Agricultural droughts and under-five mortality in Côte d’Ivoire: Differential impacts across social and demographic groups

The individual-level data come from the CDHS rounds of 1994, 1998–1999, and 2011–2012. The DHS Program collects nationally representative data on population health, child mortality, and socioeconomic characteristics. Importantly for this study, the CDHS reports the GPS latitude and longitude coordinates of clusters where women and their children reside, which enables linking them to their geographic location at survey time.² I use their geographic location at survey time as a proxy for their geographic location during gestation. Using the latitude and longitude data in the DHS, I match each child to the weather grid cell corresponding to their location to construct variables of in-utero exposure to agricultural droughts.

The analytical sample includes children born in the 5 years prior to the survey, covering 13,545 children born to 10,275 mothers living in over 727 geographic locations (see Fig. 1 for a map of the geographic locations).

The climate data used to calculate the variable for agricultural droughts, as measured by SPEI, are taken from the high-quality ERA-Interim reanalysis dataset (Dee et al., 2011). The ERA-Interim dataset contains information on climate for every 6 hours from 1 January 1979, on a global grid of parallels and meridians at a 0.75 × 0.75° resolution (about 80 km by 80 km at the equator).³ This dataset has been previously used in studies exploring the effects of extreme weather on life-course transitions, infant mortality, and conflict incidence (Andriano & Behrman, 2020; Flatø & Kotsadam, 2014; Harari & La Ferrara, 2018).

In order to calculate SPEI, I use data up to April 2012⁴ on monthly mean daily net solar radiation, monthly mean daily minimum and maximum temperatures, monthly mean daily wind speeds, monthly mean daily dewpoint temperature, elevation above sea level, latitude, and monthly total precipitation. I elaborate on how these weather parameters are used to calculate SPEI in the methods section below.

The advantages of the ERA-Interim data are that (i) they are reanalysed using various sources, such as satellites, radiosondes, pilot balloons, aircraft, wind profilers, ships, drifting buoys, and land stations,⁵ and (ii) they have less bias in rainfall outcomes in the arid and semi-arid areas of Africa compared to gauge data (Zhang et al., 2013).

In this study, I introduce a novel measure of in-utero exposure to agricultural droughts, focusing primarily on its impact on maternal malnutrition during pregnancy. It is important to note that this measure does not isolate maternal malnutrition as the sole pathway through which agricultural droughts may affect early childhood mortality. For example, agricultural droughts could lead to maternal malnutrition which may impair a newborn’s growth. Alternatively, this impaired growth itself can increase the newborn’s susceptibility to infections, further impairing growth. This measure still offers an understanding of the broader effects of agricultural droughts on early childhood mortality.

To measure agricultural drought conditions, I use the SPEI. SPEI is a standardised, multiscalar index measuring drought severity, intensity, and duration. Unlike precipitation-based drought indices, it allows for comparisons of drought severity across space and time since it can be calculated over a wide range of climates (Vicente-Serrano et al., 2010a, b). The SPEI also has the advantage of being multiscalar, meaning it can be computed at different time scales over which water deficits accumulate to reflect different drought types. Short time scales refer to soil water content and river discharge, and long time scales relate to changes in groundwater storage (Vicente-Serrano et al., 2010a, b).

For this analysis, I use the SPEI at a 3-month time scale (SPEI3) because it reflects short- and medium-term moisture conditions, providing a seasonal estimation of climate relevant for agriculture (WMO, 2012). Agricultural droughts are associated with a shortage of water for plant growth and are characterised by insufficient soil moisture to replace evapotranspirative losses (WMO, 1975), which negatively affects crop production (Boken, 2005; Sivakumar et al., 2010; Wu et al., 2011). Several indices have been developed to measure agricultural droughts, but the SPEI has been identified as the most complete and robust agricultural drought index in Africa (WMO, 2012) (see Sivakumar et al. (2010) for a review of all agricultural drought indices).

SPEI is expressed in units of standard deviation from the cell average and has mean 0 by construction, where negative values below −1 indicate dry conditions and positive values above +1 indicate wet conditions. Dry conditions can be categorised into moderately dry \(\left(-1 \le \text{SPEI}<-1.5\right)\), severely dry \(\left(-1.5 \le \text{ SPEI}<-2\right)\), and extremely dry \(\text{SPEI} \le -2\). Similarly, wet conditions are categorised into moderately wet \(1 \le \text{ SPEI}<1.5\), very wet \(\left(1.5 \le \text{ SPEI}<2\right)\), and extremely wet \(\left(\text{SPEI }\ge 2\right)\) (WMO, 2012). In additional analyses, I leverage these pre-established categorisations of SPEI to explore the effects of severe and extreme droughts (\(\text{SPEI} \le -1.5\)), which I refer to as severe droughts in the remainder of the paper.

To create the measure of in-utero exposure, I follow three steps. First, I use the monthly weather data to compute the potential evapotranspiration (PET), representing the amount of water that could be evaporated and transpired if a sufficient water source were available. This calculation is performed for each ERA-I cell and month up to April 2012. Second, I use the monthly PET and monthly total precipitation to derive the SPEI3 for each ERA-I cell and month up to April 2012.⁶ (Additional information on all the weather parameters and calculations used for the derivation of PET and SPEI is available in the Supplementary Material.) In the third step, I calculate the average SPEI3 value during the 12 months preceding the child’s date of birth. This measure based on SPEI3 likely captures the weather conditions affecting maternal nutrition during pregnancy.

Figure 2 shows the SPEI3 values across Côte d’Ivoire over time for the period of analysis between 1992 and 2012. The figure shows areas where SPEI3 values are negative (yellow to red squares) or positive (light blue to dark blue). The patterns observed in Fig. 2 are compatible with the historical records of drought in Côte d’Ivoire. For example, the data used for this analysis show that the country was hit by drought in 2007, an event caused by the delayed onset of the wet season (NASA, 2007).

A fixed-effects linear probability model is used to investigate the effect of in-utero exposure to SPEI on under-five mortality. The model specification reads as follows:where \({p}_{imygs}\) is a binary variable for under-five mortality of child \(i\) born in month \(m\) and year \(y\) residing in ERA-I cell \(g\) from survey year \(s\); \({\text{spei}}_{myg}\) is the average SPEI3 value during the 12 months before the child’s date of birth.⁷ The vector of child, parental, and community characteristics, \({{\varvec{x}}}_{i}\), includes child’s sex, child’s birth order (first birth; 2–4 births; 5 + births), maternal education attainment (no formal education; primary education; secondary and higher education), paternal education attainment (no formal education; primary education; secondary and higher education; missing/don’t know; the ‘missing/don’t know’ category implicitly describes a woman’s marital status because the question about paternal education was not asked to women who were not in a union at the time of the survey), mother’s age at birth, mother’s age at birth squared, household wealth (wealthy or poor),⁸ and place of residence (rural or urban).

The vectors of month-of-birth fixed effects, year-of-birth fixed effects, and ERA-I cell fixed effects are denoted as \({\boldsymbol{\alpha }}_{m}\), \({\boldsymbol{\alpha }}_{y}\), and \({\boldsymbol{\alpha }}_{g}\), respectively. The use of these fixed effects allows to account for ERA-I cell-specific, temporal, and seasonal trends, enabling a causal interpretation of estimates. The survey-year fixed effects, \({\boldsymbol{\alpha }}_{s}\), capture potential effects from pooling data across surveys, controlling for any unobserved differences between survey years which may bias the results such as survey design, implementation, and population characteristics. I also use survey-year weights, which adjust the data to reflect how many people were interviewed as a proportion of the population in that year.⁹ This weighting ensures results are representative of the individuals living in Côte d’Ivoire over time (ICF International, 2012).

To investigate whether key demographic and socioeconomic factors can mitigate the effect of in-utero exposure to agricultural droughts on under-five mortality, I estimate a new set of fixed-effects linear probability models. These models include an interaction term between the variable of severe agricultural droughts and a categorical variable which captures the key demographic and socioeconomic factors. Population subgroups are distinguished by the child’s sex, the mother’s level of education, the father’s level of education, the household wealth, and the place of residence. I also combine parents’ level of education with household wealth and place of residence into two categorical variables to pinpoint potential population subgroups most vulnerable to climatic shocks such as the lower educated and poor (Dimitrova & Muttarak, 2020). The first categorical variable has the following categories: ‘no formal education and poor’, including children whose parents have no formal education and whose households are in the lowest two wealth quintiles; ‘formal education and poor’, including children whose parents have at least primary education and whose households are in the lowest two wealth quintiles; ‘no formal education and wealthy’, including children whose parents have no formal education and whose households are in the highest three wealth quintiles; and ‘formal education and wealthy’, including children whose parents have at least primary education and whose households are in the highest three wealth quintiles. The second categorical variable has the following categories: ‘no formal education and rural’; ‘formal education and rural’; ‘no formal education and urban’; and ‘formal education and urban’.

Table 1 presents the summary statistics of the variables used in this analysis for the overall sample and by survey. The average under-five mortality rate in the sample is 10.4%, increasing from 8.6% in 1994 to 14.2% in 1998–1999 and then decreasing to 8.2% in 2011–2012, a trend in line with other national estimates (Garenne & Gakusi, 2005). Children in the sample are evenly distributed by sex and born to mothers aged 26 years on average. The percentage of children varies by birth order, with 22.4% being firstborn, 46.9% being of birth order 2–4, and 30.7% being of birth order 5 or higher. 64.7% of the children live in rural areas, and 45% are born in poor households. A large percentage of children have parents with no formal education. Notably, more than 19% of children were exposed to severe agricultural droughts during gestation.¹⁰ This relatively high prevalence of drought exposure highlights the importance of investigating potential health outcomes associated with severe agricultural droughts.

Table 2 shows the results of the model described in Eq. (1).¹¹ Due to significantly higher under-five mortality in the 1998–1999 survey, primarily attributable to HIV/AIDS (Garenne & Gakusi, 2005), and the relatively low prevalence of drought exposure among children in the survey, I have excluded children from the 1998–1999 survey from the main analysis. This exclusion is intended to enhance the efficiency of the estimator. It is worth noting that the results remain consistent even when children from the 1998–1999 survey are included (Table 1 in the Supplementary Material).

Findings suggest that SPEI3 is negatively linked to under-five mortality (Model 1). A one-unit decrease in SPEI3 is associated with a 3.4 percentage point increase in the probability that a child dies before the age of five, an effect that is significant at the 1% level and that corresponds to a 41% increase in the under-five mortality rate given an average mortality rate of 8.4%.

In the next analysis I investigate whether more recent in-utero exposure to agricultural dry conditions should have larger or smaller effects on under-five mortality than less recent one. This analysis helps disentangle the stage in pregnancy in which drought influences under-five mortality. In particular, a less recent drought would affect intrauterine growth around conception and in the first period of pregnancy (i.e. embryogenesis and initial foetal development), as opposed to a more recent drought that would influence intrauterine growth during the final foetal period.

I therefore calculate SPEI3 in the 12-to-24-month period before birth to measure less recent in-utero exposure to agricultural dry conditions and add it to the model. The coefficients presented in Model 2, Table 2, indicate that SPEI3 is negatively linked to under-five mortality; however, this negative relationship is statistically significant for more recent exposure, while the coefficient for less recent exposure is also negative but not statistically significant. This finding suggests that exposure to agricultural dry conditions during the final foetal period is detrimental to the survival chances of children. The most plausible explanation for this finding is that we are capturing a nutrition effect given that maternal nutritional shocks primarily impact foetal growth during the later stages of gestation (Darling, 2017; Lawton et al., 1988; Lumey, 1998; Painter et al., 2005; Strauss & Dietz, 1999).

I next investigate potential non-linear relationships between SPEI3 and under-five mortality. For this analysis, I include the quadratic polynomial of the SPEI3 variable in the regression model (Model 3, Table 2). Figure 3 shows the predicted values of under-five mortality at different levels of SPEI3 derived from this model.

The results indicate children exposed to droughts in-utero are considerably more likely to die. As SPEI3 decreases from 0 to − 1.5, the probability that a child dies before the age of five rises from 7.4 to 11.5%, ceteris paribus. Importantly, the data also shows that the risk of mortality decreases at positive values of SPEI (≥ 2), but the estimates are imprecise, as indicated by wide confidence intervals.¹² Such extreme SPEI values indicate exposure to very wet conditions. While climatic shocks often have negative consequences for local livelihoods, there are also instances when climatic shocks could be beneficial for livelihoods. A study in Senegal found that positive rainfall shocks reduced infant mortality, suggesting a positive income shocks and food supply from a plentiful harvest (Pitt & Sigle, 1998).

I next investigate the impact of in-utero exposure to droughts on under-five mortality. I classify a child as being exposed to ‘severe agricultural droughts’ in-utero if the average SPEI3 value during the 12 months before the child’s date of birth is ≤ − 1.5, and as ‘non-droughts’ otherwise.

The results shown in Table 3 indicate that in-utero exposure to severe agricultural droughts significantly increases under-five mortality (Model 1). Specifically, being exposed to severe agricultural droughts is associated with a 2.3 percentage point increase in the probability that a child dies before the age of five, which corresponds to a 28% increase in the under-five mortality rate given an average mortality rate of 8.4%. Again, I only find evidence of a relationship between more recent droughts and under-five mortality (Model 2), and no significant relationship is found between exposure to very wet conditions and under-five mortality (Models 1 and 2).

In the next analysis, I examine the impact of in-utero exposure to severe agricultural droughts on infant mortality to understand whether previous studies may have underestimated the impact of droughts on early childhood mortality (Table 4). My findings indicate that droughts are positively linked to infant mortality, but the effect is not statistically significant (Model 1). I then explore the impact on child mortality, defined as death between 12 and 59 months old, and find that the impact is positive and significant (Model 2). These findings suggest that the detrimental effects of in-utero drought exposure may manifest more significantly as children grow older, perhaps due to impaired growth in-utero that increases their susceptibility to infections and diseases as they grow.

I now provide a detailed discussion about the several potential channels that may bias this effect. One potential source of bias in the estimate is selection in-utero. If a weaker foetus is more likely to die in-utero due to drought exposure, then the stronger foetuses are more likely to survive. Previous research finds that maternal exposure to climate extremes increases pregnancy loss (Davenport et al., 2020). It is worth noting that this will likely underestimate the true effect of in-utero exposure to agricultural droughts on early childhood mortality.

Another source of bias is migration after conception. Because the DHS data contain women’s geographic location at survey time, it is difficult to determine that women and children are exposed to the weather conditions in the same area (i.e. ERA-I grid cell) in the 12 months before birth. The direction of the bias is unclear. For example, if a woman moves to an area with more favourable climatic conditions within the year preceding her child’s birth and the time of the survey, the estimate may potentially underestimate the true impact of in-utero exposure to agricultural drought on early childhood mortality.

Unfortunately, the CDHS does not contain information on women’s migration history.¹³ The country experienced significant displacement as a result of the political instability between 1999 and 2011. The conflict, however, varied widely across the country, and some regions were more affected by the conflict than others. I use spatial data on conflict available from xSub (Zhukov et al., 2019), a repository of microlevel, subnational event data on armed conflict and contention to be able to identify the areas that were most affected by conflict in the 2000s. Figure 4 shows that the western, eastern, southern, and to a lesser extent the central parts of the country were most affected. I therefore estimate the same model excluding children living in areas that were highly affected by the conflict, that is, all children from the 2012 DHS living in 19 regions. Excluding children living in the highly affected areas from the sample does not change the sign and significance level of the coefficient (Model 3, Table 3). If anything, the magnitude of the coefficient increases, suggesting a potential underestimation of the true effect of in-utero exposure to agricultural drought on under-five mortality.

In the following analysis, the empirical model is extended by including an interaction term between the droughts variable and key demographic and socioeconomic factors. Above, I identified that droughts in the 12 months prior to birth are critical to child survival. I therefore focus on droughts experienced during this timeframe.

Figure 5 shows the results of the models with an interaction term. First, gender differences in climate-related vulnerability can be observed: droughts are associated with an increased risk of mortality for boys, while marginally significant effects for girls are found (Model 1). This disparity can be attributed to male vulnerability in early life (DiPietro & Voegtline, 2017), particularly related to nutritional factors. Male foetuses generally have higher growth rates, which result in greater nutritional demands. During droughts, the reduced availability of nutrients can disproportionately affect male foetuses, making them more susceptible to health issues later in childhood.

Second, the results show that children born to less educated mothers, regardless of the family’s wealth status, are particularly susceptible to mortality due to drought exposure (Model 6). Interestingly, there is no evidence that higher wealth status provides the same level of protection as maternal education (Models 4 and 6). The negative consequences of in-utero exposure to severe agricultural droughts are mainly concentrated among children whose mothers have no formal education.

One plausible explanation for the finding that maternal education mitigates the effects of in-utero exposure to droughts on under-five mortality regardless of the family’s wealth status is that highly educated mothers typically reside in urban areas. In urban areas, they may benefit from less direct dependence on agriculture and better access to diverse nutritional options, which can improve their ability to manage risks associated with droughts during pregnancy. They may also have greater access to information and support services, important for maintaining health during drought conditions. However, this is not what my findings suggest. I find no evidence that urban living shields children from drought-related mortality (Models 5 and 8). The negative impact of agricultural drought on under-five mortality is statistically insignificant for children of highly educated mothers, regardless of whether they live in rural or urban areas, yet remains significant for children of less educated mothers in both settings (Model 8). These findings suggest that the protective effects of maternal education against the impacts of droughts on under-five mortality are not dependent on the context in which mothers live. They could instead be linked to the specific knowledge, behaviours, and practices that highly educated mothers employ.¹⁴

Father’s education seems to protect against drought-related mortality in a similar way as mother’s education. As expected, children born to less educated and poor fathers are the most vulnerable (Model 7). Interestingly, the protective effect of paternal education is only evident in urban settings, where the impact of in-utero exposure to droughts on under-five mortality is statistically insignificant (Model 9). Unlike for mothers, the benefits of education for fathers are more context dependent. In urban settings, access to diverse and stable job markets that are less affected by environmental shocks may amplify the benefits of paternal education, allowing fathers to better support their families.