2.1 Climate change and vulnerability

Farm households throughout Sub-Saharan Africa are particularly exposed to weather induced risks, due to the preponderance of rain-fed production and imperfect market conditions. Climate change exacerbates these risks by increasing the probability and severity of adverse weather conditions. Furthermore, severe climatic events such as droughts, floods, and heat waves are expected to increase in frequency and intensity over time (Nelson and van der Mensbrugghe, Reference Nelson and van der Mensbrugghe2013; IPCC, 2014). In the absence of measures to reduce the vulnerability of farmers to these events, significant negative impacts on food security are expected (Roy et al., Reference Roy, Tschakert, Waisman, Abdul Halim, Antwi-Agyei, Dasgupta, Hayward, Kanninen, Liverman, Okereke, Pinho, Riahi and Suarez Rodriguez2018). Hence, climate change not only represents a threat to incomes today, but also makes them less predictable by changing the probability distributions in ways that are difficult for households to incorporate into decision making (Thornton and Lipper, Reference Thornton and Lipper2014).

In most cases, extreme weather events increase vulnerability of rural households through their effects on crop production and income (Banerjee, Reference Banerjee2007; Dercon and Christiaensen, Reference Dercon and Christiaensen2007; Mueller and Quisumbing, Reference Mueller and Quisumbing2010; Wineman et al., Reference Wineman, Mason, Ochieng and Kirimi2017; Hill and Fuje, Reference Hill and Fuje2018; McCarthy et al., Reference McCarthy, Kilic, de la Fuente and Brubaker2018; Michler et al., Reference Michler, Baylis, Arends-Kuenning and Mazvimavi2019). Although limited, empirical evidence suggests that households subject to severe climate events often experience increasing levels of vulnerability related to large losses in agricultural income.

Households can adopt risk management practices – such as SLM practices that moderate negative impacts of weather extremes on crop yields (FAO, 2001, 2007). They can also implement risk-coping strategies ex post, such as labour reallocation, sales of durables and livestock, and access to transfers from friends and relatives. Yet, despite adoption of risk management and coping strategies, the empirical evidence suggests that these mechanisms are never more than partial, and that consumption shortfalls remain high when rural households face extreme shocks (Alderman and Paxson, Reference Alderman, Paxson and Bacha1994; Dercon, Reference Dercon2005; Baez and Mason, Reference Baez and Mason2008).

2.2 Sustainable land management practices and diversification strategies

In Zambia, maize is both the primary crop grown by small-scale producers and the national staple food. As both a cause and a consequence, agricultural policy in the recent past has focused predominantly on the maize sector. This includes significant public expenditure on output market and input subsidies, as well as frequent use of maize trade restrictions to affect prices (Sitko et al., Reference Sitko, Chamberlin, Cunguara, Muyanga and Mangisoni2017).

Large efforts, advocacy and investments have been made in the country to promote the adoption of farming practices such as conservation farming,Footnote ² which is a set of SLM practices, agroforestry and improved fallows in order to improve and stabilize maize yield while offering income benefits through diversification (Chidumayo, Reference Chidumayo1987; Umar et al., Reference Umar, Aune, Johnsen and Lungu2011; Arslan et al., Reference Arslan, McCarthy, Lipper, Asfaw and Cattaneo2014). SLM practices incorporating MSD, crop rotation, intercropping, residue management, agroforestry, and soil and water conservation structures, are meant to generate climate adaptation benefits through impacts on improved water retention capacity and soil nutrients, and reduce erosion. For instance, the Zambia National Farmers Union started promoting Conservation Agriculture (CA) in 1995 through the Conservation Farming Unit. In 2004, CA became a national priority and this focus was echoed by a number of initiatives and projects supported and implemented by various NGOs, as well as international agencies and organizations including the Food and Agriculture Organization (FAO) and the World Bank, among others (Arslan et al., Reference Arslan, McCarthy, Lipper, Asfaw and Cattaneo2014).

Several studies show that CF, when implemented fully on experimental plots, has the potential to mitigate the negative effects of climatic shocks by increasing water productivity, water infiltration and soil moisture buffering capacity (Giller et al., Reference Giller, Witter, Corbeels and Tittonell2009; Chikowo, Reference Chikowo2011). Despite these benefits, disadoption of these practices at the household level is common (Arslan et al., Reference Arslan, McCarthy, Lipper, Asfaw and Cattaneo2014), although aggregate adoption levels are reportedly increasing over time (Baudron et al., Reference Baudron, Mwanza, Triomphe and Bwalya2007; Umar et al., Reference Umar, Aune, Johnsen and Lungu2011). The stubbornly low, partial and volatile adoption levels are related to multiple constraints, including the time it takes until positive returns are obtained in low productivity settings (up to 10 years), competition from livestock, labour constraints as well as cessation of other incentives (input support) provided by some promoters (Giller et al., Reference Giller, Witter, Corbeels and Tittonell2009; Nkala et al., Reference Nkala, Mango, Corbeels, Veldwisch and Huising2011).

In addition to these practices, diversification strategies in terms of crop, income and livestock are considered important measures to diversify and manage risks on income. Diversification can be adopted by agricultural households ex ante as a risk-management and income smoothing strategy (Smit and Wandel, Reference Smit and Wandel2006), as well as after a shock (i.e., ex post) to cope with the negative effects it generates (Davies, Reference Davies1993; Murdoch, Reference Murdoch1995). Climate change not only decreases incomes when weather shocks occur, but also makes them less predictable in ways that are difficult for households to incorporate into their decision making (Thornton and Lipper, Reference Thornton and Lipper2014). Empirical evidence shows that diversification may help farmers deal with droughts and other weather shocks (Di Falco and Chavas, Reference Di Falco and Chavas2009; Cavatassi et al., Reference Cavatassi, Lipper and Narloch2011; Macours et al., Reference Macours, Premand and Vakis2012; Arslan et al., Reference Arslan, Cavatassi, Alfani, McCarthy, Lipper and Kokwe2018). Analyses to test the effects of diversification under extreme weather events such as El Niño, however, are rare, with the exception of Maggio and Sitko (Reference Maggio and Sitko2019), who use the ENIAS data to assess the adoption of adaptive cropping strategies in response to seasonal forecast information.

3.1 Sampling frame

This study was conceived while Zambia was in the midst of the El Niño crisis, and the sampling frame and the empirical strategy were designed to assess the direct impacts of El Niño on smallholders’ maize yields and total incomes, as well as to identify relevant interventions to guide policy in the wake of such crises. These features set this study apart from others in the literature that rely on pre-existing data to identify shocks, which are hard to predict and hence difficult to mobilize in the midst of an unfolding crisis to establish panel data.

The starting point for this analysis is the nationally representative household data from the 2015 wave of the RALS collected by the Central Statistics Office (CSO, 2015) in collaboration with Michigan State University and the Indaba Agricultural Policy Research Institute (IAPRI). The survey is designed to be representative of rural farm households at national and province levels and covers a sample of 7,934 households.Footnote ³ RALS includes detailed information on agricultural (crop and livestock) production and sales, off-farm activities and other income sources, along with household demographic characteristics and social capital indicators.Footnote ⁴

RALS 2015 provided a rich background for the design of the ENIAS sample and questionnaire, which was initiated in response to the delayed onset of the rainy season due to the El Niño at the beginning of the 2015–2016 rainy season. The FAO-EPIC programme of work, in collaboration with FAO Zambia office and IAPRI, implemented the ENIAS to analyse the impacts of El Niño on maize yields, and to identify agricultural and livelihood strategies that successfully improve farmers’ resilience to droughts, as well as to investigate the types of policies and institutions needed to improve resilience to such shocks.

The sampling frame for ENIAS was defined by using PSM at the Standard Enumeration Area (SEA) level in order to match severely affected areas in the RALS 2015 data with those that were not severely affected to ensure that the sample has enough households in both areas for analysis. The definition of ‘severely affected areas’ was based on the most recent Zambia Vulnerability Assessment Committee (ZVAC) Situation Report at the time, which was released in January 2016.

Given the fact that the northern parts of the country were experiencing normal or above normal rainfall, all of Luapula, Northern and North-Western and most of Copperbelt and Muchinga provinces were excluded from the sampling frame. This choice was also driven by the significant differences between the agro-ecological and cropping systems of the excluded areas and the severely affected areas, meaning they provide limited opportunities for matching. Out of the 35 severely affected districts, 22 that were covered in the RALS 2015 surveys were used to create a sampling frame for ENIAS using PSM. From these districts, 149 SEAs were selected comprised of 60 severely affected (treatment) and 89 not severely affected (control) SEAs, and a random sample of 9–10 households from the RALS 2015 roster was interviewed in each SEA, yielding a final sample of 1,311 households.Footnote ⁵ Figure 1 shows the 35 severely affected districts (in red) as reported in the ZVAC (2016) report together with the 28 districts included in the ENIAS sample, which are marked with blue squares.Footnote ⁶

Figure 1. Severely affected districts as reported in the ZVAC (2016) report and sampled districts.

Note: The 35 severely affected districts as reported in the ZVAC Report are coloured in red, whereas the blue squares indicate the 28 districts included in the ENIAS sample.

Source: ZVAC, 2016.

The distribution of the households in the final sample across provinces is provided in table 1.

Table 1. Distribution of interviewed households by province and sample type

Note: Given that the RALS sample was much larger than the ENIAS sample, a randomly selected list of replacement households were provided to enumerators for each selected SEAs. In cases of no response/contact after three trials, households from this list were interviewed.

Source: Authors’ elaboration.

The resulting household panel data are merged with rainfall information at the ward level using geo-referenced ward boundaries (there are 136 wards in the final sample).Footnote ⁷ Rainfall data are from ARC2 of the National Oceanic and Atmospheric Administration's Climate Prediction Center for the period of 1983–2016. ARC2 data are available on a daily basis and have a spatial resolution of 0.1 degrees (~10 km).Footnote ⁸ We use these data to construct our shock variable, which is defined as a dummy variable that identifies wards in which rainfall between November 2015 and February 2016 fell below the minimum of long-run average rainfall of this period.Footnote ⁹

We created other rainfall variables to trace historical trends in rainfall variation that are closely linked with agricultural production as well as the adoption of livelihood strategies with implications for vulnerability and welfare of small farmers. This novel dataset provides a unique opportunity to understand the impacts of shocks like El Niño that are expected to become more frequent and severe in Zambia using a robust empirical methodology detailed in the next section. It also facilitates a thorough understanding of the agricultural practices and livelihood strategies that can buffer household production and welfare from the impacts of such shocks to drive policy recommendations.

3.2 Empirical strategy

In order to identify direct impacts of El Niño on smallholders, we define two estimating equations, one for maize yield and one for total gross income per capita, as follows:

(1)\begin{equation}{Y_{it}} = \alpha + \beta E{N_{i16}} + \gamma {R_{kt}} + \delta {X_{it}} + \varphi {P_{it}} + \; \vartheta {P_{it}}\ast E{N_{i16}} + {\varepsilon _{it}},\end{equation}

where ${Y_{it}}$ is the outcome variable (maize yield in kg/ha, or the value of total gross income per capita, both in logarithms) for the $i\textrm{th}$ household $(i = 1, \ldots ,n)$ at time t (t = 2015, 2016); EN represents the El Niño drought shock which is equal to 1 if between November 2015 and February 2016, total rainfall in each ward was below the minimum of long-run average rainfall; ${R_{kt}}$ are the rainfall variables at the ward levelFootnote ¹⁰ (k = 1,…,136); ${X_{it}}$ is a vector of household level variables including socio-demographic characteristics, wealth and social capital indicators at time t; ${P_{it}}$ are practice, diversification and policy variables that capture the potential ex ante measures, diversification strategies and relevant policies that are expected to ameliorate the impact of the shock on the outcomes; and the ${P_{it}}\mathrm{\ast }E{N_{i16}}$ are interaction terms between these variables and the shock indicator.Footnote ¹¹ The error term ${\varepsilon _{it}}$ is composed of a normally distributed term independent of the regressors $({u_{it}})$, and time-invariant unobserved effects ${\nu _i}$.

We use the Hausman test to assess whether fixed effects (FE) or random effects (RE) should be used to model time-invariant heterogeneity (Wooldridge, Reference Wooldridge2002, Reference Wooldridge2009). We reject the hypothesis that RE models, which consider unobservables as a random variable (uncorrelated with covariates) whose probability distribution can be estimated from data, are consistent. Because FE models prevent the use of time-invariant variables, some of which are critical for our model, we use the correlated random effects (CRE) model, also referred to as the quasi-FE model. The CRE controls for possible additional correlations between time-varying explanatory variables and RE by including the means of the time-varying characteristics as regressors in the analysis, parameterizing the distribution of ${\nu _i}$ and allowing the ${X_{it}}$ and ${P_{it}}$ to be correlated with ${\nu _i}$ (Mundlak, Reference Mundlak1978; Chamberlain, Reference Chamberlain, Griliches and Intriligator1984; Wooldridge, Reference Wooldridge2002, ch. 16; Wooldridge, Reference Wooldridge2009).Footnote ¹²

In addition to analysing the average impact of the El Niño shock on yields and incomes, we test the following hypotheses to guide future policies:

1. (i) Agricultural practices adopted have no effect on maize productivity under average shock exposure conditions ($\hat{\varphi }$ = 0).

2. (ii) These practices do not have a different effect on productivity under extreme shock conditions posed by El Niño ($\hat{\vartheta }$ = 0).

Mathematically, the same hypotheses are tested in income models, where the main focus is on the average effect of diversification and policy variables, and their interactions with the El Niño shock. We cluster the error terms at the ward level for all models to control for potential correlation across households in the same ward.