Infrastructure mitigates the sensitivity of child growth to local agriculture and rainfall in Nepal and Uganda

Significance

Incorporating agriculture into nutrition policy requires an understanding of how agricultural performance, rainfall, and the economic and physical environments in which children reside relate to linear growth and weight gain. This paper combines anthropometric data from children below the age of 5 y in Nepal and Uganda with rainfall data and other information to measure these connections. Anthropometric outcomes are positively correlated with rainfall prior to birth, during the first year, and during agricultural growing seasons preceding child measurement. High rainfall is found to be deleterious to child growth in some settings. Evidence points to the need for agricultural adaptation to low rainfall, as well as broadly based economic development, including continued investments in health and transport infrastructure, to help improve child nutrition.

Abstract

This paper investigates linear growth and weight gain among 11,946 children below the age of 5 y in Nepal and Uganda, testing the hypothesis that child growth is sensitive to precipitation during key periods in a child's early life. The paper also tests the importance of the economic and physical environments in which children reside. Outcomes are not completely explained by agricultural performance or the observed characteristics of children or their households. Associations between height-for-age z-score (HAZ) and weight-for-height z-score (WHZ) and rainfall are generally positive, but patterns are heterogeneous. At the mean, an increase of 1 SD in agricultural season rainfall is associated with a 0.05- to 0.25-point higher z-score, which translates into increases of roughly 4–13% for HAZ and 1–7% for WHZ. Nutrition sensitivity to rainfall is greater in Nepal, where rainfall is lower on average and wider ranging, than in Uganda. Health and transport infrastructure help to buffer children from the deleterious nutritional effects of precipitation shortfalls, underscoring the role of broadly based economic development in promoting child nutrition.

Persistent malnutrition among young children severely hinders physical and cognitive development (1) and increases the risk and severity of illness, further deepening malnutrition (2). In recognition of the overall importance of child nutrition to a country’s development prospects, finding ways to improve nutrition has been labeled a “quintessential sustainable development goal” (3). Evidence-based policy making in pursuit of this goal requires a better understanding of the mechanisms leading to nutrition improvements. Isolating and clearly identifying causal pathways is an empirical challenge, however, because many potential determinants of child health and nutrition are hidden or covary. In this paper, I confront this challenge by combining a wide range of observed data, including precipitation data, to study patterns of linear growth [height-for-age z-score (HAZ)] and wasting [weight-for-height z-score (WHZ)] among children below the age of 5 y (U5s) in Nepal and Uganda. I measure the magnitude and strength of correlations between anthropometric measurements and precipitation at key stages of early life, taking particular care to distinguish clearly between variation in precipitation within locations and variations between locations. I also control for agricultural performance and market and health infrastructure, and test whether these factors are correlated with growth. I find that greater road density and improved access to health facilities mitigate a child’s sensitivity to variations in precipitation.

Nepal and Uganda are emblematic of nutrition shortcomings in South Asia and Sub-Saharan Africa. Demographic and Health Survey (DHS) data show that a very large proportion of U5s suffer from malnutrition: 41% of Nepalese U5s were stunted in 2011 (HAZ < −2.0) and 11% were wasted (WHZ < −2.0). In the same year, 33% of Ugandan U5s were stunted and 5% were wasted. Nepal and Uganda face familiar development challenges, including chronic and widespread food insecurity, poor infrastructure, and weak governance. Nepal’s most recent Global Hunger Index (GHI) score is 22.2, placing it 58th of 104 countries ranked in 2015 (4). Population growth, low per-capita investments in agriculture, and food grain deficits contribute to food insecurity (5). Even before the 2015 earthquakes, more than 3.5 million people in Nepal were food-insecure. Many reside in geographically isolated areas dependent on food aid (6).

Uganda is also among the least well-nourished countries in the world. In 2005, a comprehensive study showed that 6% of Ugandan households were food-insecure and 21% were at risk for food insecurity (7). Uganda’s GHI score is 27.6 (75th of 104 countries) (4). Malnutrition accounts for 40% of child deaths (8), and childhood anemia prevalence exceeds 60% (9). Uganda’s food situation has been complicated by influxes of refugees and asylum seekers, many of whom cannot produce or purchase food. Although Uganda as a whole does not lack food, the typical Ugandan diet lacks diversity and fails to provide sufficient micronutrients (10). Seasonal food insecurity is also a problem (11).

Food entitlements have long been recognized as important to nutrition and health (12, 13). The 2008 and 2013 Lancet maternal and child nutrition reviews underscore the additional importance of microlevel factors for child growth, including breastfeeding and intakes of vitamin A, zinc, iron, and folic acid (1, 14). Access to safe water, sanitation, and health information and facilities are key (15), especially because immature gut biota and subclinical enteropathy may be causal factors in undernutrition (16, 17). Poor indoor air quality and maternal tobacco smoking are risk factors in child undernutrition, morbidity, and mortality (18–20). Aflatoxin exposure also contributes to impaired growth (21). At broad scales, war, civil unrest, and drought undermine child nutrition, with consequences for educational attainment (22).

Agricultural capacity, diversity, and performance may be especially important to nutrition where households pursue subsistence strategies or are poorly integrated into markets (23). Associations between nutrition outcomes and both crop diversity and specific farm practices, such as production of animal protein or fruits and vegetables, are positive (24, 25), but because production is highly sensitive to weather and the climate characteristics of locations, nutrition outcomes may track environmental variability (26, 27). For example, strong mortality effects from season of birth have been found in The Gambia (28), India (29), Indonesia (30, 31), Mexico (32), Nigeria (33), and Vietnam (34). Problematically, the mechanisms by which weather variation affects a population may be masked by beliefs that drive individuals to adapt and adjust to environmental change, as well as temporal and spatial delays and displacements in effects (35). In addition, environmental conditions potentially influence nutrition via multiple indirect pathways, including food prices (36–39) and market access (40, 41). Because seasonality and climate anomalies can affect a range of interrelated factors, among which are crop performance, prices, labor markets, and disease vectors, empirical analysis must control for time- and location-specific covariates, and also isolate within-location weather variability from more substantial between-location variability. When used alone, standard anthropomorphic datasets are insufficient for this task. Here, I combine data from a range of sources to reduce potential omitted variable bias and to measure the sensitivity of linear growth and weight gain to precipitation during key periods, conditional on birth timing, location, and local situation.

Materials and Methods

Smith and Haddad (42) proposed a model of child nutrition in which growth is driven by fixed and random effects observed at three levels, arranged hierarchically: (i) immediate determinants (food intake and health status), (ii) underlying determinants (household food security and health access), and (iii) basic determinants (physical, cultural, and institutional features of communities and nations). Accordingly, I connect two standard growth indicators for U5s to a range of factors aligned along the health and nutrition pathways at different levels. HAZ reflects the impacts of early life events and the accumulation of health and nutrition inputs over time; WHZ is more sensitive to recent consumption and health (43, 44). These z-scores measure dispersions in growth and are calculated as $Z_i=\left(x_i-\overline{x}\right)/{\sigma }_x$, where $x_i$ is the individual observation and $\overline{x}$ and ${\sigma }_x$ are the median and SD of the WHO’s reference population (45).

Nationally representative HAZ and WHZ data come from the 2006 and 2011 DHSs administered in Nepal and Uganda. These DHSs provide 11,946 observations on U5s (7,572 in Nepal, 4,374 in Uganda). The DHS includes detailed information about children, their mothers, and their households. Basic determinants potentially include local weather anomalies and climate characteristics of locations, and must be accounted for to obtain reliable estimates of other relationships. I therefore supplement the DHS with data from other sources, matching information using latitude, longitude, and calendar dates, or, where such matching is not possible or appropriate, by date and district or DHS cluster. These data include historic and contemporaneous monthly precipitation and temperature (Figs. S1–S3), variables from World Bank Living Standard Measurement Surveys, and other public data. The data, their sources, and matching procedures are described in SI Materials and Methods. No institutional review board approval was required; anonymized personal data were previously collected with informed consent.

Fig. S1.

Nepal data points overlaid with a temperature grid. Data are from Nepal 2006 and 2011 DHSs (n = 7,572). Temperature data are from the Berkeley Earth Project. Details are provided in SI Materials and Methods.

Fig. S3.

Local rainfall and temperature anomalies, by month, in Nepal. DHS data correspond to birth year for children, which was measured in 2006 and 2011 (n = 7,572). Rainfall was standardized using local mean and SD over the period 1998–2012. Rainfall anomalies of ±5 units (n = 93) have been removed from the figure. Solid symbols (red) indicate 2006 DHS data (birth years 2000–2006; n = 5,165); hollow symbols (green) indicate 2011 data (birth years 2005–2011; n = 2,314).

I estimate linear regressions and multilevel (mixed-effects) regressions of the form:

$$Z_i={\beta }_{0j}+{\beta }_{1j}{X}_{ij}+e_{ij},$$

$${\beta }_{0j}={\gamma }_{00}+{\gamma }_{01}{W}_j+u_{0j},$$

$${\beta }_{1j}={\gamma }_{10}+{\gamma }_{11}{W}_j+u_{1j},$$

where Z_i is the HAZ or WHZ of the ith child (the level-1 unit), X denotes variables observed at the child level, and W represents variables to account for higher level variance. The j subscripts indicate that separate level-1 regressions are estimated for each level-j unit. The expanded variance terms allow one to account for variance arising at multiple levels. For empirical estimation, I extend the nesting to include households, districts, and regions.

SI Materials and Methods

Data come from published sources, including official government documents and reports. The data and programs used in this paper have been archived in the Interuniversity Consortium for Political and Social Research (ICPSR) data repository (doi.org/10.3886/E100387V1).

The Nepal and Uganda DHSs.

The 2006 and 2011 Nepal and Uganda DHSs (DHS datasets; www.dhsprogram.com/) provide the primary data on child growth and child and household characteristics. In Nepal, the DHS data were collected by trained enumerators under the supervision of the Ministry of Health and Population (MOHP). In Uganda, the DHSs were conducted by the Uganda Bureau of Statistics (UBOS). In both countries, the surveys were designed as comprehensive and nationally representative household surveys focusing on men aged 15–59 y, women aged 15–59 y, and U5s. DHS samples are selected using a stratified two-stage cluster design. DHS data provide key health measurements and indicators across all ecological zones and development regions. A total of 5,237 U5s are included from the 2006 Nepal survey, 2,335 are included from the 2011 Nepal survey, 2,370 are included from the 2006 Uganda survey, and 2,004 are included from the 2011 Uganda survey, for a total sample size of 11,946. The DHS data are georeferenced, but to protect the confidentiality of DHS respondents, data from the same enumeration area are aggregated to a single point coordinate and the coordinate is then masked using a GPS coordinate displacement process. Urban clusters are displaced up to 2 km, and rural clusters are displaced up to 6 km, with a randomly selected 1% of rural clusters displaced up to 10 km (55).

The DHSs do not define urban/rural locations but, instead, adopt the country-specific designations made by national governments and their national statistical agencies. For this reason, the way that the “urban” indicator is assigned to households in the DHS varies somewhat across countries. In some cases, the designation is population-based, and it is based on political or other criteria in some cases. The International Labor Office maintains a database (56) describing how various countries define urban areas. For Nepal, the designation is made by the Ministry of Local Development based on administrative and legal considerations. For Uganda, the UBOS defines urban areas as gazetted cities, municipalities, and towns with populations over 2,000 persons.

For Uganda, the 2006 DHS includes 205 children from households with members designated as internally displaced people (IDP). These households were living in United Nations refugee camps that had been established to protect people displaced by fighting between the rebel Lord’s Resistance Army and the Ugandan army. According to the United National High Commission on Refugees (UNHCR) (57), at the conflict’s peak in 2005, there were 1.84 million IDP living in 251 camps across 11 districts of northern Uganda. The UNHCR activities ceased at the end of 2011, but hostilities had already been reduced starting in 2006 and many households had returned home by the time of the 2011 DHS. For 2011, there were no IDP households represented in the DHS. In total, IDP children in the sample represent 4.7% of all observations. These children came from the districts of Apac (n = 13), Gulu (n = 66), Kitgum (n = 44), Lira (n = 35), and Pader (n = 47). As a group, IDP children in the DHS sample have average WHZs and average HAZs that are not significantly different from the average WHZs and HAZs of non-IDP children (WHZ: 0.03 vs. −0.04, t = 0.84; HAZ: −1.50 vs. −1.47, t = 0.26).

Nepal Living Standards Survey and Uganda National Household Survey.

The Nepal Living Standards Survey (NLSS) and Uganda National Household Survey (UNHS) data come from World Bank Living Standard Measurement Surveys (LSMS datasets; www.worldbank.org/en/research). NLSS data are derived from two nationally representative household surveys conducted in 2004 and 2010. These NLSS surveys were conducted by the Central Bureau of Statistics, Nepal. UNHS data come from surveys conducted by the UBOS in 2005 and 2010. All surveys followed the methodology of the World Bank’s Living Standard Measurement Survey using a two-stage stratified random sampling technique. These surveys asked questions related to agriculture, food consumption, and expenditure; farm and off-farm income; migration; labor; access to facilities and market infrastructures; and other measures at individual and household levels. Agriculture-related variables were extracted from the surveys and summarized at the district level (n = 75 for Nepal, n = 56 for Uganda). These variables were merged onto corresponding DHS observations based on year and district (2004 and 2010 NLSSs were matched to 2006 and 2011 Nepal DHSs, respectively; 2005 and 2010 UNHSs were matched to 2006 and 2011 Uganda DHSs, respectively). These added district-level variables provide information on the agricultural and economic environment in which households operate.

Rainfall and Temperature.

For Nepal, monthly rainfall data (in millimeters) from January 2002 to December 2010 were obtained from the Department of Hydrology and Meteorology, Nepal. These data cover 280 meteorological stations and all districts of Nepal. Rainfall data (in millimeters) for each survey cluster sampled in the Uganda DHSs in 2006 and 2011 come from the Climate Hazards Group InfraRed Precipitation with Station (CHIRPS) data (CHIRPS datasets; chg-wiki.geog.ucsb.edu/wiki/CHIRPS_FAQ). CHIRPS was developed by the University of California, Santa Barbara to aid in understanding and monitoring of droughts, and it provides a global rainfall dataset from 1981 to near the present (58). Rainfall data provided by CHIRPS are monthly gridded, satellite-based precipitation estimates from the National Aeronautics and Space Administration and National Oceanic and Atmospheric Administration produced at a spatial resolution of 0.05° × 0.05°. For Uganda, each child in a cluster is matched with cluster-specific rainfall data. For Nepal, standardization for each month and location is based on mean rainfall (in millimeters) for each month and location is observed over 15 y (1998–2012). For Uganda, it is based on mean rainfall (in millimeters) for each month and location observed over 34 y (1981–2014). The standardized monthly rainfall anomalies are computed as $a_{it}=\left(x_{it}-{\overline{x}}_{it}\right)/{\sigma }_{it}$, where $a_{it}$ is the rainfall anomaly for location i in month t, $x_{it}$ is the observed rainfall at location i in month t, ${\overline{x}}_{it}$ is the monthly mean rainfall for location i in month t computed across all years observed, and ${\sigma }_{it}$ is the calculated SD of observed monthly rainfall values for location i in month t across all years observed.

As a complement to rainfall, temperature is important to crop growth and health (59–61). For many locations, including Nepal and Uganda, including temperature in spatial analysis is problematic because of altitude and varied terrain. Datasets of ground-recorded temperatures with full and dense coverage are not readily available for either country. To account for temperature, a global gridded temperature data product is used (Berkeley Earth Project; berkeleyearth.org). These temperature data are reported as unstandardized local anomalies. Each data element measures the local temperature anomaly, in degrees celsius, for the corresponding location and time. Anomalies are computed relative to a baseline average calculated from January 1951 to December 1980.

The methods used to generate the temperature data are described by Rohde et al. (62). The procedure incorporates both short and discontinuous temperature records, and uses Kriging to interpolate data from stations to arbitrary locations on the Earth. In addition, a weighting procedure is used to reduce the influence of outliers. These data have not been bias-corrected for solar irradiance or for altitude, and may therefore be inaccurate for locations with varied altitude and terrain. Although Kriging allows for a continuously defined temperature field, in practice, the field is sampled on an equal-area grid with a resolution of 1.6° at the equator. This sampling procedure results in 20 unique temperature centroid reference points for Nepal and 26 unique temperature centroid reference points for Uganda. For Nepal, the 20 temperature data points are matched to 547 DHS clusters defined by unique combinations of latitude and longitude. For Uganda, the 26 temperature data points are mapped to 713 DHS clusters. The results for both countries are illustrated in Figs. S1 and S2. In Fig. S1, the three blue cells for Uganda represent Lake Victoria. A value is reported as missing if the grid cell is less than 5% land. As a result, 11 observations from islands in Lake Victoria are assigned values from the immediately adjacent cells in the same latitude. In general, the amount of locally resolved information provided by the temperature product is somewhat limited. Incorporating temperature data is also complicated by the fact that the baseline reference period against which temperature anomalies are constructed differs from the time period covered by the rainfall data. Fig. S3 presents a series of bivariate scatter plots of the temperature and rainfall anomalies, by month, for Nepal. These scatterplots indicate no strong local monthly pattern of correlation between average annual temperature anomalies and average annual rainfall anomalies.

Fig. S2.

Uganda data points overlaid with a temperature grid. Data are from Nepal 2006 and 2011 DHSs (n = 4,374). Temperature data are from the Berkeley Earth Project. Details are provided in SI Materials and Methods.

HDI and Other Data.

District-level HDI data for Nepal and Uganda come from various United Nations Development Program publications available at www.un.org. The 2001 and 2014 Nepal HDIs are reported in the 2005 and 2014 Nepal Human Development Reports, respectively. The 2003 and 2005 Uganda HDIs are reported in the 2003 and 2007 Uganda Human Development Reports, respectively. District-level data on total food storage warehouse capacity in Nepal were obtained from the offices of the Nepal Food Corporation in Kathmandu. Data correspond to 2014, although capacity has varied little over the period covered by this analysis. Data for Nepal on distance to hospitals and number of health facilities, by district, are for calendar year 2010 and come from Nepal’s Human Resources Development and Information System, MOHP, Kathmandu. Open-defecation data come from project reports for Nepal’s Community-Led Total Sanitation program. District-level data on road density in Nepal were obtained from various issues of Nepal Road Statistics, sourced from the Department of Road of the Ministry of Physical Planning, Works, and Transport Management, Kathmandu. Remaining district-level data for Nepal come from the Central Bureau of Statistics of Nepal.

Results

Fig. 1 places nutrition outcomes in an overall development context by displaying the bivariate relationships between basic living conditions and child growth in Nepal and Uganda. Each x axis measures the district-level Human Development Index (HDI) in an initial period (2001 for Nepal, 2003 for Uganda); y axes measure subsequent district-average nutrition outcomes in 2006 and 2011 (5 and 10 y later for Nepal, 3 and 8 y later for Uganda). The HDI is an indicator of basic living conditions, constructed as the weighted average of health (life expectancy), education (mean years of schooling), and income (per capita purchasing power parity). The z-score averages are unweighted, but applying population weights does not alter the basic shapes of the figures. Fig. 1 shows that initial conditions are positively correlated with subsequent nutrition outcomes and illustrates that the gains in child growth that have accompanied HDI improvements are especially pronounced for Uganda at a low HDI, and flatten out at a high HDI. In this sense, Uganda’s curves bear a striking resemblance to Preston’s curves for per capita income and life expectancy (46). Fig. 1 also shows that z-scores have increased in both countries over time, but patterns differ: At the highest HDI values, HAZs have diverged in Nepal, suggesting proportionately larger nutrition gains in high-HDI districts. Uganda displays the opposite pattern: The largest gains in linear growth occurred in low-HDI districts, suggesting convergence of outcomes. In sum, higher levels of literacy, basic health, and economic well-being are associated with nutrition improvements over time, but with a lag: Nutrition gains were greater after a decade than after 3–5 y.

Fig. 1.

District-level human development indicators and district-level nutrition outcomes in Nepal and Uganda.

Fig. 2 displays district-level bivariate relationships for several potential drivers of HDI in Nepal, among which are crop yield (median = 2,862 kg/ha; r = 0.39, P < 0.01), extent of agricultural commercialization (mean = 42% of households; r = 0.24, P < 0.05), distance to market (median = 192 min; r = −0.24, P < 0.05), distance to hospital (median = 771 min; r = −0.29, P < 0.01), improved sanitation in the district (mean = 11% of villages open-defecation-free; r = 0.20, P < 0.10), density of health infrastructure (mean = 0.95 facilities per 100,000 persons; r = 0.14, P < 0.25), and road density (mean = 0.20 km/km²; r = 0.52, P < 0.01). Although all factors are important, assigning independent causality to any single one is problematic: They tend to move and work in concert.

Fig. 2.

District-level HDI (2014) and district-level development indicators (2010) in Nepal. Crop yield is measured as the total quantity of district agricultural production in kilograms divided by total planted area in hectares; agricultural commercialization is the percentage of households in the district selling crops and/or livestock; distance to market is the average distance to a haat bazaar (open-air market) in minutes by foot; distance to hospital is the average distance to a hospital in minutes by foot; improved sanitation represents the percentage of villages declared open-defecation-free by the government of Nepal; the intensity of health infrastructure is defined as the number of health facilities (clinics and hospitals) per 100,000 persons; and road density is measured as the total number kilometers of earthen, gravel, or sealed roads per square kilometer.

Focusing on agriculture, the idea that consumption tracks production, and therefore that favorable agricultural performance will lead to desirable nutritional performance, is attractive. Table 1 compares district-level agriculture and nutrition performance in Nepal and Uganda. Districts are assigned to one of four categories depending on whether they exhibit yields for the main staple in the district that are above or below the median for all districts, and whether they exhibit above- or below-average z-scores. Observations in the upper left and lower right cells of Table 1 are consistent with a view that favorable agricultural performance accompanies desirable nutritional outcomes. Roughly two-thirds of districts conform to this pattern in Nepal (for HAZ) and Uganda (for WHZ). However, a substantial proportion of districts do not conform: 28% of Nepalese districts and 27% of Ugandan districts are positive deviants for WHZ (above-average nutrition outcomes accompany subpar agricultural performance). Furthermore, 17% and 20% are negative deviants (poor nutrition but strong agricultural performance). These patterns are robust to a change in data (2006 in place of 2011, rural districts only) or variables (medians instead of averages, proteins or starches in place of grains, kilograms per person in place of kilograms per hectare). Of course, district-average agricultural yield may not be the best indicator of average household food security in a district, in part, because productivity is unlikely to be decisive where diets lack diversity or essential nutrients. Nevertheless, the fact that a nonnegligible proportion of districts are found in the off-diagonal cells of Table 1 suggests that other factors may augment and offset the agriculture–nutrition relationship. For example, Nepal and Uganda have been the focus of multiple development projects. These projects could have had differential effects across locations and over time that are correlated with nutrition drivers and outcomes in ways that cannot be easily identified or accounted for in data.

Table 1.

District-level HAZ, WHZ, and crop yields in Nepal and Uganda

	Crop yield in Nepal		Crop yield in Uganda
Indicator	Below average, %	Above average, %	Below average, %	Above average, %
HAZ
Below average	35	17	25	20
Above average	18	31	48	7
WHZ
Below average	25	25	46	4
Above average	28	22	27	23

HAZ and WHZ were computed from unweighted 2011 DHS data. Yields were computed from the 2010 Nepal Living Standards Survey (75 districts) and 2010 Uganda National Household Survey (56 districts).

A further complication is that weather in the form of temperature and precipitation differs across locations and time. Measuring the influence of spatial variations in temperature or wetness on nutrition poses serious identification challenges, in part, because both may influence human health and crop growth in potentially confounding ways, and through channels that may respond to or codetermine nutrition. For example, although adequate rainfall tends to be good for crops, agricultural productivity is not uniformly responsive to variation in precipitation; some regions and some periods of the crop cycle suffer from excessive rainfall, and others from insufficient rainfall. Even where a high amount of rainfall might favor agriculture, it may undermine child health due to water-borne disease transmission. Moreover, to uncover local effects, one must isolate within-location weather variability from between-location variability, which tends to be much larger. Fig. 3 isolates weather variability for rainfall by plotting standardized local rainfall anomalies against the explained portion of short-term child growth. These local associations are particularly heterogeneous and nonlinear in Nepal. In the Terai, WHZ moves strongly in concert with growing season rainfall. In contrast, weight gain in the mountains is especially sensitive to departures from the local norm. Overall, linear regressions for Nepal (Table S2) suggest positive correlations, on average, between growing season rainfall (while in utero, during the year of birth, and preceding the survey year) and HAZ and WHZ. These results hold up to the inclusion of fixed effects for birth year, birth month, and location. Patterns in Uganda (Table S3) are less definitive when birth month, birth year, and location fixed effects are included as regressors, but they do point to a positive correlation between HAZ and growing season rainfall during the birth year. In terms of magnitudes, increases of 1 SD in agricultural season rainfall (110 mm in Nepal, 135 mm in Uganda) are associated with 0.05- to 0.25-point higher z-scores at the mean (increases of roughly 4–13% for HAZ and 1–7% for WHZ). Nutrition sensitivity to rainfall is greater in Nepal, where growing season rainfall is lower on average (mean = 281 mm) and wider ranging (minimum = 15 mm, maximum = 837 mm), than in Uganda (mean = 455 mm; minimum = 222, maximum = 1,022 mm). The cluster-level correlation between WHZ and contemporaneous rainfall is negative for rural U5s in Uganda, a pattern that could be driven by the pernicious effects of rainfall on diarrheal or other diseases. As evidence, Fig. 4 plots the explained portion of WHZ against contemporaneous rainfall anomalies (locally standardized) for three Ugandan groups: urban U5s, rural U5s who presented with diarrhea during the 2 wk before measurement (27% of the rural sample), and rural U5s who did not. Urban U5s appear to be buffered from high rainfall, but positive rainfall anomalies are associated with lower WHZ, especially for rural U5s with diarrhea. Nearly the same pattern arises when a fever indicator (45% of rural U5s) replaces the diarrhea indicator.

Fig. 3.

Growing season rainfall in year of birth and WHZ for Nepalese children 0–60 mo of age in 2001–2011 (locally weighted regression plots, n = 4,561). Values of WHZ used in the figure are predictions from an ordinary least-squares (OLS) regression that includes child- and household-level covariates as well as control variables for location and birth month (Table S1). Values on the x axis are growing season rainfall anomalies in the child’s year of birth, standardized by the local mean and SD of rainfall. Standardization is based on rainfall observed over 15 y (1998–2012) at the closest rainfall reporting station (n = 276) to the child. Details are provided in SI Materials and Methods.

Table S2.

Results for Nepal HAZ and WHZ linear rainfall regressions incorporating birth month, birth year, and location fixed effects

Variables	HAZ	HAZ	HAZ	WHZ	WHZ	WHZ
Growing season rainfall
In birth year	0.0006	0.0012				0.0006
	(0.0001)***	(0.0001)***				(0.0001)***
While in utero			0.0011
			(0.0001)***
During survey period				0.0006	0.0004
				(0.0001)***	(0.0001)***
Fixed effects
Birth month	No	Yes	Yes	No	Yes	Yes
Birth year	No	Yes	Yes	No	Yes	Yes
Ecological zone	No	Yes	Yes	No	Yes	Yes
R2	0.01	0.15	0.15	0.01	0.04	0.05
No. of observations	7,572	7,572	7,572	7,572	7,572	7,572

^***Significantly different from zero at the 1% level.

Table S3.

Results for Uganda HAZ and WHZ linear rainfall regressions incorporating birth month, birth year, and location fixed effects

Variables	HAZ	HAZ	HAZ	WHZ	WHZ	WHZ
Growing season rainfall
In birth year	0.0003 (0.0002)*	0.0002 (0.0003)				0.0001 (0.0002)
While in utero			0.0001 (0.0003)
During survey period				−0.0004 (0.0001)***	0.0001 (0.0002)
Fixed effects
Birth month	NO	YES	YES	NO	YES	YES
Birth year	NO	YES	YES	NO	YES	YES
Ecological zone	NO	YES	YES	NO	YES	YES
R2	0.01	0.09	0.09	0.01	0.09	0.09
No. of observations	4,345	4,345	4,345	4,345	4,345	4,345

^*Significantly different from zero at the 10% level; ***significantly different from zero at the 1% level.

Fig. 4.

WHZ for Ugandan U5s (measured in 2006 and 2011) and average annual rainfall in household cluster (locally weighted regression plots, n = 4,374). Values of WHZ are the predictions from an OLS regression that includes child- and household-level covariates as well as control variables for location and season of birth (Table S1). Values on the x axis are rainfall anomalies observed in the month of anthropometric measurement, standardized by the local mean and SD of rainfall. Standardization is based on rainfall observed over 34 y (1981–2014) for the sample clusters matched to the CHIRPS dataset (n = 560). Details are provided in SI Materials and Methods.

Table S1.

Results for OLS regressions for WHZ in Nepal and Uganda

Variables	Nepal WHZ	Uganda WHZ
Child’s age	0.0062	0.0113
(mo)	(0.0008)***	(0.0012)***
Sex	0.0189	−0.0208
(0/1; 1 = female)	(0.0240)	(0.0361)
Twin birth	−0.4689	−0.5377
(0/1; 1 = twin)	(0.1223)***	(0.1216)***
Home delivery	−0.1087	−0.0834
(0/1; 1 = home)	(0.0333)***	(0.0420)**
Vaccinations	−0.0235	−0.0519
(0/1; 1 = one or more)	(0.0306)	(0.0393)
Diarrhea	−0.1668	−0.1915
(0/1; 1 = past 2 wk)	(0.0363)***	(0.0424)***
Mother’s age	−0.0027	−0.0066
(y)	(0.0032)	(0.0049)
Mother’s education	0.0535	0.0478
(y)	(0.0178)***	(0.0336)
Total births	−0.0080	0.0185
(no.)	(0.0108)	(0.0134)
Mother smokes	0.1012	−0.1938
(0/1; 1 = yes)	(0.0381)***	(0.1037)*
Safe water source	0.0061	−0.0578
(0/1; 1 = yes)	(0.0333)	(0.0698)
Dependency ratio	0.0072	−0.0173
[(<15 + >65)/(16–64)]	(0.0145)	(0.0177)
Flush toilet	0.0762	−0.1505
(0/1; 1 = yes)	(0.0395)*	(0.0897)*
Biomass fuel	0.0462	0.0301
(0/1; 1 = yes)	(0.0579)	(0.0781)
Wealth	0.0008	0.0009
(index factor score)	(0.0003)**	(0.0004)**
Urban/rural	0.0747	0.0906
(0/1; 1 = urban)	(0.0356)**	(0.0830)
Altitude	0.0002	−0.0002
(meters above sea level)	(0.0000)***	(0.0001)*
	−0.8421	0.0200
Constant	(0.1631)***	(0.4148)
R2	0.09	0.09
No. of observations	7,572	4,345

All regressions include fixed effects for month of birth and district. *Significantly different from zero at the 10% level; **significantly different from zero at the 5% level; ***significantly different from zero at the 1% level.

To gauge the importance of environmental conditions to child growth, Table S4 provides results for multilevel regressions estimated with children nested within households, households nested within districts, and districts nested within regions. The child-level components of the regressions contain the child, the mother, and household control variables; for Table S4, the upper-level random-effect components are naive regressions with no predictors. Table S5 summarizes the results. The first part of Table S5 contains intraclass correlation coefficients (ICCs). The second part of Table S5 contains variance shares for each level. Although the ICCs are relatively small for upper levels, even modest amounts of intracluster correlation suggest similar treatments or outcomes within groups, underscoring the importance of direct or indirect interventions targeted at households, districts, or regions. Approximately 15–20% of total z-score variance (75–80% of explained variance) occurs at the household level, with relatively modest shares at higher aggregation. After controlling for child- and household-level observables, differences across children within households are far smaller than differences across children at district or regional levels.

Table S4.

Results for multilevel regressions for HAZ and WHZ in Nepal and Uganda

	Nepal		Uganda
Variable	HAZ	WHZ	HAZ	WHZ
Child’s age	−0.0207	0.0068	−0.0207	0.0116
(mo)	(0.0009)***	(0.0008)***	(0.0013)***	(0.0011)***
Sex	0.0061	0.0236	0.1736	−0.0110
(0/1; 1 = female)	(0.0270)	(0.0235)	(0.0434)***	(0.0354)
Twin birth	−0.4893	−0.4599	−0.8656	−0.5624
(0/1; 1 = twin)	(0.1479)***	(0.1288)***	(0.1547)***	(0.1263)***
Home delivery	−0.2002	−0.1008	−0.0966	−0.0820
(0/1; 1 = home)	(0.0383)***	(0.0333)***	(0.0514)*	(0.0418)**
Vaccinations	−0.2185	−0.0455	0.0523	−0.0532
(0/1; 1 = one or more)	(0.0349)***	(0.0303)	(0.0484)	(0.0395)
Diarrhea	—	−0.1749	—	−0.1956
(0/1; 1 = past 2 wk)		(0.0359)***		(0.0421)***
Mother’s age	0.0141	−0.0012	0.0244	−0.0067
(y)	(0.0038)***	(0.0033)	(0.0062)***	(0.0051)
Mother’s education	0.1601	0.0575	0.0950	0.0511
(y)	(0.0208)***	(0.0181)***	(0.0422)**	(0.0345)
Total births	−0.0434	−0.0117	−0.0142	0.0165
(no.)	(0.0128)***	(0.0112)	(0.0169)	(0.0138)
Mother smokes	−0.2754	0.0984	0.0194	−0.2595
(0/1; 1 = yes)	(0.0453)***	(0.0394)**	(0.1266)	(0.1038)**
Safe water source	0.0623	−0.0032	−0.0181	−0.0652
(0/1; 1 = yes)	(0.0386)	(0.0333)	(0.0866)	(0.0696)
Dependency ratio	−0.0259	0.0022	−0.0268	−0.0133
[(<15 + >65)/(16–64)]	(0.0175)	(0.0152)	(0.0228)	(0.0186)
Flush toilet	0.0181	0.0816	0.2032	−0.1584
(0/1; 1 = yes)	(0.0470)	(0.0409)**	(0.1118)*	(0.0910)*
Biomass fuel	−0.0256	0.0156	0.0636	0.0403
(0/1; 1 = yes)	(0.0685)	(0.0596)	(0.0987)	(0.0802)
Wealth	0.0017	0.0008	0.0021	0.0010
(index factor score)	(0.0003)***	(0.0003)***	(0.0005)***	(0.0004)**
Urban/rural	0.0063	0.0678	0.2192	0.0821
(0/1; 1 = urban)	(0.0424)	(0.0368)*	(0.1025)**	(0.0831)
Altitude	−0.0002	0.0002	−0.0005	−0.0001
(meters above sea level)	(0.0000)***	(0.0000)***	(0.0001)***	(0.0001)
Constant	−1.0699	−1.1154	−1.3865	0.0023
	(0.1278)***	(0.1146)***	(0.2350)***	(0.1957)
Random effects (SD)
Region	0.0162	0.0722	0.1057	0.1723
	(0.0860)	(0.0354)	(0.0452)	(0.0460)
District	0.1987	0.1384	0.0385	0.0005
	(0.0270)	(0.0210)	(0.1356)	(0.0013)
Household	0.5574	0.4961	0.5837	0.4737
	(0.0261)	(0.0230)	(0.0467)	(0.0366)
Residual	1.0541	0.9114	1.3269	1.0825
	(0.0144)	(0.0127)	(0.0217)	(0.0188)
LR test vs. linear regression	230.3***	222.7***	75.3***	93.5***
Wald χ2	1,695.0	345.7	443.1	224.3
No. of observations	7,572	7,572	4,345	4,345

LR, likelihood ratio test. *Significantly different from zero at 10% level; **significantly different from zero at 5% level; ***significantly different from zero at 1% level. All regressions include fixed effects for month of birth.

Table S5.

ICCs for HAZ and WHZ in Nepal and Uganda

	Nepal		Uganda
Level variables	HAZ	WHZ	HAZ	WHZ
ICCs
Household	0.21	0.22	0.16	0.16
District	0.24	0.24	0.17	0.17
Region	0.25	0.26	0.18	0.19
Variance shares for each level
Household	21%	22%	16%	16%
District	3%	3%	1%	1%
Region	1%	1%	1%	3%
Unexplained	75%	74%	82%	80%

Statistics were estimated using 2006 and 2011 DHS data. Nesting contains 13 regions, 75 districts, and 5,332 households for Nepal (n = 7,572) and 10 regions, 56 districts, and 2,770 households for Uganda (n = 4,345). The ICC is computed as ${\sigma }_e^2/\left({\sigma }_e^2+{\sigma }_u^2\right)$, where subscripts identify the variances from explained (i.e., group level) and unexplained portions of the total residual. By construction, ICCs range from 0 to 1. Values close to 0 indicate the grouping conveys relatively little information about total variance in the sample, whereas values near 1 indicate similarity of observations in the grouping.

Variance shares in Table S5 are similar when administrative regions are replaced by agroecological zones or sample clusters. In other words, suprahousehold group effects are important, perhaps due to spatial spillovers in markets or health provisioning. Support for this conjecture comes from reestimating the WHZ regression as a two-level model (child and district), adding rainfall before the survey and birth month fixed effects, while, at the same time, incorporating district-level variables for road density (kilometers per square kilometer) and median distance to the nearest health facility (minutes by foot). The first two columns of Table S6 indicate that roads and health facilities are both statistically important in explaining district-level variance in WHZ. When included as fixed factors at the child level, they are positively correlated with WHZ and unambiguously reduce the importance of precipitation: The magnitude of the point estimate on rainfall falls by roughly one-third, and the statistical significance of the coefficient declines.

Table S6.

Results for multilevel regressions for WHZ in Nepal incorporating district variables, rainfall, and temperature

Variable	WHZ 1	WHZ 2	WHZ 3	WHZ 4
Child’s age	0.0063	0.0062	0.0064	0.0063
(mo)	(0.0008)***	(0.0008)***	(0.0008)***	(0.0008)***
Sex	0.0192	0.0215	0.0219	0.0233
(0/1; 1 = female)	(0.0239)	(0.0239)	(0.0238)	(0.0238)
Twin birth	−0.4816	−0.4752	−0.4858	−0.4769
(0/1; 1 = twin)	(0.1216)***	(0.1217)***	(0.1213)***	(0.1214)***
Home delivery	−0.0875	−0.0910	−0.0664	−0.0672
(0/1; 1 = home)	(0.0331)***	(0.0332)***	(0.0334)**	(0.0335)***
Vaccinations	−0.0337	−0.0296	−0.0445	−0.0427
(0/1; 1 = one or more)	(0.0303)	(0.0304)	(0.0306)	(0.0306)
Diarrhea	−0.1616	−0.1664	−0.1658	−0.1703
(0/1; 1= past 2 wk)	(0.0361)***	(0.0362)***	(0.0361)***	(0.0361)***
Mother’s age	−0.0013	−0.0014	−0.0020	−0.0019
(y)	(0.0032)	(0.0032)	(0.0032)	(0.0032)
Mother’s education	0.0531	0.0566	0.0463	0.0487
(y)	(0.0175)***	(0.0175)***	(0.0175)***	(0.0176)***
Total births	−0.0134	−0.0115	−0.0114	−0.0106
(no.)	(0.0107)	(0.0107)	(0.0107)	(0.0107)
Mother smokes	0.1081	0.1009	0.1147	0.1080
(0/1; 1 = yes)	(0.0374)***	(0.0375)***	(0.0374)***	(0.0374)***
Safe water source	0.0032	−0.0054	−0.0148	−0.0196
(0/1; 1 = yes)	(0.0319)	(0.0317)	(0.0374)	(0.0318)
Dependency ratio	0.0096	0.0076	0.0086	0.0073
[(<15 + >65)/(16–64)]	(0.0144)	(0.0144)	(0.0144)	(0.0144)
Flush toilet	0.0644	0.0592	0.0291	0.0262
(0/1; 1 = yes)	(0.0394)	(0.0392)	(0.0399)	(0.0396)
Biomass fuel	0.0609	0.0609	0.0838	0.0890
(0/1; 1 = yes)	(0.0576)	(0.0576)	(0.0580)	(0.0579)
Wealth	0.0009	0.0009	0.0012	0.0012
(index factor score)	(0.0003)***	(0.0003)***	(0.0003)***	(0.0003)***
Urban/rural	0.0881	0.0768	0.0803	0.0665
(0/1; 1 = urban)	(0.0352)**	(0.0352)**	(0.0351)**	(0.0351)**
Altitude	0.0002	0.0002	0.0002	0.0002
(meters above sea level)	(0.0000)***	(0.0000)***	(0.0000)***	(0.0000)***
Growing season rainfall	0.0007	0.0004	—	—
(mm, in survey year)	(0.0002)***	(0.0002)**
Rainfall anomaly	—	—	−0.0171	−0.0165
(mm, in survey year)			(0.0178)	(0.0146)
Temperature anomaly	—	—	−0.3068	−0.2845
(°C, in survey year)			(0.3047)	(0.0496)**
Rainfall anomaly x	—	—	0.0420	0.0437
Temperature anomaly			(0.0223)*	(0.0200)**
Road density	—	0.1851	—	0.1851
(km/km2)		(0.0794)**		0.1642
Distance to health facility	—	0.0006	—	(0.0798)**
(min by foot)		(0.0003)**		−0.0001
Constant	−1.2743	−1.2423	−0.8891	−0.9089
	(0.1262)***	(0.1270)***	(0.1171)	(0.1212)***
District random effects (SD)
Constant	0.1461	0.1587	0.1525	0.1593
	(0.0219)	(0.0206)	(0.0215)	(0.0202)
Road density	0.1211	—	0.1083	—
(km/km2)	(0.0734)		(0.0774)
Distance to health facility	0.0003	—	0.0002	—
(min by foot)	(0.0001)		(0.0001)
Residual	1.0301	1.0342	1.0291	1.0319
	(0.0085)	(0.0084)	(0.0084)	(0.0084)
LR test vs. linear regression	112.1***	74.0***	116.6***	83.0***
Wald χ2	338.8	351.0	366.76	385.7
No. of observations	7,572	7,572	7,572	7,572

^*Significantly different from zero at 10% level; **significantly different from zero at 5% level; ***significantly different from zero at 1% level. All regressions include fixed effects for month of birth.

To address the potential importance of temperature as an additional driver of agriculture and health, and hence nutrition, the final two columns of Table S6 report regression results for Nepal that incorporate both temperature and precipitation anomalies. For Nepal, local temperature anomalies over the period covered by the DHS were generally positive, and tended to be larger in 2011 than in 2006. Additional research is required to ascertain the relationship between temperature and child growth. However, inclusion of the temperature anomaly as a regressor reduces the statistical significance of the rainfall variable, leaving signs and magnitudes of other point estimates unchanged. Evidence also indicates a positive contemporaneous nutrition association with the combination of rainfall and temperature (as an interaction), underscoring the likely importance of the agriculture/nutrition pathway for child growth.

High collinearity among variables precludes including a large complement of district-level regressors in these models. Instead, factors operating at higher levels are elucidated for Nepal in Table 2. The entries are constructed by first assigning all HAZ and WHZ observations to their respective quintiles and then summarizing associated district-level data on rainfall, altitude, and five indicators of public investment: roads, health facilities, clean water, markets, and public grain storage. Average growing-season rainfall increases across the HAZ and WHZ quintiles. Average altitude decreases with HAZ and increases with WHZ. In terms of the five infrastructure variables, isolation is negatively correlated with linear growth and weight gain. With few exceptions, children in higher growth quintiles are found in districts with greater road and health facility densities, better water sources, more proximate agricultural markets, and greater public grain storage capacity. Patterns are strongest for HAZ, echoing Fig. 1, where overall development was seen to magnify nutrition improvements over time. Intrinsic to sharp gains in linear growth are greater clinic density (22% higher in the top HAZ quintile vs. the bottom quintile), food storage capacity (41% higher), road density (58% higher), and nearness to markets (61% higher). Claims of independent causality are problematic because these investments tend to move together, but the patterns of association cast light on a key policy domain where data and findings are, at present, extremely scarce.

Table 2.

Environment and infrastructure by HAZ and WHZ quintiles in Nepal

	HAZ quintiles					WHZ quintiles
Variable	1	2	3	4	5	1	2	3	4	5
Z-score	−3.70	−2.59	−1.93	−1.23	0.06	−2.31	−1.30	−0.77	−0.23	0.68
Rainfall	299	307	314	318	311	274	279	280	285	290
Altitude	977	902	802	770	676	675	783	817	913	943
Roads	0.19	0.21	0.22	0.24	0.30	0.23	0.24	0.25	0.28	0.33
Clinics	4.9	5.0	5.1	5.4	6.0	4.7	5.0	5.0	5.6	6.0
Water	0.68	0.71	0.76	0.76	0.80	0.76	0.74	0.74	0.73	0.74
Markets	297	245	188	225	180	200	208	224	245	253
Storage	2.7	3.2	3.5	3.4	3.8	3.4	3.3	3.4	3.4	3.3

Z-scores are from the 2006 and 2011 DHSs (n = 7,572). Rainfall (millimeters) was measured at the DHS cluster level for growing season in the year before birth (HAZ) and in the growing season immediately preceding child measurement (WHZ). Altitude (meters above sea level) is from the DHS. Roads (kilometers per square kilometer), clinics (health facilities per 100,000 persons), water (percentage of households self-reporting access to a safe source of drinking water at the time of child measurement), markets (average distance in minutes of walking to a weekly bazaar; excluding observations from Kathmandu), and public food storage capacity (in kilograms per person) were measured at the district level in the year of the child’s birth.

In developing countries, urban residency improves access to a wide range of services, serving as a proxy for many unobservables. Urban/rural differences in child growth are substantial and significant in Uganda: On average, urban HAZ in 2011 was more than one-half of 1 SD higher than rural HAZ (−0.96 vs. −1.56; t = −9.4, P < 0.001), and WHZ was one-sixth of 1 SD higher (0.10 vs. −0.07; t = 3.25, P < 0.001). These differences cannot be fully explained by observable differences between urban and rural children, mothers, or households. An urban environment may confer health and nutrition advantages through many channels, including food security, dietary diversity, access to health services, and improved water and sanitation services. It might also reduce exposure to some environmental risks, including extreme weather. As evidence, Fig. 5 displays HAZ and WHZ for Ugandan U5s by rainfall decile. Fig. 5 is constructed by computing two quantile indicators for precipitation: growing season rainfall during the year of birth (compared with HAZ) and growing season rainfall during the survey year (compared with WHZ). Fig. 5 shows (i) rainfall exposure differs across the landscape such that rural U5s have greater exposure to extreme rainfall, both at the lower and the upper ends of the rainfall distribution, than urban U5s; (ii) rainfall in Uganda is weakly but negatively correlated with WHZ, consistent with the pattern displayed in Fig. 4 and the conjecture that rainfall-related health outcomes influence nutrition in rural areas; and (iii) Ugandan children born during periods of extremely high or low rainfall display comparatively lower linear growth than cohorts born during less extreme periods.

Fig. 5.

Rainfall deciles and HAZ and WHZ for Ugandan U5s (n = 4,374) as measured in 2006 and 2011. HAZ (blue) uses rainy season rainfall in year of birth; WHZ (green) uses rainy season rainfall in year of growth measurement. Symbol size is proportional to the percentage of urban children in each group.

Discussion

Many drivers of linear growth and weight gain in Nepalese and Ugandan children are found at the child and household levels, but the overall explanatory power of these variables is incomplete. Although higher level factors alone are insufficient for understanding patterns of individual child growth, they provide complementary and confirmatory evidence that investments in market and health infrastructure pay dividends over time. Previous work from Nepal suggests a positive relationship between household welfare and investments in irrigation, agricultural extension, and rural roads (47), as well as sanitation (48). Roads and bridges, for example, improve market performance, reducing both the level and variability of food prices (49). For Uganda, investments in agricultural research and rural roads have boosted agricultural productivity and reduced poverty (50), despite the fact that the effectiveness of public investments in Uganda remains low (51).

Roughly three-quarters of Nepalis and Ugandans are dependent on agriculture for their livelihoods. Changes in the frequency, intensity, timing, and total amount of rainfall; temperature changes; and general variability and unreliability of growing conditions are already being felt in the form of seasonal droughts and, in the case of Nepal, changes in surface hydrology (52). Where households are isolated from markets and health infrastructure, the nutrition effects of extreme climate and weather are likely to be more pronounced, consistent with global evidence linking economic underdevelopment to high rates of climate-driven mortality and morbidity (53). Encouragingly, a nascent second “Green Revolution” has the potential to develop and disseminate new crops with enhanced climate sensitivity (54). Nevertheless, to achieve large and lasting nutritional gains at scale, these agricultural innovations will need to be coupled with economy-wide improvements in health and market infrastructure to buffer exposure, and institutional strengthening to maintain and sustain investments over time. The findings of this study underscore the need to maintain a holistic approach to combating child malnutrition.