Cognitive development and educational attainment in a drought prone - - PDF document

1

Cognitive development and educational attainment in a drought prone region: Evidence from Wajir, Kenya

John A. Maluccio,1 Laura Nubler,1,2 and Karen Austrian3

1 Middlebury College; 2 St Andrews University; 3 Population Council, Kenya

Preliminary Draft September 29, 2017 Abstract: There is growing evidence that early life conditions are important for outcomes during adolescence including cognitive development and education. Economic conditions at the time children enter school also can be important for such outcomes. In low-income pastoral settings, where rainfall patterns influence household, and consequently child outcomes, programs are often implemented to mitigate the potential effects of rainfall shocks. One such program in Northern Kenya is the Hunger Safety Net Program (HSNP). We use historical rainfall patterns as exogenous shocks to 1) examine their effects at different ages on a broad set of children’s cognitive and educational outcomes and 2) explore whether any of these effects are mitigated by HSNP. Using a cluster random sample of over 2000 girls collected in Wajir County, combined with historical rainfall and HSNP program delivery data at the cluster level, preliminary findings indicate that indeed, rainfall shocks have a negative effect on girls’ cognitive development and educational

• attainment. However, living in an area receiving the HSNP cash benefits mitigates the deleterious

effects of the drought on girls’ educational attainment. Key words: cognitive development; achievement; adolescent girls, schooling; drought; Hunger Safety Net Program, Kenya Acknowledgments: We thank Marcos Barrozo Filho for excellent GIS research assistance and preparation of maps and Mohamed Hussein for additional research assistance in preparation

• f the data. All remaining errors are our own.

Funding: This research has been funded by UK aid from the UK government. The views expressed do not necessarily reflect the UK government’s official policies.

2

• I. Introduction

Investment in women’s education and the development of their human capital is a highly important factor in a country’s economic and social advancement. Development Economist Jejeebhoy states that “at the national level, educating women results in improved productivity, income, and economic development, as well as a better quality of life, notably a healthier and better nourished population” (Jejeebhoy, 1995). Nevertheless, a family’s investment in a girl’s education often remains low as parents expect insufficient returns for their investment (Alderman & King, 1998). In countries where women’s educational and economic opportunities are limited, the advancement of women is an integral factor in developing and accumulating human capital, generating economic activity and transitioning to more advanced stages of industrialization (Jejeebhoy, 1995). Many cultural, social, environmental and financial factors may contribute to low educational attainment among girls and women, and literature suggests that there are certain periods in a child’s early development where these factors are important determinants of its cognitive abilities, educational attainment and professional advancement later in life (Currie & Almond, 2011). In sub-Saharan African communities, many of which rely heavily on rain fed agriculture, droughts can act as a significant, multi-faceted shock disrupting a girl’s development through a multitude of financial and environmental pathways. The effects of drought on early childhood development and health have been studied extensively in sub-Saharan Africa and parts of Latin-America, where a lack of rainfall has been linked to an increase in malnutrition and civil conflict, and a decrease in childhood health indicators and cognitive attainment (Barham, Macours, & Maluccio, 2013; Miguel, Satyanath, & Sergenti, 2004; Molina Millán, 2014). However, the effect of shocks in later childhood and adolescence on educational progression and attainment has not been investigated to the same

• extent. Research shows that drought shocks can affect physical and cognitive development of

children in utero and throughout early childhood (Barham et al., 2013; Hoddinott & Kinsey, 2001). These effects may be exacerbated as a variety of social, economic and cultural barriers further limit a girl’s development and educational attainment. Among others, early marriage and child- bearing, gender expectations, low prioritization of women’s education and financial limitations pose a significant threat to the development of adolescent girls, and many of these factors become exacerbated during times of drought (Schroeder, 1987). Children’s high vulnerability to the wide-

3 reaching impacts of drought shocks are great cause for concern, as meteorological models predict that, as global temperatures continue to rise, droughts and other extreme weather events will increase across most of the African continent. This puts populations that are already vulnerable to the impacts of drought at an increased risk of cognitive and educational losses (IPCC, 2014). In

• rder to protect a girl’s cognitive development and education, thereby generating benefits not only

for the girls but also the national economy, it may be necessary to prioritize scarce aid resources by designing interventions that efficiently target the age groups most vulnerable to drought or other shocks. This research paper aims to determine at which ages throughout childhood girls are most vulnerable to the impacts of a drought shock affecting their cognitive development and educational

• utcomes by investigating a sample of adolescent girls and their households in the rural, highly

underdeveloped Wajir County in Kenya. We find evidence that a drought shock during the early childhood development period and during the school start period around the age of 7 leads to lower school enrollment, slower educational progression, and reduced cognitive and academic abilities in my sample. In addition, we seek to understand if a means-tested unconditional cash transfer being administered in Wajir during the drought shocks mitigated these effects. We also find evidence that the social safety next program was able to mitigate the effects of the drought on education and cognitive outcomes. This research will add to the evidence base about the importance of early life and school transition years for a girl’s development, will inform interventions for girls susceptible to these sorts of shocks, and may help prevent the serious economic and social costs associated with the loss of cognitive capacities and human capital development (Currie & Almond, 2011).

• II. Background
• A. Climate change and the macro-level effects of drought

The effects of rainfall shocks and droughts may pose a serious challenge in the future. Climate models project with high confidence that precipitation and soil moisture will decrease in most of the African continent and several other regions as a result of progressing global warming

4 and climate change over the next 50 years, but project, albeit with low confidence, an increase in rainfall in Eastern Africa (Dai, 2013; IPCC, 2014). Nevertheless, rainfall in Eastern Africa is still expected to decrease March through May which may adversely affect crop yields of maize, millet, sorghum and soy, as well as most legumes, some vegetables and other staple crops maturing and ripening during this period (FAO, 2017; IPCC, 2014). This also makes Eastern Africa prone to intermittent droughts, food shortages and malnutrition, even if this effect may be more pronounced in Southern and Western Africa (IPCC, 2014). Since rainfall is highly exogenous, a drought provides an exogenous shock, and any correlation observed between rainfall and cognitive and educational outcomes is caused by the indirect effects of the shock operating through economic and political conditions (Miguel et al., 2004). There are a multitude of pathways through which a drought shock causes economic and political shocks that can disrupt cognitive development and education. In sub-Saharan Africa, only 4% of crop land is irrigated, and the well-being of these economies is highly dependent on rain- fed agriculture, with 27% of GDP and 62% of employment coming from the agricultural sector (Bank, 2017). Therefore, personal and national income in a given year may vary significantly with crop yields and the success of other agricultural operations (Miguel et al., 2004). In Kenya, 30.3%

• f GDP comes from agriculture and only 0.38 % of agricultural land is irrigated, which suggests

that the country would very vulnerable to the effects of weather shocks (Bank, 2017; FAO, 2017). Using precipitation to instrument for economic growth in 41 African countries, Miguel et

• al. (2004) found a strong relationship between rainfall shocks and GDP. The first-stage regression

results show that a 10% decrease in rainfall to the previous year was consistently and significantly associated with a 0.5 percentage point decrease in the income growth rate in the given year, and a 0.03 percentage point decrease in the following year. The second stage regression also finds some evidence that a higher income growth rate reduces the incidence of civil conflict in the given and the subsequent year. Dercon et al. (2005), studying several Ethiopian villages between 1999 and 2004, found that households who experienced a drought shock in the previous 3-5 years saw a 13% decrease in household consumption, and those who experienced a shock in the previous 2 years saw a 17% decrease in household consumption. The decline was particularly significant and pronounced in

5 female-headed households, which saw a 40% decrease in consumption following a drought shock (Dercon, Hoddinott, & Woldehanna, 2005).

• B. The effects of shocks during early childhood

There is solid evidence that a child’s cognitive development and educational progression is heavily affected by factors like nutrition and physical health in early childhood. Drought shocks are a major cause of crop failure and the resulting malnutrition, and may adversely affect a child’s early development. Lavy et al. (2016) exploit a natural experiment resulting from the Ethiopian famine of 1991, during which the Israeli government evacuated most Ethiopian Jews to Israel within a matter

• f days. Since most Israeli health care facilities provided nutritional supplements during pre-natal

care while most Ethiopian ones did not, the move essentially served as a nutritional intervention for pregnant women at different stages (trimesters) of their pregnancy. The authors find that, although the intervention did not result in significant differences in birth weight, significant impacts on cognitive abilities could be observed in these individuals 20 years later. Results show that children whose mothers immigrated during the first trimester of pregnancy and therefore received nutritional supplements in utero had significantly higher cognitive and educational

• utcomes than those children whose mothers arrived during the 2nd or 3rd trimester. Children in

this first group were 12 percentage points less likely to repeat a grade, 7 percentage points less likely to drop out of school, and 12 percentage points more likely to obtain a high school diploma. On average, these children also earned over 50% more English and Mathematics credits than the later cohorts. In addition, the study shows that the effects of in-utero nutrition on cognitive development were observed primarily among girls, while the effects on boys were both smaller and not statistically significant (Lavy, Schlosser, & Shany, 2016). Similarly, Fields et al. (2009) link improved pre-natal nutrition with improved educational outcomes in later life, finding that sufficient pre-natal iodine in pregnant women in Tanzania resulted in their children obtaining an average of half a year of additional schooling by the age of 13, with the impact being more pronounced for girls than for boys. Although these interventions refer to pre-natal supplements in

6 particular, the results of these studies highlight the importance of nutritional factors in utero that determine a girl’s cognitive development and educational attainment. After birth, infants and young children continue to be highly sensitive to shocks, especially during the first 1000 days of their lives (Barham et al., 2013). Akresh et al. (2011) found that Rwandan children born during a period of civil conflict, and girls born during a period of crop failure, two likely results of a drought, presented with significantly lower height-for-age z-scores that those who were not. Hoddinott and Kinsey (2001) estimate that children who experienced drought during the first 12 months of their lives lose between 1.5 and 2 cm of growth in the year following a drought. Children who experienced drought at the ages 12-24 months did not experience a significant reduction in physical growth. The authors note that children from poor households and children whose mother is the daughter of a household head are particularly vulnerable to this effect (Hoddinott & Kinsey, 2001). Alderman et al. (2004), using the incidence of droughts and civil conflict to instrument for malnutrition, and investigate the effects of malnutrition on health, education, and later-life

• utcomes. They estimate that preschoolers affected by the civil war and drought shocks of the late

1970s and early 1980s in rural Zimbabwe suffered reduced height as adults. Additionally, they estimate that the children who experienced malnutrition and stunting also started school later, completed fewer years of education by the time they reached adulthood, and suffered a 14% reduction in lifetime earnings compared to those who did not. Barham et al. (2013) link malnutrition during the first 1000 days of boys lives to a reduction in long-term cognitive abilities. They find that boys who were exposed to nutritional interventions in utero and during the first two years of their lives had significantly better cognitive outcomes 7 years later than those who only received the intervention after the age of two, while the HAZ-scores were not affected. These studies indicate that a long-term loss of cognitive skills can result from in utero and childhood malnutrition even if there is no long-term physical stunting. While improved nutrition after a period of malnutrition can help stunted children catch up in terms of physical development, this does not appear to be the case with cognitive skills. Considering that cognitive damage can already occur as a result of early childhood malnutrition even when physical stunting doesn’t, it is highly likely that children presenting with low HAZ-scores due to malnutrition also suffer from

7 cognitive capability losses. Since the evidence on gendered impacts of drought and malnutrition

• n early childhood and cognitive development is inconclusive, it is possible that gendered

vulnerability to drought may depend on other, possibly unobserved, factors such as parental and social bias towards protecting and advancing one gender over another.

• C. The effects of shocks during schooling ages

Jensen (2000) investigates the effects of rainfall-based income shocks on investment in children in Cote d’Ivoire, where families receive roughly 50% of their household income from crop agriculture and are also highly dependent on rainfall. Studying the school enrollment rates of school children aged 7-15, the author observes a decline of 14 and 11 percentage points among boys and girls (respectively) in regions that were adversely affected by rainfall shocks, although the difference between these two rates was not statistically significant. Furthermore, results show that malnutrition increased in all regions affected by the adverse shock, while use of health services (given that the child is ill) decreased by roughly 30% in shock-affected areas, even though incidences of illness did not differ significantly between affected and unaffected regions. Jensen also notes that children who were affected by rainfall shocks were somewhat more likely to be sent elsewhere, for example to live with relatives. While adverse rainfall shocks cause temporary disturbances in schooling, health care reception, and nutrition, the author notes that it is unclear whether these effects have permanent implications, as the study does not follow up on these children at a later point in their lives (Jensen, 2000).

• III. Context and Data

We combine historical rainfall data with: 1) individual girl-level data from a 2015 cross- sectional survey carried out for AGI-K and, separately, 2) individual girl-level data from a 2014 census of Wajir carried out by the Hunger Safety Net Programme (HSNP). A. Adolescent Girls Initiative-Kenya (AGI-K) The Adolescent Girls Initiative-Kenya (AGI-K) is a study led by the Population Council that aims to test which package of multi-sectoral stacked interventions, namely violence prevention, education, health, and wealth creation, most improve educational outcomes and

8 general well-being of vulnerable adolescent girls. Implemented as a randomized trial, AGI-K targets girls aged 11-14 in two marginalized areas of Kenya, the Kibera slums in Nairobi and the rural Wajir County near the Somalian border (Austrian et al. 2016). This paper uses the baseline survey data from Wajir, where drought effects are more relevant. Wajir County in Northeastern Kenya is a sparsely populated region bordering Somalia with a predominantly Muslim population of approximately 800,000 in 2015, an average of only 12 persons per square kilometer though most live in towns or small settlements. The primary economic activity in the county is pastoralism. For example, livestock and livestock products are dominant in the local economy, and in 2013 were valued at nearly 50 times crop production (ASDSP, 2014). Wajir has the lowest literacy rate (26%) of all Kenyan counties, and also the lowest education rate among women (Kenya National Bureau of Statistics et al., 2015). The median years of schooling reported for women are 0 and for men 4.7 (Kenya National Bureau of Statistics et al., 2015). This makes the educational gender gap in Wajir one of the highest in the

• country. Furthermore, Wajir County has the highest fertility rate and the poorest access to health

services, and a median marriage age of 18.1 (Kenya National Bureau of Statistics et al., 2015). Less than 1% of parents in the sample stated that they want their girl to finish secondary school before getting married and 28% of parents indicated they would prioritize a boy’s education over a girl’s, given financial constraints. Furthermore, only two-thirds of girls surveyed believed that girls are just as intelligent boys. In the sample, nearly 90% of households reported owning livestock, primarily goats, sheep, chicken, and in some instances, cows and highly valuable camels, while only 23% reported owning agricultural land. The pastoral-based livelihood of this population may make it highly sensitive to the impact

• f droughts, as family members may need to travel further than usual to find viable pastures and

watering sites for their herds, or even relocate altogether. Most previous work using rainfall shocks has been done in environments where there is a high dependence on rain-fed crops (Björkman- Nyqvist, 2013; Dercon et al., 2005; Hoddinott & Kinsey, 2001). There is much less evidence on the effects of drought in pastoral societies, where the influence of rainfall is equally if not more important.

9

• B. AGI-K Baseline Survey 2015

AGI-K identified all village settlements in Wajir County with a public primary school, necessary to ensure that the girls in the trial had access to the regular AGI-K group meetings, which are a key component of the intervention. A small number of villages were excluded either because they were an urban settlement with more than one public primary school, had fewer than 8 eligible girls in them, were inaccessible, or were located in zones with significant security risks near the Somali border. In total, 80 village clusters were identified. In villages with less than 40 households with eligible girls, all households and girls were interviewed, while in villages with more than that, 40 households within the cluster were selected at random, and when there was more than one eligible girl in a household, one of them would be selected at random to be included in the baseline survey (though all were included in the program). Girls who were enrolled in boarding school at the time of the baseline survey were excluded from the sample as they could not participate in the trial (Austrian et al., 2016). A comprehensive survey at both the household and individual-girl levels was administered by trained female enumerators in 2015, prior to program start. The short household survey was conducted with the head of household or guardian adult providing consent at the time of the interview, and included household characteristics and assets, attitudes toward education, and the global positioning system (GPS) location of the residence. The individual-girl level survey was administered after written permission was obtained both by the respondent’s parent/guardian (consent) and by the respondent herself (assent, given her age). The study protocol was approved by the Population Council Institutional Review Board and the AMREF Ethical and Scientific Review Committee. In addition, the protocol was reviewed by the Kenyan National Commission for Science, Technology and Innovation to obtain research permits for study

• investigators. The individual-girl level survey collected information on socio-demographic

characteristics, schooling history, and educational attainment, among other things. Girls also completed three tests that assessed literacy in Swahili and English, mathematics (using a portion

• f the Kenya National Learning Assessment tool (UWEZO, 2014)), and nonverbal cognition

(using a subset of Raven’s Coloured Progressive Matrices (Raven, 1984)). The survey was translated into Swahili, pilot-tested, and revised based on feedback from interviewers prior to data collection.

• C. Hunger and Safety Net Program

10 The Hunger and Safety Net Program (HSNP) is an unconditional cast transfer program in four counties Northern Kenya that seeks to alieve poverty and by providing regular cash transfers to beneficiaries (Merttens et al., 2013). During its pilot phase, from 2009-2012, HSNP provided payments every other month that could be collected from a range of small shops via a Smartcard that was provided to beneficiaries. When the pilot started the value of each payment was KES 2,150 (~US$ 21.50) and increased periodically and stood at KES 3,500 (~US$ 35) at the end of the pilot period. Approximately 300,000 beneficiaries were targeted in 60,000 households. The

• verall goal was to reduce poverty, good insecurity and malnutrition, as well as asset accumulation

and retention, however it was also hypothesized that there would be a positive impact on a wider range of indicators including resilience to shocks, health and education (Merttens et al., 2013). Currently, in its second phase, HSNP serves almost 100,000 households in the four counties, including about 19,200 households in Wajir (HSNP, 2017).

• D. HSNP Census 2014

In preparation for its second phase, the HSNP program carried out a census in four districts

• f eastern Kenya including Wajir. The instrument included household roster and education

information for all members as well as selected household level characteristics. Data available to us includes information on nearly 65,000 girls aged 7 to 14 inclusive, residing in 42,000 households across 275 sublocations of Wajir.1 While the education measures are limited only to grades attained, the large sample size increases power and the potential to control for village and even household-level fixed effects in the analyses.

• E. Rainfall Data and Construction of Shocks

To explore how early life and early childhood economic shocks influence later cognitive and educational outcomes, we need exogenous, and economically meaningful, shock indicators measured throughout the child’s life (Burke, Gong, & Jones, 2015). Following others (Björkman- Nyqvist, 2013; Miguel et al., 2004), we use variation in rainfall. Often, this approach is justified by the critical role of rainfall in crop agriculture and the important role of agriculture in rural

1 Presently, we have a pending data request for all individuals in the household (including school-age boys) and

additional household level variables. This is one reason results reported here are preliminary.

11 livelihoods throughout much of sub-Saharan Africa (Davis et al., 2010) in semi-arid Wajir County, where pastoralism rather than crop agriculture is the main economic activity, rainfall patterns also play a critical role in economic and other outcomes (ASDSP, 2014).2 We use the Climate Hazards Group InfraRed Precipitation with Station data (CHIRPS), a dataset spanning latitudes between 50°S and 50°N and includes monthly precipitation values (in millimeters) from 1981 to the present. CHIRPS data consist of monthly raster files, each containing a grid with 0.05 x 0.05 degree spatial resolution. That is, each data point (or pixel) represents an approximately 5km x 5km square. Using geographical positioning system (GPS) locations of the villages in the survey, we imputed monthly rainfall levels for each village based on the nearest pixel (Funk et al., 2014). For girls who previously lived elsewhere, we approximated the location

• f their previous residence and generated similar information, with estimates available for about

half of the girls who had moved. Consequently, rainfall data is available starting 20 years before the birth of every girl in the primary sample we analyze. Starting from the monthly rainfall measures, historical long-term averages from 1981–1999 were constructed for each grid location and then used to construct standardized z-scores for each year after 1999 in each grid location. Standardization at the grid, rather than the region level ensures that differences in local average rainfall are accounted for, as individuals living in areas with lower average rainfall are likely used to, and prepared for, such lower levels and they do not therefore represent unusual situations or shocks in the way that deviations from local averages do. This approach yields an observation for every grid in every

• year. Each girl is then linked to the rainfall data based on the grid in which she resided each year
• f her life (using GPS location information collected in the baseline and historical information on

previous residence). In total, the 2,147 girls for whom we have information currently resided in 80 sample clusters, but about one-third previously lived elsewhere.3

2 In future work we will explore links between recent rainfall and current economic outcomes in the sample to examine

evidence in this population on the importance of current shocks.

3 The same methodology was used for girls and their locations in the HSNP census. In that data, however, there is no

information on previous locations of residence so we assign to each girl all previous rainfall shocks in the current residence.

12 Last, using the z-score deviation from rainfall a set of dummy variables was constructed indicating whether the grid in which the girl was living experienced a drought shock in each year, beginning the year before she was born up until 2016, the year after the survey. A drought shock was defined as an annual average precipitation measure falling 1 or more standard deviations below the local historical mean (from 1981-1999). Table 1A presents average z-scores (based on historical means) and the prevalence of drought shocks at the survey cluster level by year through 2010 for both samples. Drought shocks as we measure them are uncommon and therefore possibly

• important. Moreover, shocks are clearly concentration in specific calendar years, though this of

course means they affect girls at different ages. Consideration of correlation across the years shows neither high nor consistent correlation, with both positive and negative year-to-year correlation, i.e., there is only minimal persistence over time within clusters (not shown).

• III. Methodology

The general model incorporates multiple past shocks as determinants of current educational

• utcomes for girls, putting the spotlight on the effect of shocks occurring at specific ages in the

girl’s life. We estimate the following econometric model: (1) 𝑧𝑗𝑘𝑢 = 𝛾0 + 𝛾1𝑏𝑗𝑘𝑢 + ∑ 𝜀𝑏

𝑈 𝑏=𝑏0

𝑇

𝑘,𝑢−𝑏𝑗+𝑏 + 𝒀𝒋𝒌𝒖𝜸𝟑 + 𝑣𝑗𝑘𝑢

where 𝑧𝑗𝑘𝑢 is the outcome variable for child i in cluster j in year t (2014 for the HSNP census and 2015 for the AGIK baseline survey), aijt is the child’s age when measured, 𝑇

𝑘,𝑢−𝑏𝑗𝑘+𝑏 is an indicator

for whether there was a drought shock in cluster j in calendar year (t – ai + a), that is, when child i was age a, with a ranging from a0 to T. In the most comprehensive specifications, a ranges from a = –1 (i.e., one year prior to birth or, equivalently, the year of conception and in utero) to age T = 8. For example, for a girl aij= 14 years old at the time of the baseline survey in 2015, and thus born in 2001, 𝑇

𝑘,𝑢−𝑏𝑗𝑘−1 is the drought shock indicator in her cluster j in 2000, the year before her

• birth. In similar fashion, for a = 0, 𝑇

𝑘,2015−𝑏𝑗𝑘−0 is the drought shock indicator for the year 2001

(the year of her birth), 𝑇

𝑘,2015−𝑏𝑗𝑘+1 for 2002 (the second year of her life) and so on up to

𝑇

𝑘,2015−𝑏𝑗𝑘+8 for 2009. 𝒀𝒋𝒌𝒖 include parental and household level controls or fixed effects

(described for each sample below) and 𝑣𝑗𝑘𝑢 represents an assumed idiosyncratic error term. The main parameters of interest are the 𝜀𝑏, which capture the impact of a drought shock at each age a

13

• n 𝑧𝑗𝑘. Equation (1) represents the most general form estimated, with annual drought indicators.

Other specifications, limiting to fewer years or combining multiple years (approximately corresponding to specific periods of child development) also are considered. Finally, we add to the models an indicator for whether HSNP was available in the location at different child ages. One quarter of the 80 baseline survey clusters were at least partially covered by HSNP by 2012, with approximately 5 clusters entering each year from 2009 to 2011, and 11 starting in 2012. Approximately half of the 274 HSNP sub-locations were partially covered by HSNP, with 30 entering in 2009 and 2010, and 40 entering in 2011 and 2012. The HSNP indicators are captured in similar fashion to the drought shocks, with indicators of whether or not the program was available in the geographic cluster at given child ages. Importantly, we do not condition on individual level participation or receipt of the program, but rather use administrative data from HSNP to assess whether anyone in the cluster was receiving the program for each age from 3 to 8.4 (2) 𝑧𝑗𝑘𝑢 = 𝛾0 + 𝛾1𝑏𝑗𝑘𝑢 + ∑ 𝜀𝑏

𝑈 𝑏=𝑏0

𝑇

𝑘,𝑢−𝑏𝑗+𝑏 + ∑

𝜄ℎ

𝐼 ℎ=ℎ0

𝐼𝑇𝑂𝑄

𝑘,𝑢−𝑏𝑗+ℎ + 𝒀𝒋𝒌𝒖𝜸𝟑 + 𝑣𝑗𝑘𝑢

In this model, the parameter of interest is 𝜄ℎ, the association of HSNP presence at age h on the

• utcome. Under the assumption that presence of HSNP is not associated with the outcomes

(conditional on the controls), this enables exploration of whether girls’ education benefits from presence of HSNP, offsetting possible detrimental effects of drought shocks. Unless otherwise indicated, standard errors are clustered at the level of the 80 survey clusters or 275 sublocations, according to the source.

• IV. Results

Table 2 presents summary statistics for the adolescent girls in the AGI-K sample, who averaged 12.4 years old (standard deviation [SD]: 1.1). Approximately 4 percent report their mother is deceased and 7 percent, their father (but very few are double orphans). Approximately 90 percent of girls whose mother is alive reside with her and 86 percent whose father is alive reside

4 Because HSNP did not begin until 2009 there was no availability at younger ages.

14 with him. Despite short-term residency patterns in the district, 90 percent have lived in their current location for at least five years and consequently most of their schooling has taken place while residing there. Less than 5 percent of their fathers and 2 percent of their mothers have ever attended

• school. More than 92% of girls identify as Somali. Ownership of important assets and reported

difficulty meeting food security or hypothetical expenditure needs confirm that it is a poor

• population. More than half of the girls live in households in which they had gone without food for

a day in the previous month. Moreover, 50 percent live in households in which they did not have enough savings or something to sell readily to meet an expenditure of 1000 Kenyan Shillings (KES) or approximately 10 U.S. dollars.5 The average current grade level of girls in the sample is 2.8, already well ahead of their

• parents. Less than 15 percent, however, have completed grade 6 or higher. We measure schooling

progress by the number of grades a girl is “ahead” of where she should be given her age, had she begun primary school at age 6 and completed one grade each year. A zero indicates the girl is on schedule for her age, 1 that she is a grade ahead, and -1, a grade behind, and so on. For example, a 12 year old in grade 6 would have a measure of zero, while her friend of the same age in grade 5 a measure of -1. On average, girls are on schedule, with one-third a grade or more behind and

• ne-third a grade or more ahead.

Trained female enumerators administered in the home three education-related tests. The first is a literacy test assessing the girl’s ability to read aloud completely without error two sentences in Swahili and two sentences in English.6 The average number correct was 1.6 and one- third of the girls read all four sentences correctly. The second is a numeracy test assessing the girl’s facility with basic addition, subtraction, division, and multiplication. The test includes 26 mathematics questions based on the Kenyan grade 2 level curriculum (UWEZO, 2014). The average score was 16.0 (SD 10.1). Twenty percent earned a perfect score, with an additional 20 percent answering at least 23 out of 26 correctly.

5 In mid-2015, the official exchange rate was 98.5 KES to the U.S. dollar. 6 The two Swahili sentences are: 1) Ukulima ni kazi ngumu; 2) Mtoto anasoma kitabu. The two English sentences, also administered in the 2008–09 Kenyan Demographic and Health Survey, are: 1) Parents love their children; 2) Farming is hard work.

15 The final assessment was the Raven’s Progressive Matrices test (Raven, 1984), a nonverbal multiple-choice assessment of cognitive ability where the respondent identifies the missing element that completes a pattern. The test measures one’s ability to make sense out of confusing

• r complex data and the ability to perceive new patterns and relationships, rather than achievement
• n a schooling related test. Every other problem from the Raven Progressive Matrices AA, AB,

and BB was administered, for a total of 18 problems; the average number correct was 7.2 (SD 3.0) and while no girl earned a perfect score, about 10 percent answered 11 or more correctly. Table 3 presents results based on the AGI-K baseline survey, for which there are more detailed educational outcomes. Sample sizes prohibit precise estimation of all annual drought shocks individually, so they are combined into three groupings, age 0-1, 2-6, and 7-8 corresponding to the early life period that is potentially important for cognitive development, early childhood, and school-starting ages. For each outcome we present two sets of results (both of which control for age-cohort by subdistrict fixed effects). The first examines the effect of the drought shocks alone and the second, after controlling for parental and household characteristics the association of HSNP. Drought shocks in at least one of the three periods negatively affects every outcome considered, with early life and school-age periods being the most important. Early life (0-1) negatively affects all but whether the girl has ever attended school. Ages 2-6 only affect grade for age and the math test scores. School-entry age shocks hinder all but the reading test and

• Raven. The pattern of shocks for the Raven is consistent with its interpretation as a cognitive test

not solely related to schooling, as early life shocks (0-1) are the only significant period. Examining the role of HSNP, similarly combined to enhance power (5-7), we see that while generally positive, estimated associations are only significant in the case of math tests. Analysis of the AGI-K survey sample indicates the important role of drought shocks in girls’ educational development but is less conclusive regarding the potential role of HSNP to “mitigate” the impacts of such shocks. Using the much larger HSNP census, we are able to shed further light on this latter question for the indicator of grades ahead for age. In Table 4 we present two versions each of equations 1 and 2 with annual drought shock and HSNP presence indicators. All models include household level fixed effects (results are nearly identical if instead village level fixed effects are used) and calculation of the standard errors accounts for clustering at the household level. Identification of the estimated effects is based on within-household comparisons

16

• f girls of different ages and accounts for all household (and village) level persistent heterogeneity.

Our preferred estimates are those in columns (1) and (3), given the very low percentage of shocks and HSNP availability for higher ages as shown in Table 1B which may explain the large point estimates for drought shocks in ages 6-8. Overall, the evidence confirms the negative effect of drought shocks on grades ahead for age seen with the AGI-K sample. Shocks in early life substantially reduce grades and shocks in later childhood have even larger effects. These results also demonstrate, however, that availability of HSNP in the village during childhood significantly improves outcomes, suggesting HSNP can play a role in offsetting the negative impact of rainfall shocks.

• V. Conclusions

Using data from a the AGI-K baseline survey, which is representative of young adolescent girls living in rural areas in three districts of Wajir County, combined with historical rainfall data, examined the associations between drought-induced shocks during different stages of childhood and educational and cognitive outcomes in early adolescence. Using 2014 HSNP census data, we explored whether an unconditional cash transfer program was able to offset the effects of rainfall shocks on schooling relative to a girls age. Preliminary evidence suggests that early life shocks, in particular in the first few years of life and then again at ages when children should be starting primary school, negatively influence adolescent girls educational and achievement outcomes, including reading assessments, mathematical assessments and a non-verbal cognitive skills test. The latter are significantly affected by shocks in early life, a period important for brain development, but not significantly affected in later childhood. However, the presence of an important relief program that has operated in the region since 2009, the Hunger and Safety Net Program, targeted to some, but not all of the study clusters, somewhat mitigates those effects. The initial evaluation of the HSNP program found that children in beneficiary households were not more likely to enroll in or attend school (Merttens et al., 2013). However, these findings show that in the context of a drought, the HSNP payments may help keep girls in school. Given

17 that the second phase of HSNP makes provisions for additional payments to a wider set of beneficiaries during emergencies (drought included), these findings suggest that making using of this platform may have the potential to cushion the negative effects of drought on girls education. Further research should be done integrating the AGI-K data with the HSNP census data to understand the role of HSNP payments together with the AGI-K intervention components, as well as a continued mitigating factor during droughts or other emergencies.

18

• VI. Tables

Table 1A - Rainfall in Wajir, by year and location Census sublocations AGI-K clusters z-score Shock (%) z-score Shock (%) 2000

• 0.959

45.8%

• 0.977

36.3% (0.225) (0.156) 2001

• 0.315

0.0%

• 0.394

0.0% (0.226) (0.247) 2002 0.364 0.0% 0.478 0.0% (0.339) (0.287) 2003 0.550 0.0% 0.354 0.0% (0.331) (0.369) 2004

• 0.140

0.0%

• 0.210

0.0% (0.214) (0.201) 2005

• 0.997

54.2%

• 0.990

51.3% (0.191) (0.165) 2006 0.985 0.0% 1.063 0.0% (0.323) (0.394) 2007

• 0.195

0.0%

• 0.190

0.0% (0.285) (0.297) 2008

• 0.209

0.0%

• 0.282

0.0% (0.254) (0.289) 2009

• 0.295

0.4%

• 0.300

0.0% (0.234) (0.260) 2010

• 0.539

5.8%

• 0.539

3.8% (0.245) (0.228) N 274 274 80 80

Notes: Shock (%) is percent of locations where z-score <-1.0 that year. Location-specific z-score calculated using location mean and standard deviation calculated from 1980-1999; Standard deviations in parentheses.

19

Table 1B - Rainfall shocks in Wajir, by age Census sublocations AGI-K clusters Shock (%) Shock (%)

• 1 (year before birth)

18.5% 7.6% 0 (first year of life) 10.6% 2.8% 1 11.8% 12.8% 2 4.2% 14.5% 3 9.2% 11.6% 4 6.4% 8.9% 5 6.1% 0.3% 6 0.9% 0.7% 7 0.4% 1.2% 8 0.6% 0.7% 9 0.4% 0.5% 10 0.4% 0.0% N 66,158 2,147

Notes: Shock (%) is percent of locations where z-score <-1.0 that year. Location-specific z-score calculated using location mean and standard deviation calculated from 1980-1999; Standard deviations in parentheses.

20

Table 2 - Summary Statistics for AGI-K baseline survey 11 12 13 14 All N=565 N=630 N=489 N=463 N=2147 Ever enrolled 0.772 0.752 0.722 0.767 0.754 Currently enrolled 0.735 0.713 0.687 0.724 0.715 Enrolled by age 8 0.319 0.338 0.266 0.371 0.324 Enrolled by age 9 0.519 0.492 0.425 0.495 0.484 Enrolled by age 10 0.692 0.627 0.581 0.618 0.632 0.772 0.710 0.677 0.691 0.714 Grades completed 1.986 2.616 3.070 3.907 2.832 (1.672) (2.111) (2.443) (2.835) (2.365) Number grades behind 1.986 2.616 3.070 3.907

• 1.564

for age (1.672) (2.111) (2.443) (2.835) (2.304) Math test (# correct 14.237 15.995 16.078 18.274 16.042 [N=2135]

• ut of 26)

(9.573) (9.961) (10.287) (10.058) (10.047) Read Swahili correctly 0.279 0.407 0.424 0.563 0.411 [N=2135] Read English correctly 0.215 0.337 0.370 0.533 0.355 Raven (# correct out 6.881 7.226 7.103 7.630 7.195 [N=2030]

• f 18)

(2.695) (2.971) (3.139) (3.191) (2.999) Notes: AGI-K baseline survey

21

Table 3 - Rainfall shocks, HSNP, and educational outcomes (AGIK Baseline survey) VARIABLES

Ever attended Started by age 7 Grades ahead for age Reading test-all correct Math test Z-score Raven test Z-score

Shock ages 0-1

• 0.044
• 0.050
• 0.093***
• 0.086**
• 0.603***
• 0.593***
• 0.127***
• 0.117***
• 0.242**
• 0.246**
• 0.150*
• 0.148*

(0.042) (0.042) (0.034) (0.033) (0.190) (0.189) (0.034) (0.033) (0.097) (0.095) (0.080) (0.083)

Shock ages 2-6

• 0.056
• 0.060*
• 0.030
• 0.027
• 0.418*
• 0.427*
• 0.026
• 0.029
• 0.169**
• 0.167**
• 0.017
• 0.011

(0.036) (0.036) (0.026) (0.027) (0.223) (0.221) (0.033) (0.033) (0.081) (0.082) (0.077) (0.078)

Shock ages 7-8

• 0.222**
• 0.250***
• 0.214***
• 0.231***
• 1.811***
• 1.881***
• 0.100
• 0.116
• 0.430***
• 0.547***
• 0.189
• 0.259

(0.100) (0.092) (0.036) (0.037) (0.565) (0.496) (0.110) (0.106) (0.162) (0.156) (0.172) (0.183)

HSNP ages 5-7

0.068 0.030 0.194 0.036 0.232* 0.126 (0.063) (0.046) (0.265) (0.056) (0.130) (0.102)

Mother any schooling

0.150* 0.082 0.643 0.266** 0.581***

• 0.053

(0.079) (0.109) (0.474) (0.122) (0.164) (0.294)

Father any schooling

0.035 0.161** 0.780*** 0.171*** 0.191* 0.058 (0.046) (0.067) (0.295) (0.060) (0.109) (0.133)

Wealth index

• 0.004

0.006

• 0.022

0.000 0.013 0.004 (0.010) (0.008) (0.055) (0.008) (0.021) (0.017)

Constant

0.812*** 0.800*** 0.231*** 0.216***

• 2.237***
• 2.313***

0.272*** 0.249***

• 0.045
• 0.096

0.057 0.033 (0.044) (0.045) (0.037) (0.037) (0.191) (0.189) (0.040) (0.042) (0.099) (0.102) (0.072) (0.075)

Observations

2,147 2,125 2,147 2,125 2,147 2,125 2,135 2,113 2,147 2,125 2,030 2,008

R-squared

0.012 0.015 0.024 0.031 0.058 0.061 0.072 0.082 0.043 0.052 0.021 0.022

Notes: Ever attended, started by age 7, and reading test (all four correct) are dummy variables. *** indicates significance at p<0.01, ** at p<0.05, and * at p<0.10. All models include controls for single age subdistrict dummies for the the four age groups and three subdistricts. Standard errors are calculated allowing for clustering at the (80) survey cluster level.

22

Table 4 - Rainfall shocks, HSNP, and grades ahead for age (HSNP census) (1) (2) (3) (4)

Shock in utero

• 0.514***
• 0.700***
• 0.357***
• 0.408***

(0.055) (0.054) (0.055) (0.054) Shock age 0

• 0.891***
• 1.076***
• 0.567***
• 0.681***

(0.055) (0.054) (0.058) (0.057) Shock age 1

• 1.626***
• 1.773***
• 1.175***
• 1.056***

(0.054) (0.052) (0.057) (0.057) Shock age 2

• 2.156***
• 2.330***
• 1.708***
• 1.486***

(0.069) (0.068) (0.072) (0.074) Shock age 3

• 2.592***
• 2.795***
• 2.196***
• 1.823***

(0.066) (0.062) (0.067) (0.070) Shock age 4

• 2.608***
• 3.012***
• 2.322***
• 2.086***

(0.075) (0.075) (0.074) (0.080) Shock age 5

• 0.881***
• 2.407***
• 0.919***
• 1.752***

(0.214) (0.197) (0.198) (0.184) Shock age 6

• 3.038***
• 2.154***

(0.172) (0.160) Shock age 7

• 2.766***
• 1.923***

(0.232) (0.211) Shock age 8

• 4.561***
• 3.449***

(0.190) (0.182) HSNP age 3 0.571*** 0.344*** (0.143) (0.130) HSNP age 4 0.303*** 0.369*** (0.094) (0.092) HSNP age 5 1.193*** 0.419*** (0.064) (0.074) HSNP age 6 0.433*** (0.065) HSNP age 7 0.239*** (0.071) HSNP age 8 0.990*** (0.064) Constant

• 0.761***
• 0.557***
• 1.104***
• 1.492***

(0.026) (0.026) (0.029) (0.039) Observations 60,125 60,125 60,125 60,125 R-squared 0.147 0.178 0.174 0.228 Number of unique households 40,857 40,857 40,857 40,857 Notes: *** indicates significance at p<0.01, ** at p<0.05, and * at p<0.10. All models include household-level fixed effects. Standard errors are calculated allowing for clustering at the household level.

23

• VII. References

Alderman, H., & King, E. M. (1998). Gender differences in parental investment in education. Structural Change and Economic Dynamics, 9(4), 453-468.

• ASDSP. (2014). Household Survey Baseline Report: Wajir County. Retrieved from Nairobi,

Kenya: Austrian, K., Muthengi, E., Mumah, J., Soler-Hampejsek, E., Kabiru, C. W., Abuya, B., & Maluccio, J. A. (2016). The Adolescent Girls Initiative-Kenya (AGI-K): study protocol. BMC public health, 16(1), 210. Bank, W. (2017). Kenya Fact Sheet. Retrieved from http://www.worldbank.org/en/country/kenya/overview Barham, T., Macours, K., & Maluccio, J. A. (2013). Boys' cognitive skill formation and physical growth: Long-term experimental evidence on critical ages for early childhood

• interventions. The American Economic Review, 103(3), 467-471.

Björkman-Nyqvist, M. (2013). Income shocks and gender gaps in education: Evidence from

• Uganda. Journal of Development Economics, 105, 237-253.

Burke, M., Gong, E., & Jones, K. (2015). Income shocks and HIV in Africa. The Economic Journal, 125(585), 1157-1189. Currie, J., & Almond, D. (2011). Human capital development before age five. Handbook of labor economics, 4, 1315-1486. Dai, A. (2013). Increasing drought under global warming in observations and models. Nature Climate Change, 3(1), 52-58. Davis, B., Winters, P., Carletto, G., Covarrubias, K., Quiñones, E. J., Zezza, A., . . . DiGiuseppe,

• S. (2010). A cross-country comparison of rural income generating activities. World

development, 38(1), 48-63. Dercon, S., Hoddinott, J., & Woldehanna, T. (2005). Shocks and consumption in 15 Ethiopian villages, 1999-2004. Journal of African economies, 14(4), 559.

• FAO. (2017). Crop Calendar. Retrieved from

http://www.fao.org/agriculture/seed/cropcalendar/welcome.do Funk, C. C., Peterson, P. J., Landsfeld, M. F., Pedreros, D. H., Verdin, J. P., Rowland, J. D., . . . Verdin, A. P. (2014). A quasi-global precipitation time series for drought monitoring (2327-638X). Retrieved from Hoddinott, J., & Kinsey, B. (2001). Child growth in the time of drought. Oxford Bulletin of Economics and statistics, 63(4), 409-436.

• HSNP. (2017). HSNP Dashboard. Retrieved from http://www.hsnp.or.ke/#
• IPCC. (2014). Climate Change 2014–Impacts, Adaptation and Vulnerability: Regional Aspects:

Cambridge University Press. Jejeebhoy, S. J. (1995). Women's education, autonomy, and reproductive behaviour: Experience from developing countries. OUP Catalogue. Jensen, R. (2000). Agricultural volatility and investments in children. The American Economic Review, 90(2), 399-404. Kenya National Bureau of Statistics, Ministry of Health/Kenya, National AIDS Control Council/Kenya, Kenya Medical Research Institute, Population, N. C. f., & Development/Kenya. (2015). Kenya Demographic and Health Survey 2014. Retrieved from Rockville, MD, USA: http://dhsprogram.com/pubs/pdf/FR308/FR308.pdf

24 Lavy, V., Schlosser, A., & Shany, A. (2016). Out of africa: Human capital consequences of in utero conditions. Retrieved from Merttens, F., Hurrell, A., Marzi, M., Attah, R., Farhat, M., Kardan, A., & MacAuslan, I. (2013). Kenya Hunger Safety Net Programme Monitoring and Evaluation Component. Impact Evaluation Final Report. Oxford Policy Management. Miguel, E., Satyanath, S., & Sergenti, E. (2004). Economic shocks and civil conflict: An instrumental variables approach. Journal of political Economy, 112(4), 725-753. Molina Millán, T. (2014). Regional Migration, Insurance and Economic Shocks: Evidence from Nicaragua. Raven, J., Court, JH, and Raven, J. (1984). Coloured progressive matrices. Manual for Raven’s Progressive Matrices and Vocabulary Scales. Schroeder, R. A. (1987). GENDER VULNERABILITY T0 DROUGHT: A CASE STUDY OF THE HAUSA SOCIAL ENVIRONMENT.

• UWEZO. (2014). Kenya National Learning Assessment: Test Booklet. Retrieved from Nairobi,

Kenya: