2. Economic shocks and individual outcomes

Substantial literature documents the importance of rainfall and weather patterns for economic outcomes in low-income settings, including in much of sub-Saharan Africa (Dercon et al., Reference Dercon, Hoddinott and Woldehanna2005; Björkman-Nyqvist, Reference Björkman-Nyqvist2013). Most commonly the link is explained by the critical role of rainfall in crop agriculture and the role of agriculture in rural livelihoods in the region (Davis et al., Reference Davis, Winters, Carletto, Covarrubias, Quiñones, Zezza, Stamoulis, Azzarri and DiGiuseppe2010). Links through civil conflict resulting from economic stress have also been investigated (Miguel et al., Reference Miguel, Satyanath and Sergenti2004). Additionally, there are aspects of disease transmission related to weather that can influence individual health outcomes (Rabassa et al., Reference Rabassa, Skoufias and Jacoby2014; Berazneva and Byker, Reference Berazneva and Byker2017). In settings where pastoralism rather than crop agriculture is the main economic activity, such as I semi-arid Wajir County, Kenya, rainfall patterns play a vital role through similar mechanisms (Fafchamps et al., Reference Fafchamps, Udry and Czukas1998; Kazianga and Udry, Reference Kazianga and Udry2006; ASDSP, 2014; Maystadt and Ecker, Reference Maystadt and Ecker2014).

Building on the crucial role of rainfall, investigators use exogenous rainfall measures to examine the causal effect of economically meaningful shocks during childhood on contemporaneous or later human capital outcomes. In addition to exogenous variation, a key element of the research design for this research is that historical rainfall measures provide shock indicators at different points throughout the life of the child. In such research, the correlation observed between rainfall deviations and the outcome is interpreted to be caused by the indirect effects of the rainfall shocks operating through economic or other conditions (Burke et al., Reference Burke, Gong and Jones2015). Research has examined human capital outcomes measured in childhood (typically anthropometrics), adolescence (schooling, achievement and cognition), and adulthood (completed schooling and marriage, fertility and labor market outcomes) (Currie and Vogl, Reference Currie and Vogl2012; Frankenberg and Thomas, Reference Frankenberg, Thomas, Barrett, Carter and Chavas2018). Studies typically consider shocks in early life (e.g., the first two or three years) or shocks during later childhood or school-age years, but only a few incorporate both periods into a single framework as done in this paper.

Research examining rainfall or other shocks in early life demonstrates that they can have negative consequences on physical and cognitive outcomes during this critical developmental period (Glewwe and King, Reference Glewwe and King2001; Hoddinott and Kinsey, Reference Hoddinott and Kinsey2001; Grace et al., Reference Grace, Davenport, Funk and Lerner2012; Rabassa et al., Reference Rabassa, Skoufias and Jacoby2014). Moreover, effects can persist, translating into longer-term impacts in adolescence and adulthood (Alderman et al., Reference Alderman, Hoddinott and Kinsey2006; Maccini and Yang, Reference Maccini and Yang2009; Currie and Vogl, Reference Currie and Vogl2012; Thai and Falaris, Reference Thai and Falaris2014; Dinkelman, Reference Dinkelman2017). Leight et al. (Reference Leight, Glewwe and Park2015) provide evidence of the effects of early life shocks at different stages in life using panel data from China; they find negative effects of early life shocks on early childhood anthropometry and on later primary school age cognitive skills. Their findings also indicate partial ‘catch-up’ for those experiencing early life shocks, demonstrating some recovery, possibly the result of compensatory behaviors and investments within households. Akresh et al. (Reference Akresh, Bagby, de Walque and Kazianga2012) use early childhood shocks in Burkina Faso as instrumental variables for ability in relations modeling adolescent schooling and find evidence that the shocks reduce human capital directly through reductions in ability, and indirectly through consequent differential investments across siblings. The studies underscore how observed effects of past shocks capture the net cumulative effect of the initial shocks as well as subsequent influences and behaviors undertaken in response.

Research on shocks during later childhood up to early school ages points to the potential importance of that period as well (Beegle et al., Reference Beegle, Dehejia and Gatti2006). Moreover, effects may differ for girls who in some settings face different social, economic and cultural barriers to their development and educational attainment than boys. Gender expectations, low prioritization of female education, early marriage and childbearing can all limit adolescent development, and the influence of many of these factors may be exacerbated during periods of low rainfall or drought. Björkman-Nyqvist (Reference Björkman-Nyqvist2013) finds that shocks had more detrimental effects on educational outcomes of school-age girls than boys in Uganda. Shocks during childhood may also lower the age at which girls marry in settings like Wajir where it is common for the groom and his family to pay a bride price; such early marriage might even be a mechanism for household consumption smoothing (Corno et al., Reference Corno, Hildebrandt and Voena2017).

A smaller set of studies examines shocks both in early life and in subsequent childhood periods in a single framework. In addition to the mechanisms described above, rainfall shocks may operate differently at different periods of individuals' lives because of the timing of physical development and also because, in settings where low rainfall affects wages, there is another pathway to human capital investment through changes in the opportunity cost of time for children. Shah and Steinberg (Reference Shah and Steinberg2017) demonstrate in India that positive rainfall shocks in early life improved adult educational outcomes but positive shocks during school-age years lowered them. They argue that the latter reduction results from pulling students out of school when there are positive productivity shocks and associated higher wages.

In all analyses using historical shocks, observed outcomes capture the cumulative effect of: (1) the historical pattern of the shocks themselves, (2) previous anticipatory behaviors, and (3) any subsequent actions and investments undertaken by families. The latter, for example, include the investments highlighted by Leight et al. (Reference Leight, Glewwe and Park2015) and Akresh et al. (Reference Akresh, Bagby, de Walque and Kazianga2012). Frankenberg and Thomas (Reference Frankenberg, Thomas, Barrett, Carter and Chavas2018) find that subsequent investments undertaken by families have the potential to substantially limit the longer-term effects of large shocks in Indonesia; they also argue that such behavioral responses will likely vary by context, making it difficult to reach general conclusions about the effect of shocks in different settings. Last, in addition to individual- or household-level factors, the outcomes are also conditioned by the surrounding program and policy environment (Frankenberg and Thomas, Reference Frankenberg, Thomas, Barrett, Carter and Chavas2018).

3. Wajir County context and data

Situated in northeastern Kenya on the Somali border, Wajir County is a sparsely populated region with a predominantly Muslim population. It has about 800,000 residents and on average only 14 persons per square kilometer (km), though most residents are clustered together and live in small concentrated settlements or towns. In 2005 the extreme poverty rate was 58 per cent, and overall poverty was 84 per cent (Merttens, Reference Merttens2018). The region has experienced insecurity in recent years related to instability in Somalia, especially near the international border (ASDSP, 2014).

Wajir has a hot, semi-arid desert climate and the primary economic activity is pastoralism. Only a few thousand of its 57,000 square km are arable, and only a fraction of that cultivated. Consequently, livestock and related products dominate the local economy. In 2013 livestock products were valued at nearly 50 times crop production (ASDSP, 2014).

The pastoral-based livelihoods of the population make it sensitive to low rainfall and drought conditions. These conditions can have substantial effects on livestock prices and economic wellbeing (Kazianga and Udry, Reference Kazianga and Udry2006). Additionally, during drought family members may need to travel farther, or even relocate altogether, to find suitable watering sites or pasture for their herds (MoALF, 2017; Merttens, Reference Merttens2018). These factors make Wajir a particularly relevant setting for examining some of the longer-term consequences of local rainfall shocks.

The pastoral-based livelihoods in Wajir also have implications for educational outcomes in the county. The ethnolinguistic groups in Kenya that commonly practice pastoralism, the Somali and Maasai, are the least likely to send their children to school, which is related to factors including cultural norms, language and opportunity costs of time (Vimefall et al., Reference Vimefall, Andrén and Levin2017). Consistent with this pattern, the largely Somali Wajir County has the lowest adult literacy rate (26 per cent) and the lowest average female education in the country (KNBS, 2015). Median grades completed are zero for adult women and 4.7 for men (KNBS, 2015). Notably, the educational gender gap in Wajir is one of the highest in the country. Low levels of female education are accompanied by early marriage and high fertility; the median age at first marriage was 18 and the total fertility rate 7, the highest in Kenya (KNBS, 2015).

We combined remotely sensed, local historical rainfall data with: (1) individual-level data from a 2015 survey carried out in Wajir for the Adolescent Girls Initiative-Kenya (AGI-K) and, separately, (2) individual-level data from a 2013 census of Wajir carried out by the Hunger Safety Net Programme (HSNP). Because communities in Wajir depend on rainfall for agricultural and pastoral livelihoods, we use rainfall anomalies to capture lower than expected precipitation.

3.1 Adolescent Girls Initiative-Kenya (AGI-K): 2015 baseline survey

The first data source is the 2015 baseline survey collected for the AGI-K study of girls ages 11–14. AGI-K identified 80 rural villages with a primary school in areas of Wajir County except for Wajir North (figure 1a), which was excluded because for security reasons the implementing partner did not operate there. In villages with fewer than 40 households with eligible girls, all households and girls ages 11–14 were interviewed. In larger villages, 40 households were randomly selected and within each one girl aged 11–14 was interviewed (Austrian et al., Reference Austrian, Muthengi, Mumah, Soler-Hampejsek, Kabiru, Abuya and Maluccio2016).

Figure 1. (a) AGI-K 2015 survey villages, (b) HSNP census survey locations.

Female enumerators implemented a comprehensive survey at both the household and individual girl levels. The household survey was conducted with the household head or guardian adult in the language of their choice and included household characteristics, assets and parental attitudes toward education. The individual girl level survey collected information on schooling history and educational attainment, among other things. We construct binary indicators of ever enrolled (whether an individual had ever enrolled in any level of primary schooling), and whether the individual had ever enrolled by age eight. We characterize schooling grade progress by the total number of completed grades.

Enumerators also administered three education-related tests in the home, covering literacy in the two languages used in public school instruction, mathematics and nonverbal cognitive skills. To avoid selection bias since nearly 25 per cent of girls were not enrolled in school, all girls regardless of schooling status took the test. The literacy test assessed the ability to read aloud completely without error two sentences in Swahili and two sentences in English. The mathematics test assessed numeracy and facility with basic arithmetic including 26 questions based on the Kenyan grade 2 level mathematics curriculum using the Kenya National Learning Assessment tool (UWEZO, 2014). The third assessment was the Raven Coloured Progressive Matrices test, a nonverbal multiple-choice assessment of cognitive skills in which the respondent identifies the missing element completing a pattern (Raven et al., Reference Raven, Court and Raven1984). Rather than directly assessing schooling-related achievement, the Raven test measures the ability to make sense out of confusing or complex data and perceive new patterns and relationships.

The AGI-K survey also collected information on schooling attitudes and expectations for the girls which enables exploration of underlying mechanisms. Girls were asked about the importance they ascribed to schooling and how confident they felt about being able to attain their schooling objectives. Girls also provided self-reported health assessments. And, although there was no comprehensive expenditure or income module at the household level, the survey collected an array of indicators related to household economic wellbeing or wealth. The final sample size for the individual AGI-K girl-level survey was 2,147. Some girls resided in the same household, so at the household level the final sample size was 2,030.

3.3 Rainfall data and construction of rainfall shocks

To construct historical rainfall measures we use Climate Hazards Group InfraRed Precipitation with Station data (CHIRPS), satellite-based high resolution gridded data spanning latitudes 50°S to 50°N that capture daily and monthly precipitation from 1981 to the present (Funk et al., Reference Funk, Peterson, Landsfeld, Pedreros, Verdin, Shukla, Husak, Rowland, Harrison, Hoell and Michaelson2015). We used the aggregated data consisting of monthly raster files, each containing a grid with 0.05 × 0.05 degree spatial resolution. In Kenya, each data point (or pixel) represents an approximate 5 km × 5 km square. Using the global positioning system (GPS) coordinates for each location where sample individuals lived, we imputed average monthly rainfall for a 10 km radius buffer, averaging the rainfall values across grid cells that overlaid the buffer. The 10 km buffer is intended to approximate the area a pastoralist would typically range with their herd from a home base, though 5 km buffers yielded similar results. The rainfall data are available starting in 1981, nearly 20 years before the birth of the oldest children surveyed. Using the monthly measures, we constructed historical long-term averages and standard deviations (SD) from 1981–1999 for each location and then used the measures to construct standardized z-scores for each location in each year after 1999. Standardization at the local 10 km radius area, rather than a larger one, controls for differences in local average rainfall. Because individuals living in areas that have historically low average rainfall may employ anticipatory strategies in response, typical low levels of rainfall they experience are not exogenous shocks in the way that deviations from local averages can be. The procedure yields annual measures of the z-score for every location where sample individuals lived each year from 2000–2015.

In the next step, we use the z-scores to construct a binary indicator for whether the location experienced a local rainfall shock each year. To ensure they are substantial, we define a shock as average annual precipitation one or more SD below the local historical average. Online appendix table A1 presents average z-scores and the prevalence of local rainfall shocks across locations by year from 2000 through 2015 for each sample. There is substantial variation in rainfall deviations over time, reflected in the differences in average z-scores across years, and minimal correlation over time. In 11 out of 16 years, average z-scores were negative. One or more SD rainfall shocks are uncommon in recent years in Wajir and concentrated in only two calendar years: 2000 and 2005. In all other years, less than 5 per cent of the study locations experienced shocks. Calendar year shocks, however, are experienced by children living in the same location, or even the same household, at different ages. We map average annual rainfall patterns for these years in figure 2a. In 2000, low rainfall was widespread in Wajir and affected most locations throughout the county (including more than 90 per cent of the study locations) whereas in 2005 low rainfall affected only about half of the locations. The latter year underscores the variation across different locations in Wajir County. A comparison with the subsequent calendar years 2001 and 2006 shows the variation over time. Rainfall shocks in this paper are not measured by overall rainfall, but rather by deviations from local historical averages. In figure 2b we present the z-score deviations from the historical averages for the same calendar years (with lighter areas representing more negative z-scores), demonstrating there is similarly substantial variation in z-scores across time and space.

Figure 2. Wajir County rainfall (in 2000, 2001, 2005, 2006). (a) Average annual rainfall in mm, (b) Z-score deviations from historical average rainfall.

The final step uses age and resident location to link each sample individual to the rainfall shock data so that for every individual we have indicators for whether they experienced a rainfall shock in their resident location in each year of their lives. In table 1 we summarize exposure to rainfall shocks by age for all children in each sample, beginning with the year the child was born. Because the most recent year with substantial rainfall shocks was 2005, the percentage of children in the sample with rainfall shocks declines with their age and is rare at the older ages.

Table 1. Rainfall shocks in Wajir County, by location and age

Notes: Percentage of individuals experiencing rainfall shock at each age, based on calendar year shocks reported in online appendix table A1.

4. Methodology

Drawing from the literature, we theorize that rainfall shocks affect adolescent education primarily through their effects on household-level resources and the value of activities often done by children such as herding. These effects can differ at each stage of child development.

Negative economic shocks to household income or wealth during early childhood can limit household resources available for nutrition, health and care, directly affecting early life physical and cognitive development. Offsetting behavioral responses (such as selling assets in the short term) may not fully prevent reductions in child investments and also may result in persistently reduced household resources. Consequently, shocks during early years can also influence later life outcomes, either through persistently reduced household resources available to invest in education during school ages or through persistent negative effects on the young child's physical or cognitive preparedness to advance in school. There might also be persistent effects on schooling attitudes and expectations (Wydick et al., Reference Wydick, Glewwe and Rutledge2013).

Similarly, rainfall shocks when children are older can reduce household resources available for child investment at that point as well. Later shocks might also change the opportunity cost (or value) of time of the now older children who often assist in carrying out household chores and herding, a particularly important activity for children in the Wajir context. In the extreme, when herding animals locally is no longer possible, children might need to travel long distances with the herd, making it infeasible for them to attend school. In such instances, child schooling might be interrupted because of short-term economic necessity.

A final relevant aspect of mechanisms evident in the literature is the potential for gendered effects resulting from possible differences in physical development, societal attitudes, initial schooling differences, economic returns and the willingness of parents to invest in children. Some literature suggests that impacts of shocks might be larger for girls than for boys. When asked about the importance of investments in child schooling, nearly one-third of parents in the AGI-K survey sample indicated they would prioritize male over female education when facing financial constraints, suggesting investments in females would be more sensitive to negative shocks. In Wajir, evidence from this study indicates girls are commonly responsible for herding so that rainfall shocks can affect their opportunity costs. Given the survey design, in this paper we primarily analyze girls and are unable to comprehensively assess gender differentials.

Turning to the econometric specifications, we first estimate a reduced form cross-sectional econometric model that incorporates multiple past annual rainfall shocks from birth through early primary school ages as exogenous determinants of educational outcomes. The base model includes rainfall shocks at each age:

(1)\begin{equation}{y_{ij}} = {\beta _0} + \mathop \sum \limits_{a = {a_0}}^T {\delta _a}{S_{ij,a}} + {\boldsymbol{X}_{\boldsymbol{ij}}}{\boldsymbol{\beta }_1} + {u_{ij}},\end{equation}

where ${y_{ij}}$ is the outcome for child i in resident location j. ${S_{ij,a}}$ is a binary indicator for whether there was a rainfall shock in child i's location j when she was age a, with age ranging from a ₀ to T. For the analysis using the AGI-K survey sample, a ranges from 0 (the first year of life) to 7, after which few girls experienced shocks (table 1). For a 14-year-old at the time of the survey in 2015 and therefore born in 2001, ${S_{ij,0}}$ is the binary rainfall shock indicator for her location j in 2001, when she was a = 0 years old; ${S_{ij,1}}$ is the rainfall shock indicator for when she was a = 1 year old in 2002; and so on up to ${S_{ij,7}}$ for when she was a = 7 years old in 2008. The rainfall shock indicators for children of different ages when interviewed in 2015 are constructed similarly. Accurate measurement of both age and resident location are important for the analysis.Footnote ¹ The negative rainfall shocks also may have influenced mortality, resulting in a selective sample with healthier individuals more likely to have survived. Both possible mismeasurement and selective mortality would likely lead to attenuation bias in the estimation of (the negative) shock parameters, making them conservative. The set of controls in vector ${\boldsymbol{X}_{\boldsymbol{ij}}}$ includes age-in-year fixed effects, a binary female indicator in the HSNP census analyses which includes both girls and boys, and geographic-level fixed effects described below.Footnote ² ${u_{ij}}$ represents an assumed idiosyncratic error term. We estimate similar relationships for educational outcomes for girls and boys using the HSNP census data. Because the HSNP census was carried out two years earlier, the timing of the controls in econometric models for those analyses are modified accordingly.Footnote ³

For the AGI-K survey sample of girls, all analyses include age-cohort by subcounty (West, East and South) fixed effects, controlling for age-cohort effects as well as for factors common to all girls in each subcounty, including similarity in rainfall patterns within those geographic areas as reflected in figures 2a and b. For the HSNP census, all analyses include location level fixed effects for the 274 locations so that identification of the causal impacts of rainfall shocks exposure is strengthened further.

To shed light on the mechanisms underlying the estimated effects on educational outcomes, we use the AGI-K survey to examine schooling expectations and health-related indicators, employing the same individual-level specification in equation (1). This demonstrates whether the mechanism is linked to persistence of effects on these related outcomes for the girls.

The other possible mechanism we explore is household-level economic status, considering the role of prior shocks on a set of currently available household-level resources including livestock ownership and housing conditions. Unlike for the individual child level analyses where single calendar year rainfall shocks occur at different child ages (table 1), at the household level this source of variation is unavailable (online appendix table A1). Put differently, while identification still rests on the exogenous rainfall shock at the location level, there is no variation between households in the timing of those shocks within a single location. Consequently, the only practical shock lag we can consider is whether there was a shock in 2005 in the community of residence. For household-level outcomes, therefore, we estimate

(2)\begin{equation}{y_{hj}} = {\beta _0} + {S_{j,2005}} + {\boldsymbol{X}_{\boldsymbol{hj}}}{\boldsymbol{\beta }_1} + {u_{hj}},\end{equation}

where ${y_{hj}}$ is the outcome for household h in resident location j. ${S_{j,2005}}$ is a binary indicator for whether there was a rainfall shock in household h's location j in 2005, 10 years prior to the survey when shocks were widespread in Wajir (online appendix table A1 and figure 2b). The set of controls in vector ${\boldsymbol{X}_{\boldsymbol{hj}}}$ includes subcounty (West, East and South) fixed effects and controls for parental education and whether the parents were alive at the time of the survey. ${u_{hj}}$ represents an assumed household-level idiosyncratic error term.

All econometric models are estimated using ordinary least squares. Analyses with binary outcomes are interpreted as linear probability models. For models using the AGI-K survey sample, standard errors are clustered at the level of the 80 AGI-K survey villages and for models using the HSNP census, standard errors are clustered at the household level.

5. Summary statistics

Both samples reflect the low education and high poverty levels that characterize Wajir County.

5.1 Adolescent Girls Initiative-Kenya survey sample

The AGI-K survey sample indicates that nearly 90 per cent of households own livestock, primarily goats, sheep and chickens, and for some, cows and highly valuable camels. Less than one-quarter own any agricultural land, including pasture. Assessment of various factors related to poverty confirm it is a poor population. More than half of the households had gone without food for at least one day in the previous month. Fifty per cent report that they did not have enough savings or something that could be readily sold to meet an expenditure need of 1,000 Kenyan Shillings (KSh) (~US$10) and a quarter could not meet an expenditure need of KSh5000 (~US$50). Less than 1 per cent of households have electricity or running water in the home. Formal adult education is minimal, with fewer than 2 per cent of mothers and 5 per cent of fathers ever having enrolled in school. Despite generally high residential mobility in the county, 90 per cent had lived in their current location for at least five years and consequently most child schooling has taken place while there. 92 per cent identify as Somali.

A full quarter of the girls had never enrolled in school (table 2). Of the three-quarters who had, many did so late, not starting until ages 8, 9 or 10. The average grade level of girls was 2.9, already higher than their parents. Less than 15 per cent, however, had completed grade 6 or higher. On average girls were more than three grades behind for their age compared to where they would be if they had begun primary school at the recommended starting age of 6 and completed one grade each year. By that measure, less than 15 per cent were on or ahead of schedule. Despite the relatively low levels of schooling, however, among girls who had not yet completed primary school more than 85 per cent indicated that finishing primary school was important or very important to them. Of those, about two-thirds assessed the probability that they would complete primary school as high, greater than 50–50.

Table 2. Summary statistics for 2015 AGI-K survey sample (girls 11–14)

Notes: Standard deviations shown in parentheses for non-binary variables. Math and reading tests (N = 2,135) and Raven (N = 2,030) have some missing outcomes. Every other problem from the Raven Matrices AA, AB and BB was administered, for a total of 18 problems. Finishing primary variables conditional on not having already completed primary (N = 2,043).

One-third of the girls read all four of the literacy test sentences correctly and the average number correct was 1.6. The average score on the mathematics assessment was 16.0 out of 26 (SD 10.0). Twenty per cent earned a perfect score, with an additional 20 per cent answering at least 23 out of 26 correctly. The average number correct on the Raven test was 7.2 (SD 3.0) out of 18 and about 10 per cent answered 11 or more correctly. Self-assessment of health was also high, above 8 (SD 1.7) on a scale of 1–10. Nevertheless, nearly 50 per cent reported an illness in the past month. The most common illnesses were fever and night sweats, with smaller percentages reporting fatigue or diarrhea, coughing or vomiting. In a separate module of the questionnaire asking about different concerns the girls may have, more than one-third reported being worried about their health.