2.1 Literature review

A nascent literature examines the impact of temperature on learning outcomes. For example, hotter school days (Park et al., Reference Park, Goodman, Hurwitz and Smith2020), and test date temperature in high stakes exams have been associated with cognitive decline (see Cho, Reference Cho2017; Park, Reference Park2020; Graff Zivin et al., Reference Graff Zivin, Song, Tang and Zhang2020; Roach and Whitney, Reference Roach and Whitney2022) in Korea, the United States and China and have negligible effects on test scores in Brazil (Li and Patel, Reference Li and Patel2021). However, most of this research uses the learning outcomes from tests administered at school and in exam settings. The few studies that have examined the effect of high-temperature exposure on cognition in household settings have found adverse effects of test date heat on teenage and adult populations on math scores in China (Zhang et al., Reference Zhang, Chen and Zhang2024) and the United States (Graff Zivin et al., Reference Graff Zivin, Hsiang and Neidell2018). Also, using similar test measures of learning, Garg et al. (Reference Garg, Jagnani and Taraz2020) find that the previous year's temperature affects children's math and reading scores in India.

Nonetheless, some important questions remain unanswered: first, how generalizable are the findings on the effect of contemporaneous heat and long-run climate change on learning in the African context, where the climate is primarily tropical and where ongoing non-climate vulnerabilities to poverty exacerbate vulnerability to climate? For instance, World Bank Group (2023) estimates that 45 per centFootnote ³ of people have access to electricity in Uganda, compared to 100 per cent in China and the United States. Furthermore, the African context is peculiar because it is characterized by limited uptake and use of formal climate information services in adaptation and response to climate shocks. In addition to socioeconomic differences from the rest of the world, the climate in Africa is peculiar. Although it is the region with the least greenhouse gas emissions, it is predicted to suffer the most disruptions due to climate change (Fonjong et al., Reference Fonjong, Matose and Sonnenfeld2024). For example, In recent decades, the East African region has experienced a lot of variation in precipitation, including alternating seasons of droughts and floods. The study context, Uganda, has a tropical climate and experiences moderate temperature variations within a year; however, data from the EM-DAT shows that it experienced 15 floods and five droughts between 1990 and 2010. Moreover, research on other outcomes has shown that data from wealthier countries tend to underestimate climate impacts on poorer regions (Carleton et al., Reference Carleton, Jina, Delgado, Greenstone, Houser, Hsiang, Hultgren, Kopp, McCusker, Nath, Rising, Rode, Seo, Viaene, Yuan and Zhang2022).

Second, there is no evidence about how ex-ante coping strategies to boost children's learning mitigate the effect of exposure to climate shock. Third, the paper that the short-term analysis in this work most closely relates to, based on data from China (Zhang et al., Reference Zhang, Chen and Zhang2024), finds no effect on verbal scores measured by word recognition tests and excludes children younger than ten years old, thereby raising the question of how heat affects the cognition of the youngest children and other forms of literacy. This paper contributes to some of these gaps.

Likewise, a rich literature examines the effect of long-term shocks on children's human capital. For instance, variable rainfall (Cornwell and Inder, Reference Cornwell and Inder2015; Kien and My, Reference Kien and My2021), and floods in utero (Rosales-Rueda, Reference Rosales-Rueda2018) have been shown to have adverse effects on children's health and cognitive outcomes. Also, using a similar measure of cognition as in this paper, Shah and Steinberg (Reference Shah and Steinberg2017) finds that positive rainfall shock experienced in utero up to age two increases cognition as measured by test scores in English and mathematics. Nonetheless, the findings on long-term climate effects are mixed. Studies have found that long-term effects are relatively muted compared to short-term impacts (Graff Zivin et al., Reference Graff Zivin, Hsiang and Neidell2018; Garg et al., Reference Garg, Jagnani and Taraz2020). Moreover, this literature mostly comes from Asian countries and focuses on educational attainment (Maccini and Yang, Reference Maccini and Yang2009; Randell and Gray, Reference Randell and Gray2019; Le and Nguyen, Reference Le and Nguyen2023). In Kenya, Nübler et al. (Reference Nübler, Austrian, Maluccio and Pinchoff2021) examine the effect of childhood rainfall shocks on adolescent girls in a pastoralist setting and find that it decreased achievement in math and English. This research expands the coverage demographically by including primary-aged children, boys, and all regions in Uganda in the analysis. Additionally, this work is the first to examine the effect of childhood heat shocks and rainfall shocks on learning outcomes in an African setting.

2.2 Mechanisms

There are different mechanisms through which heat could affect children in the short and long term. In the short term, heat stress on the test date has been shown to have a cognitive impact on children. This could be due to the effect of heat on physiological factors, including arterial blood oxygen saturation level (Lan et al., Reference Lan, Tang, Wargocki, Wyon and Lian2022) and brain temperature (Yablonskiy et al., Reference Yablonskiy, Ackerman and Raichle2000), which is related to cognitive performance. In the long term, young children, particularly those below the age of five, may be affected by exposure to extreme weather shocks via physiological, socioeconomic, environmental, and parental pathways. In the case of prolonged exposure to high temperatures, children can suffer heat strokes, which are especially harmful to growth and development during critical periods of vulnerability, including gestation, when vital physiological systems are developing, and early childhood, when the immune and central nervous systems are developing (Bennett and Friel, Reference Bennett and Friel2014). Weather shocks could affect children's cognition via low or damaged crop yields in agricultural settings (Schlenker and Lobell, Reference Schlenker and Lobell2010), which could lead to changes in household income (Matyas and Silva, Reference Matyas and Silva2013), food availability and related undernourishment, which has been shown to affect later life outcomes (Maccini and Yang, Reference Maccini and Yang2009). This mechanism may especially be salient in Uganda, where over 80 per cent of the population is engaged in agricultural activities, and the Agricultural sector contributes up to a quarter of the country's aggregate output.

With regard to precipitation, in tropical African countries, high and heavy rainfall is associated with a rise in vector-borne diseases, such as dengue fever and malaria, which is associated with cognitive declines in some cases (Boivin, Reference Boivin2002; Carter et al., Reference Carter, Mung'ala-Odera, Neville, Murira, Mturi, Musumba and Newton2005).Footnote ⁴ Also, high and low rainfall extremes have been associated with a rise in diarrhoeal disease (Hashizume et al., Reference Hashizume, Armstrong, Hajat, Wagatsuma, Faruque, Hayashi and Sack2007). In addition to direct effects, children could be indirectly affected by extreme climate shocks through intergenerational effects because of their dependence on adults. For instance, mothers who themselves suffer any adverse impact of climate shock could transmit the effects genetically to their unborn children or through the quality of nurture they provide their young children. Although this paper does not examine mechanisms empirically, the mechanisms lead one to expect that the impacts of transient test date temperature, which primarily affects children physiologically, are likely to be different from exposure to adverse climate shocks in childhood, which could affect children via multiple pathways.

3.1 Learning outcomes data

The human capital and socioeconomic measures come from the child- and household-level data from the UWEZO East Africa citizen-led assessment, covering Uganda from 2010 to 2015. UWEZO is an adaptation of ASER, an education survey developed by Pratham in India that conducts learning assessments nationally. In Uganda, in rounds 2012 and 2015, the tests were also administered in four local languages, including Ateso, Leblango, Luganda, and Runyoro (ACER, 2014).Footnote ⁵

The data represents a cross-section of students aged 6 to 16 who were examined using a grade 2-level curriculum and home surveys. Grade 2-level curriculum is used as it coincides with the age at which children are expected to have mastered basic literacy and numeracy skills. For all the relevant years, enumerators visited all census districts in each county. They sampled 30 villages within each district using probability proportional to size, with 20 households sampled in every village using systematic random sampling. Then, each child within the target population in a household is surveyed and tested at the child's home, regardless of whether the child is enrolled in school or not.

Children are typically visited by test-takers recruited from close communities and tested in math, English, and local language literacy. A child is assigned a score on a test based on the number of questions on a test that they answer correctly. In math, the competencies tested are counting, numbers, addition, subtraction, multiplication, and division. In the English language, the levels are letter, word, paragraph, and story. Similarly, in the local language, the levels are syllables, words, paragraphs, and stories. Other child variables collected include age, gender, and type of school attended. In addition, household information collected has parents’ age and years of education completed, household size, number of children within a household, asset ownership, and gender of the household head.

The sample used in this paper includes children for whom reliable test dates and GPS information are available, i.e., for years 2010, 2011, 2012, and 2015.

3.2 Weather and climate data

Geographically, Uganda has the following spatial classifications in descending order: region, district, county, sub-county, parish, and village. In 2014, the National Population and Housing Census of Uganda revealed that the average population in a district was about 240,000. More than half of the sub-counties had a population of fewer than 25,000 persons, while parishes have about 4,000 people on average. Therefore, we use unique parish, sub-county, and district names to identify the longitude and latitude of children at the parish level. The GPS information for parishes was then used to obtain the daily maximum temperature, relative humidity, and rainfall data measured in degrees Celsius and millimeters, respectively, from CHIRTS and CHIRPS daily data from the Climate Hazard Centre (CHC) at UC Santa Barbara. The CHC data has a spatial scale of about 0.05$^\circ$x 0.05$^\circ$, representing about 5 km by 5 km.

3.3 Data overview

Table 1 shows an overview of the outcome variables and the child and household controls used in the analysis. The average child in the data is 10.6 years old and is in the $3^{rd}$ grade. Sixty-one per cent and 11 per cent of households have telephones and electric assets, respectively. The average child scores 4 in math and about 3 in English and local language tests, respectively, corresponding to the ability to add and read words.

Table 1.

Summary table

Notes: SD: standard deviation. Math, English, and local Language are based on tests administered by UWEZO to children at home. Mother's education is coded 0 “None,” 1 “Primary,” 2 “Secondary,” 3 “More than Secondary.”.

3.4 Construction of variables

The learning outcomes measured are children's performance in math, English, and local language literacy tests. Also, the long-run estimations look at the probability of a child being “ontrack” in school. Ontrack is defined as the difference between a child's age and grade being at most 6; one reason for using this definition is that it can be seen as a stock variable of human capital that shows how enrolment and advancement have evolved until a child's current grade. The short-run analysis of the effect of test date temperature uses the maximum daily temperature as the measure of temperature. The maximum daily temperature ranged from about 19 to 39$^\circ$C in the years in the sample.

Long-run temperature and rainfall shocks for a child are based on deviations from the annual long-run average temperature and rainfall over almost three decades from 1985 to 2015. A parish is designated as having a heat shock in a given year if the annual temperature is in the $80^{th}$ percentile of the long-run temperature in the parish. In contrast, a cool year is defined as having a yearly temperature in the $20^{th}$ percentile of the long-run temperature. Hence, the variable Temperature Shock at period $t$ is coded one (1) if the annual temperature is in the $80^{th}$ percentile of the distribution of the parish's long-run temperature and coded $-1$ if the annual temperature is in the $20^{th}$ percentile. Any other case is coded as zero. Climate shocks are defined in 6 periods, i.e., t=-1,0,1,2,3,4, where t=0 is the year a child was born, that is, the birth year. The birth year variable is constructed by subtracting the survey year from a child's age. Hence, the in utero period is the year preceding the birth year.

3.5 Empirical framework

3.5.1 Short-term effects of test date temperature

To empirically determine the effect of high temperatures on test scores, the analysis exploits the plausible exogeneity of test date temperature given that the tests are administered without prior notice of the exact day to the households. It uses a fixed effect regression to compare outcomes between children living in the same parish and tested in the same year but who take the tests on different days. Hence, it utilizes variations in children's test dates within a parish in a year. The primary treatment variable is test date temperature, while the outcomes are test results in math, English, and local language literacy for children aged 6–15, measured at the individual child level. This estimation strategy implicitly assumes that the daily maximum temperature of the test date indicates the temperature around the time the child took the test, which is plausible given that the tests are administered during the day. Realistically, households have priors about their location's climate (Ortiz-Bobea, Reference Ortiz-Bobea2021), and the average monthly maximum temperature is around 30$^\circ$C in Uganda (see figure 1 for the annual average maximum temperatures recorded for the locations in the dataset) year round.Footnote ⁶ However, the daily weather fluctuations via temperatures we measure will likely be unexpected. Hence, the regressions include parish-year fixed effects for the local region's climate and socio-political conditions in a given year, as well as the month in which a parish is tested in a given year.

Figure 1.

Average maximum daily temperatures by year.

The regressions also control for test date rainfall and relative humidity. The regression equation is:

(1)\begin{equation} Y_{ijt}=\sum_{k=1}^K \beta_k T_{ijt}^k +X_{ijt}\beta +\tau_{tj}+\epsilon_{ijt},\end{equation}

where $Y$ is an outcome for child $i$ in parish $j$ in survey year $t$, $T_{ijt}^k$ is the $k^{th}$ bin of daily maximum temperature. The test date temperature is binned into the following five bins: $<=28,\, (28,\,30],\,(30,\,32],\, (32,\,34],$ and $>34^\circ$C. By utilizing bins, the estimation departs from a linearity assumption of the effects of temperatures on human capital. Child characteristics, including the age, grade, and gender of the child, the age and gender of the household head, the mother's education, and household assets, are included in the vector $X$. At the same time, parish-year fixed effects are denoted as $\tau _{tj}$. The coefficient $\beta _k$ on the temperature bins $T_{ijt}^k$ can be interpreted as the effect of testing on a day with temperature within a specific bin compared to an omitted bin. In the regressions, the omitted bin is the (23,28]$^\circ$C bin, chosen because it includes the typical daily temperature in Uganda and the optimal temperature for work (López-Sánchez and Hancock, Reference López-Sánchez and Hancock2018).

To examine the heterogeneous effect of temperature on human capital, the temperature dummies interacted with variables $C_i$ representing the heterogeneous dimensions of the child or household, i.e., age, gender, household socioeconomic status, and participation in extracurricular instruction. In this case, the regression equation becomes

(2)\begin{equation} Y_{ijt}= \sum_{k=1}^K \beta_k T_{ijt}^k +\sum_{k=1}^K \beta_c T_{ijt}^k*C_i +X_{ijt}\beta +\tau_{tj}+\epsilon_{ijt}.\end{equation}

Here, one would interpret the coefficient on the temperature bins $\beta _k$ as the effect of exposure to high temperature without the individual characteristic $C_i$. In contrast, the interaction term $\beta _c$ coefficients will signify the differential impact of temperatures within a particular bin compared to days with maximum temperature in the (23,28]$^\circ$C category on those with that characteristic or asset compared to children without the characteristic.

One concern with identification is that high temperatures may affect the performance of the test administrators, who may misreport student performance in these tests. However, this measurement error source is unlikely since data collectors follow rigorous training, visit households in pairs, and are supervised by a field supervisor who validates data entries periodically.

3.5.2 Long-term effects of childhood rainfall and temperature

To examine whether climate shocks in early childhood have long-term effects, we determine whether a child experienced abnormal rainfall or heat in early childhood, from the in utero period until age 4. Thus, we use lagged climate shocks to assess the effect of early-life rainfall and heat on current human capital outcomes. A climate shock in each of those years is defined as experiencing a childhood period in a year with average annual rainfall in the left ($20^{th}$) and right ($80^{th}$ percentile) tails of the parish's long-run rain (calculated over about three decades) and temperature distribution. Following Shah and Steinberg (Reference Shah and Steinberg2017), we denote a positive rain shock as $1$ if rainfall is in the $80^{th}$ percentile and above, as $-1$ if rainfall is in the $20^{th}$ percentile, while any other case is coded as 0. Temperature shocks are defined similarly. We denote a heat shock $1$ if the average annual temperature is in the $80^{th}$ percentile and above, as $-1$ if the temperature is in the $20^{th}$ percentile, while any other case is coded as 0.

The results of experiencing temperature and rainfall shocks in early childhood are presented separately, given that studies have demonstrated their different impact on outcomes in the African context (Burke et al., Reference Burke, Miguel, Satyanath and Lobell2009).Footnote ⁷ We check if experiencing temperature shocks in each period is correlated with experiencing a rainfall shock and find that, on average, there is a weak negative correlation ($-0.3$) (see table A10 in the online appendix) between experiencing positive rain shock and high heat in the same year. Also, we do not find statistical evidence that experiencing the same type of shock in a year strongly correlates with experiencing the same shock the following year. The effects of long-term climate shocks are estimated using the following equation:

(3)\begin{equation} Y_{ijyt}=\beta_l\theta_{jk} + X_{ijyt}\beta+\tau_t+\delta_y+ \eta_j+\epsilon_{ijyt},\end{equation}

where $Y$ is an outcome for child $i$ in parish $j$ , born in year $y$ and surveyed in year $t$, $\theta _{jk}$ are rainfall or temperature shocks from in utero to age $4$ ($k=-1,\,0,\,1,\,2,\,3,\,4$). All variables remain as defined in equation (1) except for $\delta _y$, which represents birth year fixed effects. Our primary variable of interest, $\beta _l$, shows the impact of rainfall shock relative to a typical year, and $2\beta _l$ is the impact of experiencing a rainfall shock relative to a drought year. In the case of temperature shocks, $\beta _l$ shows the effect of experiencing a heat shock relative to a normal year, and $2\beta _l$ is the impact of experiencing a high heat relative to a cool year. The standard errors are clustered at the parish level. Parish fixed effects account for a parish's average rainfall and unchanging geographic characteristics. In contrast, birth year fixed effects control for average rainfall experienced by all the children born in a specific year in Uganda. Thus, identification relies on within-parish variation in rainfall from the long-run rainfall in each parish. Further, as a robustness check, we estimate a version of the regressions using household fixed effects instead of parish fixed effects. Regarding household fixed effects, identification came from households with children born in different years or birth cohorts.