Does rainfall variability explain low uptake of agricultural credit? Evidence from Ethiopia

Abstract

1. Introduction

Access to credit is widely acknowledged to be a key means of transforming the livelihoods of poor rural households in developing countries. Previous empirical studies have shown that access to credit can stimulate investments in modern agricultural inputs (Giné and Yang, 2009; Zerfu and Larson, 2010; Abate et al., 2016), facilitate the start-up of new enterprises (De Mel, McKenzie and Woodruff, 2008) and, at least in some cases, reduce poverty (Berhane and Gardebroek, 2011).¹ Such evidence has generally encouraged the expansion of microfinance institutions (MFIs) in many developing countries over the last few decades, making credit accessible to traditionally unbanked environments. The outcomes that microcredit is presumed to engender are thus conceptually well founded. That is, given the marginal returns to capital investments exceed the costs of accessing credit markets, after adjusting for risk preferences and the stochastic nature of agricultural investment outcomes, we would expect farmers to be net demanders of agricultural credit markets (De Mel, McKenzie and Woodruff, 2008; Fafchamps et al., 2014; Duflo et al., 2011). The fact that the significant expansion of MFIs and credit services in much of sub-Saharan Africa has generally not been accompanied by high rates of credit uptake suggests that the costs and risks of credit market participation by African smallholders deserve more scrutiny (Karlan and Morduch, 2009; Karlan, Morduch and Mullainathan, 2010; Banerjee, 2013).

An important consideration in the uptake of credit for agricultural investments in predominantly rainfed farming systems is farmers’ exposure to agroclimatic uncertainty, particularly as manifested through variability of seasonal rainfall outcomes. The increasing frequency in recent years of catastrophic extreme weather events associated with climate change in the horn of Africa means that the region’s farm households are facing increasingly significant weather-related uncertainties with far-reaching implications to their subsequent production and input use decisions. In the absence of insurance, lenders also suffer when poor crop outcomes induce increased default rates. In contexts where lenders have limited information about potential borrowers, and where borrowers generally do not have the collateral wealth required to effectively guarantee loans, lenders are likely to shift much or all of the contractual risk onto the borrower to the extent that the borrower voluntarily withdraws from the credit market (Boucher, Carter and Guirkinger, 2008; Guirkinger and Boucher, 2008; Boucher, Guirkinger and Trivelli, 2009). Such risk rationing adds to the traditional price (due to high interest rate) and quantity (limits in loan sizes) rationing, which ultimately reduces the uptake of actual credit. As such, uninsured rainfall risk can severely undermine investments in productive farm activities, discouraging the adoption of improved and more productive farm technologies, thereby impeding efforts to reduce poverty. For these reasons, lenders have been attracted to recent experimentations on bundling of credit with rainfall indexed insurance. However, a key lesson from recent studies is that bundling credit with index insurance does not necessarily lead to increased credit uptake (Giné and Yang, 2009; Karlan et al., 2011). Hence, it is not clear to what extent variabilities in rainfall under different contexts can explain credit uptake.

The contribution of this paper is to provide such scrutiny, focusing on the effect of rainfall variability on rural households’ credit uptake, and using Ethiopia as an empirical case study. Despite some attempts to identify the effect of weather risks on credit risk in developing countries (e.g. Miranda and Gonzales-Vega, 2011; Von Negenborn, Weber and Musshoff, 2018), there exists limited evidence on whether rainfall variability explains the low uptake of credit in sub-Saharan Africa. In this paper, we revisit this question, with a slightly different framing. Smallholder agriculture in Ethiopia—like much of the rest of the region—is predominantly rainfed and constitutes a fundamentally risky livelihood context. Theoretically, the presence of uninsured production and consumption risk can contribute to credit market imperfections that may lead to credit rationing, effective demand being less than potential demand, in rural economies. In Ethiopia, lenders usually require collateral, which smallholder farmers lack, or rely on other lending schemes (e.g. group lending) to alleviate, among others, asymmetric information problems.

Nationally representative data sets show that more than half of Ethiopian rural households are not borrowing due to the risk of default, i.e. fear of being in debt or distress of losing assets in the case of default (Mukasa, Simpase and Salami, 2017; Berhane and Abay, 2019). These pieces of evidence highlight the potential of risk rationing in Ethiopian rural credit market as the risks involved in participation in the credit market are perceived to be prohibitive in the form of indebtedness or potential loss of collateral in case of default. Thus, we also investigate the impact of rainfall uncertainty on alternative forms of credit rationing, including risk rationing. We finally explore whether the impact of rainfall uncertainty and associated responses vary across livelihood and production systems, mainly across the arid and semi-arid lands (ASAL) and the non-ASAL (or highlands) of Ethiopia.

We employ three waves of the Living Standard Measurement Study-Integrated Surveys on Agriculture (LSMS-ISA) for Ethiopia: 2011/12, 2013/14 and 2015/16. These nationally representative data sets cover a large sample of rural households throughout the country and hence provide substantial variation in rainfall variability as well as credit uptake and credit market landscape. The LSMS-ISA data provide GPS coordinates of households’ residences, enabling us to merge these longitudinal data with time-series rainfall data. We define several measures of rainfall variability, including the coefficient of variation (CV) and the standard deviation of rainfall, over a 10-year period as well as over 29-year period.

We find that inter-annual rainfall variability strongly and negatively affects credit uptake. This result holds under alternative measures of rainfall uncertainty and remains robust to a battery of robustness checks. As expected, we find that these effects are only significant in rural areas where livelihoods are heavily reliant on rainfed agriculture. We provide suggestive evidence showing that the effect works by introducing risk rationing. The impact of rainfall uncertainty in discouraging demand for credit is more pronounced in ASAL of Ethiopia, where livestock herding remains a dominant source of livelihood and these regions are believed to be vulnerable to recurrent weather shocks (e.g. Thornton et al., 2002, 2006; Abay and Jensen, 2020). These results are noteworthy in terms of understanding the implications of weather risk on rural households’ demand for credit and associated agricultural investments as well as in view of identifying rural farmers’ adaptive strategies to climate change. These results speak to the empirical puzzle of low uptake of credit and underinvestment in agricultural technologies documented in many African countries. Our findings highlight the importance of interventions aimed at both relaxing smallholders’ credit rationing while also reducing their production risk.

2. Conceptual framework

Taking the derivative of these with respect to the credit constraint |$\tilde a$|⁠, we see that |${{d{x^*}} \over {d\tilde a}} \lt 0$|⁠, i.e. if credit constraints are binding, then the optimal input use |${x^*}$| is increasing in the amount of credit available. When farmers are not credit constrained, |$cov\left( {\,f{^{\,\,^{\prime}}}\left( x \right)\!,u{^{^{\prime}}}\left( {c_s^1} \right)} \right) \lt 0 $| and |${\lambda _a} = 0$|⁠. This has two important implications: first, the presence of risk reduces input x; second, production risk reduces the demand for credit. We see the latter by noting that in the case of no credit constraints, input use is lower in the second period and so is the marginal utility of consumption for any level of borrowing |$a$|⁠. The implication of this is that exposure to uninsured risk will reduce the demand for credit by risk-averse farmers.

3. Data and descriptive statistics

3.1. Context and data

Ethiopia’s smallholder agriculture is predominantly rainfed, with the main rainy season known as Meher covering from June to September (accounting for over 95 per cent of total crop production) and a second season covering some areas known as the Belg spanning from February to May. With limited access to water and irrigation structures, the rest of the year for the most part remains dry with no possibility to grow crops. Seasonal patterns vary by region but in most of the regions the meher rains are more important than the rainfall in the belg season. The most visible distinction in seasonal and livelihood patterns is between ASAL and non-ASAL areas.² In the ASAL regions of Ethiopia, livestock production remains the largest source of livelihood and these regions are believed to be most vulnerable to weather shocks (e.g. Thornton et al., 2002, 2006; Abay and Jensen, 2020). In these non-ASAL (highlands) areas, mixed farming remains the dominant livelihood strategy, both as source of income and employment. We will explore potential differential responses across these two livelihood zones and farming systems.

The primary source of data for this study is LSMS-ISA for Ethiopia, a collaborative project between the Central Statistics Agency of Ethiopia (CSA) and the World Bank (CSA and World Bank, 2017). The Ethiopian LSMS-ISA data are longitudinal data sets collected every 2 years, covering a wide range of topics related to agricultural production decisions, including a detailed module on agricultural credit use and reasons for borrowing.

The first round of the LSMS-ISA data for Ethiopia were collected in 2011/12. The first round covers only rural and small towns and hence is only representative of the rural population in Ethiopia.³ The first wave included 3,776 households from 333 enumeration area across all regional states. The second and third rounds, collected in 2013/14 and 2015/16, respectively, include urban areas and are designed to be nationally representative. For example, the third round covers 433 enumeration areas of the same regions (including major towns and cities), reaching about 5,262 rural and urban households. In the interest of constructing nationally representative longitudinal data and to focus on agricultural credit, we focus on the rural sample and hence exclude major towns and cities. Our final sample is only therefore representative of the Ethiopian rural population.

The LSMS-ISA households are georeferenced, allowing us to merge them with other geospatial data sets.⁴ Our key rainfall measures of interest are derived from the POWER (Prediction Of Worldwide Energy Resource) Data Archive (Release 8) of the National Aeronautics and Space Administration (NASA).⁵ These spatial time-series data are available globally at a 0.5 decimal degree spatial resolution, in daily time steps. The data are widely used to evaluate and model agricultural crop yields (e.g. Bai et al., 2010; Van Wart, Grassini and Cassman, 2013, 2015). We construct rainfall data time series spanning 29 years, for each enumeration area in the survey, and from these we construct localized measures of rainfall variability. Our key measure of rainfall variability, the CV of inter-annual total rainfall, is constructed for the 10 years prior to each survey period. Intuitively, it may be difficult for farmers to recall rainfall spells of the distant past, arguably, going beyond 10 years. However, as a robustness check, we also evaluate the sensitivity of our results using measures based on longer (29 years) and shorter (5 years) time spans.

We note that rainfall variability is just one dimension of the weather risk that rural farmers in sub-Saharan Africa face. Thus, rainfall variability only captures one aspect of production risk.⁶ Our focus on rainfall variability is motivated by the notion that rainfall variability is arguably the most important indicator of overall weather variability, particularly from a production perspective in low-input rainfed farming systems, such as those found in our study area. For this reason, rainfall variability is a key metric commonly employed in both policy and analysis of farmer decision-making and investment decisions (e.g. Alem et al., 2010; Bezabih and Di Falco, 2012; Auffhammer et al., 2013; Colmer, 2020). Rainfall is also a major source of income (and livelihood) in Ethiopia and some recent studies have exploited variations in rainfall to study the impact of income on inter-temporal decision-making (discounting) behaviour (e.g. Di Falco et al., 2019).

3.2. Descriptive statistics

The LSMS-ISA surveys for Ethiopia elicit households’ credit uptake over the 12 months prior to enumeration, both from formal and informal sources. All credit-related outcomes in all rounds of the Ethiopian LSMS data were collected at similar times of the year (January–April). Thus, the 12-month period covers the most recent main production season, up until January of the previous year. We expect that a good share of the credit taken over the 12 months prior to enumeration would be spent on inputs for the most recent production season. We confirm this in our next empirical exercises (see Table 2). For reducing some heterogeneity in credit sources, we focus on formal credit, meaning credit from commercial banks, rural microfinance organizations, and rural savings and cooperatives.

In Table 1, we report households’ credit market participation for the three waves. This table shows that rural credit participation is low, with less than 10 per cent rural households taking credit. These participation rates are consistent with those reported in Berhane and Abay (2019), which come from different data sources. A more exhaustive list of household characteristics and summary statistics are given in Table A1 in the Appendix.

Rural households’ credit market participation rates are the results of an interaction of demand- and supply-side attributes of rural credit markets. The supply-side attributes can be measured by availability of credit in each village, in the form of availability and accessibility of financial institutions. In the context of Ethiopia, despite the need for further progress, several MFIs and cooperatives have been established to resolve households’ access to financial services. Berhane and Abay (2019) show that many rural villages have at least one microfinance or credit and saving association.

In terms of identifying the impact of rainfall variability on uptake of agricultural credit, the distribution of these rural MFIs and their relationship with rainfall uncertainty have important implications. For instance, if MFIs are targeting high agricultural potential areas or those areas less susceptible to rain failure, we are more likely to overestimate the impact of rainfall variability on credit uptake. In the context of Ethiopia, where regional and federal governments have stakes in most MFIs, such systematic distribution of MFIs and services is less likely to be the case.⁷ To empirically confirm that rainfall variability is uncorrelated with access to microfinance, we run simple regressions characterizing farmers’ access to microfinance as a function of rainfall variability. Farmers’ access to microfinance is defined at village level and assumes a value of 1 if a village has one or more microfinances operating and 0 otherwise. These regressions are given in Table A2 in the Appendix. These results clearly show that access to microfinance and hence the distribution of microfinance services is not significantly related to our measure of production risk, rainfall variability.

Table 2 summarizes households’ major reasons and purposes for obtaining loans from formal sources. We aggregate the reasons into those mentioned in Table 2. More than half of the credit taken from formal sources is targeted for agricultural investments, i.e. the purchase of inputs. This is consistent with the Ethiopian government’s agenda of expanding access to credit for smallholders geared towards boosting agricultural investments. An important proportion of credit taken over the prior 12 months was used for agricultural investments in the season. The next important reason appears to be starting up or expanding businesses. The purposes of credit remain comparable across years.

A large literature attributes rural financial market failure to two fundamental problems in rural credit market (Stiglitz and Weiss, 1981; Hoff and Stiglitz, 1990), adverse selection and moral hazard.⁸ The coexistence of these two hurdles in rural financial markets, the information asymmetry among lenders and borrowers and the poor’s lack of collateral to pledge, leads to credit rationing, a situation where effective demand for credit is less than potential demand for credit (Conning and Udry, 2007; Guirkinger and Boucher, 2008; Boucher, Carter and Guirkinger, 2008; Boucher, Guirkinger and Trivelli, 2009; Armedariz and Morduch, 2010). Previous studies have employed direct elicitation methods to understand and quantify the prevalence and types of credit rationing. More specifically, households not participating in credit markets mention that they face one or more of the following types of credit rationing: quantity rationing, transaction cost rationing and risk rationing (Guirkinger and Boucher, 2008; Boucher, Carter and Guirkinger, 2008; Boucher, Guirkinger and Trivelli, 2009). Quantity rationing involves a situation where households’ effective demand exceeds existing supply due to lack of collateral to access supplied loans or due to credit limits put forth by the lender.⁹ Transaction cost (or price)–rationed households are those for whom credit market participation is sufficiently expensive due to high interest rates or the potential cost of processing, screening and monitoring loans. Risk rationing is demand driven and voluntary exclusion from the credit market for reasons associated with the risks involved in credit participation, including the risk of indebtedness or potential loss of collateral in case of default.

In Table 3 we report households’ reasons for not participating in rural credit markets, with responses grouped into four categories, following the above literature on eliciting credit constraints and types of credit rationing. We classify households to be unconstrained if they do not need credit or have enough resource for their business. Those households are not borrowing either due to fear of being in debt or risk of default and hence loss of collateral is categorized as risk rationing. Those households for whom credit market participation is expensive due to high interest rates or the potential cost of accessing, processing, screening and monitoring loans are classified as transaction cost or price rationing. We then have an additional category for those mentioning other reasons, including personal and religious reasons. As also shown from other data sources for Ethiopia (Mukasa, Simpase and Salami, 2017; Berhane and Abay, 2019; Abay et al., 2019), the most important constraint limiting rural households’ credit uptake is risk rationing. About half the households are not taking credit for reasons related to the production risk they face, which may involve fear of loss of assets or additional risk of having to bail out fellow group members in the case of group lending.

4. Estimation and identification strategy

We initially estimate equation (5) without additional controls but also evaluate specifications that include additional covariates related to alternative mechanisms through which rainfall uncertainty may affect demand for credit. For instance, without controlling for contemporaneous or recent rainfall spells, |${\beta _1}$| captures both the effect of weather uncertainty as well as the effect of recent rainfall spells, which in turn can have a negative or positive impact on households’ demand for credit. In extending the specification in equation (5), we consider and employ detailed household and community-level characteristics that may explain either the demand or supply side of credit. Credit market participation is a result of the interaction between the demand and supply sides of the credit market. Thus, we control both for demand- and supply-side attributes of rural credit markets. We capture a long list of household characteristics, including gender and education of household heads as well as different types of assets the household owns. These factors are likely to affect demand for credit. For instance, male-headed households may be more likely to take agricultural credit because of better access (Hansen and Rand, 2014) to credit services or risk-taking behaviour (e.g. Charness and Gneezy, 2012). From a supply point of view, we control for households’ access to microfinance and distance to nearest markets, which are likely to affect the supply and availability of credit.

Our measure of inter-annual rainfall variability (CV of 10 year’s rainfall prior to the reference period for each survey round) is expected to vary little across 2-year panel waves. Thus, including household fixed effects in such estimations can be expected to attenuate the true effect of rainfall uncertainty and hence can be viewed as a conservative approach. Furthermore, rainfall variability is reasonably taken as exogenous and hence estimates can be expected to be robust to inclusion of these household characteristics and household fixed effects. In estimating equation (5), we have two sources of correlation in error terms (unobserved effects), arising from spatial and temporal dimensions. Our panel structure follows households across years. Similarly, our key variable of interest (CV) varies at the enumeration area level, where households living in the same enumeration area share similar unobserved effects. Thus, we are reporting results with standard errors clustered at the enumeration area level.¹⁰ Because of this clustering challenges in binary outcome models, we estimate linear probability models, even though our key outcome variables are measured as binary indicators.

To examine heterogeneous responses to rainfall uncertainty across different farming and livelihood systems, we estimate the specification in equation (5) both for the ASAL and non-ASAL regions. We also test the robustness of our results in several ways: (i) We estimate our empirical models using two alternative measures of rainfall: the CV and standard deviation in annual rainfall over the last 10 years (the latter results are reported in the Appendix); (ii) We test alternative measures of CVs and standard deviations computed over 5 or 29 years instead of 10 years; (iii) We split and construct our rainfall variability for each of highland Ethiopia’s two main growing seasons: meher (long rainy season) and belg (short rainy season). Finally, to confirm our expectation that the effect of rainfall uncertainty is more pronounced in contexts where livelihoods depend on rainfed agriculture or livestock, we also estimate the effect of rainfall risk focusing on urban households, a sample which is not part of our main analyses.

Where |${U_{hct}}$| can be interpreted as the degree of constraint that a household (h) faces in his/her quest for accessing agricultural credit associated with each type of credit constraint/credit rationing (c) and time (t). The first term in equation (6) captures constant terms for each type of credit rationing (see Table 3) while the second term is our key variable of interest, and the parameters |${\beta _{1c}} $| capture the impact of rainfall uncertainty on the different types of credit rationing. X_ht stands for other observable factors that affect households’ level and type of credit rationing. |${\varepsilon _{hct}} $| stands for other unobserved factors that may influence households’ demand for credit. As households report the most important constraints and reason for not participating in credit market, the expression in equation (6) can be thought of as a framework guiding farmers’ reporting behaviour and imply that households report a particular type of credit rationing that imposes the highest constraint. Assuming that the error terms |${\varepsilon _{hct}}$| follow extreme value distribution, households’ credit rationing type can be modelled as a standard multinomial logit probability distribution.

5. Estimation results and discussion

5.1. The effect of rainfall uncertainty on credit uptake

In Table 4 we provide benchmark estimates of our main model of interest (equation (5) in Section 4), based on CV as measure of rainfall uncertainty. In column 1 of Table 4 we regress credit uptake as a function of only inter-annual rainfall variability. In column 2 we additionally control for last year’s rainfall to account for income or related effect of recent rainfall. The CV of rainfall is aggregated over the last 10 years and hence is correlated with recent rainfall. Then, in the third and fourth columns we extend the empirical specification by adding household- and community-level characteristics that may explain demand for agricultural credit. The estimates associated with rainfall variability in columns 2–4 remain stable, suggesting that rainfall uncertainty is almost uncorrelated with these household- and community-level characteristics. Finally, in column 5 we provide estimates controlling for household fixed effects.

Table 4 also provides estimates of the impact of prior year’s rainfall on credit uptake. Theoretically, the effect of lagged rainfall on uptake of agricultural credit is ambiguous and depends on the purpose of credit. If the purpose of credit is mainly for purchase of agricultural inputs (e.g. fertilizer and improved seed), as seems the case for Ethiopian households (see Table 2), favourable rainfall spells may lead to increased productivity, which can be used to purchase agricultural inputs and hence reduce demand for credit. Similarly, in situations where credit is mainly used for consumption smoothing purposes, then positive rainfall shocks can reduce demand for consumption credit while negative rainfall shocks may boost demand for credit. In our context, the most dominant purpose of credit seems to be the purchase of agricultural inputs (see Table 2). Consistent with this reality, the estimates in Table 4 show that recent rainfall has negative impact on demand for credit.

In Table 5 we consider inter-season rainfall variability, instead of inter-annual rainfall variability. This involves computing rainfall variability across seasons rather than across years. Panel A of Table 5 provides results based on inter-season variability for meher season, while Panel B provides corresponding results for belg season. Consistent with those estimates associated with inter-annual rainfall variability, the estimates in Table 5 show that inter-season rainfall variability negatively affects demand for agricultural credit. The effects are consistent across both seasons and the magnitude of the effect is broadly comparable across both seasons. However, rainfall variability in the belg season appears to have more pronounced effect. This is plausible in the context of Ethiopia where most credit is used for input use in the main production (meher) season that immediately follows the belg season. Thus, rainfall spells and rainfall uncertainty during the belg season can give important signal on the favourability of the forthcoming production season on which farmers may rely on their input use and associated demand for credit.

The impact of rainfall uncertainty and associated responses to weather risks can vary across farming systems. Rainfall spells may also evolve differently across these livelihood zones and effects of weather conditions depend on households’ resilience to weather shocks. Households with ample opportunities and mechanisms to finance agricultural investments may not respond to rainfall variability and weather shocks in their quest for agricultural credit. In the interest of uncovering potential heterogeneities in the impact of rainfall uncertainty on households’ demand for credit, we separately estimate responses for the ASAL and non-ASAL regions of Ethiopia. The results in Table 6 show stark differences in the impact of rainfall uncertainty on demand for credit across the ASAL and non-ASAL region households. Those households living in the ASAL regions, who are usually pastoralists relying on livestock production and prone to recurrent shocks, are more responsive in reducing uptake of agricultural credit in response to high rainfall variability. This is intuitive as households in Ethiopian highlands (non-ASAL regions) practice mixed livelihood and farming systems, which may offer them relatively diversified options to manage and cope with rainfall uncertainty and weather shocks. Those households with alternative and mixed livelihood strategies seem to be less sensitive to rainfall uncertainty while those households in the ASAL regions of Ethiopia are sensitive to rainfall variability. These differential impacts and responses to rainfall variability are noteworthy for several reasons. First, the heterogeneous impacts of weather risk among farming systems and livelihoods highlight the need for more tailored credit and insurance products. Second, the evidence that rainfall uncertainty is more impactful in reducing demand for credit and hence productivity-enhancing investments of already vulnerable ASAL households imply that such types of production risk may perpetuate poverty traps (e.g. also Zimmerman and Carter, 2003; Barnett, Barrett and Skees, 2008; Dercon and Christiaensen, 2011).

5.2. Rainfall uncertainty and credit rationing

In this section, we aim to empirically investigate the implication of rainfall variability in explaining credit rationing. Given the share of households reporting to be risk rationed, we particularly aim to explore the impact of rainfall uncertainty on risk rationing in their demand for agricultural credit. We hypothesize that the production environment rural households in Ethiopia operate and associated production risk can contribute to risk rationing in credit market. We directly test this hypothesis by grouping and characterizing the sample of households who are not participating in credit markets. Following the direct elicitation method commonly employed in the literature, we classify households’ reasons for not participating in credit market (see Table 3). As shown in equation (6), we then run simple multinomial logit regression characterizing households’ credit market participation and credit rationing.

The marginal effects shown in Table 7 indicate that rainfall variability significantly and positively predicts risk rationing. Those households exposed to substantial rainfall variability are more likely to be risk rationed in their quest for credit market participation. This confirms our hypothesis that production risk may lead to a specific type of credit rationing, which forces households to be out of credit market and hence potentially engage in low-risk but low-return agricultural investments. Intuitively, these results corroborate our main estimates characterizing households’ demand for credit by highlighting potential mechanisms for the low demand for credit. That is, given rural households’ aversion to borrowing, rainfall uncertainty discourages demand for agricultural credit. These results corroborate recent attempts and interventions that relax rural households’ credit constraints while also addressing production risks rural households face (e.g. Giné and Yang, 2009).

5.3. Summary of robustness checks

We run several empirical checks to probe the robustness of our results. Our main results are based on rainfall variability computed over 10 years. As a robustness check, we also compute rainfall variability computed over 29 and 5 years (which are given in Tables A3 and A4 in the Appendix). Instead of CV, we also repeat these empirical exercises using standard deviation of inter-annual rainfall as an alternative measure. Results are reported in Table A5 in the  Appendix, using similar empirical specifications as in Table 4. Both sets of estimates show that rainfall uncertainty negatively and significantly affects uptake of agricultural credit. These effects are large and consistent across different ways of constructing rainfall uncertainty. The sizes of the effects are broadly stable and comparable across all specifications. The sizes of these effects suggest that agricultural credit uptake is non-trivially responsive to rainfall variability. For instance, the estimates in Table A5 show that a 1 per cent increase in the variability of rainfall (measured by standard deviation of rainfall) decreases uptake of (demand for) agricultural credit by about 6–10 per cent points.

Although our current analysis focuses on the rural sample of households, our data have an additional relatively ‘urban’ sample, those households living in enumeration areas defined as medium and major towns by the CSA of Ethiopia. Thus, we can employ these data to probe the robustness of our results and hence as falsification test. A priori, we would expect the demand for credit to be more risk elastic in the rural sample where production and livelihood rely on rainfed agriculture. To confirm this, we estimate similar empirical models focusing on the urban sample. These results are given in Table A6 in the Appendix. These results confirm that the effect of rainfall uncertainty is statistically insignificant in the urban sample, implying that their demand for credit is unaffected by the rainfall variability in urban households environment.

6. Concluding remarks

Despite many theoretical discussions of the role of production risk in deterring demand for agricultural credit and discouraging investments with positive expected net returns, there exists limited empirical evidence for these linkages. In this paper, we show that rainfall variability is associated with credit risk rationing and is a key determinant of the observed low uptake of agricultural credit in Ethiopia. This relationship is the strongest amongst those households living in the ASAL areas of Ethiopia, which are most vulnerable to recurrent weather shocks. Our results suggest that exposure to rainfall uncertainty may explain the existing low uptake of agricultural credit in many other sub-Saharan African contexts, which is likely to have important implications for smallholders’ productive agricultural investments.

The evidence in this paper speaks both to the literature on agricultural credit and to the literature on the uptake of agricultural credit and profitable agricultural technologies in sub-Saharan Africa (Morris et al., 2007; Alem et al., 2010; Rashid et al., 2013; Sheahan and Barrett, 2017; Abay et al., 2017). Our results support the credit literature’s general assertion that smallholders’ production risks, coupled with the absence of well-functioning insurance markets, stifle demand for credit by introducing risk rationing. Given the covariance of production risk in rainfed agriculture, this type of credit rationing mainly arises due to lenders’ requirement for collateral or the additional risk to bail out fellow group members in most of the group lending schemes that most MFIs rely on. At the same time, the inherent riskiness of rainfed agriculture poses a serious challenge to the sustainability and hence an existential threat to MFIs often mandated to serve the rural poor.

While the problem we address in this research is of very broad relevance in the region, our analysis has focused on a single country. Claims about the external validity of our findings would benefit from additional empirical work in a broader range of production and marketing conditions elsewhere in the region. Furthermore, it is important that we acknowledge that we focus on only one measure of production uncertainty, i.e. annual or seasonal rainfall variability. Future work in this area may address other sources of production risk, including other stochastic biophysical conditions (including drought, disease and pests) as well as uncertainty around market prices or other socio-economic factors affecting agricultural livelihood outcomes. For example, examining the implication of downside risk (e.g. Antle, 1983, 1987; Kim et al., 2014) on demand for agricultural credit can provide a complete picture of the impact of the overall production risk in discouraging agricultural credit and investment. Furthermore, our measure of weather risk, rainfall variability, does not capture the frequency and timing of precipitation.

Intuitively, the implication of weather risk for credit demand, combined with the latter’s role in enabling agricultural investments, suggests that weather risk may be associated with spatial poverty traps (see also Zimmerman and Carter, 2003; Barnett, Barrett and Skees, 2008; Dercon and Christiaensen, 2011). This is particularly plausible as the adverse impact of rainfall variability is more pronounced among the already vulnerable households living in the ASAL region of Ethiopia. These households facing substantial and recurrent production risk are less likely to make profitable investments and more likely to remain poor. For instance, Dercon (1996) shows that rural household in Tanzania are willing to forgo up to 20 per cent of their income to avoid the production risk associated with some crops. Our study suggests the importance of complementary interventions that relax rural farmers’ credit constraints while also addressing the production risks they face. Such complementary interventions are more challenging to design and implement, but the available evidence suggests that their pay-offs may be particularly high in areas of erratic rainfall. Efforts to catalyse a rural transformation in sub-Saharan Africa’s primarily rainfed smallholder production systems should incorporate these insights into the design and targeting of complementary investments in agricultural technologies and credit and insurance markets. For example, the introduction of risk contingent credit—one that combines insurance and credit—might be one way to address this problem (Shee and Turvey, 2012; Miranda and Gonzales-Vega, 2011; Von Negenborn, Weber and Musshoff, 2018; Shee, Turvey and Woodard, 2015).

Conflict of Interest

The authors declare that they have no conflict of interest.

Footnotes

References