lists the estimates of food expenditure per adult equivalent, comparing the results for the IV and two-stage PS estimators.24 In each case, we consider a linear functional form (1) and then a non-linear one (2) that uses a third-degree polynomial, as suggested by the LP estimator. When the results obtained from the two-step PS estimator are analysed, the non-linear form is introduced in f(X), as in equation (4). Table 3: Impact of CSG on total food expenditure per adult equivalent Instrumental variable Propensity score (1) (2) (1) (2) Child Support Grant (CSG) 55.337*** 54.202*** 52.357*** 53.964*** (20.040) (19.589) (19.547) (19.176) Constant term 1046.445*** 1232.174*** 1221.254*** 1432.057*** (184.109) (78.246) (77.864) (187.934) Fixed effects Yes Yes Yes Yes Linear and quadratic trends Yes Yes Yes Yes Selected covariates Yes Yes Yes Yes Cragg–Donald F statistic 1359.645 499.282 (0.000) (0.000) Kleibergen–Paap F test statistic † 1433.471 510.756 (0.000) (0.000) R2 0.676 0.676 0.681 0.682 Adjusted R2 0.658 0.658 0.663 0.664 No. of observations 3399 3399 3399 3399 Instrumental variable Propensity score (1) (2) (1) (2) Child Support Grant (CSG) 55.337*** 54.202*** 52.357*** 53.964*** (20.040) (19.589) (19.547) (19.176) Constant term 1046.445*** 1232.174*** 1221.254*** 1432.057*** (184.109) (78.246) (77.864) (187.934) Fixed effects Yes Yes Yes Yes Linear and quadratic trends Yes Yes Yes Yes Selected covariates Yes Yes Yes Yes Cragg–Donald F statistic 1359.645 499.282 (0.000) (0.000) Kleibergen–Paap F test statistic † 1433.471 510.756 (0.000) (0.000) R2 0.676 0.676 0.681 0.682 Adjusted R2 0.658 0.658 0.663 0.664 No. of observations 3399 3399 3399 3399 Notes: In all specifications, dependent variable is total food expenditure per adult equivalent (see the list of the selected covariates in Section 2). Robust standard errors, clustered at household level, are shown in brackets. Asterisks: p-value levels ( *p<0.1;**p<0.05;***p<0.01). For each model, linear (1) and non-linear specification (2) is shown. When two-step PS procedure is applied, non-linearities are introduced only in first-stage regression. Tests for weak instrument hypothesis and first-stage F statistics and Wald statistics based on Cragg and Donald (1993) and Kleibergen and Paap (2006). †Confidence intervals for Kleibergen–Paap F test statistic follow Bazzi and Clemens (2013). View Large Table 3: Impact of CSG on total food expenditure per adult equivalent Instrumental variable Propensity score (1) (2) (1) (2) Child Support Grant (CSG) 55.337*** 54.202*** 52.357*** 53.964*** (20.040) (19.589) (19.547) (19.176) Constant term 1046.445*** 1232.174*** 1221.254*** 1432.057*** (184.109) (78.246) (77.864) (187.934) Fixed effects Yes Yes Yes Yes Linear and quadratic trends Yes Yes Yes Yes Selected covariates Yes Yes Yes Yes Cragg–Donald F statistic 1359.645 499.282 (0.000) (0.000) Kleibergen–Paap F test statistic † 1433.471 510.756 (0.000) (0.000) R2 0.676 0.676 0.681 0.682 Adjusted R2 0.658 0.658 0.663 0.664 No. of observations 3399 3399 3399 3399 Instrumental variable Propensity score (1) (2) (1) (2) Child Support Grant (CSG) 55.337*** 54.202*** 52.357*** 53.964*** (20.040) (19.589) (19.547) (19.176) Constant term 1046.445*** 1232.174*** 1221.254*** 1432.057*** (184.109) (78.246) (77.864) (187.934) Fixed effects Yes Yes Yes Yes Linear and quadratic trends Yes Yes Yes Yes Selected covariates Yes Yes Yes Yes Cragg–Donald F statistic 1359.645 499.282 (0.000) (0.000) Kleibergen–Paap F test statistic † 1433.471 510.756 (0.000) (0.000) R2 0.676 0.676 0.681 0.682 Adjusted R2 0.658 0.658 0.663 0.664 No. of observations 3399 3399 3399 3399 Notes: In all specifications, dependent variable is total food expenditure per adult equivalent (see the list of the selected covariates in Section 2). Robust standard errors, clustered at household level, are shown in brackets. Asterisks: p-value levels ( *p<0.1;**p<0.05;***p<0.01). For each model, linear (1) and non-linear specification (2) is shown. When two-step PS procedure is applied, non-linearities are introduced only in first-stage regression. Tests for weak instrument hypothesis and first-stage F statistics and Wald statistics based on Cragg and Donald (1993) and Kleibergen and Paap (2006). †Confidence intervals for Kleibergen–Paap F test statistic follow Bazzi and Clemens (2013). View Large All specifications include province and district dummies, and also linear and quadratic trends to take into account the non-linear patterns of food expenditure (Pieroni et al., 2013, Pieroni and Salmasi 2015), together with household and head of household characteristics (see Section 2). Robust standard errors, clustered at household level, are shown in brackets.25 For each specification, we test for weak instruments. In more detail, we run a test for weak instruments and present first-stage F statistics and Wald statistics based on the Cragg and Donald (1993) and Kleibergen and Paap (2006) generalisation to non-independently and non-identically distributed errors, together with the p-values (Bazzi and Clemens 2013). The under-identification and weak instrument tests show that the assignment variable accounts for the entire endogeneity appearing in the IV estimations. We find that the non-parametric and parametric estimations show very similar variations due to the CSG policy change. That is, from the linear (1) and non-linear (2) specifications of the IV estimator, we find a variation in food expenditure per adult equivalent that ranges between R54 and R55, and this variation is significant at 1%. Considering the full sample (see Table 1), this variation corresponds to a mean increase of 10% in total food expenditure per adult equivalent. This result is stronger than the findings of Coetzee (2013) for the CSG and almost in line with other evaluations of the impact of social grants on several indicators of food security in other countries of sub-Saharan Africa (Pellerano et al., 2014, Burchi et al., 2016, Davis and Handa 2016, Pellerano et al., 2016).26 Moving to columns 3 and 4, we use the two-step PS estimator to check for the robustness of the IV results. It should be noted that the two-step PS is used to analyse the robustness of the results when the identification mechanism is not completely known. In the present case, it shows that the multiple sources of variations in participation in the CSG, caused by the change in the income cutoff and by problems in access related to administrative burdens, do not affect the IV results, since there are no differences in the standard deviations between the IV and the two-step PS estimates. Table 4 extends the previous analysis to the shares of food groups in the total food expenditure per adult equivalent and to the synthetic index of dietary diversity. In depth, the table compares the results from a linear specification of the IV and two-step PS estimations.27 Table 4: Impact of CSG on food expenditure by food group and on dietary diversity index Carbohydrates Dairy products Proteins Vitamins Dietary diversity IV PS IV PS IV PS IV PS IV PS Child Support Grant (CSG) 2.267** 2.022** 0.808** 0.753 0.196 −0.233 1.230** 1.019 −0.074 −0.043 (1.150) (1.030) (0.405) (0.483) (1.216) (1.176) (0.557) (0.588) (0.176) (0.171) Constant term 32.921*** 32.716*** 5.768*** 5.685*** 9.164 9.270 14.986*** 14.898*** 9.506*** 9.506*** (7.440) (7.359) (2.156) (2.185) (5.772) (5.774) (4.379) (4.391) (0.897) (0.895) Fixed effects Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Linear and quadratic trends Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Selected covariates Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Cragg–Donald F statistic 866.207 866.207 866.207 866.207 848.632 (0.000) (0.000) (0.000) (0.000) (0.000) Kleibergen–Paap F test statistic † 930.615 930.615 930.615 930.615 913.688 (0.000) (0.000) (0.000) (0.000) (0.000) R2 0.361 0.361 0.090 0.094 0.222 0.222 0.097 0.106 0.384 0.385 Adjusted R2 0.321 0.321 0.034 0.038 0.174 0.174 0.042 0.051 0.346 0.346 No. of observations 2193 2193 2193 2193 2193 2193 2193 2193 2175 2175 Carbohydrates Dairy products Proteins Vitamins Dietary diversity IV PS IV PS IV PS IV PS IV PS Child Support Grant (CSG) 2.267** 2.022** 0.808** 0.753 0.196 −0.233 1.230** 1.019 −0.074 −0.043 (1.150) (1.030) (0.405) (0.483) (1.216) (1.176) (0.557) (0.588) (0.176) (0.171) Constant term 32.921*** 32.716*** 5.768*** 5.685*** 9.164 9.270 14.986*** 14.898*** 9.506*** 9.506*** (7.440) (7.359) (2.156) (2.185) (5.772) (5.774) (4.379) (4.391) (0.897) (0.895) Fixed effects Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Linear and quadratic trends Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Selected covariates Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Cragg–Donald F statistic 866.207 866.207 866.207 866.207 848.632 (0.000) (0.000) (0.000) (0.000) (0.000) Kleibergen–Paap F test statistic † 930.615 930.615 930.615 930.615 913.688 (0.000) (0.000) (0.000) (0.000) (0.000) R2 0.361 0.361 0.090 0.094 0.222 0.222 0.097 0.106 0.384 0.385 Adjusted R2 0.321 0.321 0.034 0.038 0.174 0.174 0.042 0.051 0.346 0.346 No. of observations 2193 2193 2193 2193 2193 2193 2193 2193 2175 2175 Notes: The dependent variables are shares of food groups in total food expenditure per adult equivalent. In all specifications, dependent variable is total food expenditure per adult equivalent (see the list of the selected covariates in Section 2). Robust standard errors, clustered at household level, are shown in brackets. Asterisks: p-value levels ( *p<0.1;**p<0.05;***p<0.01). For each model, linear (1) and non-linear specification (2) is shown. When two-step PS procedure is applied, non-linearities are introduced only in first-stage regression. Tests for weak instrument hypothesis and first-stage F statistics and Wald statistics based on Cragg and Donald (1993) and Kleibergen and Paap (2006). †Confidence intervals for Kleibergen–Paap F test statistic follow Bazzi and Clemens (2013). View Large Table 4: Impact of CSG on food expenditure by food group and on dietary diversity index Carbohydrates Dairy products Proteins Vitamins Dietary diversity IV PS IV PS IV PS IV PS IV PS Child Support Grant (CSG) 2.267** 2.022** 0.808** 0.753 0.196 −0.233 1.230** 1.019 −0.074 −0.043 (1.150) (1.030) (0.405) (0.483) (1.216) (1.176) (0.557) (0.588) (0.176) (0.171) Constant term 32.921*** 32.716*** 5.768*** 5.685*** 9.164 9.270 14.986*** 14.898*** 9.506*** 9.506*** (7.440) (7.359) (2.156) (2.185) (5.772) (5.774) (4.379) (4.391) (0.897) (0.895) Fixed effects Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Linear and quadratic trends Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Selected covariates Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Cragg–Donald F statistic 866.207 866.207 866.207 866.207 848.632 (0.000) (0.000) (0.000) (0.000) (0.000) Kleibergen–Paap F test statistic † 930.615 930.615 930.615 930.615 913.688 (0.000) (0.000) (0.000) (0.000) (0.000) R2 0.361 0.361 0.090 0.094 0.222 0.222 0.097 0.106 0.384 0.385 Adjusted R2 0.321 0.321 0.034 0.038 0.174 0.174 0.042 0.051 0.346 0.346 No. of observations 2193 2193 2193 2193 2193 2193 2193 2193 2175 2175 Carbohydrates Dairy products Proteins Vitamins Dietary diversity IV PS IV PS IV PS IV PS IV PS Child Support Grant (CSG) 2.267** 2.022** 0.808** 0.753 0.196 −0.233 1.230** 1.019 −0.074 −0.043 (1.150) (1.030) (0.405) (0.483) (1.216) (1.176) (0.557) (0.588) (0.176) (0.171) Constant term 32.921*** 32.716*** 5.768*** 5.685*** 9.164 9.270 14.986*** 14.898*** 9.506*** 9.506*** (7.440) (7.359) (2.156) (2.185) (5.772) (5.774) (4.379) (4.391) (0.897) (0.895) Fixed effects Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Linear and quadratic trends Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Selected covariates Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Cragg–Donald F statistic 866.207 866.207 866.207 866.207 848.632 (0.000) (0.000) (0.000) (0.000) (0.000) Kleibergen–Paap F test statistic † 930.615 930.615 930.615 930.615 913.688 (0.000) (0.000) (0.000) (0.000) (0.000) R2 0.361 0.361 0.090 0.094 0.222 0.222 0.097 0.106 0.384 0.385 Adjusted R2 0.321 0.321 0.034 0.038 0.174 0.174 0.042 0.051 0.346 0.346 No. of observations 2193 2193 2193 2193 2193 2193 2193 2193 2175 2175 Notes: The dependent variables are shares of food groups in total food expenditure per adult equivalent. In all specifications, dependent variable is total food expenditure per adult equivalent (see the list of the selected covariates in Section 2). Robust standard errors, clustered at household level, are shown in brackets. Asterisks: p-value levels ( *p<0.1;**p<0.05;***p<0.01). For each model, linear (1) and non-linear specification (2) is shown. When two-step PS procedure is applied, non-linearities are introduced only in first-stage regression. Tests for weak instrument hypothesis and first-stage F statistics and Wald statistics based on Cragg and Donald (1993) and Kleibergen and Paap (2006). †Confidence intervals for Kleibergen–Paap F test statistic follow Bazzi and Clemens (2013). View Large When the share of carbohydrates is analysed, we find a variation of about 2% in expenditure produced by the policy change. Again, this result is similar to that found by the LP estimator and does not change in the two-step PS estimation. Less robust but still significant results are found when the shares of dairy products and vitamins are accounted for in the IV specification: the beneficiary families increase the share of expenditure on diary products by 0.8% and that on vitamins by 1.2%. However, these results become non-significant at the 10% level when the two-step PS estimators are used. Instead, no significant variations are found when the share of proteins in total expenditure per adult equivalent or the synthetic index of dietary diversity are examined. It should be emphasised that carbohydrates are the major food item consumed by the treated households (on average, 23% of total food expenditure per adult equivalent), and these results thus indicate that the programme grants were not sufficient to allow the beneficiary households to make significant changes in their dietary habits. However, we must use caution because we are using the restricted sample (2008–2011) when considering food shares and dietary diversity whereas, in the case of total food expenditure, we use the full sample (2008–2012). We now consider the robustness tests of our estimates. Appendix C presents the results for a restricted sample including only 2008 of the same estimations as were run for Tables 3 and 4. Indeed, since the policy change took place in 2010, we can use these estimates to run placebo tests showing that the estimated variations are due to the extension in the age eligibility of the CSG and not to the self-selection of individuals across different age cohorts. These tests show that the estimated average treatment effect is not only valid for the population of compliers but can be generalised to the entire population since, for each outcome variable, we do not find any statistically significant effect due to the extension of the CSG. Following the outline described in Section 3, Table 5 presents the two formal tests based on the CPD and TED to assess the validity of the RD design. The CPD estimates are presented for the two sample 2008–2012 and 2008–2011, at the bottom of the table. Comparing the two estimated parameters, we find statistically significant evidence of a certain instability in the population of compliers. As shown in Figure 2, the probability of being enroled in the CSG at the cutoff is about 2% (7% when we consider the sample 2008–2011), whereas we find that it is about 4.4% (9.2% in the second sample) when we consider a birth cohort that is a month older. As shown in the upper part of the table, we do not find any statistically significant estimate of the TED for each outcome variable. Together, these results indicate that despite the fact that the complier population is unstable, this result does not affect the constancy of the average treatment effect. Thus, the external validity assumption still holds for each specification. Table 5: Impact of CSG on food expenditure by food group and on dietary diversity index Food expenditure Carbohydrates Diary products Proteins Vitamins Dietary diversity TED −380.548 −2.415 0.053 −3.333 −0.014 0.329 (387.730) (2.647) (0.773) (3.217) (0.905) (0.368) No. of observation 3399 2183 2183 2183 2183 2183 Sample Sample (2008–2012) (2008–2011) CPD 0.022*** 0.019*** (0.003) (0.003) No. of observation 3399 2183 Food expenditure Carbohydrates Diary products Proteins Vitamins Dietary diversity TED −380.548 −2.415 0.053 −3.333 −0.014 0.329 (387.730) (2.647) (0.773) (3.217) (0.905) (0.368) No. of observation 3399 2183 2183 2183 2183 2183 Sample Sample (2008–2012) (2008–2011) CPD 0.022*** 0.019*** (0.003) (0.003) No. of observation 3399 2183 Notes: Bootstrapped Standard errors based on 500 simulations are reported in brackets. Asterisks: p-value levels ( *p<0.1;**p<0.05;***p<0.01). View Large Table 5: Impact of CSG on food expenditure by food group and on dietary diversity index Food expenditure Carbohydrates Diary products Proteins Vitamins Dietary diversity TED −380.548 −2.415 0.053 −3.333 −0.014 0.329 (387.730) (2.647) (0.773) (3.217) (0.905) (0.368) No. of observation 3399 2183 2183 2183 2183 2183 Sample Sample (2008–2012) (2008–2011) CPD 0.022*** 0.019*** (0.003) (0.003) No. of observation 3399 2183 Food expenditure Carbohydrates Diary products Proteins Vitamins Dietary diversity TED −380.548 −2.415 0.053 −3.333 −0.014 0.329 (387.730) (2.647) (0.773) (3.217) (0.905) (0.368) No. of observation 3399 2183 2183 2183 2183 2183 Sample Sample (2008–2012) (2008–2011) CPD 0.022*** 0.019*** (0.003) (0.003) No. of observation 3399 2183 Notes: Bootstrapped Standard errors based on 500 simulations are reported in brackets. Asterisks: p-value levels ( *p<0.1;**p<0.05;***p<0.01). View Large To complete the analysis, Figure 4 shows the elasticity measures for total food expenditure per adult equivalent and the share of carbohydrates. The elasticity measures are based on equation (4) (Section 3) and are obtained by jointly considering the interaction term between the estimated PS and the dummy variables, describing, one by one, five subgroups of the population which may present substantial differences in the potential outcomes. We do not present the other panels (for dairy products, proteins and vitamins and for the index of dietary diversity) since the estimated parameters are not robust across differing estimation methods (see Table 4). Figure 4 shows the results for the total population and some population subgroups, which are the households under both the food poverty line and the poverty line, households in which the head is an African and households living in rural and urban areas. It should be recalled that the main targets of the CSG programme are the subgroups most vulnerable to poverty and food insecurity, i.e. households living under the food poverty line, Africans, and those living in rural areas. The elasticity measure allows us to compare the various subgroups with the entire population in terms of the percentage effect of the variation in their participation in the programme after the policy change and to check its homogeneity. Figure 4: View largeDownload slide Elasticity Measures: Total Food Expenditure Per Adult Equivalent and Share of Carbohydrates. (a) Total food expenditure per adult equivalent and (b) Share of carbohydrates. Notes: The elasticity measures are based on the the two-step PS estimated parameters (Table 3 specification 2 and in Table 4 specification 2). Standard errors are estimated by Delta-Method. Asterisks: p-value levels ( *p<0.1;**p<0.05;***p<0.01). Figure 4: View largeDownload slide Elasticity Measures: Total Food Expenditure Per Adult Equivalent and Share of Carbohydrates. (a) Total food expenditure per adult equivalent and (b) Share of carbohydrates. Notes: The elasticity measures are based on the the two-step PS estimated parameters (Table 3 specification 2 and in Table 4 specification 2). Standard errors are estimated by Delta-Method. Asterisks: p-value levels ( *p<0.1;**p<0.05;***p<0.01). Figure 4 panel (a) shows that although some differences between the subgroups and full-sample estimates are found, all the subgroups elasticities are plotted in the same confidence interval as the full-sample elasticity. In this case, the analysis shows that a 1% variation in participation in the programme produces a 3% increase in total food expenditure per adult equivalent, at a significance level which is less than 5%. We find a higher variability when the rural/urban subgroups are analysed. Figure 4 shows that a 1% increase in participation in the programme for the households living in urban areas produces a positive variation in the food expenditure per adult equivalent, which is double that estimated for rural areas. This result may be explained by the fact that rural areas in South Africa are geographically isolated and marginalised, so that food availability is a serious problem (Kirsten 2012). Several studies have found a higher percentage of food insecurity in rural areas in South Africa (Kirkland et al., 2013). However, we must also recall that, as Aliber (2009) stresses, rural households spend less than their urban counterparts on food purchases because they have their own production. In principle, measures of food expenditure are designed to capture this information, but the imputed value is probably smaller than the true one. In this case, the effect of the CSG on food security in rural areas in our analysis would be underestimated. When the share of carbohydrates is analysed, we find more variability between the full-sample elasticity measure and those for the population subgroups, but the results are significant at the 5% level only for households under the food poverty line or ones in which the head is African. In more detail, the elasticity measure for beneficiaries under the food poverty line is about double that of the full sample, even though it falls in the same confidence interval. This result is in line with the existing literature, which states that poor dietary variety mainly concerns the poorest people, who consume far more cereals and less fruit, vegetables, dairy products, and meat, than other population subgroups (Labadarios et al., 2011, South African Department of Health 2013). In the case of the subgroup of households with an African head, the elasticity is very close to that of the full sample. Overall, these results indicate that the CSG has proved to be effective in supporting the food expenditure of its beneficiary households. However, the strategy of providing a basic grant to a large share of the poor population has not been effective in producing a significant change in the dietary habits of the very poor beneficiaries or at guaranteeing a nutritionally varied food basket. This result is not unexpected since the grant money is used to meet the broader needs of the families, including education, in low-income families. However, this outcome should not be overlooked since increasing food consumption without modifying the dietary habits may shift poor households from undernutrition to malnutrition, worsening their health condition (Tzioumis and Adair 2014). This suggests that ancillary social services, starting with the improvement of mother’s and children’s nutritional education, should complement the CSG in strengthening food security impacts (Heinrich et al., 2012, Burchi et al., 2016). Our findings also show that the CSG has failed to improve the food security of the most disadvantaged groups to a greater extent than for the others. This result, that the CSG is not effectively targeting those most in need, may be partially explained by several administrative constraints in the implementation of the programme, serving as an operational barrier to reaching the most vulnerable caregivers and children (Delany et al., 2008, Heinrich et al., 2012, Heinrich and Brill 2015, Davis and Handa 2016, Samson et al., 2016). This study suggests that an integrated approach and nutrition-sensitive social protection programme, specifically targeting additional resources and complementary social services to the most food-insecure households, would contribute more effectively to increasing the food and nutrition status of the very poor households in South Africa. 5. Conclusions This paper estimates the impact of the South African Child Support Grant (CSG) on food security. We used the dataset provided by the National Income Dynamics Study covering 2008, 2010–2011, and 2012, and carried out a Regression Discontinuity design to estimate the effect of the programme on total food expenditure per adult equivalent, as well as on dietary diversity. Our results show that the transfers provided by the CSG significantly increased the total food expenditure per adult equivalent. Even if our findings must be treated with caution, given that they are based on a non-experimental study, we show that the results are very robust, being confirmed by a comparison between non-parametric, semi-parametric, and parametric estimations. Our analysis also shows that the estimated parameters are constant when the cutoff diverges, which means that it provides plausible casual effects of the programme. When dietary diversity is analysed, we find robust positive results only when the share of carbohydrates in the total food expenditure per adult equivalent is included. Since this is the largest food group consumed by the treated households (23% of total food expenditure), this result indicates that the CSG has not been effective in making significant improvements in the dietary habits of the beneficiary households. The estimates of the elasticity measures for the total food expenditure and the share of carbohydrates do not show any significant difference between various population subgroups. This means that the statistical results are homogeneous. It also suggests that the CSG, at variance with its stated objective, has not been effectively targeted to the poorest population subgroups who are the most vulnerable to food insecurity. Overall, the policy implication of our study is that the current design of the CSG, which only provides a small grant for each beneficiary child, and its strategy of gradually widening the eligible population, has not been sufficient for guaranteeing a significant reduction of deprivation for the most vulnerable households. A more effective approach would be to integrate this policy with a specific, comprehensive strategy to reduce food insecurity and deliver additional grants to those most in need. In addition, ancillary social services, including investments in mother’s and children’s nutritional education, should be introduced to improve the food habits of poor households. Lastly, income-generating programmes and the enhancement of small-scale agricultural activities remain crucial in increasing household access to food in the most poverty-stricken areas, which are the rural ones. Acknowledgements We would like to thank Ingrid Woolard, J. Paul Dunne and the participants at the SITES/IDEAs 2nd Annual Conference for their insightful comments. We thank SALDRU (Southern Africa Labour & Development Research Unit), University of Cape Town, for its support. The authors wish also to acknowledge the Editor and two anonymous reviewers of this journal for their helpful suggestions. The usual disclaimers apply. References Agüero J. , Carter M. R. , May J. ( 2007 ) ‘ Poverty and iNequality in the First Decade of South Africa’s Democracy: What Can Be Learnt From Panel Data From KwaZulu-Natal? ’, Journal of African Economies , 16 ( 5 ): 782 – 812 . Google Scholar CrossRef Search ADS Agüero J. , Carter M. , Woolard I. ( 2010 ) The impact of unconditional cash transfers on nutrition: The South African Child Support Grant. International Poverty Centre Discussion Papers No. 39, United Nations Development Programme (UNDP), New York. Alderman H. ( 2014 ) Can transfer programs be made more nutrition-sensitive? International Food Policy Research Institute (IFPRI) Discussion Paper No. 1342, IFPRI, Washington D.C. Aliber M. A. ( 2001 ) Study of the incidence and nature of chronic poverty and development policy in South Africa: An overview. Chronic Poverty Research Centre Working Papers No.1, University of Manchester, Manchester. Aliber M. ( 2009 ) ‘ Exploring Statistics South Africa’s National Household Surveys as Sources of Information About Household-level Food Security ’, Agrekon , 48 ( 4 ): 384 – 409 . Google Scholar CrossRef Search ADS Angrist J. D. ( 2004 ) ‘ Treatment Effect Heterogeneity in Theory and Practice ’, The Economic Journal , 114 ( 494 ): 52 – 83 . Google Scholar CrossRef Search ADS Bassett L. ( 2008 ) Can conditional cash transfer programs play a greater role in reducing child undernutrition? World Bank Social Policy Discussion Papers No. 835, The World Bank, Washington D.C. Bazzi S. , Clemens M. A. ( 2013 ) ‘ Blunt Instruments: Avoiding Common Pitfalls in Identifying the Causes of Economic Growth ’, American Economic Journal: Macroeconomics , 5 ( 2 ): 152 – 86 . Google Scholar CrossRef Search ADS Becker S. O. , Egger P. H. , von Ehrlich M. ( 2013 ) ‘ Absorptive Capacity and the Growth and Investment Effects Of Regional Transfers: A Regression Discontinuity Design With Heterogeneous Treatment Effects ’, American Economic Journal: Economic Policy , 5 ( 4 ): 29 – 77 . Google Scholar CrossRef Search ADS Bor J. ( 2013 ) Cash transfers and teen pregnancy in South Africa: Evidence from a natural experiment. Northeast Universities Development Consortium Conference, Harvard University, Cambridge, MA. Burchi F. , Scarlato M. , d’Agostino G. ( 2016 ) Addressing food insecurity in sub-Saharan Africa: The role of cash transfers. German Development Institute Discussion Papers No. 17, German Development Institute Discussion, Bonn. Calonico S. , Cattaneo M. D. , Titiunik R. ( 2014 a) ‘ Robust Data-driven Inference in the Regression-Discontinuity Design ’, Stata Journal , 14 ( 4 ): 909 – 46 . Calonico S. , Cattaneo M. D. , Titiunik R. ( 2014 b) ‘ Robust Non-parametric Confidence Intervals for Regression-discontinuity Designs ’, Econometrica: Journal of the Econometric Society , 82 ( 6 ): 2295 – 2326 . Google Scholar CrossRef Search ADS Cameron A. C. , Miller D. L. ( 2015 ) ‘ A Practitioner’s Guide to Cluster-Robust Inference ’, Journal of Human Resources , 50 ( 2 ): 317 – 72 . Google Scholar CrossRef Search ADS Cerulli G. , Dong Y. , Lewbel A. , Poulsen A. ( 2017 ) ‘Testing Stability of Regression Discontinuity Models’, in Cattaneo M. D. , Escanciano J. C. (eds) , Regression discontinuity designs, Advances in econometrics , Vol. 38 . Bingley (UK) : Emerald Publishing Limited , pp. 317 – 39 . Google Scholar CrossRef Search ADS Coetzee M. ( 2013 ) ‘ Finding the Benefits: Estimating the Impact of the South African Child Support Grant ’, South African Journal of Economics , 81 ( 3 ): 427 – 50 . Google Scholar CrossRef Search ADS Cragg J. G. , Donald S. G. ( 1993 ) ‘ Testing Identifiability and Specification in Instrumental Variable Models ’, Econometric Theory , 9 ( 2 ): 222 – 40 . Google Scholar CrossRef Search ADS Davis B. , Handa S. ( 2016 ) ‘Implementing rigorous evaluations in the real world: the quantitative approach to evaluation design in the transfer project’, in Davis B. , Handa S. , Hypher N. , Winder Rossi N. , Winters P. , Yablonski J. (eds) , From Evidence to Action. The Story of Cash Transfers and Impact Evaluation in Sub-Saharan Africa . Oxford : Oxford University Press . Google Scholar CrossRef Search ADS Deaton A. , Grosh M. ( 2000 ) ‘Consumption’, in Grosh M. , Glewwe P. (eds) , Designing Household Survey Questionnaires for Developing Countries: Lessons from 15 Years of the Living Standards Measurement Study , vol.1 . Washington, DC : The World Bank , pp. 91 – 133 . Delany A. , Ismail Z. , Graham L. , Ramkissoon Y. ( 2008 ) Review of the Child Support Grant: Uses, Implementation and Obstacles . Johannesburg : United Nations Children’s Fund . del Ninno C. , Mills B. ( 2015 ) ‘Introduction’, in Del Ninno C. , Mills B. (eds) , Safety Nets in Africa . Washington D.C : The World Bank . Department of Social Development ( 2010 ). An Overview of South Africa’s Social Security System . Republic of South Africa , Pretoria . Dong Y. , Lewbel A. ( 2015 ) ‘ Identifying the Effect of Changing The Policy Threshold In Regression Discontinuity Models ’, Review of Economics and Statistics , 97 ( 5 ): 1081 – 92 . Google Scholar CrossRef Search ADS Du Toit D. ( 2011 ) Food Security. Department of Agriculture, Forestry and Fisheries . Pretoria : Republic of South Africa . Eyal K. , Woolard I. ( 2013 ). School Enrolment and the Child Support Grant: Evidence from South Africa. Southern African Labour and Development Research Unit (SALDRU) Working Papers No. 12, University of Cape Town, Cape Town. Hahn J. , Todd P. , Van der Klaauw W. ( 2001 ) ‘ Identification and Estimation of Treatment Effects With a Regression-discontinuity Design ’, Econometrica: Journal of the Econometric Society , 69 ( 1 ): 201 – 9 . Google Scholar CrossRef Search ADS Heinrich C. J. , Brill R. ( 2015 ) ‘ Stopped in the Name of the Law: Administrative Burden and its Implications For Cash Transfer Program Effectiveness ’, World Development , 72 ( C ): 277 – 95 . Google Scholar CrossRef Search ADS Heinrich C. J. , Hoddinott J. , Samson M. , Mac Quene K. , van Niekerk I. , Bryant R. ( 2012 ) The South African Child Support Grant impact assessment. Prepared for the Department of Social Development (DSD), the South African Social Security Agency (SASSA) and the United Nations Children’s Fund (UNICEF) . Pretoria : Republic of South Africa . Heinrich C. J. , Hoddinott J. , Samson M. ( 2017 ) ‘ Reducing Adolescent Risky Behaviors In A High-risk Context: The Effects Of Unconditional Cash Transfers in South Africa ’, Economic Development and Cultural Change , 65 ( 4 ): 619 – 52 . Google Scholar CrossRef Search ADS Hoddinott J. , Bassett L. ( 2008 ) Conditional Cash Transfer Programs And Nutrition In Latin America: Assessment Of Impacts And Strategies For Improvement . Washington DC/Rome : United Nations (UN) and Food and Agriculture Organization (FAO) . Hoddinott J. , Yohannes Y. ( 2002 ) Dietary Diversity as a Food Security Indicator. Food and Nutrition Technical Assistance Project . Washington DC : Academy for Educational Development . Jacob R. T. , Zhu P. , Somers M.-A. , Bloom H. S. ( 2012 ) A Practical Guide to Regression Discontinuity . New York : MDRC . Kennedy G. , Ballard T. , Dop M. ( 2011 ) Guidelines for Measuring Household and Individual Dietary Diversity . Rome : Food and Agriculture Organization of the United Nations (FAO) . Kirkland T. M. , Kemp R. J. , Hunter L. M. , Twine W. M. ( 2013 ) ‘ Toward imProved Understanding Of Food Security: A Methodological Examination Based in Rural South Africa ’, Food, Culture & Society , 16 ( 1 ): 65 – 84 . Google Scholar CrossRef Search ADS Kirsten J. F. ( 2012 ) The political economy of food price policy in South Africa. United Nations University, World Institute for Development Economics Research (UNU-WIDER) Working Papers No.102, UNU-WIDER, Helsinki. Kleibergen F. , Paap R. ( 2006 ) ‘ Generalized Reduced Rank Tests Using The Singular Value Decomposition ’, Journal of Econometrics , 133 ( 1 ): 97 – 126 . Google Scholar CrossRef Search ADS Labadarios D. , Steyn N. P. , Nel J. ( 2011 ) ‘ How Diverse is the Diet of Adult South Africans? ’, Nutrition Journal , 10 ( 33 ): 1 – 11 . Google Scholar PubMed Lee D. S. ( 2008 ) ‘ Randomized Experiments From Non-random Selection in U.S. House Elections ’, Journal of Econometrics , 142 ( 2 ): 675 – 97 . Google Scholar CrossRef Search ADS Lee D. S. , Lemieux T. ( 2010 ) ‘ Regression Discontinuity Designs in Economics ’, Journal of Economic Literature , 48 ( 2 ): 281 – 355 . Google Scholar CrossRef Search ADS Leibbrandt M. , Levinsohn J. ( 2011 ) Fifteen years on: Household incomes in South Africa. NBER Working Papers No. 16661, NBER, Cambridge Mass. Leibbrandt M. , Wegner E. , Finn A. ( 2011 ) The policies for reducing income inequality and poverty in South Africa. Southern African Labour and Development Research Unit (SALDRU) Working Papers No. 64, University of Cape Town, Cape Town. Lund F. , Noble M. , Barnes H. , Wright G. ( 2008 ) ‘ Is There a Rationale For Conditional Cash Transfers For Children in South Africa? ’, Transformation: Critical Perspectives on Southern Africa , 70 : 70 – 91 . Google Scholar CrossRef Search ADS Manley J. , Gitter S. , Slavchevska V. ( 2013 ) ‘ How Effective Are Cash Transfers At Improving Nutritional Status? ’, World Development , 48 ( issue C ): 133 – 55 . Google Scholar CrossRef Search ADS Manyamba C. , Hendriks S. , Chilonda P. , Musaba E. ( 2012 ) Factors contributing to inequalities in food security in South Africa: implications for agricultural policy. Paper presented at conference ‘Strategies to overcome poverty & inequality’, University of Cape Town, Cape Town. McEwen H. , Kannemeyer C. , Woolard I. ( 2009 ) Social assistance grants: Analysis of the NIDS wave 1 dataset. NIDS Discussion papers No.10, National Income Dynamics Study (NIDS). Southern African Labour and Development Research Unit (SALDRU) and University of Cape Town, Cape Town. Pauw K. , Mncube L. ( 2007 ) Expanding the social security net in South Africa: opportunities, challenges and constraints. Development Policy Research Unit (DPRU) Working Paper No. 127, University of Cape Town, Cape Town. Pellerano L. , Daidone S. , Davis B. , Farooq M. , Homayoun M. , Kardan A. , Masasa M. , Niang O. , Ramirez B. , Safi N. ( 2016 ) ‘Does Evidence Matter? Role of the Evaluation of the Child Grants Programme in the Consolidation of the Social Protection sector in Lesotho’, in Davis B. , Handa S. , Hypher N. , Winder Rossi N. , Winters P. , Yablonski J. (eds) , From Evidence to Action. The Story of Cash Transfers and Impact Evaluation in sub-Saharan Africa . Oxford : Oxford University Press . Pellerano L. , Moratti M. , Jakobsen M. , Bajgar M. , Barca V. ( 2014 ) Child Grants Programme Impact Evaluation (Follow-up Report) . Oxford : Oxford Policy Management . Pieroni L. , Lanari D. , Salmasi L. ( 2013 ) ‘ Food Prices and Overweight Patterns in Italy ’, The European Journal of Health Economics , 14 ( 1 ): 133 – 51 . Google Scholar CrossRef Search ADS PubMed Pieroni L. , Salmasi L. ( 2015 ) ‘ Does Cigarette Smoking Affect Body Weight? Causal Estimates From the Clean Indoor Air Law Discontinuity ’, Economica , 82 ( 328 ): 671 – 704 . Google Scholar CrossRef Search ADS Ravallion M. ( 1998 ) Poverty Lines in Theory and Practice . Washington DC : The World Bank . Google Scholar CrossRef Search ADS Ruel M. T. ( 2003 ) ‘ Operationalizing Dietary Diversity: A Review of Measurement Issues and Research Priorities ’, The Journal of Nutrition , 133 ( 11 ): 3911 – 26 . Google Scholar CrossRef Search ADS Samson M. , Heinrich C. J. , Hoddinott J. , Laryea-Adjei G. , Buthelezi T. , Cluver L. , Jehoma S. , Mogotsi M. , Stevens T. , van Nieckerk I. , Nyokangi E. ( 2016 ) ‘The impact of a promise realized’, in Davis B. , Handa S. , Hypher N. , Winder Rossi N. , Winters P. , Yablonski J. (eds) , From Evidence to Action. The Story of Cash Transfers and Impact Evaluation in sub-Saharan Africa . Oxford : Oxford University Press . Samson M. , Heinrich C. , Williams M. , Kaniki S. , Muzondo T. , Mac Quene K. , Van Niekerk I. ( 2008 ) Quantitative Analysis of the Impact of the Child Support Grant. Report to the Department of Social Development, South Africa, Statistics South Africa and United Nations Children’s Fund (UNICEF) . Pretoria : Republic of South Africa . Slater R. , Holmes R. , Mathers N. ( 2014 ) Food and nutrition (in)security and social protection. OECD Development Co-operation Working Papers No. 15, OECD, Paris. South African Department of Agriculture ( 2002 ) The Integrated Food Security Strategy for South Africa . Pretoria : Republic of South Africa . South African Department of Health ( 2013 ) South African National Health and Nutrition Examination Survey (SANHANES1) . Pretoria : Republic of South Africa . South African Social Security Agency, SASSA ( 2015 ) A statistical summary of social grants in South Africa. Fact sheet: Issue No. 2, Pretoria. Statistics South Africa, StatSA ( 2007 a) A Discussion Note: Constructing Comparable Household Survey Data for the Analysis Of Poverty in South Africa (1995–2000) . Pretoria : Statistics South Africa . Statistics South Africa, StatSA ( 2007 b) A National Poverty Line for South Africa . Pretoria : Statistics South Africa . Statistics South Africa, StatSA ( 2014 ) Poverty Trends in South Africa. An Examination of Absolute Poverty between 2006 and 2011 . Pretoria : Statistics South Africa . Tzioumis E. , Adair L. S. ( 2014 ) ‘ Childhood Dual Burden of Under- And Over-nutrition In Low- And Middle-income Countries: A Critical Review ’, Food and Nutrition Bulletin , 35 ( 2 ): 230 – 43 . Google Scholar CrossRef Search ADS PubMed van der Berg S. , Siebrits F. K. , Lekezwa B. ( 2010 ) Efficiency and equity effects of social grants in South Africa. Stellenbosch Economic Working Papers No.15, Stellenbosch University, Cape Town. van der Klaauw W. ( 2002 ) ‘ Estimating the Effect of Financial Aid Offers on College Enrolment: A Regression-discontinuity Approach ’, International Economic Review , 43 ( 4 ): 1249 – 87 . Google Scholar CrossRef Search ADS Vorster H. H. , Badham J. B. , Venter C. S. ( 2002 ) ‘ An Introduction to the Revised Food-based Dietary Guidelines for South Africa ’, South African Journal of Clinical Nutrition , 26 ( 3 ): 1 – 164 . Williams M. J. 2007 . The social and economic impacts of South Africa’s Child Support Grant. Economic Policy Research Institute (EPRI) Working Paper No. 39, EPRI, Cape Town. Woolard I. , Harttgen K. , Klasen S. ( 2011 ) ‘ The History and Impact of Social Security in South Africa: Experiences and Lessons ’, Canadian Journal of Development Studies , 32 ( 4 ): 357 – 80 . Google Scholar CrossRef Search ADS Woolard I. , Klasen S. ( 2005 ) ‘ Determinants of Income Mobility and Household Poverty Dynamics in South Africa ’, Journal of Development Studies , 41 ( 5 ): 865 – 97 . Google Scholar CrossRef Search ADS World Bank ( 2015 ) The State of Social Safety Nets . Washington DC : The World Bank . You J. ( 2013 ) ‘ The Role of Microcredit in Older Children’s Nutrition: Quasi-experimental Evidence from Rural China ’, Food Policy , 43 ( C ): 167 – 79 . Google Scholar CrossRef Search ADS Footnotes 1 South Africa has several types of social grants targeted to children, older persons and people with disabilities, amounting to 3.5%–4% of GDP (Department Social Development, 2010). By March 2015, these programmes had reached almost 16.5 million people, representing more than 25% of the population (South African Social Security Agency, SASSA, 2015). 2 Although food expenditure is only a proxy for food security, in the case of low-income households it estimates information closely related to access to food, which is an important dimension of food security (Burchi et al., 2016). See also Coetzee (2013) for a discussion of the shortcomings related to this index. 3 The 2008 wave can be used as baseline for running placebo experiments because the policy change in the age eligibility was introduced in 2010. 4 The grant is given for each beneficiary child up to a maximum of six children per caregiver. The transfer was fixed at a level of R100 per month in 1998, but this has increased over the years, reaching R280 in 2012 (and R320 in 2014). As from 2008, the amount of the grant is now adjusted every year for inflation. 5 The primary caregiver is defined as the person who takes primary responsibility for meeting the daily care needs of the child, without payment. In 98% of the cases, the caregiver is a woman of the household in which the child lives (Agüero et al., 2010). 6 Similarly, Heinrich and Brill (2015) showed that some children who should have been enroled in the CSG following the increase in the eligible age were disconnected from the programme because of lack of knowledge of the social welfare offices regarding the policy changes, or because of misinformation among households and high burdens placed on families in the re-application process which discouraged participation of the poorer households. 7 The NIDS is the only South African dataset that provides some detailed information (e.g., CSG receipts, reliable date of birth of beneficiary children) which are crucial for evaluating the impact of the CSG on food expenditure at the national level. In particular, the dates of birth of children younger than 15 years are written down by the interviewers using birth certificates, and a computer check allows comparing the answers of the three waves. See also Eyal and Woolard (2013) and McEwen et al. (2009). 8 Note that the attrition rate in NIDS is very low or negative. Indeed, 7,296 households successfully completed the interview during the NIDS wave 1, 6,787 households were re-interviewed during wave 2, and more households than in wave 2 (8,040) were interviewed again during wave 3. 9 The variable is the result of the aggregation of 4 separate sources of food expenditure: (i) expenditure for food items; (ii) value of food items received as gifts; (iv) value of food items received as payment; (v) value of self-produced food items. To obtain the total food expenditure, the survey considers 32 food items in 10 major categories: (i) cereals; (ii) meat; (iii) fish; (iv) dairy products; (v) fats; (vi) fruits; (vii) vegetables; (viii) sweets; (ix) beverages; (x) other food expenses. In order to yield expenditure in terms of constant prices, it was adjusted according to the monthly and provincial Consumer Price Index (CPI) (December 2012=100). The household food expenditure was also adjusted according to a per adult equivalent scale, to take into account economies of scale at the household level. Following Woolard and Klasen (2005), we applied the formula commonly used for poverty and welfare analyses in South Africa, thus obtaining the total food expenditure per adult equivalent at constant prices (AdultEquivalentScale=HouseholdIncome(Adult+0.5*Children)0.9 ). 10 The corresponding sample is composed of 1,331 households with one beneficiary child, for a total of 2,673 observations. 11 Some caution is necessary when interpreting this index. The recall period of the NIDS was fixed at 30 days to reduce the bias related to over-reporting of low frequency food purchases (Deaton and Grosh 2000). Nonetheless, recall errors in the surveys are still possible for those households that have not been purchasing certain goods. 12 For an extensive discussion of dietary diversity in South Africa, see Labadarios et al. (2011) and South African Department of Health (2013). 13 For the definition of the two poverty lines, see del Ninno and Mills (2015) and Appendix A. 14 The main shortcoming of the LP approach concerns the choice of the most appropriate bandwidth. We follow Calonico et al. (2014a) for the best choice. 15 One of the assumptions behind the RDD is that the only observed discontinuity should be on the access to CSG. 16 As stressed by Jacob et al. (2012), more sophisticated functional forms may also be used as robustness checks of the linear formulation. 17 We apply to food expenditure the procedure used by Pieroni et al. (2013) and Pieroni and Salmasi (2015) to capture non-linear patterns in food consumption. 18 The degree of the polynomial in equation (3) is derived from the polynomial degree used by the LP estimator. 19 A second robustness check is run by using placebo experiments to assess the reliability of the internal validity assumption. Since the policy change did not occur before 2010, we can use the 2008 wave to determine if and to what extent our identification strategy is able to remove the selection bias in our estimates. 20 See also Dong and Lewbel (2015). The CPD uses a linear approximation to estimate the relation between the treatment and running variable, whereas the TED uses a linear approximation to estimate the relation between the outcome and the running variable. The corresponding treatment effect is given by the ratio of the parameters estimated in the two equations. Furthermore, we take the derivative of the estimated ratio and obtain a formal test for the stability of the complier population and of the treatment effect. 21 We use a local-polynomial regression-discontinuity point estimator with robust confidence intervals proposed by Calonico et al. (2014a). 22 We do not perform the continuity test on the province and district dummy variables. 23 As showed in Table 2, the treatment and control samples are not well balanced with respect to several covariates. However, when there is no evidence of discontinuity at the cutoff, the RDD provides unbiased estimated results. 24 The results using food expenditure in per capita terms, available upon request from the authors, do not show any statistically significant difference from those presented in Table 3. 25 Using the clustering method at household level may involve a large number of clusters and, in turn, this could produce wrong standard errors. In this case, we obtain robust results also when clustering the standard errors using the provincial and district variables. Comparing different clustering methods, allows us to be confident with the estimated standard errors. For a discussion of the different clustering procedures, see Cameron and Miller (2015). 26 Note that Coetzee (2013), as most of the literature, uses food expenditure in per capita terms, whereas we are using food expenditure in terms of adult equivalents. However, we do not find any statistical difference when assessing the impact of the CSG using per capita food expenditure. 27 Omitted non-linear specifications are available from the authors. Appendix A. Poverty line boundaries We examine the Statistics South Africa (StatsSA) money metric measure of poverty in the country (Statistics South Africa, StatSA 2007a, 2007b) following a ‘cost of basic needs’ approach, as reported by Ravallion (1998). This approach determines a consumption bundle considered adequate for basic consumption needs and its estimated cost. Following del Ninno and Mills (2015), we examine two poverty lines: food poverty line (FPL): the level of consumption below which individuals are unable to purchase sufficient food to provide them with an adequate diet; poverty line (PL): the level of consumption which allows individuals to purchase both adequate food and non-food items. The FPL is calculated as the cost of satisfying the daily energy requirement for an average person for one month, which, according to the South African Medical Research Council, is 2,261 calories per person (Statistics South Africa, StatSA, 2007b). The food basket is composed of the food items commonly consumed by all expenditure-ranked household groups and usually recommended for a balanced diet. Median quantities of the reference food basket, as purchased by reference households, were then derived from data on food expenditure at the household level, with the CPI for food in September 2000. The PL includes non-food expenditure, such as accommodation, electricity, clothing, schooling for children, transport, and medical services, among other things. The two poverty lines are measured in per adult equivalent terms, assuming that resources are equally shared in the households without any differences based on age, gender, or spousal status. The two poverty lines are also expressed in rand per month and are annually adjusted according to CPI data, which track the rate of change in the price of goods and services purchased by consumers (see Table A1). Table A1: Poverty line boundaries Year Food poverty line Poverty line 2008 259 507 2010 307 594 2011 321 620 2012 339 655 Year Food poverty line Poverty line 2008 259 507 2010 307 594 2011 321 620 2012 339 655 Notes: Poverty line boundaries extracted from Statistics South Africa (StatsSA) (Statistics South Africa, StatSA 2014). View Large Table A1: Poverty line boundaries Year Food poverty line Poverty line 2008 259 507 2010 307 594 2011 321 620 2012 339 655 Year Food poverty line Poverty line 2008 259 507 2010 307 594 2011 321 620 2012 339 655 Notes: Poverty line boundaries extracted from Statistics South Africa (StatsSA) (Statistics South Africa, StatSA 2014). View Large Appendix B. Falsification tests: balance of selected covariates Figure B1. Figure B1: View largeDownload slide Figure B1: View largeDownload slide Appendix C. Falsification tests: impact of CSG on selected outcome variable, year 2008 Food expenditure Carbohydrates Dairy products IV PS IV PS IV PS Child Support Grant (CSG) 19.274 21.748 0.941 1.243 −0.032 0.044 (34.971) (33.668) (1.692) (1.512) (0.764) (0.672) Constant term 1103.221*** 1018.377*** 21.772*** 21.135*** 2.961* 2.955* (88.104) (106.131) (6.099) (6.095) (1.641) (1.608) Cragg–Donald Wald F statistic 413.206 413.206 413.206 Kleibergen–Paap Wald rk F statistic 460.191 460.191 460.191 R2 0.652 0.653 0.453 0.452 0.095 0.095 Adjusted R2 0.628 0.630 0.416 0.415 0.034 0.034 No. of observations 1139 1139 1139 1139 1139 1139 Food expenditure Carbohydrates Dairy products IV PS IV PS IV PS Child Support Grant (CSG) 19.274 21.748 0.941 1.243 −0.032 0.044 (34.971) (33.668) (1.692) (1.512) (0.764) (0.672) Constant term 1103.221*** 1018.377*** 21.772*** 21.135*** 2.961* 2.955* (88.104) (106.131) (6.099) (6.095) (1.641) (1.608) Cragg–Donald Wald F statistic 413.206 413.206 413.206 Kleibergen–Paap Wald rk F statistic 460.191 460.191 460.191 R2 0.652 0.653 0.453 0.452 0.095 0.095 Adjusted R2 0.628 0.630 0.416 0.415 0.034 0.034 No. of observations 1139 1139 1139 1139 1139 1139 Notes: The dependent variables are shares of food groups in total food expenditure per adult equivalent. In all specifications, dependent variable is total food expenditure per adult equivalent (see the list of the selected covariates in Section 2). Robust standard errors, clustered at household level, are shown in brackets. Asterisks: p-value levels ( *p<0.1;**p<0.05;***p<0.01). For each model, linear (1) and non-linear specification (2) is shown. When two-step PS procedure is applied, non-linearities are introduced only in first-stage regression. Tests for weak instrument hypothesis and first-stage F statistics and Wald statistics based on Cragg and Donald (1993) and Kleibergen and Paap (2006). †Confidence intervals for Kleibergen–Paap F test statistic follow Bazzi and Clemens (2013). Food expenditure Carbohydrates Dairy products IV PS IV PS IV PS Child Support Grant (CSG) 19.274 21.748 0.941 1.243 −0.032 0.044 (34.971) (33.668) (1.692) (1.512) (0.764) (0.672) Constant term 1103.221*** 1018.377*** 21.772*** 21.135*** 2.961* 2.955* (88.104) (106.131) (6.099) (6.095) (1.641) (1.608) Cragg–Donald Wald F statistic 413.206 413.206 413.206 Kleibergen–Paap Wald rk F statistic 460.191 460.191 460.191 R2 0.652 0.653 0.453 0.452 0.095 0.095 Adjusted R2 0.628 0.630 0.416 0.415 0.034 0.034 No. of observations 1139 1139 1139 1139 1139 1139 Food expenditure Carbohydrates Dairy products IV PS IV PS IV PS Child Support Grant (CSG) 19.274 21.748 0.941 1.243 −0.032 0.044 (34.971) (33.668) (1.692) (1.512) (0.764) (0.672) Constant term 1103.221*** 1018.377*** 21.772*** 21.135*** 2.961* 2.955* (88.104) (106.131) (6.099) (6.095) (1.641) (1.608) Cragg–Donald Wald F statistic 413.206 413.206 413.206 Kleibergen–Paap Wald rk F statistic 460.191 460.191 460.191 R2 0.652 0.653 0.453 0.452 0.095 0.095 Adjusted R2 0.628 0.630 0.416 0.415 0.034 0.034 No. of observations 1139 1139 1139 1139 1139 1139 Notes: The dependent variables are shares of food groups in total food expenditure per adult equivalent. In all specifications, dependent variable is total food expenditure per adult equivalent (see the list of the selected covariates in Section 2). Robust standard errors, clustered at household level, are shown in brackets. Asterisks: p-value levels ( *p<0.1;**p<0.05;***p<0.01). For each model, linear (1) and non-linear specification (2) is shown. When two-step PS procedure is applied, non-linearities are introduced only in first-stage regression. Tests for weak instrument hypothesis and first-stage F statistics and Wald statistics based on Cragg and Donald (1993) and Kleibergen and Paap (2006). †Confidence intervals for Kleibergen–Paap F test statistic follow Bazzi and Clemens (2013). Proteins Vitamins Dietary diversity IV PS IV PS IV PS Child Support Grant (CSG) 0.900 0.717 1.249 1.092 0.008 −0.074 (1.679) (1.451) (0.887) (0.807) (0.233) (0.206) Constant term 19.652*** 19.194*** 8.747*** 8.080** 9.502*** 9.525*** (4.160) (4.166) (3.268) (3.322) (0.684) (0.682) Cragg–Donald Wald F statistic 413.206 413.206 407.654 Kleibergen–Paap Wald rk F statistic 460.191 460.191 452.837 R2 0.215 0.215 0.109 0.117 0.394 0.394 Adjusted R2 0.162 0.162 0.049 0.058 0.353 0.353 No. of observations 1139 1139 1139 1139 1133 1133 Proteins Vitamins Dietary diversity IV PS IV PS IV PS Child Support Grant (CSG) 0.900 0.717 1.249 1.092 0.008 −0.074 (1.679) (1.451) (0.887) (0.807) (0.233) (0.206) Constant term 19.652*** 19.194*** 8.747*** 8.080** 9.502*** 9.525*** (4.160) (4.166) (3.268) (3.322) (0.684) (0.682) Cragg–Donald Wald F statistic 413.206 413.206 407.654 Kleibergen–Paap Wald rk F statistic 460.191 460.191 452.837 R2 0.215 0.215 0.109 0.117 0.394 0.394 Adjusted R2 0.162 0.162 0.049 0.058 0.353 0.353 No. of observations 1139 1139 1139 1139 1133 1133 Notes: The dependent variables are shares of food groups in total food expenditure per adult equivalent. In all specifications, dependent variable is total food expenditure per adult equivalent (see the list of the selected covariates in Section 2). Robust standard errors, clustered at household level, are shown in brackets. Asterisks: p-value levels ( *p<0.1;**p<0.05;***p<0.01). For each model, linear (1) and non-linear specification (2) is shown. When two-step PS procedure is applied, non-linearities are introduced only in first-stage regression. Tests for weak instrument hypothesis and first-stage F statistics and Wald statistics based on Cragg and Donald (1993) and Kleibergen and Paap (2006). †Confidence intervals for Kleibergen–Paap F test statistic follow Bazzi and Clemens (2013). Proteins Vitamins Dietary diversity IV PS IV PS IV PS Child Support Grant (CSG) 0.900 0.717 1.249 1.092 0.008 −0.074 (1.679) (1.451) (0.887) (0.807) (0.233) (0.206) Constant term 19.652*** 19.194*** 8.747*** 8.080** 9.502*** 9.525*** (4.160) (4.166) (3.268) (3.322) (0.684) (0.682) Cragg–Donald Wald F statistic 413.206 413.206 407.654 Kleibergen–Paap Wald rk F statistic 460.191 460.191 452.837 R2 0.215 0.215 0.109 0.117 0.394 0.394 Adjusted R2 0.162 0.162 0.049 0.058 0.353 0.353 No. of observations 1139 1139 1139 1139 1133 1133 Proteins Vitamins Dietary diversity IV PS IV PS IV PS Child Support Grant (CSG) 0.900 0.717 1.249 1.092 0.008 −0.074 (1.679) (1.451) (0.887) (0.807) (0.233) (0.206) Constant term 19.652*** 19.194*** 8.747*** 8.080** 9.502*** 9.525*** (4.160) (4.166) (3.268) (3.322) (0.684) (0.682) Cragg–Donald Wald F statistic 413.206 413.206 407.654 Kleibergen–Paap Wald rk F statistic 460.191 460.191 452.837 R2 0.215 0.215 0.109 0.117 0.394 0.394 Adjusted R2 0.162 0.162 0.049 0.058 0.353 0.353 No. of observations 1139 1139 1139 1139 1133 1133 Notes: The dependent variables are shares of food groups in total food expenditure per adult equivalent. In all specifications, dependent variable is total food expenditure per adult equivalent (see the list of the selected covariates in Section 2). Robust standard errors, clustered at household level, are shown in brackets. Asterisks: p-value levels ( *p<0.1;**p<0.05;***p<0.01). For each model, linear (1) and non-linear specification (2) is shown. When two-step PS procedure is applied, non-linearities are introduced only in first-stage regression. Tests for weak instrument hypothesis and first-stage F statistics and Wald statistics based on Cragg and Donald (1993) and Kleibergen and Paap (2006). †Confidence intervals for Kleibergen–Paap F test statistic follow Bazzi and Clemens (2013). © The Author(s) 2017. Published by Oxford University Press on behalf of the Centre for the Study of African Economies, all rights reserved. For Permissions, please email: journals.permissions@oup.com This article is published and distributed under the terms of the Oxford University Press, Standard Journals Publication Model (https://academic.oup.com/journals/pages/about_us/legal/notices)

Journal

Journal of African EconomiesOxford University Press

Published: Aug 1, 2018

There are no references for this article.

You’re reading a free preview. Subscribe to read the entire article.


DeepDyve is your
personal research library

It’s your single place to instantly
discover and read the research
that matters to you.

Enjoy affordable access to
over 18 million articles from more than
15,000 peer-reviewed journals.

All for just $49/month

Explore the DeepDyve Library

Search

Query the DeepDyve database, plus search all of PubMed and Google Scholar seamlessly

Organize

Save any article or search result from DeepDyve, PubMed, and Google Scholar... all in one place.

Access

Get unlimited, online access to over 18 million full-text articles from more than 15,000 scientific journals.

Your journals are on DeepDyve

Read from thousands of the leading scholarly journals from SpringerNature, Elsevier, Wiley-Blackwell, Oxford University Press and more.

All the latest content is available, no embargo periods.

See the journals in your area

DeepDyve

Freelancer

DeepDyve

Pro

Price

FREE

$49/month
$360/year

Save searches from
Google Scholar,
PubMed

Create lists to
organize your research

Export lists, citations

Read DeepDyve articles

Abstract access only

Unlimited access to over
18 million full-text articles

Print

20 pages / month

PDF Discount

20% off