1
MSc in Economics
Faculty of Economics and Business
University of Amsterdam
The Effect of Cash Transfer Programs on
Height-For-Age and Weight-For-Age of
Infants in Rural Ecuador: A Quantile
Regression Approach
Mauricio Chavez Gomez (10397337)
Master Thesis
2
Table of Contents
1 Introduction ... 3
2 Literature Review... 4
2.1 Cash Transfer Programs ... 4
2.2 Quantile Regression for Assessing Anthropometric Measures ... 5
2.3 Height-for-Age, Weight-for-Age and its Implications for Children ... 7
3 Empirical Approach ... 8
3.1 Intention-to-treat (ITT) ... 10
3.2 Quantile Treatment Effect (QTE) ... 11
4 Data ... 13
5 Results ... 17
6 Conclusion ... 24
3
1 Introduction
Cash transfer programs have been used as a strategy for reducing poverty for more than 15 years already. Other goals of these programs include health, schooling and nutrition improvements for children (Fiszbein et al. 2009). Departing from these goals, it is natural that one may ask: how effective are cash transfer programs in improving these conditions? The present thesis addresses one dimension of this question by asking: What is the Effect of Cash Transfer Programs on Height-for-Age and Weight-for-Age of Infants in Rural Ecuador?
More than 100 million people benefit from cash transfer programs around the world (only in Brazil and Mexico, beneficiaries add up to more than 60 million people) with coverage proportions that can reach the 40% of the population for some countries. In terms of money, this intervention can cost up to a 0.5% of the country GDP as in Brazil and Ecuador (Fiszbein et al. 2009). From this perspective, it is important to evaluate if this policy is having the expected results in the dimensions it is intended to improve, as significant amounts of resources are allocated to it. One of the dimensions in which positive results are expected is child nutrition. Undernutrition is a major health problem around the world (as major as to be included in the Millenium Development Goals) due to the negative consequences that it can have for children, as for example: higher mortality rates (Caulfield et al. 2004), reduced mental development (Ruel & Hoddinott 2008), poorer school achievement (Grantham-McGregor & Ani 2001), among others. Unicef et al.(2012) estimate that more than 150 million children around the world suffer from low Height-for-Age (stunting), while more than 100 million suffer from low Weight-for-Height-for-Age (underweight). If policy makers expect that cash transfers can reduce the incidence of undernutrition, then it is important to thoroughly evaluate these programs. Proper evaluation will allow deciding if complementary interventions are actually needed for fighting against this problem.
There is an extensive body of literature evaluating the impact of cash transfer programs. Fiszbein et al. (2009) survey this literature and for the specific outcomes, results are mixed. Positive impacts on children’s height are found in Colombia and Mexico whereas in Brazil, Honduras, Nicaragua and Ecuador no impact is found. These evaluations have been carried through the use of Ordinary Least Squares (OLS) in a Randomized Controlled Trial (RCT) design; except for the one in Colombia which used a Differences-in-Differences with Propensity Score Matching approach. The contribution of the current work on the reassessment of the Ecuadorian cash transfer program lies on the utilization of Quantile Regression Methods. Quantile Regression (QR) has the advantage of being robust to outliers and more informative than OLS. This approach was originally proposed by Koenker & Bassett (1978) as the Least Absolute Deviations method, in which the minimization of the absolute values of the errors yields the median. While the mean of a distribution will move along with outlying observations, the median will remain in the same place. For this reason, the optimality of the mean estimation is dubious when the
4 distribution is non-Gaussian, making QR more robust and superior in this case. From this, it is straightforward that the normality assumption is not needed to consider the QR estimation unbiased, in contrast to OLS which heavily relies on it. The use of Quantile Regression (QR) allows analyzing the impact of the program at different points of the conditional distribution of the outcome, whereas with Ordinary Least Squares only the mean impact is obtained. This gives a better overview of the results, and the analysis of different quantiles is highly useful in assessing health related outcomes. For this type of indicators, it is common to find the most vulnerable individuals in the tails of the distribution; therefore, raising an interest on having information about the impact of the intervention on different quantiles. Although the main interest is the effect of the cash transfer program, this empirical tool also helps to have a more detailed idea of the associations of the outcome variables with the control variables for the different points of the distribution. This allows not only answering the question Does it matter? but as well
For whom and how much does it matter? In addition, Quantile Treatment Effects (QTE) are
also estimated. QTE gives the chance of instrumenting the variable of interest in order to eliminate endogeneity. In this case, self-selection is the potential source of endogeneity since in spite of being randomized into the treatment group, only 84% of the subjects actually took the program.
The empirical exercise shows that the program has no effect on Height-for-Age at any quantile. In contrast, for the Weight-for-Age indicator, the estimation shows that the program has a negative effect of roughly 13% of a standard deviation at the 75th percentile. With respect to the set of covariates, positive associations are found between the outcome variables and the education years of the mother whereas negative associations are found between the outcomes and the amount of people in the household. In addition, signs of vulnerability for younger children are observed since being in the age group of 0-12 months in baseline is associated with having up to approximately 40% and 50% of a standard deviation less in Height-for-Age and Weight-for-Age respectively. The remainder of this thesis is organized as follows. The next section provides a literature review on the relevant issues concerning this topic. Section 3 describes the empirical approach. Section 4 the data used. Section 5 shows the main findings and section 6 gives a final summary and conclusions.
2 Literature Review
2.1 Cash Transfer Programs
By the year 2010, at least three regions in the world – namely Sub-Saharan Africa, South Asia and Latin America and Caribbean- had implemented cash transfer programs in order to confront several problems related to poverty and the associated vulnerability of the poorest. These three regions include more than 20 countries with most of them located in Latin America and Caribbean with programs like: Bolsa Familia for Brazil; Familias en
5 Acción for Colombia; Bono de Desarrollo Humano for Ecuador; Oportunidades / Progresa for Mexico; Juntos for Peru, among others (ELLA 2011).
Cash transfer programs have been implemented with the aim of reducing inequality, stopping intergenerational transmission of poverty and also improving health, schooling and nutrition for children (Fiszbein et al. 2009). The pathway through which these objectives are attained partially lies on the conditionality that can be attached to these cash transfer programs. More specifically, these programs are usually designed with conditions that need to be accomplished in order for the households to receive the transfer. Examples of these conditions are: children attending a minimum of school days, health checkups and receiving dietary supplements; pregnant women attending perinatal controls; parents attending meetings in relevant topics for development (nutrition, for example) and others. If these conditions actually contribute to reach the desired outcomes is not subject of this thesis, though some interesting positions can be found in Baird et al. (2011); de Brauw & Hoddinott (2011) and Freeland (2007).
In the specific health case, the evidence of mean impacts of cash transfer programs on anthropometric measures is mixed. In the World Bank’s Policy Research Report “Conditional Cash Transfers Reducing Present and Future Poverty” prepared by Fiszbein et al. (2009), Mexico and Colombia show a positive impact of their program on Height-for-Age while the programs in Brazil, Honduras and Nicaragua show no impact on this outcome variable. For the rural Ecuador case, Fernald & Hidrobo (2011) show that the cash transfer has no effect on Height-for-Age and hemoglobin levels for children aged between 12 and 35 months. Likewise, on assessing the effect of the Ecuadorian program on physical measures (hemoglobin, height and fine motor control), Paxson & Schady (2010) find no effect; although, when the sample is disaggregated by income and age, there are some “modest” effects as the authors mention.
As seen in this sub-section, the impact of cash transfer programs on anthropometric measures such as Height-for-Age and Weight-for-Age is mixed; and, in the particular case of study (Ecuador) it is rather inexistent. Nonetheless, all the treatment estimations are only mean estimates while the attempt of the present thesis is to find estimations at different quantiles of the outcome variable. The advantage of using the Quantile Regression methodology would be to verify if – at least for the Ecuadorian case- there is no impact at the tails of the weight and height distributions. This would be of better interest than just a mean effect, considering the severe health problems that can arise from being located in the tails of these distributions.
2.2 Quantile Regression for Assessing Anthropometric Measures
As many authors coincide, more than 90% of all of the empirical research lies in the estimation of mean effects. Therefore, it is difficult to find studies using methodologies as Quantile Regression (QR), in which an impact is analyzed along the whole distribution of
6 the outcome variable. The fields of application of QR are varied. Yu et al. (2003) survey several topics in which the methodology in its different forms can be or is applied: reference charts in medicine – as for example growth charts-; survival analysis – for example expected time of survival of a heart transplant patient-; financial value at risk; income studies in labour economics; hydrology – modeling rainfalls for water provision, for instance-; and, detecting heteroskedasticity. Other applications relate to improvements in the education field and many others use this empirical tool in order to assess inequality – usually income inequality among different groups.
From the few studies addressing a somewhat similar research question and using the QR methodology, I mention two of them. In Variyam et al. (2002) the authors try to find the association between sociodemographic variables like income, race, or region; the intake of macronutrients and anthropometric measures for a sample of the US population. The motivation for using QR is the fact that the most affected individuals by problems derived from the intake of macronutrients are not the ones around the mean of the distribution but those on the tails – high fat consumption or low protein consumption, for example. As the authors also explain, an OLS type of method would be appropriate if the impact of the explanatory variables were the same for every individual in the outcome variable; but usually this can’t be assumed with medical related issues.
Aturupane et al. (2008) find a similar motivation to use the methodology as the one used in this thesis and in the previously mentioned article. In their study, they try to find the determinants of child weight in Sri Lanka. The authors state that there are several interventions that may not impact the “average” child, hence the importance to assess them for the ones that are at the bottom of the nutritional distribution (who usually are the ones with more risk). They also state that the QR method does not simply answer if policy intervention can affect nutrition on children but more specifically and more relevant, the method allows them to answer who are the ones that are actually being influenced by the intervention. In their preferred specification they use variables such as child’s age, sex, birth order, expenditure, among others; and, their outcomes of interest are Height-for-Age and Weight-for-Age for children under 5 years old. Although the study addresses an interesting research question, the way the methodology is used would be analogous to just using OLS on a cross-section with potentially endogenous variables (this method is also used in the study for comparison purposes). Consequently, the study does not look for a causal relation but only for correlations of the explanatory variables with the outcomes at different quantiles of the distribution.
Although there are good reasons – which will be presented in section 3- for using QR to address the present research question, to my own knowledge – and as the scarce literature shows-, cash transfer programs have not been analyzed under this approach.
7
2.3 Height-for-Age, Weight-for-Age and its Implications for Children
Chronic malnutrition or stunting is understood as the delay or reduction on child’s potential growth over time, which can be the result of a prolonged nutritional deprivation or severe illnesses with consequences to cognitive and organ development. On the other hand, underweight corresponds to a low Weight-for-Age which can be a consequence of past or present undernourishment (Nutrition Works et al. 2011). A child is stunted if his or her Height-for-Age is two standard deviations or more below the median of the world’s referential measures. Likewise, a child is underweighted if his or her Weight-for-Age is 2 standard deviations below the median. In 2011 about 165 million children under 5 years old around the world suffered from stunting with the 90% of this group being concentrated in Asia and Africa; roughly 100 million children around the world suffer from underweight (UNICEF et al. 2012).
The usual mechanisms through which stunting or underweight can cause serious impairments in the cognitive development of children, or even death, are deficiencies in vitamins and minerals. Lack of zinc is associated with deaths by diarrhea, pneumonia, malaria and measles. In a study conducted by Caulfield et al. (2004) using data from 10 cohort studies, it is stated that roughly more than half of deaths for the causes mentioned before can be attributed to undernourishment or low Weight-for-Age. In addition, Ruel & Hoddinott (2008) refer to iodine and iron deficiencies primarily affecting mental development and functioning as well as vitamin A deficiencies leading to blindness or death. Furthermore, iron deficiency leading to anemia can affect the structure of the central nervous system creating cognitive problems and inducing behavioral issues such as children isolating themselves due to a limited movement capacity and being more fearful, tense or unhappy (Grantham-McGregor & Ani 2001).
Although the consequences of undernourishment seem to be widely understood in medical terms, it is more complicated to trace down its effects on other variables of interest such as school performance or academic attainments. This is mostly due to the scarce amount of studies that can actually state causal relationships in the light of statistical inference. Many of the existing studies can’t be interpreted as causal as it is not clear, for example, if undernourished children perform worse in school because of their lack of nutrients or if it is confounded by their general disadvantaged situation (Glewwe & King 2001).
Even though many studies do not have a causal interpretation, it is worth mentioning some interesting associations that have been found. Grantham-McGregor & Ani (2001) in a review of previous studies on the effect of iron deficiency on cognitive development, show that anemic children older than 2 had poorer cognition and school achievement than their non-anemic counterparts. Furthermore, anemic children before the age of 2 do not necessarily catch up with non-anemic ones even after receiving iron supplements. As a critique to this type of studies, Pollitt (2001) states that cognitive development scales that
8 have been designed in order to search for this associations do not have the adequate characteristics to compare development of children in the first two years of life, which casts doubt on their results. Nevertheless, he recognizes that iron deficiency can indeed affect some mental functions in early and late development.
Mendez & Adair (1999) inspect the relationship between stunting in the first 2 years of life and cognitive development of the same children at the age of 8 and 11. The authors use data from the CEBU Longitudinal Health and Nutrition Study in Philippines with more than 2000 subjects. Delayed initial enrolment, higher rates of absenteeism, repetition of grades and lower cognitive test scores are found among the kids who were stunted by the age of 2. An interesting finding of this study is the fact that the gap of test scores between stunted and non-stunted children is smaller at age 11 than at age 8, showing the possibility of a catch-up. In the same vein, using a panel dataset Alderman et al. (2006) find that better Height-for-Age scores are related to higher school attainments, starting school earlier and being taller later on in life. The empirical approach of this study consists on a maternal fixed effects instrumental variables approach, exploiting the exposure of some siblings of the same family to a civil war and a drought in rural Zimbabwe.
Finally, in an unusual Randomized Controlled Trial (RCT), using a sample of 555 children in a Kenyan school, the impact of dietary intervention on cognitive development was studied by Whaley & Sigman (2003). The treatment group received meat, milk or an energy supplement during 21 months while the control group received no intervention. Children who received the meat and energy treatment had better achievements in arithmetic ability. Only-meat treated kids had the best outcomes in tests regarding comparisons, analogies and perceptual details.
Although the actual mechanisms of the effects that have been reviewed in this sub section are not completely clear; it is indeed clear that nutrition in early childhood remains to be a critical issue that needs prior attention. Failure to address the undernutrition problem can expose children to all the unwished outcomes reviewed in this section.
3 Empirical Approach
Quantile Regression was originally proposed by Koenker & Bassett (1978) as an alternative for the ordinary least squares method in the presence of outliers. As in OLS, the quantile regression method departs from the minimization problem of the loss function – namely, the minimization of the prediction errors. One of the main differences lies in the characterization of the loss function: while OLS finds the parameter that minimizes the squared of the errors, QR as presented by Koenker & Bassett (1978) minimizes the absolute value of these deviations. Therefore, the method was named as Least Absolute Deviation when the quantile of interest is the 0.50thquantile or, in other words, the median.
9 More formally, If ̂ as a function of is the estimation of , then the error of estimation or
loss function will be ( ) with ̂. As just above mentioned, OLS will minimize
the square of the error, ( ) , while the Least Absolute Deviation will minimize its absolute value, ( ) | |. In the same way that the sample mean is defined as the solution to the minimizations of the sum of squared residuals, the solution to the minimization of the sum of absolute residuals yields the median (Koenker & Hallock 2001). OLS assumes that the data is distributed in a symmetrical way, and in this case the mean and the median will be the same. In case of asymmetry, the mean of the dependent variable will move along with outlying observations therefore invalidating the inference process that is derived from the estimation of the parameters. Nonetheless, the median will remain located in the same place; this is the foundation of the robustness of the method to outliers. This also implies that no assumptions about the distributional form of the errors shall be made.
For the present thesis, it is not only the median effect to be analyzed, but also the corresponding effects at different quantiles than the 0.50th. More generally, my interest is on , the Conditional Quantile Function (CQF) at different quantiles given a set of predictors , which is represented as:
( | ) ( | ) (1)
where the infimum requires that the cumulative distribution function of given , ( | ) is equal or greater than the quantile of interest . Assuming integrability and not necessarily a linear relationship of the outcome variable with the regressors -as noted in Angrist et al. (2006)-, the CQF problem is solved through:
( ) (2)
With representing the weighting function of the errors which – in contrast to the OLS weighting function- is asymmetrical except when the quantile of interest is the 50th as it is defined as:
( ) {
( )| |
( )| | (3)
so that a weight of ( ) is used in overestimation, and similarly a weight of ( ) for underestimation.
10 The use of quantile regression is justified if it is expected that the effect of the regressors on the response variable is not the same at different points of the conditional distribution of the outcome. If it were expected that the regressors have the same effect along the whole distribution, then a mean estimation suffices. This is known in the quantile regression terminology as a “location shift” as the regressors move the average, and the other parts of the distribution move in the same way, only “shifting” the complete distribution. For the anthropometric indicators under analysis, it is expected that the regressors have different effects along the distribution of the outcomes. For example, the cash transfer may not have a significant impact at the center or right tail of the distribution (the well-nourished children) as these children might have attained their growth potential already; in contrast, it may have a significant effect for those located in the left tail of the distribution (the undernourished children) as they haven’t attained their appropriate height or weight. In the same way, it can be the case that an extra member in the household is more problematic for those in the lower quantiles than for those in the upper ones. Indeed, a QR approach is more suitable in this case than just a mean estimation approach, for all the relevant information that is not obtained when using the latter.
3.1 Intention-to-treat (ITT)
As mentioned before, I am interested on the effect of the cash transfer program on Height-for-Age and Weight-for-Age of children in rural Ecuador. It must be specified that since roughly 84% of the assigned to treatment group actually took the program, the effect sought is the Intention-to-Treat (ITT). ITT estimates become of interest when – as in this case- there is partial compliance of the treatment. This means that due to dubious exogeneity of actually receiving treatment – and in order to avoid making unfounded assumptions-, the only causal effect that can be for sure interpreted is that of offering treatment to individuals. This effect is assumed to be exogenous given baseline characteristics. Consequently, the relationship to be modeled in the solving minimand of the Conditional Quantile Function , is defined:
(4)
where represents the outcome of interest for the th child in the sample, refers to the randomization of individual to treatment or control group, the set of covariates, and the error term. As it is seen, it is not the status of actually receiving the treatment which is used for the estimation, but if the individual was initially assigned to treatment or control group.
The th quantiles of interest are 0.05, 0.25, 0.50, 0.75 and 0.95. The first two are relevant since for the sample used they are close to the cutoff points of three and two standard deviations below the mean of the anthropometric measure respectively. Two standard deviations below the mean is indicated as moderate undernutrition while three standard
11 deviations below the mean as severe undernutrition. The 0.50th quantile or median estimator is useful in this case to be compared with an OLS or mean estimator. To use an OLS estimator as a benchmark will allow to have an idea on the strength of outlying observations for biasing the estimation. In addition, to compare the mean estimator with estimations at different quantiles will help to understand the usefulness of the method. If all the coefficients obtained through QR are very similar to the ones obtained through OLS, then there is not really useful information lost if only a mean estimation were used. On the other hand, if the coefficients differ across quantiles and from the mean estimator, then QR is very useful to have a better overview of the relationship between predictors and outcomes.
An important consideration must be mentioned with respect to the interpretation of results. As Angrist & Pischke (2008) point out, the estimations obtained through the use of Quantile Regression refer to “effects on distributions” rather than “effects on individuals”. For example, if the cash transfer had a positive significant effect on the fifth percentile of the conditional Weight-For-Age distribution, it is not necessarily that those infants are now better nourished. It is more accurate to state that those who are badly nourished receiving the treatment, are not as badly nourished as they would be in the case of not having received it. Thus, it is possible to say that they are better off after the treatment, but not that they are now in an upper quantile of the distribution due to the intervention.
3.2 Quantile Treatment Effect (QTE)
The advantages of QR over OLS do not prevent it from yielding biased estimations in presence of endogeneity. The basic QR has been analogously compared to be an OLS type of estimation in the sense that its interpretation can’t be causal if there are endogenous variables. Partial compliance is a relevant source of endogeneity. One can’t rule out the possibility that the group of actual treated and non-treated is not as if randomly assigned. Consequently, differences in unobservables could be biasing the estimation, invalidating any causal interpretation of the effect. For example, it could be that the children in actual beneficiary households were more (less) prone to catch up in height or weight than their non-beneficiary households’ counterparts, overestimating (underestimating) the true effect. For the present case, a take-up of about 84% of the initially-assigned-to-treatment group could cast doubt on considering that the estimation obtained by regular OLS is unbiased and consistent. The voluntary component in program take-up could be translated in obtaining estimates that reflect the effect of the treatment plus systematic differences between the treated and non-treated groups.
Abadie et al. (2002) propose a Quantile Treatment Effect (QTE) estimation analogous to the Instrumental Variable (IV) estimation in order to deal with endogeneity. As a matter of fact, the QTE estimator makes use of the same foundations as the Local Average Treatment Effect (LATE) estimator: independence of the instrument from the outcome, relevance of the instrument in explaining the endogenous variable, and monotonicity
12 (which means that there are no defiers in the experimental population). The QTE problem departs from assuming that there exist a and a such that,
( | ) , (5)
where is a binary indicator of the actual treatment status; and the treatment
status given that the instrument is equal to 1 or 0 respectively. Equation (5) states that it is possible to find a that represents the difference between the potential outcome given that the treatment was received, , and the potential outcome given that the treatment was not received, , at the different conditional quantiles for the compliant population. Regarding this , an important point is made in Abadie et al. (2002): although a difference in average is the same as an average difference, the parameter of interest in equation (5) does not represent the conditional quantile of the distribution of ( ). In other words, the Quantile methodology allows examining the difference in the conditional distributions of the potential outcomes given an intervention, not really the conditional distribution of the effect of the intervention.
The estimation of (5) would be straightforward if the compliant population would be identified; nevertheless, since one can never observe the counterfactual, the true population of compliers is not identified. More specifically, the compliers are the ones for whom the treatment status and the instrument take the same value, but one never gets to observe the treatment status for the same individual if the instrument had taken the opposite value. Under the IV framework, the QTE estimator is built upon the Abadie Kappa theorem which gives an approximation of the causal effect of the treatment for the compliant population by the use of a weighted minimization of errors including covariates. The weighting scheme or kappas takes into consideration that the unconditional average treatment effect is a weighted average of the effects on compliers, always takers and never takers and that the complying group can be somehow found throughout these weights; hence, making it possible to estimate the parameter of interest for the compliers. The compliant population is the center of interest of this estimation, since for them the treatment status (which is potentially endogenous) can be replaced by the instrument (which is assumed to be exogenous conditional on the X covariates), so that the estimation has a causal interpretation as the treatment would be no longer related to the potential outcomes.
Following Angrist & Pischke (2008), the QTE estimator results from the minimization of the weighted errors given the monotonicity assumption of the IV framework. This problem can be transformed into the minimization of double weighted errors: the QR weights and the Abadie Kappa weights . Formally expressed,
13 (7) ( ) (6) with ( ) ( | ) ( ) ( | )
When = 1 and = 0 the second term of (7) is different than zero; this represents the always-takers, as although the randomization did not assign treatment, they took it. In the opposite case, the third term of (7) is different than zero representing the never takers, as even though the randomization assigned them to treatment, they did not take it. The third case of this weighting scheme is when = 1 and = 1 or = 0 and = 0 in which the second and third term will be zero yielding a weight of 1 for these observations as they represent the complying group. The denominators of the second and third terms represent the probability of being randomized to control and treatment groups given the set of covariates.
For the available data, there is a very low proportion of treated individuals that were initially randomized to control group (about 3%). Due to this very low proportion, it can be assumed that there are no always-takers. Consequently, this weighting scheme in the QR minimand will yield the Quantile Treatment Effect on the Treated at the different th quantiles. The quantiles of interest are the same as the ones denoted in the previous section. Finally, IV estimation will also be conducted as a benchmark for the QTE estimation.
4 Data
Ecuador established its cash transfer program in 1998 with the name Bono Solidario in the middle of an upcoming economic crisis that would lead to the collapse of a substantial part of the banking system and finally adopting the US currency instead of the national one. This cash transfer program was redesigned in 2003 when it changed its name to Bono
de Desarrollo Humano.
The beneficiaries were means-tested in order to qualify for the program. This redesign gave the chance to implement an impact evaluation as families which had not received the cash transfer before became eligible – as they belonged to the first two quintiles of the poverty index calculated for redesigning the program, called the Selben- and it would be made available progressively. Six provinces were chosen in order to carry the impact evaluation: 3 in the coast and 3 in the highlands containing 77 rural and 41 urban parishes. Baseline survey was conducted between the last and the first quarter of years 2003 and 2004 respectively, while first follow up was conducted between the last third and first month of the years 2005 and 2006.
14 I had access to the rural sample of the survey conducted for evaluating the program – hence, all the analysis hereafter refer to the rural group-, which served as a basis for the final sample that Paxson & Schady (2010) used for estimating the effect of the cash transfer in Ecuador. More specifically, while the sample contains children up to 6 years old in baseline, Paxson & Schady (2010) only use children between 36 and 72 months in baseline. This is due to the fact that one of their main interests is the effect of the cash transfer on cognitive tests scores, and children younger than 36 months could not be administered with the corresponding tests. It is important to remark – as they do as well- that since the sample is restricted to households that have no kids older than 72 months and that had not been beneficiaries of the program before, it can’t be considered as representative of all the actual beneficiaries of the Bono.
The original rural sample contains 2806 children no older than age 6 whose household is classified in the first two quintiles of the Selben Poverty Index designed by the government for improving the targeting of the cash transfer. As mentioned before, these children have no siblings older than 72 months and the household had not received the cash transfer before. Beside the 2806 observations, there are 593 that were born between baseline and first survey and are not considered as part of this study.
Failure to re-interview the household on first survey accounts for roughly the 7% of the sample. In addition, there are problems with missing information and failure to measure height and/or weight for some observations. In order to lose the least possible amount of observations for the analysis, one sample for each outcome is used. The Height-for-Age sample contains 2221 observations with full information in the outcome measure and in the respective covariates. Likewise, the Weight-for-Age sample contains 2375 subjects with the required measures.
Table 1 describes baseline characteristics for the observations in the height and weight samples. The upper panels of the table show the comparison among the ones that were originally randomized into treatment and control groups for each sample. The lower panels depict the comparison made among the ones that were actually treated and not treated. As treatment take-up was not mandatory for those who were randomized into it, balancing tests are performed taking into consideration not only initial randomization (which is supposed to be completely exogenous) but also the actual treatment status. From this information it must be remarked that on average, all the children in the height sample (left panels) have low height-for-age – about 1.2 standard deviations below the mean. It shall be remembered as well that the cutoff for considering this measure a real threat to health is 2 standard deviations below the mean. Mothers of children in the sample are relatively young (around 23 years old), with approximately primary education completed and roughly a 75% of them living with their husbands or partners. I must precise that the variable “Log of imputed per capita expenditure” was not part of the survey but instead calculated in Paxson & Schady (2010). They regressed log expenditure
15 on measures of household characteristics and goods ownership from another survey to eligible households in different parishes. These coefficients were then used to impute the monthly expenditure for the present dataset.
Except for the child gender in which there is a significant difference at a 10% level of confidence, randomization seems to have actually produced balanced control and treatment groups. Likewise, the actual treated and non-treated groups seem to be well balanced, except for a significant difference in the log of per capita expenditure at a 10% level of confidence.
It is logical to expect that people who need the intervention the most, will actually receive it if there is some voluntary component for participating in the program. The latter finding is not considered troublesome, as when using a probit model for the balancing of baseline characteristics (not reported), the p-value of the imputed log per capita expenditure exceeds conventional significance levels. Regarding the significant difference in the child’s gender, it must be stated that one advantage of using Z-scores is that they are “sex-independent” as they have been already standardized with respect to their reference population. Hence, there is no inconvenient with having a difference in gender between treatment and control groups.
The right panels of Table 1 are analogous to its left panels: the same type of information is found with the difference that it draws it from the sample used for the Weight-for-Age analysis. The indicators are quite similar to the ones in the Height-for-Age sample. Children have on average a Weight-for-Age almost one standard deviation below the mean. Their mothers are relatively young with roughly primary education completion, living with their husbands or partners in about 3 out of 4 cases. In the same way as with the Height-for-Age sample, treatment and control groups seem to be well balanced. Significant differences at 5 and 10% levels for child gender and log of imputed per capita expenditure appear when the groups are analyzed by randomization and by actual treatment status respectively.
The variables presented as baseline characteristics – except for the anthropometric measures in baseline, mother living with her partner and the log of per capita expenditure- will also serve as the set of controls for the different specifications. Children, mother and household related information will control for initial characteristics that can affect the outcome in any way and could be confounded with the effect of the program. Although, this – in principle- is not very likely to happen as randomization has proven to be independent of any other characteristic. Successful randomization rules out any omitted variable bias that could arise (at least theoretically).
16
TABLE 1
Baseline Descriptive Statistics for Treatment and Control Groups in Height and Weight Samples
Height Sample Weight Sample
(1) Treated (2) Control (3) Difference (Std. Error) (1)-(2) (4) Treated (5) Control (6) Difference (Std. Error) (1)-(2) A. By Randomization Height-for-Age -1.176 -1.175 -0.001 -0.872 -0.854 -0.018 (0.1251) (0.1587) Child is male 0.491 0.528 -0.037* 0.491 0.529 -0.038** (0.0191) (0.0185)
Child's age in months 34.145 33.574 0.571 33.454 32.954 0.500
(0.8106)
(0.7094)
Mother's Education Years 6.756 6.700 0.056 6.779 6.665 0.115
(0.3048) (0.2989) Mother's Age 23.494 23.739 -0.246 23.476 23.770 -0.295 (0.2765) (0.2687)
Mother lives with husband 0.738 0.744 -0.006 0.742 0.737 0.004
(0.0294)
(0.0294)
Log of imputed per capita exp 3.483 3.503 -0.021 3.486 3.497 -0.011
(0.0377) (0.0382) Total HH members 4.901 4.858 0.043 4.899 4.903 -0.004 (0.1654) (0.1707) Obs. 1484 737 1591 784 B. By Actual Treatment Height-for-Age -1.156 -1.201 0.045 -0.889 -0.836 -0.053 (0.1017) (0.1493) Child is male 0.495 0.514 -0.019 0.497 0.512 -0.014 (0.0200) (0.0196)
Child's age in months 34.276 33.532 0.744 33.494 33.019 0.476
(0.6609)
(0.6012)
Mother's Education Years 6.722 6.758 -0.035 6.751 6.729 0.023
(0.2704) (0.2635) Mother's Age 23.574 23.578 -0.004 23.560 23.590 -0.030 (0.2472) (0.2477)
Mother lives with husband 0.743 0.736 0.007 0.748 0.729 0.019
(0.0276)
(0.0268)
Log of imputed per capita exp 3.464 3.523 -0.058* 3.462 3.526 -0.064*
(0.0336) (0.0356) Total HH members 4.917 4.846 0.071 4.919 4.875 0.044 (0.1523) (0.1557) Obs. 1264 957 1351 1024
Notes:A different sample is used for each outcome variable in order to make the most of the available observations. *,** and *** indicate 10%, 5%,
17 Controls in randomized trials are mostly used in order to improve the preciseness of estimation. The basic specification of the model includes the above mentioned variables as controls. In addition, a second specification is analyzed replacing the child’s age by a dummy indicating the age group of the children in baseline (0-12, 13-24, 25-36, 37-72). Including different age groups is of particular interest since children up to 36 months are more vulnerable to undernutrition than their older counterparts. This is because growth rates until this age are specially high (Martorell 1999) and therefore undernutrition will have a more notable effect at this stage rather than in another one. In addition, stature attained by the age of 3 (or 36 months) has shown to be a strong determinant of stature attained in adulthood (Hoddinott & Kinsey 2001), thus making this age disaggregation even more relevant.
5 Results
Table 2 presents the results of the Intention-to-Treat estimates obtained through OLS and QR for the Height sample using two different specifications. The basic specification on panel A uses controls for children’s characteristics (age and gender), mother’s characteristics (age and education) and household’s characteristics (quantity of family members). Treatment effects for this specification are not significant and their magnitudes do not exceed a 7.1% of a standard deviation. OLS and the median estimator (50th percentile) for the treatment effect differ substantially: the former more than doubles the latter, which could be interpreted as having heavy tails in the distribution downwardly biasing the mean estimator obtained through OLS. On the other hand, since both coefficients are not significant and their confidence intervals partially overlap, the latter statement is only a possibility. Although not significant, it is in a certain way puzzling to find the “wrong” signs in the estimations (except at the 95th percentile in column (6)): it would be expected that a cash transfer program has a positive effect on an anthropometric indicator rather than a negative one.
From the variables containing children related information, the gender of the child is never significant, while the age of the child is significant at the 5th, 25th, and 95th percentiles. The first two percentiles of this variable have a positive sign and the last one a negative sign. Normally, one would expect positive signs in the age of the child coefficient as it is well known that height increases with age, specially and rapidly in the early stage of life (this could be the case at the 5th and 25th percentiles). Children in the 95th percentile have a very high Height-for-Age; thus, this negative association between the child’s age and the outcome variable could be a sign that these kids receive less attention since they look “healthy”. Therefore, as they age they show a slight decrease in their growth rates. With respect to the variables indicating mother related information, the education of the mother is always significant at a 1% level with mean and median estimators quite close to each other, with only a difference of 0.001 units. Coefficients for this variable are up to almost 8.5% of a standard deviation. This is a substantial magnitude considering that a
18 mother with complete high school (12 years of education) will be associated with having a child with a Height-for-Age about one standard deviation higher than a child with an uneducated mother. The age of the mother is never significant except at the 25th percentile in which the coefficient is significant at a 10% level. From this variable, it is interesting to see that the coefficients have negative signs except at the 75th percentile. One would expect to have a positive association between the age of the mother and the anthropometric measure, perhaps since an older mother could mean a more mature person or with more experience or awareness to properly take care of a child. Nevertheless, it must also be taken into consideration that the mothers from the sample are quite young (23 years old on average), and that an older mother could also mean a mother with more children; hence, with less resources to appropriately take care of them. Finally, the variable indicating the amount of household members is always significant (except at the 5th percentile) with the expected negative sign. A negative sign is expected in this variable since more members in the household can be translated into fewer resources for everyone, unless most of its members actually have a source of income to share with the household; but, this does not seem to be the case.
Panel B uses the same controls as Panel A, but replaces the age of the child by a set of 3 dummy variables indicating the age group to which the child belongs: 0-12 months, 13-24 months and 25-36 months. As mentioned before, the inclusion of these dummies is of special relevance in order to check if younger children are indeed more vulnerable to undernutrition as the literature states. The excluded group for this age disaggregation is that of the children aged 37-72 months in baseline; hence, the coefficients for these dummy variables shall be interpreted with respect to this group.
The treatment effect for this specification does not change substantially with respect to the specification on Panel A: mean estimator (OLS) is about twice as high in magnitude than the median estimator (50th percentile), the coefficients remain insignificant and with negative signs for every quantile, and all of them show a lower magnitude than the coefficients estimated with the basic specification. Likewise, mother and household related variables show almost the same results as in the basic specification: the education of the mother is always significant with a positive association with the outcome variable and similar coefficients to Panel A; the age of the mother is only significant at the 25th percentile with a negative sign and the same magnitude in the coefficient; and, the amount of members in the household is always significant (except at the 5th percentile) with negative signs and coefficients with similar magnitudes to those previously estimated.
19
TABLE 2
OLS and Quantile Regression Intention-to-Treat Estimates of Cash Transfer on Height-for-Age A. Basic Specification Quantile Variable (1) OLS (2) 0.05 (3) 0.25 (4) 0.50 (5) 0.75 (6) 0.95 Treatment Effect -0.043 -0.071 -0.053 -0.019 -0.012 0.061 (0.0447) (0.0926) (0.0581) (0.0563) (0.0602) (0.0818)
Child's age in months 0.000 0.006** 0.004** 0.000 -0.003 -0.011***
(0.0013) (0.0026) (0.0016) (0.0016) (0.0019) (0.0029)
Child is male -0.004 -0.087 0.010 0.017 -0.021 0.080
(0.0414) (0.0872) (0.0548) (0.0533) (0.0573) (0.0829)
Mother's Education Years 0.083*** 0.069*** 0.084*** 0.081*** 0.076*** 0.076***
(0.0074) (0.0151) (0.0094) (0.0102) (0.0108) (0.0120)
Mother's Age -0.005 -0.014 -0.012* -0.009 0.001 -0.007
(0.0053) (0.0110) (0.0064) (0.0070) (0.0081) (0.0078)
Total HH Members -0.053*** -0.018 -0.039*** -0.054*** -0.057*** -0.061***
(0.0093) (0.0172) (0.0118) (0.0116) (0.0134) (0.0232)
B. Age Disaggregated Specification
Quantile Variable (1) OLS (2) 0.05 0.25(3) (4) 0.50 (5) 0.75 (6) 0.95 Treatment Effect -0.037 -0.058 -0.037 -0.014 -0.035 -0.017 (0.0447) (0.0884) (0.0583) (0.0562) (0.0593) (0.0851) Child is male -0.003 -0.025 -0.012 0.039 -0.023 0.082 (0.0413) (0.0858) (0.0546) (0.0529) (0.0565) (0.0863) Child is 0-12 months -0.147*† -0.412*** -0.289*** -0.200** -0.222** 0.206 (0.0804) (0.1550) (0.0955) (0.0911) (0.0975) (0.2319) Child is 13-24 months -0.047 -0.231* -0.213*** -0.044 0.077 0.157 (0.0590) (0.1215) (0.0808) (0.0777) (0.0805) (0.1151) Child is 25-36 months 0.136*** 0.096 0.034 0.143** 0.157** 0.263*** (0.0498) (0.1041) (0.0674) (0.0663) (0.0680) (0.0936)
Mother's education years 0.082*** 0.074*** 0.089*** 0.078*** 0.076*** 0.076***
(0.0073) (0.0140) (0.0093) (0.0102) (0.0106) (0.0121)
Mother's age -0.006 -0.011 -0.012* -0.011 0.002 -0.008
(0.0053) (0.0103) (0.0064) (0.0069) (0.0078) (0.0080)
Total HH members -0.051*** -0.016 -0.042*** -0.053*** -0.062*** -0.056**
(0.0093) (0.0171) (0.0118) (0.0115) (0.0126) (0.0240)
Notes: Dependent variable is Height-for-Age Z Score measured on first follow-up. Robust standard errors reported in parentheses. Due to sampling, clustered standard errors at parish level might be more suitable; nonetheless, only robust standard errors are used for making them comparable across methods. The use of clustered standard errors in OLS estimation does not change the significance level of any variable except for ‘†’ in which it goes from 10% to 5%. *,** and *** indicate 10%, 5% and 1% significance levels respectively.
20 With respect to the child related variables, the gender of the child is never significant (as in the basic specification) and the set of dummy variables for the different age groups shows interesting results. For the dummy indicating the 0-12 months age group, coefficients are significant for OLS and every quantile except at the 0.95th. Mean and median estimators for this variable are not very far from each other, but some noticeable differences appear across quantiles. At the 5th percentile, being in the 0-12 months age group is associated with having approximately 40% less of a standard deviation in Height-for-Age than the 36-72 months age group. In the same way, for the 25th, 50th and 75th percentiles, this associations range from 20% to roughly 29% less of a standard deviation than the reference group. Although the ones located at the 50th percentile and onwards do not suffer from malnutrition, the coefficients show that their actual growth potential has not been attained. For the dummy indicating that the child belongs to the 13-24 months age group, the coefficients are only significant at the 5th and 25th percentiles of the Height-for-Age conditional distribution, indicating an association of about 20% less of a standard deviation than their 37-72 months age group counterparts. Finally, the dummy indicating that the child belongs to the 25-36 months age group shows significant and positive coefficients for the mean estimator and at the 50th, 75th and 95th percentiles. The associations shown by these variables are in line with the vulnerability of younger children referred in the literature: the coefficients for the 0-12 months age group are the highest in magnitude with negative sign, specially for those who are initially in the worse nutritional status (5th percentile); and, as one moves to older children (i.e. the 13-24 months age group), the coefficients become smaller in magnitude and insignificant in most of the cases. More generally, if one compares the coefficients across the age groups at the different quantiles, the tendency is that the reduction in the anthropometric measure becomes smaller as the age group increases.
Table 3 is analogous to table 2 as it shows the Intention-to-Treat estimates of the cash transfer using OLS and QR but for the sample taken into consideration for the Weight-for-Age indicator analysis. For the basic specification on panel A, the treatment effect is not significant at conventional levels; mean and median estimators are close to each other and negative signs accompany all the coefficients except the one at the 25th percentile. The child’s age is always significant and positive in this specification, although the coefficients account for less than a 1% of a standard deviation. The gender of the child and the age of the mother are never significant, whereas the education of the mother is always positive and significant at 1% level for OLS estimation and for every quantile. Finally, the variable indicating the amount of members in the household is significant in the OLS estimation and for every quantile except at the 0.05th, showing a negative association between the quantity of household members and the outcome variable. Panel B of Table 3 includes the same set of controls as the ones used in panel B of table 2, with the replacement of the child’s age by the set of dummies indicating the age group of the child.
21
TABLE 3
OLS and Quantile Regression Intention-to-Treat Estimates of Cash Transfer on Weight-for-Age A. Basic Specification Quantile Variable (1) OLS (2) 0.05 0.25(3) (4) 0.50 (5) 0.75 (6) 0.95 Treatment Effect -0.054 -0.127 0.079 -0.034 -0.081 -0.081 (0.0437) (0.0924) (0.0634) (0.0542) (0.0541) (0.0812)
Child's age in months 0.005*** 0.009*** 0.007*** 0.004** 0.004** 0.003
(0.0013) (0.0024) (0.0019) (0.0015) (0.0016) (0.0026)
Child is male 0.024 0.013 -0.037 -0.013 0.023 0.055
(0.0408) (0.0917) (0.0593) (0.0490) (0.0501) (0.0796)
Mother's Education Years 0.052*** 0.051*** 0.050*** 0.052*** 0.056*** 0.068***
(0.0074) (0.0141) (0.0107) (0.0087) (0.0091) (0.0141)
Mother's Age -0.001 -0.011 -0.006 0.002 -0.002 0.001
(0.0048) (0.0114) (0.0067) (0.0060) (0.0056) (0.0093)
Total HH members -0.032*** -0.036 -0.021* -0.041*** -0.021* -0.031**
(0.0092) (0.0291) (0.0117) (0.0110) (0.0124) (0.0156)
B. Age Disaggregated Specification
Quantile Variable (1) OLS (2) 0.05 (3) 0.25 (4) 0.50 (5) 0.75 (6) 0.95 Treatment Effect -0.045 -0.039 0.092 -0.037 -0.116** -0.090 (0.0436) (0.0909) (0.0631) (0.0538) (0.0544) (0.0834) Child is male 0.018 -0.045 -0.051 -0.034 0.037 0.091 (0.0405) (0.0887) (0.0571) (0.0486) (0.0499) (0.0816) Child is 0-12 months -0.399*** -0.527*** -0.535*** -0.332*** -0.290*** -0.227 (0.0716) (0.1302) (0.1136) (0.0913) (0.0887) (0.1552) Child is 13-24 months -0.016 -0.047 -0.042 -0.015 0.010 0.021 (0.0597) (0.1269) (0.0793) (0.0685) (0.0710) (0.1263) Child is 25-36 months 0.078*¤ 0.250** 0.209*** 0.120** 0.007 -0.085 (0.0474) (0.1161) (0.0665) (0.0571) (0.0593) (0.0927)
Mother's education years 0.050*** 0.067*** 0.051*** 0.047*** 0.058*** 0.067***
(0.0073) (0.0148) (0.0098) (0.0089) (0.0091) (0.0145)
Mother's age 0.001 -0.010 -0.002 0.004 0.000 0.003
(0.0048) (0.0107) (0.0065) (0.0059) (0.0055) (0.0097)
Total HH members -0.028***† -0.058* -0.025* -0.039*** -0.019 -0.035**
(0.0093) (0.0301) (0.0130) (0.0110) (0.0127) (0.0161)
Notes: Dependent variable is Weight-for-Age Z Score measured on first follow-up. Robust standard errors reported in parentheses. Due to sampling, clustered standard errors at parish level might be more suitable; nonetheless, only robust standard errors are used for making them comparable across methods. The use of clustered standard errors in OLS estimation does not change the significance level of any variable except for ‘†’ in which it goes from 1% to 5% and ‘¤’ in which it goes from 10% to non-significant. *,** and *** indicate 10%, 5% and 1% significance levels respectively.
22 As with the specification used in panel A, the gender of the child and the age of the mother are never significant in this specification for none of the methods. The education of the mother is always significant at 1% level for both methods and at every quantile showing a positive association with the outcome variable. Likewise, the variable indicating the total of household members is always significant (except at the 75th percentile) indicating a negative association with the outcome variable that ranges from 2.5% (at the 25th percentile) to a 5.8% (at the 5th percentile) of a standard deviation less for every member in the household. The latter finding remarks the importance of analyzing this type of outcome variables at different points of the distribution: while the mean estimator shows a 2.8% decrease of a standard deviation in the outcome for an extra member in the household, the estimator at the 5th percentile –these are children who are severely underweighted – more than doubles it. In contrast with panel A, the effect of the treatment becomes significant at a 5% level at the 75th percentile which corresponds to children who are actually adequately nourished. In spite of the small magnitude of the coefficient for this quantile – which is not really going to make the difference between having or not a healthy child -, it is surprising that the effect is negative rather than positive as generally expected. It is worth mentioning that the OLS estimation is not significant and therefore using only a mean estimator would have failed to provide this significant effect of the treatment.
The variable indicating that the child belongs to the 0-12 months age group is significant in both methods and for every quantile except at the 0.95th. Mean and median estimators are close to each other for this variable. The coefficients show that being in the first year of life in baseline is associated with a decrease in Weight-for-Age that can range from a 29% to a 53.5% of a standard deviation with respect to their 36-72 months counterparts. Particularly vulnerable are the children located in the 5th and 25th percentiles (these are the ones who suffer from undernutrition in this sample according to the WHO standards), since the negative associations shown in the coefficients are 52.7% and 53.5% of a standard deviation respectively. Again, while the mean estimator would have only shown a decrease in the Weight-for-Age of roughly a 40% of a standard deviation for the sample, the quantile analysis allows seeing that the reduction is way larger for the ones in the left tail of this distribution. For the variable indicating that the child belongs to the 13-24 months age group none of the coefficients is significant. Finally, for the 25-36 months age group, positive associations are found with the outcome variable at the 5th and 25th percentiles and the mean and median estimators. As with the analysis done for the Height-for-Age indicator, the coefficients of these dummies confirm the insights given by the literature: younger children are the most vulnerable to undernutrition, particularly in uterus and in the first year of life.
23
TABLE 4
IV and Quantile Treatment Effect Estimates of Cash Transfer on Height-for-Age and Weight-for-Age A. Dependent Variable: Height-for-Age Z Score
Quantile Variable (1) IV (2) 0.05 (3) 0.25 (4) 0.50 (5) 0.75 (6) 0.95 Treatment Effect -0.046 -0.023 -0.058 -0.027 -0.024 -0.104 (0.0552) (0.1516) (0.0699) (0.0629) (0.0706) (0.1090)
B. Dependent Variable: Weight-for-Age Z Score
Quantile Variable (1) IV (2) 0.05 0.25(3) (4) 0.50 (5) 0.75 (6) 0.95 Treatment Effect -0.056 -0.073 0.073 -0.060 -0.127** -0.104 (0.0540) (0.1175) (0.0772) (0.0623) (0.0614) (0.1167)
Notes: Estimates shown correspond to the ones obtained with the “Age-Disaggregated Specification”; estimates obtained with “Basic Specification” only differ slightly in magnitude but not in significance or direction. Robust standard errors reported in parentheses. Due to sampling, clustered standard errors at parish level might be more suitable; nonetheless, only robust standard errors are used for making them comparable across methods. The use of clustered standard errors in IV estimation does not change the significance level of any variable herein reported. *,** and *** indicate 10%, 5% and 1% significance levels respectively.
Table 4 presents IV and QTE estimates for the effect of the cash transfer program on Height-for-Age (Panel A) and Weight-for-Age (Panel B) indicators. As described on section 3 of the present thesis, the proportion of always-takers (individuals randomized to control group who anyway received treatment) is very low and accounts for approximately a 3% of the sample. Therefore, assuming that this very low proportion can be neglected, the estimates obtained through this method will correspond to the Treatment Effect on the Treated for the IV method and to the Quantile Treatment Effect on the Treated at the 0.05th, 0.25th, 0.50th, 0.75th and 0.95thquantiles of the distribution of the outcome variable for the QTE method. It must also be remarked that the use of these two methods as a complement to OLS and QR lies on the fact that endogeneity due to partial compliance of the treatment can’t be completely ruled out (there is an 84% of actual take-up of the program). The results shown in both panels correspond to the “Age-Disaggregated Specification” which are only slightly different in magnitude than the ones obtained through the “Basic Specification”, but do not differ in significance or direction.
Generally speaking, the results shown with the IV and QTE estimation are not very different to the ones obtained with OLS and QR. For the Height-for-Age indicator, the treatment effect varies almost only in magnitude with respect to the estimation in Table 2 (panel B): none of the coefficients on Table 4 (panel A) is significant and all of them have a negative sign, including the one at the 95th percentile which in the QR estimation was positive. Likewise, for the Weight-for-Age indicator the treatment effects only vary in magnitude if compared to the OLS and QR estimation: only the coefficient for the 75th percentile is significant showing that receiving the cash transfer decreased the
Weight-for-24 Age of the children in this part of the distribution by almost a 13% of a standard deviation. As mentioned before (since Table 3 yielded a similar result for this percentile), the magnitude of this coefficient is not going to make the difference between having a good or badly-nourished child, but the result is still surprising. One possible explanation of this result could be parents using money from the cash transfer for consumption of food that is not necessarily the healthiest. Another possible explanation of this result could be that the parents of these children allocate fewer resources on them as they might already look “healthy” – since they are located in the 75th percentile of the Weight-for-Age distribution; hence, not really being undernourished according to this indicator.
6 Conclusion
Throughout this thesis I have presented a reassessment of the effect of the cash transfer program in Ecuador (Bono de Desarrollo Humano) on two anthropometric indicators, namely Height-for-Age and Weight-for-Age. The analysis has been carried using a rural sample with children who were aged up to 72 months in baseline and who did not have siblings older than this age. Due to problems in measuring weight and height of these children, as well as due to attrition, individual samples for each outcome have been used with the intention of using the highest amount of observations available for every anthropometric indicator. The use of different samples for each outcome has unfortunately not prevented to obtain significant differences between the attrited and non-attrited groups. Balancing tests for attrition (not reported) show that for the Height sample, attrited children are on average 4 months younger, their mothers half-of-a-year less educated and the household log per capita expenditure is smaller by 0.05. Likewise, for the Weight sample, attrited children are 2 months younger, their mothers three-quarters-of-a-year less educated and their household log expenditure is smaller by 0.07. The internal validity of the present thesis is not threatened by these small but significant differences between attrited and non-attrited groups; nevertheless, the extrapolation of the results and their external validity are subject to the judgment of the reader.
The core of the present thesis has been the use of a Quantile Regression approach in order to inspect the effect of the cash transfer on the outcome variables as well as the different associations between the controls and the outcomes. The Quantile Regression approach included estimations using Quantile Regression itself (analogous to OLS) and Quantile Treatment Effect (analogous to IV) due to dubious exogeneity of the treatment. Since both methodologies have yielded quite similar results, the latter can be considered as a robustness check of the former as the main estimates – this is the effect of the cash transfer at different quantiles – have not substantially changed across these two methods. The use of Quantile Regression methods has allowed examining different magnitudes of the associations between the predictors and outcomes for different quantiles of the conditional distribution of the latter. This is by far more informative and especially important for the outcomes analyzed in the present thesis, as it has been possible to identify which quantiles are more sensitive for the different explanatory variables.
25 For the Height-for-Age indicator the treatment effect was never significant; not in mean estimators or at any quantile. Moreover, the signs of the coefficients were mostly negative. One explanation for not finding any effect of the cash transfer on this indicator is the nature of the indicator itself. As the descriptive statistics showed, children from the sample were on average more than one standard deviation below the mean in Height-for-Age. Moreover, height is not something that can be easily or quickly changed. This delay in growth shown by the descriptives indicates that they had been suffering from inappropriate nourishment for a relevant amount of time, which creates a trend that is difficult to revert. On the other hand, previous assessment of this cash transfer program by Paxson & Schady (2010) showed no effect on Height-for-Age for the restricted sample that the authors use. Similarly, the effect of the program on Weight-for-Age was never significant, except at the 75th percentile in which the effect is (contrary to what usually expected) negative.
Experiences in other countries with cash transfers have also shown no effect of even negative ones. For example, Bassett (2008) explains that the Brazilian cash transfer program was associated with lower weight gain and even non-significant reductions of 25% and 11% of a standard deviation in Weight-for-Age and Height-for-Age respectively. In addition, the cash transfer in Honduras could not revert the trend of more underweighted and wasted children, as the proportion of these conditions increased in spite of the program. The author explains these unfavorable outcomes by stating that some parents may fear of losing the cash transfer if their children improve their anthropometric conditions “too much”, although this has not been proved. In contrast, the program in Colombia showed some favorable results and one of the reasons given by the author can give a relevant insight on how to improve the effectiveness of cash transfers. The Colombian experience included (non-compulsory) attendance to basic hygiene and nutrition courses and the estimates showed that the households were consuming more nutritive types of food. Obtaining zero or even negative estimates of the cash transfer in the present case could be a sign that the available income is not being used in purchasing healthy food. Perhaps giving complementary nutrition courses to mothers could have helped to improve these indicators as well. In addition, if these complementary interventions are also considered as a type of “education” it is very likely that they can have the desired effect: as shown in the results tables, for every estimation method and every quantile the education years of the mother are highly significant yielding a positive association with the anthropometric measure.
Finally, another interesting finding is the sign of vulnerability for younger children that the coefficients depict. For the Weight-for-Age indicator, the coefficients show that being in the 0-12 months age group is associated with having 0.3 to approximately 0.53 standard deviation less of the anthropometric measure (depending on the quantile under analysis) in comparison with children older than 36 months. Likewise, for the Height-for-Age indicator, children aged 0-12 months are associated with having from 0.20 to approximately 0.42 of a standard deviation less of the anthropometric measure than their
26 older than 36 months counterparts. As it is seen as well, this difference becomes smaller as the child ages. This goes perfectly in line with Hoddinott & Kinsey (2001) and Martorell (1999) in which they state that the youngest children are the most vulnerable, raising the importance of giving priority attention to this group.
