Drought shocks and labour reallocation in rural Africa: evidence from Ethiopia

Abstract

We study how rural households in Ethiopia adapt to droughts through labour reallocation. Using three waves of panel data and exploiting spatio-temporal variations in drought exposure, we find that households reduce on-farm work and increase off-farm self-employment in response to both short-term and persistent droughts, without abandoning family farming. Diversification into off-farm activities is driven by drought-related productivity declines in agriculture and contributes to consumption smoothing and food security. Households with better access to financial services are more likely to reallocate labour off-farm. Our results highlight the importance of strengthening the rural non-farm economy to enhance rural households’ climate resilience.

1. Introduction

Extreme weather events—such as droughts—have become more frequent with climate change and have negative impacts on farm production and income (Schlenker and Lobell, 2010; Lobell, Schlenker and Costa-Roberts, 2011; Chavas et al., 2019; Ortiz-Bobea et al., 2021). Developing countries, particularly in sub-Saharan Africa, where agriculture is the mainstay of poor people’s livelihoods, bear the brunt of these risks. The literature has looked at various ways in which rural households adapt to weather shocks, including asset sales, formal and informal insurance or adoption of climate-smart technologies. However, these adaptation strategies are often prohibitively costly, ineffective or unsustainable (Dercon and Krishnan, 2000; Giné and Yang, 2009; Karlan and Morduch, 2010; Dercon and Christiaensen, 2011). Much less is known about the extent to which households reallocate labour as a response to weather shocks, especially by shifting from farm to off-farm work, and how effective such reallocation is in protecting household welfare.

Weather shocks can prompt rural households to reallocate labour in different ways. Households may diversify their income sources. As off-farm jobs are typically less affected by weather disruptions, a certain shift from farm work to off-farm wage work is likely (Branco and Féres, 2021), to the extent that local labour markets have the capacity to absorb additional labour during times of weather shocks. If this is not the case, self-employment in non-agricultural businesses can be an alternative. Yet another alternative would be temporary or permanent migration to regions less affected by weather shocks or with better employment opportunities (Young, 2013; Rana and Qaim, 2024).

In this paper, we study labour reallocation decisions of rural households as a response to extreme weather shocks in the context of Ethiopia. We exploit spatio-temporal variation in exposure to droughts to look at the effects of short-term and persistent drought shocks on the probability of a household to be involved in farm work, off-farm wage employment and self-employment, as well as the labour time allocated to these employment categories. We also analyse to which extent the labour allocation decisions help smooth household consumption in the event of a drought shock. Ethiopia provides an interesting context for this study. First, in addition to economic vulnerability, Ethiopia exhibits a high degree of climate vulnerability, with a long history of droughts and an increasing frequency of extreme weather events (Viste, Korecha and Sorteberg, 2013; Mekonen, Berlie and Ferede, 2020). Second, Ethiopia is one of the most populous countries in Africa, and 80 per cent of its rural population are employed in agriculture (UN-DESA, 2019). Third, agriculture in Ethiopia is predominantly smallholder farming with limited access to markets and advanced production technologies, resulting in widespread poverty.

Our results suggest that exposure to both short-term and persistent droughts has two main effects. First, it increases the likelihood of off-farm self-employment (OFSE) and reduces the likelihood of farm wage employment. Second, it increases the labour hours allocated to OFSE and reduces the labour hours allocated to on-farm wage and self-employment. Our results are consistent with droughts causing lower agricultural productivity and frictions in the labour market, leading to lower economic prospects in farm wage and self-employment and limited non-agricultural wage employment opportunities. We confirm the robustness of the findings using various empirical specifications. Finally, we show that OFSE is consumption smoothing.

Our study is related to the evolving literature on climate change adaptation and economic responses to weather shocks in developing countries. For instance, Di Falco, Veronesi and Yesuf (2011) use data from smallholder farms in Ethiopia to show that the adoption of climate-smart technologies can help to increase food crop productivity. However, many smallholders are unable to adopt suitable farming innovations due to limited access to credit and information. Other studies look at links between weather shocks and labour market outcomes. For instance, Jessoe, Manning and Taylor (2018) use data from Mexico to show that in hot years, employment levels in wage work and non-farm jobs are reduced. In India, Jaychandran (2006) shows that weather-induced productivity shocks negatively affect poor rural households by significantly driving down wages. Emerick (2018) estimates that increased agricultural productivity due to abnormally high rainfall leads to an increase not only in agriculture but also in other local sectors due to sectoral linkages. On the other hand, Branco and Féres (2021) use data from Brazil to show that rural farming households increase their labour supply to non-agricultural sectors during droughts.

Colmer (2021a) finds that temperature-driven reductions in the demand for agricultural labour are correlated with increases in non-agricultural employment in India. This implies that the capacity of non-agricultural sectors to absorb workers might play a significant role in mitigating the economic impacts of negative agricultural productivity shocks. A few studies also examine links between weather shocks, child labour and education. For instance, Colmer (2021b) finds that increased rainfall variability is associated with less child labour and more schooling in rural Ethiopia, a finding the author describes is consistent with diversification strategies. However, the effects on child labour depend on a variety of socioeconomic conditions (e.g. Alam, Pörtner and Simpson, 2022; Nordman, Sharma and Sunder, 2022). Furthermore, the effects of weather shocks on agricultural labour, both child and adult labour, may also depend on the adoption of climate-smart farming technologies (Fontes, 2020).

We contribute to these evolving bodies of literature in two important ways. First, while most existing studies focus on the effects of weather shocks in one single period,¹ we look at short-term droughts and persistent droughts spanning over several years. Second, beyond our focus on labour reallocation, we also analyse effects of this reallocation on household welfare. A few previous studies investigate effects of weather shocks on welfare and interpret statistically insignificant results as evidence of successful adaptation (Emerick, 2018; Gao and Mills, 2018; Aggarwal, 2021), yet the adaptation mechanisms are not studied explicitly. In our study, we show that labour reallocation to OFSE protects households from the negative consequences of drought on consumption and food security.

The rest of this paper is organised as follows: The next section describes the conceptual framework, while Section 3 discusses the data used. We present the empirical strategy and results in Sections 4 and 5, respectively, while Section 6 concludes.

2. Conceptual framework

To study household labour allocation decisions, we apply a household production function framework, in which the household is the unit of production, consumption and decision-making (Udry, 1996). The household maximises utility by allocating available labour across different activities, such as farming, off-farm work and leisure, subject to resource constraints and the available production technology. An increase in agricultural productivity implies higher returns to agricultural inputs, thus attracting more labour into this sector (Becker, 1962). In contrast, weather shocks—such as drought—reduce agricultural productivity and income, thus lowering returns to agricultural labour and leading to a shift of labour away from farm towards off-farm economic activities (Lewis, 1954; Colmer, 2021b).

When faced with a reduction of agricultural productivity, a farm household may allocate (some of) its labour to off-farm employment. This can include wage employment—both in agricultural and non-agricultural activities—and OFSE. The availability and returns to off-farm employment depend on market wages and prices, which are influenced by local market conditions. The farm household may allocate more labour to off-farm work if this is less risky and (or) the returns are expected to be higher than in own farming.

However, out of the three off-farm employment alternatives that exist in principle, not all appear equally plausible in situations of drought. First, wage employment in agriculture is expected to be negatively affected by drought in the same way as own-farm employment. This is because weather shocks tend to be spatially concentrated and affect all local farmers at the same time. Hence, local agricultural employment opportunities and wages are expected to decline during drought, especially if out-migration is constrained (Jayachandran, 2006).

Second, at least in the short run, non-agricultural wage and self-employment are expected to be less affected by weather shocks and may, hence, offer higher returns to labour than agricultural employment. However, given widespread market failures, non-agricultural wage employment opportunities are typically scarce in rural areas, meaning that expected labour adjustments are not always feasible. Instead, households may turn to non-farm self-employment in their own small businesses, which are not always very lucrative (Haggblade, Hazell and Reardon, 2010; Davis, Di Giuseppe and Zezza, 2017). In the long run, it is also possible that non-agricultural employment is negatively affected by persistent weather shocks, as income losses in agriculture can also spill over to other local sectors.

It is also worth noting that household decisions to reallocate labour may change over time. For instance, households may gradually use on-farm adaptation measures, such as the adoption of climate-smart technologies, thus decreasing their need for extensive labour reallocation in response to weather shocks. Other households may gradually abandon their own farming, thus increasing their labour supply to non-agricultural activities over time. In summary, how farm households respond to short-term and persistent weather shocks through labour reallocation and what this means for household welfare are important empirical questions that we address in this study in the context of Ethiopia.

3. Data

We combine data from two main sources. First, we use household data from the Ethiopia Socioeconomic Survey (ESS), a part of the World Bank’s Living Standards Measurement Study. Second, we use weather data on temperature and rainfall from the National Oceanic and Atmospheric Administration (NOAA). These data are explained in more detail below.

3.1. Household data

We use data from the 2011, 2013 and 2015 waves of the ESS to construct a panel of rural households. As we use panel data regression models with fixed effects, we only include households that were surveyed at least twice, leading to 9,968 household observations.² These data are nationally representative for rural areas of Ethiopia.

The main outcome variables, i.e. farm and off-farm wage and self-employment, are constructed based on the information available in the employment module of the household survey. The employment module contains information on the employment status of all household members aged 15 years and older in the last 12 months before the survey. We aggregate this information and create two household-level measures of employment, namely a dummy variable that is equal to one if any member of the household participates in each employment category (i.e. extensive margin), and a continuous variable measuring the percentage share of the household’s weekly hours in each employment category (i.e. intensive margin). We also calculate the share of household members aged 15 years and older engaged in each employment category.

In terms of income variables, we calculate total farm and off-farm wage and business income, using data on wages, earnings from self-employment and other income sources. Variables on food and non-food consumption over the last 12 months before the survey are derived from the household expenditures modules.³ To construct farm-related variables—such as land productivity, labour productivity (agricultural output value per labour-day), hired labour, crop and livestock income—we combine information from the agriculture and livestock modules of the questionnaire. Finally, we construct a series of household control variables, including gender, education, and age of the household head, family size, total land size, tropical livestock units (TLUs), and a dummy variable indicating access to formal financial services, specifically insurance and credit.

Table 1 presents sample summary statistics. Most households (81 per cent) are self-employed on their farms. Both on-farm (2 per cent) and off-farm (9 per cent) wage employment are low. Yet, 23 per cent of the households are engaged in OFSE, which is the most common income diversification strategy in rural Ethiopia, as also pointed out by Bachewe et al. (2020). The average annual household consumption expenditure is Birr 20,280, of which food consumption accounts for 81 per cent. Such a high food expenditure share is a clear indication of the low average living standard of rural households in Ethiopia. Extended summary statistics and variable descriptions are shown in Appendix Table A1 in the supplementary data.

3.2. Weather data

We extract gridded daily rainfall and maximum and minimum temperature data from the NOAA Climate Prediction Center covering the period 1980–2022.⁴ The gridded daily rainfall in millimeters (mm) and surface temperature in degrees Celsius (°C) datasets have a spatial resolution of 0.50-degree by 0.50-degree latitude-longitude grid nodes. We leverage the enumeration area—equivalent of a village or cluster—geolocations to match the weather data with the household data.

Our main explanatory variable is drought, which we define as a continuous variable, namely as the number of dry months within the last year or, alternatively, within the last growing season before the survey. Drawing on the existing literature (Burke and Emerick, 2016; Lee, Im and Bae, 2019; Kakpo, Mills and Brunelin, 2022), we calculate this drought variable as follows.

First, for each month of the year before the survey, we generate rainfall z-scores:

where RF_cmt is the total rainfall in cluster c (same as EA) in month m of year t; |${\overline {RF} _{cm}}{\ }$|is each cluster’s 30 year (1981–2010) historical rainfall mean for a given month, while |$RF_{cm}^{SD}$| is each cluster’s historical (1980–2010) standard deviation of rainfall for a given month. This z-score corresponds to the Standardised Precipitation Index in McKee, Doesken and Kleist (1993).⁵ A z-score less than or equal to −1 indicates a drought month. Second, we sum up the number of drought months for each year before a survey wave. We refer to this variable as ‘short-term drought’, i.e. drought recorded over the last year before the survey. Additionally, we construct cumulative measures of drought by summing up the number of drought months recorded over periods of 2 and 3 years before the survey. We refer to these as measures of ‘persistent drought’.

Given that the effects of drought can vary by agricultural season (e.g. Kakpo, Mills and Brunelin, 2022), we also generate the drought variable for the crop-growing season as the aggregate of drought months within the February–September window. Our definition of the growing season draws on the classification established by the Ethiopia Meteorological Agency,⁶ whereby we combine both the short and the long rain seasons into one.⁷

Finally, to account for the fact that the occurrence and effects of drought are likely reinforced by extreme temperatures, we also generate temperature shock indicators as auxiliary weather shock proxies, which we measure as the number of hot months over the last year, the last growing season, and the last dry season before the survey. Hot months are defined as the months with temperature z-scores greater than or equal to 2, indicating the occurrence of extreme temperatures.

Panel C of Table 1 presents the summary statistics of selected weather variables. On average, households experience 1 drought month in a year and approximately 0.7 and 0.3 drought months during the growing season and the dry season, respectively. Substantial variation in drought occurrence and intensity over time and space can be seen in Figure 1.

4. Empirical strategy

4.1. Estimating labour reallocation effects

We estimate the effects of drought shocks on household labour allocation decisions at the extensive and intensive margins using the following regression model:

where i, c, d and t subscripts correspond to household, cluster, district (woreda) and time (year), respectively. L_icdt corresponds to household labour outcomes, i.e. farm and off-farm wage and self-employment dummies or, alternatively, the percentage share of weekly hours allocated to each employment category in the survey year. D_cdt−1 is our main explanatory variable and corresponds to the number of drought months in the cluster in which household i is located, measured over the last year or the last growing season prior to the survey (t − 1).

We include a vector of time-variant auxiliary weather variables at the cluster level, W_cdt (temperature shocks, monthly average temperature, monthly average rainfall), to differentiate drought shocks from other weather variations. We also control for a vector of household socioeconomic characteristics, X_icdt. We account for time-invariant unobserved heterogeneity at the district level by including district fixed effects, θ_d. Additionally, we include year fixed effects, µ_t, to account for country-wide shocks that would affect labour market conditions. We cluster standard errors at the cluster level.

We exploit spatio-temporal variation in individual households’ exposure to drought shocks for identification. Our identification strategy relies on the assumption that droughts are exogenous, and—conditional on the controls and district and year fixed effects—there are no time-varying differences to drive household labour allocation decisions other than changes in weather conditions. In Appendix Table A2 in the supplementary data, we present a balance test showing that exposure to drought is not correlated with observable household characteristics and is thus plausibly exogenous. As such, our coefficient β in equation (2) can be interpreted as the effect of 1 extra month of drought during the year (or during the growing season) on household labour allocation. We are also interested in possible heterogeneous effects by differentiating between households of different family size, market proximity, and those with differences in land ownership and access to formal financial services. Details of the analysis of heterogeneous effects are provided in the supplementary data.

As noted by Branco and Féres (2021), a crucial aspect of how drought shocks influence household decisions is the precise timing of the impacts. Plausibly, the effect of last year’s drought on household labour supply in the current year may also depend on the past distribution of droughts. Similarly, exposure to drought in the current period may have varying effects on future household labour allocation decisions. Given this context, we estimate three additional sets of regressions to examine how the effects of droughts on household labour supply evolve over time. First, we estimate regressions using our measure of persistent drought (number of drought months observed over the last 2 and 3 years combined). Second, we use separate drought months variables for periods t – 1, t – 2 and t – 3. Third, we estimate the effects of drought shocks in 2010 on outcomes in 2011, 2013 and 2015.

We also carry out several robustness checks. First, we use household fixed effects instead of district fixed effects to focus on within-household changes over time. Second, given that weather shocks are possibly spatially correlated, we re-estimate our models using Conley robust standard errors (Conley, 1999; Hsiang, 2010).⁸ We follow Hirvonen (2016) and report Conley robust standard errors at various distance cutoffs. Third, we test whether the results are robust to alternative definitions of the outcome variables. Fourth, we test whether using an alternative weather database has a major influence on the results. Further details of these robustness checks are discussed below in the results section.

4.2. Mechanisms

In addition to estimating the effects of drought shocks on labour allocation, we also explore the main underlying mechanisms. Drawing on the literature (Zhang et al., 2018; Emerick, 2018; Colmer, 2021a; Olper et al., 2021; Ibanez, Romero and Velasquez, 2022), we look at agricultural production effects and local labour market dynamics. In terms of agricultural production, we hypothesize that household labour reallocation decisions can be explained by a direct negative effect of drought shocks on agricultural productivity. If land and agricultural labour productivity significantly decline as a result of drought, households may decide to reallocate (some of) their labour away from own farming to employment activities that are less affected by drought in order to protect their incomes and consumption. To test this mechanism, we estimate:

where Y_icdt is the outcome variable for household i, such as agricultural land and labour productivity in year t. We follow the same identification strategy as in equation (2). Notice that here we measure drought in the same period in which we observe the agricultural outcomes because the effects of weather shocks on agricultural production are contemporaneous.

In terms of labour market dynamics, we test if household labour allocation can be explained by frictions in local labour markets caused by droughts, both inside and outside of agriculture. First, we hypothesise that drought shocks shrink demand for hired on-farm labour due to negative effects on agricultural productivity. If this is the case, the supply of on-farm wage jobs would significantly diminish in the presence of droughts. We test this hypothesis by estimating the direct effects of droughts on households’ demand for hired on-farm labour and corresponding daily wages paid. We expect that in response to drought, households will hire less labour, offer lower wages, or both. Second, we test whether labour demand and wages in non-agricultural activities are affected by droughts. It is difficult to predict a priori whether and to which extent non-agricultural labour demand responds to drought, as the effect will depend on the intensity of linkages between non-agricultural activities and agriculture and on local demand effects. The challenge is that we do not observe data on non-agricultural firms to directly measure their labour demand and wages over time. We therefore use non-agricultural wage income as a proxy for wages paid by firms.

4.3. Labour allocation and consumption smoothing

We estimate the following regression model to assess the effect of labour allocation on household welfare following drought:

where C_icdt is the measure of household welfare in year t. We define the outcome variable in different ways. First, we compute the value of household food and non-food consumption per adult equivalent. Second, following Swindale and Bilinsky (2006) and Kennedy, Ballard and Dop (2011), we use information on household weekly consumption of 12 different food groups to calculate the household dietary diversity score (HDDS)—a common indicator of food security. Third, given that HDDS in the 50th (median) and 75th percentiles of our sample are 5 and 7 food groups, respectively, we generate two separate dummy variables taking a value of one if the household HDDS is equal to or greater than 5 and 7. We interpret these two dummy variables as indicators of households being food-secure with medium and high levels of probability (Kennedy, Ballard and Dop, 2011). D_{cdt − 1} is the number of drought months in the previous year, and E_icdt is the number of weekly hours in off-farm employment in year t. The interaction term between D and E informs us about the extent to which off-farm employment protects household consumption against the effects of drought. We control for the same household time-variant factors and district and year fixed effects as in the other models.

5. Results

5.1. Labour reallocation: extensive margin

In this section, we discuss results from our regression models as specified in equation (2). For the dummy outcome variables, we employ a linear probability model due to its computational ease in absorbing many high-dimensional fixed effects (Guimaraes and Portugal, 2010). Panel A of Table 2 shows results for farm wage employment in columns (1) and (2), and for off-farm wage employment in columns (3) and (4). Panel B shows results for on-farm self-employment in columns (1) and (2), and for OFSE in columns (3) and (4).

The results in panel A of Table 2 show that exposure to drought in the previous year’s growing season reduces households’ probability of employment in farm wage jobs. Specifically, one extra drought month during the growing season decreases the household probability of having a farm wage job by 0.9 percentage points. We do not find any evidence of drought effects on the probability of off-farm wage employment, which may possibly be due to the scarcity of off-farm jobs in the context of rural Ethiopia.

In panel B of Table 2, we do not find statistically significant effects of drought on on-farm self-employment. However, we find evidence that an extra month of drought in the previous year increases the probability of OFSE by about 5 percentage points. The effect is amplified to 7 percentage points when the extra drought month occurs in the growing season. These findings highlight the important role of OFSE in mitigating agricultural income losses due to drought.

5.2. Labour reallocation: intensive margin

Table 3 presents results from different specifications of equation (2), where the dependent variable is the share of household weekly hours in each of the four job categories expressed in per cent. We find that one extra month of drought during the last growing season (column 2 of panel A) leads to a 0.35 percentage point decrease in the household labour share spent in farm wage employment. While this coefficient may appear small, it should be noted that the average household in the sample only spends 0.7 per cent of its labour time on farm wage labour, meaning that the drought effect is equivalent to a reduction of 50 per cent. For off-farm wage jobs, we find no significant effects of drought. Again, this may possibly be due to inadequate off-farm wage employment opportunities in the local contexts to absorb the surplus labour following drought episodes.

In panel B of Table 3, we find strong evidence that an increase in drought months significantly affects household labour allocation to both farm and off-farm self-employed activities. First, the results in columns (1) and (2) show that an additional drought month—during the year and growing season alike—leads to a 3–4 percentage point decrease in the household labour share spent in on-farm self-employment. Second, the results in columns (3) and (4) reveal that households respond to drought by allocating more labour to OFSE. We find that an additional drought month during the growing season leads to a 4.5 percentage point increase in the household labour share spent in OFSE (equivalent to a 40 per cent increase evaluated at the sample mean of the dependent variable). Taken together, the results in Table 3 suggest that rural households in Ethiopia respond to frequent drought shocks by reallocating labour away from farming to OFSE. In fact, for drought months in the growing season, the combined decrease in the share of weekly labour hours in on-farm self-employment (−4.064 percentage points) and farm wage employment (−0.345 percentage points) is very similar to the increase in the share of labour hours in OFSE (+4.492 percentage points).

Additional results on the short-term and long-term effects of droughts are summarised in Appendix Tables A3–A6 in supplementary data. The effects of persistent droughts (drought months over last 2 and 3 years) reveal two important insights (Appendix Tables A3 and A4 in supplementary data). First, they point in the same direction as the effects of short-term drought, namely a labour reallocation away from farming to OFSE. Second, the effects of persistent droughts are somewhat smaller in absolute terms than the effects of short-term drought, suggesting that in the long-run households are possibly substituting on-farm adaptation for off-farm adaptation to some extent. The results in Tables A5 and A6 (Appendix in supplementary data) with drought months in specific past years provide additional insights, namely that the effects on labour reallocation tend to decrease over time.

The analyses of heterogeneous effects are shown in Tables A7–A10 (Appendix in supplementary data). The results suggest that the labour reallocation effects from farm work to OFSE in response to drought are particularly strong among households with better access to formal financial services (Appendix Table A10 in supplementary data). This is plausible, as access to financial services allows households to make investments that may be needed for running own non-farm business activities. This result is also consistent with previous research in Africa showing that access to credit can enhance the capacity of households to adapt to rainfall shocks (Tabetando et al., 2023).

5.3. Mechanisms

We now analyse some of the main mechanisms underlying the effects of drought on labour reallocation, as explained in equation (3). The effects of drought on agricultural production are summarised in Table 4. The dependent variables are log-transformed, so the effects can be interpreted in percentage terms. As expected, drought negatively affects agricultural land productivity. One extra drought month in the year and growing season reduces land productivity by 27 per cent and 33 per cent, respectively. We also find negative effects on agricultural labour productivity. One extra drought month in the growing season reduces labour productivity by 17 per cent.

The effects of drought on labour demand and wages are summarised in Table 5. We find negative but statistically insignificant effects of drought on the use of family labour on household farms. However, households significantly reduce the demand for hired labour on their farms. One additional drought month during the growing season reduces the quantity of hired labour by about 10 per cent (panel A, column 4) and the wages paid to hired farm labour by about 13 per cent (panel B, column 2). These results, taken together with the negative effects of drought on agricultural productivity, also explain why households reduce their labour supply to farm wage employment, as shown above. We find no evidence of significant effects of drought on non-agricultural wage income (Table 5, panel B, columns 3 and 4). Given that the labour supply of households to off-farm wage employment does not change significantly in response to drought (see Table 3), any changes in wage income would primarily be driven by changes in wage rates. The insignificant estimates in Table 5 suggest that non-agricultural wage rates do not respond much to short-term drought.

5.4. OFSE and consumption smoothing

We now estimate equation (4) in order to analyse to what extent labour allocation to off-farm employment can contribute to consumption smoothing and food security. We focus on OFSE, as the results above indicate this is the main employment category that households reallocate labour to as a response to drought. We further focus our analysis on droughts occurring during the growing season, as this is the main period in which drought impacts are amplified.

The effects of drought on food and non-food consumption are summarised in Table 6. The results provide three insights. First, the drought coefficients in all models reveal that one additional drought month during the growing season leads to a 4–5 per cent reduction in household food and non-food expenditures. Second, the coefficients for OFSE are all positive and statistically significant, implying that household labour allocation to OFSE is associated with higher food and non-food consumption. Third, the coefficients of the interaction term between drought and OFSE are positive and statistically significant, at least for food consumption (column 2), suggesting that OFSE contributes to consumption smoothing during and after drought episodes.

Next, we analyse the effects of drought and OFSE on food security proxied by HDDS (Table 7). We calculate HDDS z-scores to obtain a continuous outcome variable, in addition to two dummy variables for HDDS ≥ 5 and HDDS ≥ 7. The results in all models confirm that drought is associated with a decrease in food security. In particular, 1 additional drought month during the growing season reduces HDDS by about 0.05–0.06 standard deviations (Table 7, columns 1 and 2) and also lowers the likelihood of households being categorised as food-secure (columns 3–6). Further, OFSE is associated with higher levels of food security. The coefficients of the interaction term between drought and OFSE are all positive. In column (6), the interaction term is also statistically significant, suggesting that OFSE helps to reduce the likelihood of becoming food-insecure during or after drought episodes.

5.5. Robustness checks

In this section, we highlight results from additional estimations to show that our main results are robust to alternative model specifications. Results of these robustness checks are shown in Tables A11–A20 (Appendix in supplementary data). First, we show that the same conclusions regarding reallocation of farm labour to OFSE following drought are supported when estimating the regression models with household fixed effects (Appendix Tables A11 and A12 in supplementary data). Also, the estimates of the effects of drought on household consumption and food security, and the role of OFSE for consumption smoothing during and after drought episodes, remain very similar when controlling for household fixed effects (Appendix Tables A13 and A14 in supplementary data).

Second, using various distance cutoffs, we re-estimate the models using Conley (1999) robust standard errors that account for spatial correlation (Appendix Tables A15–A17 in supplementary data). Third, we show that our results are robust to using the share of household members across the four job categories as an alternative dependent variable (Appendix Table A18 in supplementary data). Finally, we confirm that the main findings are insensitive to the use of an alternative historical weather database (Appendix Tables A19 and A20 in supplementary data).⁹

6. Conclusion

We have analysed how rural households in sub-Saharan Africa adapt to drought shocks through labour reallocation, using representative household panel data from Ethiopia. We find that households reduce their labour time in farming as a response to drought, even though they do not abandon farming altogether. At the same time, households increase their labour time in off-farm activities. This partial switch from farming to off-farm activities is plausible because droughts significantly reduce agricultural productivity. We also show that droughts have negative effects on household food security, whereas the reallocation of labour time from farm to off-farm activities helps to smooth consumption and dietary diversity. In terms of off-farm activities, households increase their labour time in self-employed business activities as a response to drought, but not their labour time in off-farm wage employment. Our interpretation is that non-agricultural wage jobs are not sufficiently available in the local rural contexts of Ethiopia to absorb the additional labour supply during and after drought episodes.

Analysis of heterogeneous effects reveals that labour reallocation to OFSE as a response to drought is particularly strong for households with access to rural financial services. In other words, households with better access to rural finance find it easier to adjust their livelihoods to weather shocks. These households are better able to overcome liquidity constraints and other typical barriers for starting or expanding non-agricultural businesses.

Differentiating between short-term droughts and persistent droughts, we find similar labour reallocation effects in general. However, interestingly, the labour adjustments are somewhat stronger for short-term droughts. These differences suggest that households may possibly improve their adaptive capacity in the longer run by implementing on-farm adaptation strategies that complement labour reallocation to off-farm activities. Even though not analysed here in more detail, on-farm adaptation strategies may include the adoption of climate-smart technological innovations, such as irrigation, more tolerant seeds, and improved agronomic practices, among others.

Our findings highlight three important takeaways for policymaking. First, labour reallocation to off-farm activities is an important strategy for farm households in Africa to cope with weather shocks. As weather extremes tend to occur more frequently with climate change, policymakers should work towards increasing the size and improving the functioning of the rural non-farm economy. The creation of non-farm wage jobs, which are currently not sufficiently available, should have high priority. This does not mean a focus only on public-sector jobs. Policies to incentivise private firms to invest more in rural regions will also be important. Second, our findings of negative effects of droughts on food security and dietary diversity point to the need to develop tailored social protection schemes that particularly target the most vulnerable and those who lack the capacity to reallocate labour to off-farm activities. Third, our finding that access to formal financial services increases household OFSE as an adaptation strategy to drought calls for financial inclusion policies in rural settings. Such policies could help households to overcome liquidity constraints that undermine their ability to not only venture into alternative non-farm jobs but also to invest in climate-smart agricultural technologies.

Our study adds to the growing climate adaptation literature and supports the idea that weather shocks are partly contributing to the unique structural transformation patterns in Africa, which are characterised by high employment on small family farms combined with a strong diversification into off-farm activities, especially self-employed activities in small non-farm businesses (Davis, Di Giuseppe and Zezza, 2017; Sen, 2019; Christiaensen and Maertens, 2022). Future research should explore how non-agricultural rural employment can be fostered, how different types of jobs influence people’s welfare and adaptive capacity and how non-agricultural employment is linked to agricultural development. Another important research direction is how smallholder farming can be made more climate-resilient through technological and institutional innovations.

Acknowledgements

The authors thank Salvatore Di Falco and two anonymous reviewers of this journal for valuable comments on an earlier version of this manuscript.

Supplementary data

Supplementary data are available at ERAE online.

Funding

Arnold L. Musungu was funded through a Ph.D. scholarship from the German Academic Exchange Service (DAAD). This research was also supported by the German Research Foundation (Deutsche Forschungsgemeinschaft - DFG) as part of the Collaborative Research Center ‘Future Rural Africa’ (Project: 328966760-TRR 228/2).

Conflict of Interest

The authors have no conflicts of interest to declare.

Footnotes

References