1. Introduction

This paper studies the effect of social transfers on human capital accumulation as it is a main driver for social mobility and long-term economic effects of investments on social protection. ^{Becker and Tomes (1986}, ¹⁹⁹⁴) argue that in the case of poor households underinvestment can be seen as a result of "poverty traps" related with households size, high opportunity costs and credit constrains, and state the role of investments on children on intergenerational mobility. Similarly, ^{Heckman and Mosso (2014}) argue on the importance of early childhood conditions on social mobility.

Social transfers support poor households to invest on human capital as they directly increase households' disposable income (i.e. income effect), but the way social transfers are implemented or conditioned may also enforce certain behaviour (i.e. non-income effect). Nevertheless, effects on human capital is conditioned on the level of coverage, quality of service and the elimination of different access barriers regarding health, education and sanitation services (^{Mideros et al., 2012}). Under this perspective, the research question this paper aims to answer is: whether and to what extent social transfers foster long-term human capital accumulation?

The contribution goes on two directions. First, the effect of social transfers on human capital accumulation over the lifecycle is empirically explored; something that the literature has not completely covered yet.¹ Second, alternative designs of social transfers are evaluated by their cost-effectiveness. Even though human capital is multidimensional the scope of this study is limited to the income effect of social transfers on schooling, measured by the years of education achieved.

A dynamic cohort microsimulation model is developed for a cohort of 5-years-old children, simulating 14 years as discrete periods, to calculate the level of human capital accumulated up to the age of 18 years old. The model uses age specific surviving rates estimated from official demographic projections and own calculations for school attendance and marriage status, while social transfers, education policies, household income, education achievements (i.e. grade promotion) are exogenous in the model. Data is from the Ecuadorian National Survey of Employment, Unemployment and Underemployment (Encuesta Nacional de Empleo, Desempleo y Subempleo, ENEMDU, by its acronym in Spanish) of the National Institute of Statistics and Censuses (Instituto Nacional de Estadística y Censos, INEC, by its acronym in Spanish) which provides necessary information about individuals, households and social transfers.

The rest of the paper is organized as follows. Section two discusses the theoretical framework to link social transfers with human capital accumulation. Section three presents the model and exposes the data and the parameters. A cost-effectiveness analysis of policy options is discussed in section four, and final remarks are elaborated in section five.

2. Social transfers and human capital

Social transfers provide additional secure income to poor households, affecting their consumption and investment patterns by allowing them to consume more goods and services of any kind, including nutritious food, health care and education, and promoting economic returns (^{Mideros et al., 2016}). In the case of education, for example, social transfers help households to cover fees, materials and transportation costs. Moreover as poverty and income insecurity "lead to distortions in inter-temporal resource allocation, forcing a focus on current consumption in preference to investment" (^{Barrientos, 2012, p. 15}), social transfers help households to cover opportunity costs such as the labour income of school age members. In addition, social transfers may include human capital conditionalities to promote non-income effects on school attendance, health care and other dimensions of human capital.

Regarding empirical evidence of the effect of social transfers in the case of education, ^{Bastagli et al. (2016}) provide an extensive literature review analysing school attendance, test scores, and cognitive and problem solving skills. By reviewing 42 studies, they report that most of them show positive effects for boys and girls on school attendance and reduction on absenteeism. For example, ^{Schady and Araujo (2008}) found an increase in school enrolment of 10% in Ecuador. Results in the same direction are presented by ^{Arnold et al. (2011}) in Pakistan (11 percentage points (p.p.)), Malawi (5 p.p.) and Cambodia (30 p.p.); and reduction in the incidence of absence and drop out of 20 and 63%, respectively, in Brazil. In the case of Mexican programme "Oportunidades", ^{Debowicz and Golan (2014}) shows that extending the cash transfer programme increases school attendance.

^{The World Bank (2015)} also presents evidence of reduction on the dropout rate in Cambodia, and improvements on cognitive outcomes in Nicaragua. Similarly, ^{UNICEF (2012)} presents evidence of an increment in the probability to complete high school of between 4 and 8 p.p. in Colombia. However, ^{Bastagli et al. (2016}) also report some studies where no effect has been found. They relate it with the baseline enrolment rate and transfer size. In the first case it is intuitive that "programmes in countries with lower baseline enrolment/attendance may deliver larger impacts compared to countries in which baseline enrolment is high" (^{Bastagli et al., 2016, p. 75}), because there is more room for improvements. In addition, if there is not enough supply of education services, any kind of effect can be expected. In the case of transfer size, the authors mention that if it is not enough to cover opportunity costs, there may be no effect especially for higher education levels (^{Bastagli et al., 2016}).

While school attendance can be intuitively related with a direct income effect which helps poor households to overcome demand side barriers (e.g. financial constraints), the effect of social transfers on test scores and cognitive development is less clear as they are mainly related with other social and environmental factors, including quality of education services (i.e. supply side policies). However some effects may be expected when social transfers promotes more regular school attendance, and due to an increase "in household expenditure resulting in better food security and nutritional status of children, which in turns may also positively affect child's cognitive ability and child's efficiency of learning while in school in the long term" (^{Bastagli et al., 2016, p. 75}). Even more in the case of positive effects on cognitive development it happens under specific conditions related with complementary policies, socio-economic context, and age range.

Looking at cash transfer design, ^{Bastagli et al. (2016}) report higher effects for conditional cash transfers than for unconditional transfers, especially for girls, younger children and lower ability children, based on studies for the cases of Morocco (^{Benhassine et al., 2013}), Burkina Faso (^{Akresh et al., 2013}) and Malawi (^{Baird et al., 2011}). Moreover, ^{Baird et al. (2010)} found that a conditional cash transfers targeted to young women in Malawi reduces early marriage, teenage pregnancy, and self-reported sexual activity. In the case of transfer size results are inconclusive.

Regarding timing, evidence suggests that "tying the transfer schedule to critical moments of the school year decision cycle can have an impact, especially on enrolment" (^{Bastagli et al., 2016, p. 118}). In addition, ^{Villa (2014}) found that a longer exposure to Colombia's Familias en Accion conditional cash transfer programme leads to a higher human capital accumulation and school registration rates. Moreover, ^{Fernald et al. (2008}, ²⁰⁰⁹) found that the combined effect of longer exposure and higher cash transfer size is positive and significant for cognitive development results.

3. The model

Microsimulation is being increasingly applied to analyse economic and social policies on individuals and households (see for example ^{Merz (1993}) and ^{Li and O'Donoghue (2013})). Dynamic cohort models simulating a single cohort over their lifetime have been applied to evaluate the effect of economic and social policies over the lifetime on income distribution (e.g.^{Baldini, 2001}), public pensions (e.g. ^{Geyer and Steiner, 2010}) and cohort earnings (e.g. ^{Van de Ven, 2006}).

In this paper a dynamic cohort microsimulation model is developed to evaluate the effect of social transfers on human capital accumulation, measured by the years of education (i.e. schooling) accumulated by a person throughout her life. The model follows a cohort from the age of 5 years-old up to the age of 18 years-old by calculating four linear equations.

Equation 1 estimates the stock of schooling at a specific age (t), as a measure of the accumulation of human capital. It is accounted for each individual (i), based on the previous level of schooling (at t - 1) plus an addition year if the individual attended school in the previous year times the grade (d)-area (a) specific probability to be promoted (see Table 4 in the following subsection).

Equation 2 estimates school attendance based on the age (t)- gender (g)-area (a) specific probability of attending school but adjusted on the social transfer's amount received (bdh_pc), past behaviour (i.e. school delay), and marriage status. Past behaviour (past _i,t ) is equal to t - (5 + schooling _i,t ).

Being married is estimated based on the age (t)- gender (g)-area (a) specific probability to get married, the effect of receiving a social transfer (bdh) as a dummy variable and the education level (schooling _i,t-1 )(equation 3).

In this sense the effect of the Bono de Desarrollo Humano (BDH) on human capital accumulation is driven by its direct effect on school attendance by covering opportunity and transaction costs, but indirectly by reducing the probability of being marriage (see ^{Baird et al., 2010}), which is also driven by school attendance (see ^{Delprato et al., 2015}).

Finally, the model considers demographic changes (i.e. ageing) by adjusting weights (equation 4) using age specific survival rates (see Table 5 in the following subsection).

The model is kept as simple as possible to highlight the effects of interest. While it means that we accept strong "everything else constant" type of assumptions, they do not affect comparison between different scenarios of social transfers (BDH) which is the aim of this paper. Exogenous variables including the probabilities of attendance, marriage and promotion, as well as average marginal effects (α, β γ, δ, ρ) are empirically calculated for the case of Ecuador on the following subsections.

Programme, data and variables

The Ecuadorian social transfer called BDH pays a flat cash transfer to poor households independent of labour conditions and the number of household members. The BDH transferred US$35 per month to each eligible household between 2009 and 2012, and US$50 since 2013. Recipient households were identified in 2009 by a proxy-mean test called Registro Social index (RS index). It is a multivariate welfare indicator estimated by non-linear principal components analysis with a value between 0 and 100. The eligibility threshold was estimated as a proxy of the consumption poverty line at a value of 36.50, mainly targeting poor households with children. Recipients are informed about their responsibility to ensure that children must attend school and visit health services, but accomplishment is not monitored.

The empirical analysis uses data from the ENEMDU, rounds of December from 2009 to 2013, of the INEC, which provides necessary information about individuals, households and the BDH.² The period correspond to the years between the global financial crisis and the local economic slowdown due to the drop of oil prices, to avoid additional external shocks. The final data set is pool data including five years of cross-sectional information.

Pooled data during period 2009-2013 includes 80.940 observations of individuals between 5 and 18 years old during the period, from which 38.7% receive the BDH. Almost all children at primary education age (i.e. between 6 and 11 years old) attend school, though the rate is lower among pre-school age children (i.e. 5 years old). The rate of school attendance decreases during secondary education (i.e. between 12 and 18 years old). Children receiving the BDH (i.e. poor and vulnerable households) have a lower probability of attending school at any age, further the decline of school attendance at secondary education age is higher for them. In addition, males and rural children have lower probabilities to attend to school than females and urban children of attending. This evidence is consistent with the accumulated level of human capital (i.e. schooling, measured by the years of education). On average, at the age of 18 years old urban females achieve 11.0 years of education, followed by urban males (10.6), rural females (9.7), and rural males (9.1). Moreover, those who receive the BDH had accumulated lower human capital than the rest of individuals (see Table 1).

Fall behind school measures the difference between the education level a child should had achieved at a specific age and her actual level of schooling. It is close to 1 year (i.e. on average, a child has one year less education than what she should had have for her age) up to 8 years old. This falling behind is mainly due to low enrolment in pre-school education (which is not compulsory in Ecuador). Later, it increases with age which may be related with children having to repeat a year and because some of them drop out school. The major increments on falling behind school happens after the age of 16, which correspond to upper secondary education and tertiary education. Falling behind school has a cumulative effect on school attendance especially at older ages as it reduces a child fit to her peers. In addition, being married which at young ages is closely related with having a child to take care of and then with school dropout, becomes higher than 1% for children around the age of 15. It is higher for females than for males, except for rural males at older ages (see Table 2).

Regarding economic status, four monthly household's per-capita income brackets are used: extreme poor (less than US$37.64), moderate poor (between US$37.64 and US$66.78), vulnerable (between US$66.78 and US$133.56) and middle-&-upper class (higher than US$133.56). Extreme poverty and poverty thresholds are based on official poverty lines at prices of 2009, while the threshold between vulnerable and middle class-&-upper is equivalent to two times the poverty line as proposed by ^{Lopez-Calva and Ortiz-Juarez (2014}). School attendance increases with income, especially for males and older ages. All this information is used at the age-specific level to parametrize the model in the next section in order to account for age, gender and socio-economic specific conditions (see Table 3).

In addition, grade and gender specific promotion rates for primary and secondary education are estimated using administrative data of the Ecuadorian Ministry of Education (see Table 4). Finally, age specific survival rates (see Table 5) are estimated as the average between 2011 and 2020 from the official population projections of the INEC.

Marginal effects

The causal effect of the BDH is calculated following ^{Ponce and Bedi (2010}) who rely on a regression discontinuity (RDD) model using instrumental variables (IV). This strategy allows identifying the effect on those close to the targeting threshold to control for non-observable characteristics. However, it cannot capture the effect on the extremely poor. The model is estimated using the 2009-2013 pooled data for children between 5 and 18 years old with a RS index, in 2009, between 31.50 and 41.50 (i.e. +/- 5 points around the RS threshold of 36.50). The RS index is estimated using administrative data from the Registro Social. The 2009's RS index uses 30 variables, but only 26 variables can be replicated in the ENEMDU. Therefore, it is necessary not only to impute the RS index in the ENEMDU, but to rescale the index due to the lack of variables. This is done by firstly estimating a partial index (RS 2009²⁶) using the available variables and official RS index's weights. Secondly, an equation to replicate the rs 2009 index is estimated using Registro Social's administrative data (equation 5).

After the estimation of the RS index in the ENEMDU the BDH'S eligibility threshold of 36.50 is localized at the percentile 41 in the ENEMDU 2009 (which corresponds to the date of the Registro Social). However, households' welfare conditions are likely to change over time. Because of this reason the eligibility threshold is estimated for the years between 2010 and 2013 as the value at the percentile 41 of the RS 2009^.

Treatment discontinuity occurs on the RS index. The probability to be eligible to receive the BDH in 2009 is 25% for those households with a RS 2009^ higher than 36.50, while it is 75% if the RS 2009^ is lower than the cut-off. If reduced to those +/- 5 points around the threshold these probabilities are 53 and 67%, respectively (see Figure 1).

Source: own calculations using ENEMDU, round of December 2009.

Figure 1 Relation between eligibility and RS index (2009) 

This non-linear relationship provides exogenous variation in treatment status. However, as it is unlikely that treatment is solely assigned by these criteria a fuzzy discontinuity is assumed, and then the evaluation strategy includes an IV approach. The first stage equation includes an instrument (T _i ) the RS index RS 2009^) and its square value RS 2009^2) as well as other variables (X _i ) value assuming an independent and identically distributed (i.d.d.) error term (ε) with mean zero (equation 6). The instrument is the assignment rule and then it is correlated with BDH eligibility, and we assume that it is not correlated with the unobserved characteristics that determine the evaluated variables (i.e. school assistance and married status).

Following this strategy, the probability of school attendance and marriage is estimated by a two stages probit model. The first stage is calculated by equation 6, while second stage estimates are obtained by:

Where Φ is the cumulative distribution function of the standard normal distribution, K _i and H _i are vectors of observables characteristics and a constant and, 𝜓 _i and (𝜔 _i are independent and identically distributed (i.d.d.) error terms with mean zero.

Household income has a negligible effect on the probability of school attendance. While it is statistically significant in the expect direction (positive effect) the coefficient is low. However, the cost of opportunity has a negative and significant effect. It means that a higher opportunity cost in terms of forgone income reduces the probability of attending school at any age. On average females have a lower probability of attending school in comparison with males. The coefficients of interest for the model show that falling behind school and being married reduces the probability of attending school. In the case of the BDH it shows a negative coefficient reducing the constant term for those who receive the transfers, while the amount received increases the probability of attending school. It means that the BDH do not necessarily increases school attendance, as it is not strongly conditioned but also because school attendance is almost universal in Ecuador, while the transfer amount is relevant to cover opportunity costs of those not going to school. In terms of average marginal effects each US$1 per month per capita of transfer increases the probability of attending school by 5.17 p.p. In the case of past behaviour each year of school delay reduces the probability by 3.03 p.p., while being married reduces it by 9.51 p.p. (see Table 6).

Child marriage is a crucial determinant of school attendance. The probability of marriage increases with age, being a female, household size and poverty conditions as for example deprivation on access to water and sanitation. On the other hand, schooling reduces the probability of marriage as well as receiving the BDH. Average marginal effects show that a person living in a BDH' recipient household has a probability 1.92 p.p. lower. In the case of schooling an addition year of education reduces the probability of marriage by 2.17 p.p. (see Table 6). This result shows that the BDH may encourage school attendance solely based on its amount, but also due to its effect on the reduction of child marriage.

4. Results: human capital accumulation

The model simulates a cohort of 1.056 children (197 892 using weights) starting at 5 years old, from which 52.3% are girls, and 40.9% are currently receiving the BDH. Data is from ENEMDU, round of December 2013. In the base line model (Scenario 1 - no social transfer), the cohort achieves 12 years of education (schooling) at the age of 18 years old, which is equivalent to incomplete secondary education³ (see Table 7). In order to analysis the cost-effectiveness of a social transfer on human capital accumulation three additional scenarios are simulated. Scenario 2 simulates the BDH in its actual design. A flat transfer of US$50 to each eligible household identified in ENEMDU 2013 up to age of 18 years old. Scenario 3 simulates a variable transfer to households on extreme income poverty. The transfer is defined as the specific household's poverty gap (up to US$37.64 per month at prices of 2009). It aims at testing a perfect targeting design on the poorest. Finally, scenario 4 transfers an amount equal to the poverty line (US$66.78 per month at prices of 2009) to poor individuals at the ages between 12 and 18, which are critical periods of school attendance. It tests targeting at critical ages.

All scenarios achieve lower rates of marriage (equation 1), higher school attendance (equation 2) and schooling (equation 3) at any age in comparison to the scenario with-out a social transfer, while cohort population declines equally for all scenarios (equation 4). Scenario 2 evaluates the BDH in its actual design. At the age of 18 years old it reduces the percentage of child marriage from 1.5 to 0.9%. School attendance increases by 4.2 p.p. at the age of 5 years old, between 0.6 and 1.1 p.p. between the ages of 6 and 11 years old, and between 2.1 and 18.4 p.p. between the ages of 12 and 18 years old. These results promote higher human accumulation of 0.4 additional years of education per person at the age of 18 years old. It means 88 551 additional years of education in terms of total cohort's human capital. Scenarios 3 and 4 have lower effects on child marriage, school attendance and schooling. It was expected as these scenarios has lower coverage (see Table 7). In general, social transfers proves to have a positive effect on human capital accumulation, but rather low in the case of Ecuador.

The cost of the transfer is between US$48.1 and US$48.5 million per year in the case of scenario 2. It is around 0.05% of Ecuadorian GDP in 2013.⁴ The annual cost of scenarios 3 and 4 are between US$32.7 and US$33.1 (0.03 of GDP), and US$57.2 and US$57.6 (0.06% of GDP), respectively (see table 8).

In order to compare scenarios in terms of their cost-effectiveness total cost is calculated adding-up the annual cost of each policy option, while total cohort's human capital accumulation is measured adding-up the years of education achieved for each person within the cohort. A cost-effectiveness indicator is then estimated as the relation between the additional cohort's total schooling in comparison with scenario 1 and the cost in US$ million. In this sense the indicator measures how many additional years of education are achieved by a cohort for each million invested on social transfers.

At the age of 18 years old, the cost-effectiveness ratio has a value of 130.7 for scenario 2, 76.0 for scenario 3, and 169.8 for scenario 4 (see Table 8). The latest is the most cost-effective. This shows that in order to promote human capital accumulation the best option is to target critical ages. However, it is also important to evaluate the effect on inequality. It is done by looking at average schooling achieved for different income brackets and its ratio with mean schooling. In the case where no social transfer is implemented the average years of schooling for extreme poor individuals is 11.3, at the age of 18 years old, while it is 12.5 for middle-&-upper class persons. In this sense, it generates a vicious circle of poverty and inequality. All policy scenarios reduce the gap, but the most efficient is scenario 3 where social transfers are designed to eradicate extreme income poverty and then promoting school attendance of the poorest of the poor (see Table 9).

5. Concluding remarks

Social transfers have the potential to promote human capital accumulation and to produce long-term economic returns by covering transaction and opportunity costs because of a higher poor households' disposable income and then promoting investments in human capital. However, the effect depends on the existence of supply side policies to guarantee coverage and quality of education services.

Results show that social transfers do promote higher levels of schooling in the case of Ecuador. At the age of 18 years old the average level of schooling is between 0.2 and 0.5 higher under social transfer scenarios. The effect (i.e. the difference with no social transfer) is higher under actual design of the BDH. However, social transfers are more cost-efficient to promote human capital accumulation if targeted at critical ages, and more efficient to reduce schooling inequality if targeted at the poorest of the poor.