Climatic conditions and household food security: Evidence from Tanzania

Abstract

Food security and adequate nutrition are critical for achieving progress toward sustainable development. Two billion people worldwide experience moderate to severe food insecurity, and rates of hunger have increased over the past several years after declining steadily for decades. The FAO attributes this increase in large part to climate change, though empirical evidence on the relationship between climatic conditions and food security remains limited. We examine this question by linking nationally representative longitudinal data from four rounds of the Tanzania National Panel Survey to high-resolution gridded climate data. We then estimate a set of fixed effects regression models to understand the linkages between recent rainy season precipitation and temperature and two indicators of household food security: Food Consumption Score (FCS) and reduced Coping Strategies Index (rCSI). We find that low rainfall—particularly dry and cool conditions—is negatively associated with food security. Moving from a typical rainfall year to a particularly dry one increases the risk of being food insecure on both measures simultaneously by 13-percentage points. This suggests that a lack of rainfall impedes households’ ability to access food, likely through reduced agricultural production and increased food prices, leading to lower dietary diversity and food shortages. Vulnerability is higher among households with fewer working age members, suggesting that households with a greater supply of labor can better withstand droughts. As climate change alters precipitation and temperature patterns over the coming decades, policies to increase resilience will be critical for improving food security, particularly among populations heavily reliant on agriculture.

1. Introduction

Food insecurity is a critical barrier to sustainable development. The inability to acquire sufficient, nutritious, safe, and affordable food negatively affects nutritional status as well as physical and mental health (Hadley and Crooks, 2012; Pourmotabbed et al., 2020). This, in turn, has implications for household labor productivity, child growth and development, and poverty reduction. Two billion people—26% of the global population—experienced moderate or severe food insecurity in 2019, and rates of hunger have increased over the past several years after declining steadily for decades (FAO et al., 2020). This increase has been felt most heavily in Sub-Saharan Africa, where the prevalence of moderate or severe food insecurity grew from 50% in 2014 to 57% in 2019 (FAO et al., 2020). In response to persistently high levels of food insecurity in many low- and middle-income countries, the United Nations has aimed a Sustainable Development Goal (SDG) at eliminating hunger and malnutrition worldwide by 2030. In addition, the UN seeks to double agricultural productivity among smallholder farmers and promote sustainable food systems, with a particular focus on assisting vulnerable populations including children, women, Indigenous peoples, and those living in poverty (United Nations, 2020).

Climate change, however, is likely to undermine progress in reducing hunger and malnutrition. The United Nations Food and Agriculture Organization (FAO) attributes the recent increase in global hunger in large part to climate change, particularly in countries heavily dependent on rainfed agricultural systems (FAO et al., 2018). Changes in rainfall and temperature over the past several decades have impacted crop production and yield variability in many food-insecure regions of the world, though effects vary spatially as well as by crop type (Agnolucci et al., 2020; Barrios et al., 2008; Iizumi and Ramankutty, 2016; Ray et al., 2019). Beyond its impacts on crop production, climate change may worsen food insecurity by affecting livestock and fisheries, non-farm income and livelihoods, transport and storage of food products, access to markets, and food safety (Vermeulen et al., 2012).

Despite the widespread concern that climate change will threaten future food security (Porter et al., 2014; Wheeler and von Braun, 2013), empirical evidence on the topic remains limited. Few studies have examined the direct linkages between recent climatic conditions and household experiences of food insecurity (Cooper et al., 2019; Randell et al., 2021; Smith and Frankenberger, 2018; Tankari, 2020). The majority of research on climate change, food systems, and nutrition has focused on the relationship between climatic conditions and other food- and nutrition-related outcomes including crop production, child anthropometry, or food expenditures (e.g., Carpena, 2019; Randell et al., 2020; Ray et al., 2019; Rosenzweig et al., 2014; Thiede and Strube, 2020; Wineman et al., 2017). While these outcomes are correlated with food insecurity, they fail to offer comprehensive insight into how households experience and respond to declines in food quantity, quality, and variety.

In this paper, we examine the relationship between recent rainfall and temperature conditions and food security in Tanzania, a country in which 31% of the population suffers from undernutrition and 65% works in agriculture (Arce and Caballero, 2015; FAO et al., 2019; United Nations, 2019). We link nationally representative household survey data from four rounds of the Tanzania National Panel Survey, conducted between 2008 and 2016, to high resolution climate data. Using fixed effects regression models, we examine how climatic conditions experienced during the most recent completed rainy season affect two indicators of household food security—the Food Consumption Score (FCS) and the reduced Coping Strategies Index (rCSI). These indicators capture multiple dimensions of food security including dietary diversity and coping strategies in response to food shortages. This study contributes to the limited literature on the linkages between climatic conditions and household experiences of food security, and the findings will help inform policies to improve health and nutrition among vulnerable populations.

2. Food Security and Climate Change

Food security exists “when all people, at all times, have physical and economic access to sufficient, safe and nutritious food to meet their dietary needs and food preferences for an active and healthy life” (World Food Summit, 1996), and is generally thought to consist of four interrelated components: availability, access, utilization, and stability (Carletto et al., 2013; Upton et al., 2016). Availability—the presence of food in a given location and time—is affected by localized food production, food stocks, trade, and food aid. Access—the ability to acquire food—is dependent on factors such as income, food prices, and distance to markets. Utilization—the capacity to obtain nourishment from food—is determined by dietary diversity, food safety, nutritional value, and illness. Stability—the lack of fluctuation in availability, access, and utilization over time—is dependent on short- and longer-term changes in environmental, social, political, and economic conditions.

Food security statistics are frequently reported at the national level, but Barrett (2010) argued that evaluating food insecurity at the household or individual scale is critical in order to identify factors associated with vulnerability and inform targeted policy interventions. Holistically assessing household food security is complex (Vaitla et al., 2017), and, in the absence of a universal measure of household- or individual-level food security, researchers have used a number of proxies including food expenditures, caloric consumption, and anthropometric outcomes. A growing body of research has examined linkages between climatic conditions and these outcomes given the strong evidence that climate change impacts each of the four components of food security (Vermeulen et al., 2012; Wheeler and von Braun, 2013). For example, extreme weather events may impact availability by decreasing crop yields (Lesk et al., 2016) or access by increasing food prices (Fanzo et al., 2018), and may affect utilization if hot weather leads to an increased risk of food contamination (Tirado et al., 2010).

Evidence also suggests that exposure to adverse rainfall conditions such as droughts, floods, or delayed monsoon onset affects child anthropometric outcomes including stunting and wasting (e.g., Bahru et al., 2019; Rodriguez-Llanes et al., 2011; Thiede and Gray, 2020), with effects that vary by gender, age at exposure, and household poverty (Bahru et al., 2019; Hoddinott and Kinsey, 2001). In some contexts, high temperatures have been linked to an increased risk of child wasting and lower adult BMI (Mueller and Gray, 2018; Thiede and Strube, 2020), but warmer conditions during early life are favorable for child nutrition in the Ethiopian highlands (Randell et al., 2020). In rural areas of Mexico, Indonesia, India and Kenya, adverse climatic conditions impact food expenditures, and the effects vary between sub-populations and by timing of exposure (Carpena, 2019; Skoufias et al., 2012; Skoufias and Vinha, 2013; Wineman et al., 2017). Taken together, these studies suggest that climatic conditions affect household food consumption and the nutrition of household members, and that vulnerability varies across sociodemographic groups and geographical contexts.

The above measures, however, are insufficient proxies for household food security. For example, food expenditures and caloric consumption require detailed recall of recent food purchases and fail to capture dietary diversity or nutrient density (Headey and Ecker, 2013; Maxwell et al., 2014). Anthropometric outcomes are often only collected for young children and may be affected by other factors aside from food security, such as infectious diseases (Headey and Ecker, 2013; Mulmi et al., 2016). Maxwell et al., (2014) argued that effectively measuring food security requires multiple indicators that capture household members’ subjective experiences of the quantity, quality, and variety of food that they are able to obtain. A diverse set of indicators has in fact been developed including the Food Consumption Score (FCS), Coping Strategies Index (CSI), and Household Hunger Scale (HHS) (Carletto et al., 2013; Maxwell et al., 2014; Vaitla et al., 2017). FCS captures dietary diversity and nutrient density, while CSI and HHS measure behavioral responses to household food shortages. The level of correlation between these indicators varies substantially, suggesting that they represent different underlying dimensions of food security (Vaitla et al., 2017).

Such indicators are commonly used by the World Food Programme (WFP), FAO, and USAID for food security assessment (Maxwell et al., 2014), but only a small number of studies have examined linkages between them and climate. Cooper et al., (2019), using the HHS, found that prolonged drought was positively associated with hunger in Ghana (where the climate is dry), while in Bangladesh (where the climate is wet) experiencing multiple years of excessive rainfall was linked hunger. Similarly, Tankari, (2020) found that in Burkina Faso, increases in rainfall were associated with better food security as measured by the Food Insecurity Experience Scale, particularly for rural households. Smith & Frankenberger, (2018) used panel data to examine how severe flooding in Bangladesh affected HHS and the number of months of adequate food, revealing that flooding was associated increased food insecurity. Lastly, Randell et al., (2021) used the Household Food Insecurity Access Scale and found that the effect of monsoon rainfall on food insecurity in Nepal varied by level of recent earthquake exposure. These studies suggest that rainfall generally has a positive effect on household food security, particularly in more arid areas dependent on rainfed agriculture, but that extreme rainfall (which causes flooding and landslides) acts as a threat. As described below, we expand this literature by examining (1) linkages between food security, rainfall and temperature, (2) two food security indicators together—Food Consumption Score (FCS) and reduced Coping Strategies Index (rCSI), and (3) how vulnerability varies across sub-populations within Tanzania.

3. The Tanzanian Context

Tanzania is the fifth most populous country in Africa, with a population of approximately 58 million (World Bank, 2019). The country has experienced rapid economic growth over the past several decades, but poverty and inequality remain key development challenges (UNDP, 2017). Food insecurity and undernutrition are prevalent, with 55% of Tanzanians experiencing moderate or severe food insecurity, 31% of the population undernourished, and 35% of children stunted (FAO et al., 2020; von Grebmer et al., 2019). The cost of a nutritious diet varies both seasonally and interannually in Tanzania (Masters et al., 2018), suggesting that households—particularly the most marginalized—may have limited access to diverse, healthy diets during periods of higher food prices.

Tanzania’s topography is highly variable, ranging from coastal plains, to a central plateau region, to highlands along the southern and northern borders including Mount Kilimanjaro (Figure 1). The climate includes tropical rainforests, arid steppes, and temperate highlands, with much of the land area experiencing tropical savannah conditions (Beck et al., 2018). Annual rainfall averages 600-800 mm (Rowhani et al., 2011). Tanzania experiences two rainfall regimes: a unimodal pattern in the central, southern, and southwestern highlands where rain typically occurs from November through April (Msimu), and a bimodal pattern in the coastal areas, northeastern highlands, and Lake Victoria basin with long rains (Masika) from March through May and short rains (Vuli) from October through December (FAO, 2016; FEWS NET, 2013).

Figure 1.

Map of Tanzania including LSMS enumeration areas, elevation, and regions with bimodal rainfall regimes.

Approximately 75% of Tanzanians live in rural areas, and 65% of the population is employed in agriculture (UNDP, 2017; United Nations, 2019). Agriculture is primarily rainfed and is dominated by smallholder production, with average farm sizes ranging from 0.2 to 2 hectares (FAO, 2016; Letta et al., 2018). Maize is grown by 85% of farmers across the country and serves as the primary food crop for most households, accounting for 50% of total calories consumed (Arce and Caballero, 2015; USAID, 2014). Maize yield in Tanzania is highly volatile from year to year and has not increased over time (Aylward et al., 2015). Short-term maize price variation in the country is primarily influenced by local weather shocks and seasonal harvest cycles (Baffes et al., 2019), with increased maize prices linked to reduced household food security (Rudolf, 2019). Unreliable rainfall as well as droughts have been identified as key risks to agricultural production in the country (Arce and Caballero, 2015). In addition, studies have found that maize is sensitive to heat, with yields decreasing with exposure to temperatures above 29 °C (Schlenker and Roberts, 2009) or growing degree days above 30 °C, particularly under drought conditions (Lobell et al., 2011). Hot temperatures have been shown to reduce consumption growth among the poorest farming households through negative effects on agricultural production (Letta et al., 2018).

Over the past several decades, Tanzania has experienced a general warming trend, with a significant increase in the maximum daily temperature across much of the country as well as a significant increase in days with a daily maximum temperature above 25 °C in the northern regions (Gebrechorkos et al., 2019). Rainfall changes have been more variable, with northeastern and southern Tanzania experiencing a drying trend and central and northwest regions becoming wetter (Conway et al., 2017).

Climate projections suggest continued warming and spatially variable changes in precipitation. An analysis of 34 General Circulation Models (GCMs) finds strong agreement that the number of days above 30 °C will increase from approximately 10 days currently to 80 days by the 2040s (Conway et al., 2017). Most GCMs predict an increase in mean annual rainfall for the country, though southern parts of the country are expected to experience rainfall declines. Further, there is high agreement that rainfall variability will increase, with fewer total rainy days but a greater amount of rainfall on each rainy day. This suggests that there will be an increased likelihood of prolonged droughts as well as heavy rainfall events. Indeed, by 2050, increasing intra-seasonal precipitation variability is projected to decrease maize, sorghum, and rice yields by 4.2%, 7.2%, and 7.6%, respectively (Rowhani et al., 2011).

4. Data and Analysis

4.1. Household Survey Data and Food Security Measures

To understand the relationship between climatic conditions and household experiences of food insecurity, we link nationally representative longitudinal household survey data from the Tanzania National Panel Survey (NPS) to high-resolution rainfall and temperature data. The NPS data were collected by the Tanzania National Bureau of Statistics with assistance from the World Bank Living Standards Measurement Study–Integrated Survey on Agriculture (LSMS-ISA). The NPS was initiated in 2008-09 with a sample of 3,265 households in 409 enumeration areas (EAs) and was designed to be representative at the national, urban/rural, and agro-ecological zone levels (National Bureau of Statistics, 2009). Multi-stage cluster sampling was used, and sampling weights are necessary to produce nationally-representative statistics. In rural areas, EAs represent entire villages, while in urban areas they represent census enumeration areas. Location information is available at the EA level, with anonymity protected by a random GPS offset of between zero and 10 km.

Follow-ups were conducted in 2010-11 and 2012-2013 with an attrition rate of ~3% per round. In 2014-2015 a nationally representative subsample of 989 households was reinterviewed from the previous rounds (called the “extended panel”), and the sample was refreshed with 3,352 new households that we exclude from this analysis. Data were collected over a year-long period in each round. Waves 2, 3, and 4 included household modules on food security, and all waves collected data on household composition and demographics, agricultural production, assets, and income. We link households across rounds using unique identifiers, and the analytic sample includes households from Wave 1 who were reinterviewed in one or more subsequent waves in the same location. Household members who left Wave 1 households were tracked in subsequent rounds, but we exclude these split-off households due to a lack of GPS information on their new location. We also exclude the 2% of original households that moved 10 or more km from their Wave 1 location because we do not have information on the timing of their move. If households moved prior to the most recent completed rainy season, including them would miss-specify their climate exposure and potentially bias our results toward the null. If households moved after the most recent completed rainy season, including them could upwardly bias our results, particularly if the most food insecure households moved in response to poor climatic conditions.

Our analysis includes three primary outcome measures. The first measure is the food consumption score (FCS), developed by the WFP (World Food Programme, 2008). FCS is a composite index of the frequency of consumption of foods from eight food groups (main staples, nuts and pulses, vegetables, fruits, meat and fish, milk and dairy, sugar, oil, and condiments) over the seven days prior to the survey (Vaitla et al., 2017). These groups are then weighted by nutrient density using food category weights established by the WFP (World Food Programme, 2008). The FCS is a proxy for food quality and dietary diversity (Jones et al., 2013), and has been validated in multiple low-income countries (Wiesmann et al., 2009). The FCS ranges from zero (most food insecure) to 112 (most food secure). We use a cutoff of 35 to create a binary indicator of food security, with households exceeding this threshold considered food secure (Vaitla et al., 2017; World Food Programme, 2008).

The second measure is the reduced coping strategies index (rCSI), a proxy for how households respond to food shortages that is comparable across different contexts (Maxwell and Caldwell, 2008). The rCSI, a more limited version of the CSI, is derived from a series of five questions on three domains of coping: dietary change, rationing, and help-seeking behavior. Respondents were asked the number of days over the prior week that household members: (1) relied on less preferred foods; (2) borrowed food or relied on help from a friend or relative; (3) limited portion sizes at meals; (4) restricted consumption by adults so that small children could eat; and (5) reduced the number of meals eaten in a day. The rCSI is created by counting the frequency of each strategy and then weighting it based on severity (Maxwell and Caldwell, 2008). The rCSI ranges from zero (most food secure) to 56 (most food insecure) and we use a cutoff of five to create a binary indicator of food security, with households below this threshold considered food secure, as suggested by Vaitla et al., (2017). The questions used in the LSMS questionnaires to construct the rCSI have been validated in Nigeria, Tanzania, and Malawi, and are shown to capture aspects of food insecurity that are overlooked by other measures, such as food expenditures (Bertelli, 2019).

Lastly, we create a composite binary indicator representing whether households were food secure in terms of both FCS and rCSI simultaneously, thus capturing two dimensions of the multidimensional concept of food security.

We then a performed a tetrachoric correlation to determine the extent to which being food secure in terms of FCS is correlated with being food secure in terms of rCSI. Food security in terms of FCS is only moderately positively correlated with food security in terms of rCSI (r=0.26). This suggests that the two indicators capture unique dimensions of food security, with rCSI representing particular phenomena (food shortages and lack of economic access) and FCS capturing different phenomena (dietary diversity and quality of food consumed).

4.2. Climate Data

We use high-resolution gridded precipitation and temperature data derived from satellite imagery. CHIRPS and the newly available CHIRTS_max from the UCSB Climate Hazards Center provide global daily data on precipitation and maximum temperature respectively from 1983-2016 on an approximately 5-km grid (0.05°) (Funk et al., 2019, 2015). CHIRPS/CHIRTS_max have been specifically designed for food insecurity assessment and forecasting in Sub-Saharan Africa as part of the FEWS-NET project, but few studies to date have linked these data to household measures of food insecurity. CHIRPS/CHIRTS_max compare favorably to other high-resolution datasets in Eastern Africa (Dinku et al., 2018; Verdin et al., 2020) and are better able to predict crop yields (Parkes et al., 2019). We link these data to the study households and extract a monthly time series of climate values to 10 km radius buffers centered on the offset EA locations, taking a spatial mean of pixels within the buffer.

We focus on precipitation and temperature conditions from October to May (which we refer to as the rainy season). This time period encompasses the Msimu rains in unimodal rainfall regions as well as the Masika and Vuli rains in bimodal rainfall regions. Climatic conditions during the rainy season are critical for crop and livestock production, and in turn for food security. To measure exposure to local climate variability, we calculate the total rainy season precipitation and average rainy season maximum daily temperature for each EA. We focus on seasonal averages given evidence that these measures serve as better predictors of agricultural yields in sub-Saharan Africa than daily extremes (Michler et al., 2021). We then link these values to households based on their EA and survey month/year. Given that surveys during each wave were conducted over a 14–16-month period, we link households to climate data based on the most recent completed rainy season. See Appendix Figure A1 for timelines of the agricultural calendar and LSMS data collection, and time periods for linking climate and survey data.

4.3. Descriptive Statistics

Table 1 presents descriptive statistics for the analytic sample, which consists of 6,169 observations from 3,029 households that were originally interviewed in Wave 1 and contained no missing values on control variables or food security outcomes. This includes 3,024 Wave 2 observations, 2,777 Wave 3 observations, and 368 Wave 4 observations. Eighty-four percent of household observations were food secure in terms of FCS, 74% were food secure in terms of rCSI, and 65% were food secure on both measures. Total rainy season precipitation averaged 9.63 dm, ranging from 3.99 dm to 22.17 dm. Mean growing season maximum daily temperature averaged 29.5 °C (85 °F), ranging from 23.1 °C to 33.6 °C (73.6 °F to 92.4 °F). At baseline, one-quarter of households were female headed and 57% of household heads had at least a primary school education. Housing quality varied considerably and 40% of households lived below the $1.90/day poverty line. Sixty percent of households earned their primary income from crops and/or livestock, with 56% of households farming between zero and two hectares of land. Lastly, 12% of household heads had lived in the EA for fewer than ten years prior to Wave 1.

Table 1.

Descriptive statistics of household observations used in analysis

	Mean	SD	Min	Max
Outcome variables:
Food secure (FCS)	0.84		0	1
Food secure (rCSI)	0.74		0	1
Food secure (FCS and rCSI)	0.65		0	1
Climate variables:
Total rainy season precipitation (dm)	9.63	2.58	3.99	22.17
Average rainy season maximum daily temperature (°C)	29.50	2.15	23.16	33.55
Baseline control variables:
Female-headed household	0.25		0	1
Age of head	45.62	15.37	18	102
Education of head:
Less than primary	0.43		0	1
Primary or greater	0.57		0	1
Number of members ages 0 to 6	1.21	1.20	0	14
Number of members ages 7 to 15	1.27	1.29	0	9
Number of members ages 16 to 64	2.47	1.43	0	22
Number of members ages 65+	0.22	0.51	0	3
Housing quality index	0.00	1.01	−1.03	2.37
Per capita consumption of under $1.90/day poverty line	0.40		0	1
Primary income source:
Crops and/or livestock	0.60		0	1
Other	0.40		0	1
Farm size:
No farm	0.18		0	1
> 0 Ha and <=1 Ha	0.32		0	1
>1 Ha and <=2 Ha	0.24		0	1
>2 Ha	0.26		0	1
Head resident in EA less than 10 years	0.12		0	1
Interview season:
Dry	0.32		0	1
Rainy	0.68		0	1
N household observations	6169

4.5. Analysis

We then estimate a set of three multivariate binary logistic regression models of the probability of being food secure in terms of FCS, rCSI, and FCS and rCSI simultaneously based on precipitation and temperature conditions during the most recent completed rainy season. To account for additional factors that affect food security, we include a set of household controls measured during Wave 1, which we describe in detail below. Further, we include sampling weights to ensure that our results are nationally representative, as well as fixed effects for EA and survey wave to account for all time-invariant community characteristics such as baseline climate as well as the national time-varying context. With this inclusion, a small number of EAs with no variation on the outcome are dropped from each analysis. We cluster the standard errors at the EA level to account for non-independence among households living in the same EA. Our results can be interpreted as comparing two household observations within the same EA but with different exposures to recent rainy season precipitation and temperature conditions.

Expressed mathematically, the probability of being food secure, Pr(Food secure_hct = 1), is given by the following equation:

$$\text{Pr}(Foodsecur{e}_{hct}=1)={\left[1+e^{-({\alpha }_0+{\alpha }_c+{\alpha }_t+{\beta }_{1}Rai{n}_{hct}+{\beta }_{2}Tem{p}_{hct}+{\beta }_3{X}_{hct})}\right]}^{-1}$$

where 0 is an intercept, c and t are fixed effects for the EA and wave, β₁ is the coefficient for precipitation, Rain_hct is the total rainy season precipitation, β₂ is the coefficient for temperature, Temp_hct is the average rainy season maximum daily temperature, β₃ is a vector of coefficients for the control variables, and X_hct is a vector of controls

Control variables measured at Wave 1 include whether the household is female headed; the household head’s age; the household head’s education (less than primary, primary or greater); the number of household members in different age categories (0-6 years, 7-15 years, 16-64 years, 65 or greater years); a housing quality index based on number of habitable rooms per capita, wall material, floor material, and roof material created using polychoric principle components analysis (Kolenikov and Angeles, 2009; Tusting et al., 2019); whether household per capita consumption was below a poverty line of $1.90/day in 2011 $PPP; whether the household’s primary cash income was derived from crops and/or livestock; the size of area farmed (no farm, >0 to 1 hectare, >1 to 2 hectares, >2 hectares); and whether the household head lived in the EA for fewer than 10 years before the time of the survey. Lastly, we include a control for whether the interview was conducted during the dry or rainy season to account for seasonal differences in food security, given that there is evidence that food prices—particularly for fruits and vegetables—vary seasonally in Tanzania (Bai et al., 2020).

Following the main analysis, we then estimate a set of regression models of economic and nutritional outcomes to provide additional insight into the relationship between rainy season climatic conditions and food security. Household-level economic outcomes measured at Waves 2-3, estimated using OLS models, include logged per capita consumption (annual consumption in 2016 Purchasing Power Parity dollars or PPP $ per household member) and logged crop income (2016 PPP $). These indicators were created from the LSMS data by the Evans School Policy Analysis and Research Group at the University of Washington (Evans School Policy Analysis and Research Group, 2019). Wave 4 is excluded from this analysis because the indicators were not constructed for households in the extended panel. For per capita consumption, precipitation and temperature refer to the most recent completed competed rainy season. For crop income they refer to the 2009-10 rainy season for Wave 2 and the 2011-2012 rainy season for Wave 3 (see Appendix Figure A1).

Individual-level anthropometric outcomes at Waves 2-3 include weight-for-height z-score (WHZ) for young children aged 6-23 months and BMI category for children aged 2-17 years and adults 18 years and older. WHZ and BMI category for children aged 2-17 years were calculated using the zanthro and zbmicat Stata commands (Vidmar et al., 2013). Child BMI category consists of six levels: grade 3 thinness, grade 2 thinness, grade 1 thinness, normal weight, overweight, and obese. Adult BMI category was calculated using the bmi Stata package and consists of eight levels: severe thinness, moderate thinness, mild thinness, normal range, pre-obese, obese class I, obese class II, and obese class III (Linden, 2019). We use OLS to estimate a set of five models: WHZ for young children aged 6-23 months, and BMI category for children aged 2 to 4 years, 5 to 9 years, and 10 to 17 years, and adults aged 18 years or older. Anthropometric outcomes are excluded for Wave 4 because the extended panel did not collect children's birth month and year (necessary to calculate age in months for young children) and only collected anthropometric data for members that were under the age of 15 or women of child bearing age (15-49 years old). In addition to the household-level control variables and fixed effects from the main model, we add individual-level controls including age (measured in months for individuals aged 6-23 months and years for individuals aged 2 years or older); age² to account for potential nonlinear relationships between age and nutritional status; sex; and relationship to the household head. For individuals aged 17 years and younger the categories include child/stepchild, grandchild, and other. For adults aged 18 years or older the categories include head, spouse, child/stepchild, and other.

We then estimate a set of models with precipitation-temperature interactions to examine whether the relationship between precipitation and food security varies across cooler and warmer temperatures. Next, we perform an exploratory, data-driven analysis to examine whether the relationship between climatic conditions and the likelihood of being food secure on both measures varies by sub-population. We add interactions, one at a time, between precipitation and temperature with control variables that were significantly associated with being food secure on both FCS and rCSI simultaneously. The control variables include whether the household is female headed, household head education, number of working age household members (ages 16-64), and housing quality. Each of these are proxies for socioeconomic status and/or human capital resources and may thus be associated with vulnerability to climate-induced food insecurity. For example, poorer households may have fewer resources to invest in climate-resilient agricultural techniques and households with fewer working age members may have more limited options for income generation during periods of adverse weather conditions. Finally, we estimate a set of supplementary models to test the robustness of our main models to alternate specifications. These supplementary models are discussed in more detail in section 5.6 below.

5. Results

5.1. Links between climatic conditions and food security

Table 2 presents results from models predicting the likelihood of being food secure on the three measures. Experiencing greater precipitation during the most recent completed rainy season is significantly positively associated with odds of being food secure in terms of FCS, rCSI, and on both measures simultaneously. Each additional dm of precipitation is associated with a 22% increase in the odds that a household is food secure in terms of FCS, an 11% increase in terms of rCSI, and an 18% increase on both FCS and rCSI. To provide additional insight into the magnitude of this relationship, moving from a year with average rainy season precipitation (10 dm) to a particularly dry year in the 10^th percentile of precipitation (6.6 dm) increases the predicted probability of being food insecure on both measures simultaneously by 13-percentage points. In addition, warmer rainy season temperatures have a marginally significant positive association with being food secure on both measures simultaneously. Other factors associated with a higher likelihood of being food secure on both measures simultaneously are being a male-headed household, having a household head with primary or greater education, having more working age household members (ages 16-64), having a higher housing quality index, and farming 2 or more hectares of land compared to >0 to 1 hectare.

Table 2.

Odds ratios from binary logistic regressions predicting the likelihood of food security based on precipitation and temperature during the most recent completed rainy season

	Food secure (FCS)	Food secure (rCSI)	Food secure (FCS and rCSI)
	Model 1	Model 2	Model 3
Climate variables:
Rainy season precipitation	1.216 ** (0.084)	1.108 * (0.052)	1.179 *** (0.054)
Rainy season temperature	2.006 (0.880)	1.343 (0.411)	1.642+ (0.491)
Baseline control variables:
Female-headed household	0.848 (0.163)	0.639 *** (0.082)	0.699 ** (0.094)
Age of head	0.989+ (0.006)	0.997 (0.005)	0.994 (0.005)
Education of head [less than primary is baseline]:
Primary or greater	1.167 (0.183)	1.518 ** (0.222)	1.384 ** (0.168)
Number of members ages 0 to 6	1.123 + (0.075)	0.936 (0.062)	1.014 (0.059)
Number of members ages 7 to 15	1.168 * (0.079)	0.937 (0.045)	1.022 (0.051)
Number of members ages 16 to 64	1.124 + (0.067)	1.159 ** (0.056)	1.174 *** (0.054)
Number of members ages 65+	1.232 (0.218)	0.881 (0.117)	0.946 (0.139)
Housing quality index	1.712 *** (0.158)	1.765 *** (0.187)	1.857 *** (0.175)
Under $1.90/day poverty line	0.929 (0.168)	0.856 (0.127)	0.896 (0.140)
Primary income source [other is baseline]:
Crops and/or livestock	1.030 (0.144)	0.912 (0.131)	0.902 (0.116)
Farm size [> 0 Ha and <=1 Ha is baseline]:
No farm	0.924 (0.311)	1.308 (0.329)	1.189 (0.305)
>1 Ha and <=2 Ha	0.790 (0.183)	1.208 (0.208)	1.028 (0.153)
>2 Ha	1.130 (0.246)	1.640 * (0.348)	1.447 * (0.271)
Head resident in EA less than 10 years	1.144 (0.261)	1.301 (0.316)	1.243 (0.267)
Interview season [dry season is baseline]:
Rainy season	0.701 (0.217)	0.950 (0.352)	0.875 (0.197)
Pseudo R2	0.16	0.19	0.19
Joint significance:
Climate variables	11.43 **	5.04 +	13.52 **

⁺p<0.10

^*p<0.05

^**p<0.01

^***p<0.001

Notes: Constant and fixed effects for enumeration area and survey wave included in the model but not shown.

Standard errors are clustered on enumeration area and shown in parentheses.

Household observations in enumeration areas with no variation in outcome variable are dropped from analysis. Analytic sample is 4,940 for Model 1, 5,557 for Model 2, and 5,915 for Model 3.

5.2. Interactions between precipitation and temperature

We then estimate a set of models that include precipitation-temperature interactions (results presented in Table S1). Significant interactions exist for being food secure in terms of FCS as well as on both measures. Figure 2 presents the predicted probability of being food secure on both measures based on precipitation and temperature conditions, holding all other variables at their means. Results indicate that the highest probability of food security occurs when conditions are both warm and wet while the lowest probability occurs when conditions during the rainy season are both dry and cool—particularly when total rainfall falls below 11 dm and average maximum daily temperatures fall below approximately 27 °C.

Figure 2.

Predicted probability of being food secure in terms of both FCS and rCSI based on precipitation-temperature interactions.

5.3. Climatic conditions, household welfare, and anthropometry

Table 3 examines a set of additional outcomes to provide greater insight into the linkages between climatic conditions and food security. We find that more precipitation is positively associated with per capita consumption, crop income, and BMI category among children aged 2-4 years. Higher temperatures are positively associated with per capita consumption.

Table 3.

Models predicting per capita consumption, crop income, and child and adult anthropometry based on recent rainy season precipitation and temperature

	Per capit consumption (logged)	Crop income (logged)	WHZ (children 6- 23 months)	BMI category
				Children 2-4 years	Children 5-9 years	Children 10-17 years	Adults 18+ years
	Model 4	Model 5	Model 6	Model 7	Model 8	Model 9	Model 10
Climate variables:
Rainy season precipitation	0.025* (0.010)	0.072* (0.029)	0.015 (0.083)	0.040* (0.020)	0.015 (0.012)	0.019 (0.013)	0.009 (0.009)
Rainy season temperature	0.117* (0047)	0.222 (0.147)	−0.003 (0.575)	−0.063 (0.116)	−0.082 (0.082)	−0.012 (0.079)	0.051 (0.062)
Individual-level variables:
Age			0.130 (0.091)	0.159 (0.254)	0.190* (0.078)	−0.333*** (0.068)	0.0244*** (0.005)
Age2			−0.003 (0.003)	−0.037 (0.042)	−0.014* (0.005)	0.012*** (0.003)	−0.0003*** (0.000)
Sex			0.050 (0.129)	−0.034 (0.038)	−0.032 (0.026)	0.043 (0.028)	0.246*** (0.031)
Relationship to household head [Child/stepchild is baseline for child models; head is baseline for adult model]
Grandchild			−0.478+ (0.266)	−0.055 (0.071)	0.000 (0.039)	−0.061 (0.055)
Other			0.125 (0.317)	0.157 (0.135)	0.006 (0.058)	−0.061 (0.055)	−0.186*** (0.053)
Child/stepchild							−0.114* (0.053)
Spouse							0.0580 (0.037)
R2	0.56	0.42	0.44	0.24	0.19	0.20	0.20
Joint significance:
Climate variables	4.96**	3.39*	0.02	2.50+	1.19	1.15	0.91
N	5,794	3,902	1,091	2,432	3,847	5,097	12,301

⁺p<0.10

^*p<0.05

^**p<0.01

^***p<0.001

Notes: Household-level control variables, constant, and fixed effects for enumeration area and survey wave included in the models but not shown.

Standard errors are clustered on enumeration area and shown in parentheses.

Age is in months for Model 6 and years for Models 7 to 10.

5.4. Sub-population differences

Next, we examine whether the relationship between climatic conditions and the likelihood of being food secure on both measures varies by sub-population. Results, presented in Table S2, indicate that the relationship between climatic conditions and food security does not vary across households by household head gender or education, or by housing quality. However, we find a significant interaction between precipitation and the number of working age household members. Figure 3 presents the predicted probability of being food secure on both measures for households with one and four working age members (the 10^th and 90^th percentile of number of members aged 16-64), holding all other variables at their means. We find a positive relationship between precipitation and food security among both groups; however, the relationship is much stronger for households with one working age member. The average within-EA standard deviation of rainy season precipitation is 2 dm. A 2-dm decrease in total rainy season precipitation from 10 to 8 dm, is associated with a 11 percentage point decline in the predicted probability of food security for households with one working age member (65% to 54%) and just a 4 percentage point decline for households with four working age members (74% too 70%). Further, when total rainy season precipitation falls below approximately 10.5 dm, households with one working age member have a significantly lower probability of being food secure than those with four working age members.

Figure 3.

Predicted probability of being food secure in terms of both FCS and rCSI simultaneously based on interaction between precipitation and number of working age household members (aged 16-64 years), including 95% confidence intervals.

5.5. Supplementary models

Lastly, we estimate a set of supplementary models to test the robustness of our findings. First, we use two alternate specifications for climate exposures: daily extremes and z-scores. The first set of models includes the number of days in the 90^th percentile or higher of rainfall for a given EA as well as the number of days in which the maximum temperature is 32 °C or above (Table S3). The second set uses seasonal z-scores for the most recent completed rainy season compared to all rainy seasons in the EA from 1984-2016 (Table S4). Results using both daily extremes and z-scores are consistent with our main models.

We then perform an analysis stratified by whether the region has a bimodal or unimodal rainfall regime (Table S5) using rainfall regime boundaries derived from the World Food Programme (2010). Results are consistent with the main models. Next, we add rainfall and temperature conditions during the most recent dry season given that conditions outside the growing season may also affect food security (Table S6). The effect of rainy season precipitation on the three food security outcomes remains positive and significant, while dry season conditions are not significantly associated with food security.

Finally, we examine alternate specifications for our food security measures. Wiesmann et al. (2009) note that an FCS cutoff of 35 underestimates food insecurity when benchmarked against caloric consumption and that no universal cutoff exists. To account for this, we estimate an OLS model (Table S7) with FCS as a continuous outcome. Results are consistent with the main models. For rCSI, we estimate an ordered logit model using a three-category variable based on Brander et al. (2021) with cutoffs of 0-4 to indicate food secure or mildly food insecure, 5-10 to indicate moderately food insecure, and >=11 to indicate severely food insecure (Table S8). In addition, because 38% of observations in the analytic sample did not engage in any coping strategies, we estimated a binary logit model with an outcome of one if rCSI is above zero (Table S9). Lastly, we estimate an OLS model with rCSI as a continuous outcome (Table S10). Results for the three rCSI specifications are consistent with the main models.

6. Discussion and Policy Implications

This study examined the linkages between recent climatic conditions and household food security in Tanzania, a country in which 65% of employed people work in agriculture and 31% of the population experiences undernutrition (United Nations, 2019; von Grebmer et al., 2019). Most studies on climatic conditions and food security have relied on indirect—and insufficient—proxies for food security such as caloric consumption or child anthropometry (Headey and Ecker, 2013; Maxwell et al., 2014; Mulmi et al., 2016). Recommended instead is the use of a combination of experiential indicators that capture different underlying dimensions of food security (Maxwell et al., 2014; Vaitla et al., 2017). We utilized two such indicators: Food Consumption Score (FCS), which serves as a measure of dietary diversity and the nutritional value of foods consumed by the household, and the reduced Coping Strategies Index (rCSI), which addresses household coping behaviors in response to food shortfalls, including dietary change, rationing, and help seeking. Among the study population, the two indicators were only weakly correlated with one another, suggesting that they indeed describe different aspects of food security.

Results indicate that greater precipitation during the most recent completed rainy season is positively associated with the likelihood of being food secure in terms of FCS, rCSI, and both FCS and rCSI simultaneously. When accounting for interactions between precipitation and temperature, we discovered that conditions that are both warm and wet are associated with the highest probability of food security, while conditions that are cool and dry are associated with the lowest probability. These results suggest that experiencing low rainfall, especially in combination with cool rainy season maximum daily temperatures, negatively affects households’ ability to access food, leading to reduced dietary diversity, food shortages, and in turn to coping strategies such as limiting portion sizes or restricting food consumption by adults in order to feed young children.

Our finding that more rainy season precipitation is beneficial for food security in Sub-Saharan Africa aligns with findings from several other studies (Cooper et al., 2019; Tankari, 2020). This is likely driven by improved crop yields, which in turn bolsters food availability and household income. Tanzania is heavily dependent on maize production, as maize is grown throughout the country and serves as the main food crop for most households (Arce and Caballero, 2015). A study of maize in Tanzania over the period 2009-2019 found that yields benefitted from greater rainfall as well as higher minimum temperatures (Laudien et al., 2020). Indeed, maize is very sensitive to frost damage and thrives when temperatures range from 28 °C to 32 °C (Sánchez et al., 2014). Maize yields are likely to benefit from warming in cooler parts of sub-Saharan Africa (Lobell et al., 2011). This suggests in Tanzania, maize production may increase over the short term from warming temperatures, especially if combined with ample rainfall. However, these gains are unlikely to be sustained over the longer term as extreme heat becomes more common (Conway et al., 2017).

In addition, we found that households with a greater number of working age members were less sensitive to the negative effects of low precipitation on food security. At low levels of rainy season precipitation, the likelihood of being food secure on both measures was 24 percentage points higher among households with four working age members (aged 16-64 years) compared to those with one working age member. This suggests that households with a greater supply of labor are better able to withstand the negative impacts of drought conditions. In response to low precipitation, members may increase labor allocation to agricultural production and/or diversify into non-agricultural income generation activities (Asfaw et al., 2019).

Lastly, we explored additional proxies for food security and discovered that greater rainy season precipitation is positively associated with per capita household consumption, crop income, and anthropometric indicators of nutrition for children aged 2-4. Higher temperatures were linked with greater per capita consumption but lower height-for-age among children aged 6-23 months. The finding that higher precipitation is beneficial for multiple indicators suggests that greater rainfall does indeed improve food availability and nutritional outcomes, at least in part through a household income mechanism. Children aged 2-4 years appear to be the most vulnerable age group to drought conditions, possibly due to high nutritional needs during this important period of growth. Infants and children under two may be buffered from the effects of drought due to breastfeeding. However, this age group is negatively impacted by higher temperatures, suggesting that nutritional outcomes among infants and young children are driven by a complex set of mechanisms that may differ from older children and adults (Randell et al., 2020).

Climate projections for Tanzania predict an overall warming trend but spatially variable changes in precipitation, with annual rainfall predicted to increase in the north and northeast but decrease in the south (Conway et al., 2017). In addition, rainfall conditions are projected to become more variable, with both an increase in the number of dry days and heavy rain events (Conway et al., 2017). Given the sensitivity of crop production in Tanzania to rainfall and temperature conditions, it is likely that climate change will further impact agriculture, and in turn food security, leading to differential effects across the country. Households in regions experiencing reduced annual rainfall, more frequent or prolonged drought conditions, and/or an increased prevalence of extreme heat are likely to be the most vulnerable.

This study has important policy implications for Tanzania as well as for other low- and middle-income countries, especially those heavily dependent on maize production. Several interventions may foster resilience to climate change, particularly in areas experiencing more frequent and severe drought conditions. Programs to buffer against climate-induced food insecurity include providing drought tolerant maize, increasing access to agricultural extension services, scaling up agricultural index insurance, improving uptake of soil and water conservation practices, and expanding actions based on drought early warning systems.

Indeed, on rainfed farms in Southern Africa that experienced moderate drought conditions, drought-tolerant maize yield was 15% higher on average than other maize varieties and resulted in net returns of between $13/ha and $95/ha (Paul, 2021). Agricultural extension services, which provide farmers with training and guidance on emergent agricultural technologies, acts as key determinants of the uptake of climate-resilient crops such as drought-tolerant maize (Acevedo et al., 2020). Further, a study in Tanzania and Mozambique found that a multipronged approach of providing farmers with drought-tolerant maize combined with index insurance improved drought resilience even further (Boucher et al., 2021). Soil and water conservation practices are associated with increased maize yields under both typical weather conditions as well as under precipitation and temperature shocks (Arslan et al., 2017). Finally, drought early warning systems can be utilized to trigger early responses that buffer households against future yield losses months before harvest. For example, a study in Kenya predicted future drought-related maize yield losses with high accuracy and found that providing farmers with ex-ante cash transfers was highly cost effective (Guimarães Nobre et al., 2019). Programs such as these, particularly when implemented alongside broader policies aimed at reducing poverty and improving agricultural production, will help buffer populations in vulnerable regions against food insecurity amid a changing climate.

Appendix

Figure A1.

Timeline of agricultural calendar, LSMS data collection, and time periods for linking climate and survey data.