Hierarchical Modelling of Small-Scale Irrigation: Constraints and Opportunities for Adoption in Sub-Saharan Africa

1. Introduction

Irrigation is an underutilized agricutural technology in Sub-Saharan Africa (SSA) with signficant potential to enhance yields and resilience to climatic shocks. Nonethless, and compared to other inputs such as inorganic fertlizers, irrigation has historically received limited policy attention in the region. Interest by donors, governments and farmers changed over the last decade, however, as a result of the food price crisis and rapidly increasing climate variability and climate change (Fuglie and Rada2012; Sheahan and Barrett2017; Ringler2021). African Heads of States have declared their commitment to strengthen existing water basin organizations and establish new ones as part of the Sirte Declaration, as well as increase the efficiency and effectiveness of water management systems through irrigation to end hunger by 2025 as part of the Malabo Declaration (African Union Commission2004, 2014). The development of sustainable land management and reliable water control systems is also one of the four pillars of the Comprehensive Africa Agriculture Development Programme (CAADP) to transform Africa’s agriculture and ensure sustainable economic development (African Union and NEPAD2003).

SSA’s share of irrigated area is among the lowest in the world (You et al.2011). Less than 3% of the total renewable water resources are utilized for productive uses in SSA, compared to 36% and 51% for South Asia and the Middle East and North Africa, respectively (Abric et al.2011). Irrigation investments in the region have historically focused on a few large-scale irrigation schemes, but the focus has recently shifted to farmer-led small-scale irrigation (SSI). This is partly due to the multifaceted challenges that are faced by large-scale dam-based schemes including high initial and maintenance costs and poor management. Unlike these large-scale, publicly-financed and often externally imposed irrigation schemes, SSI is inititated and controlled by smallholders, either individually or in groups, to irrigate relatively small plots using both traditional and modern technologies. Official statistics on irrigation use in Africa often underestimate SSI, which is shown to have signficantly higher returns compared to large-scale schemes (Woodhouse et al.2017; You et al.2011).

Despite the steady expansion of SSI in Sub-Saharan Africa, adoption is still below levels elsewhere, such as those in South Asia (Giordano and de Fraiture2014). While there is a lot of heterogeneity in what SSI entails, core characteristics include the use of fairly simple methods to transport water from its source to a field that is often small (Burney et al.2013; Burney and Naylor2012a; Kamwamba-Mtethiwa et al.2016). In most cases, water is transported using traditional methods (e.g., buckets and watering cans), while more modern technologies rely on human-powered pumps (e.g., treadle or rope and washer) and motorized pumps (e.g., diesel, electric, solar and wind) (Namara et al.2014).

The literature on adoption of SSI has identified several factors including poorly developed supply chains for irrigation technologies, weak property rights including land tenure, inadequate financing, limited access to complementary input and output markets, and access to information (Bjornlund et al.2017; Giordano and de Fraiture2014; Grimm and Richter2006; Meinzen-Dick2014; Meinzen-Dick et al.2012; Mwamakamba et al.2017; Namara et al.2014). A positive effect of social capital (including trust, norms and networks) on irrigation adoption has also been documented (Castillo et al.2021; Hunecke et al.2017).

Drawing on lessons from the psychology literature, adoption studies have also examined the role of a farmer’s locus of control-the extent to which an individual believes outcomes are determined by one’s action as opposed to by forces outside his/her control-in explaining adoption patterns (Abay et al.2017; Jonathan Malacarne2019; Taffesse and Tadesse2017). For example, farmers in Eastern Africa with a more external locus of control, i.e., those who believe weather conditions are more important in explaining maize harvest variability than their own input choices, are less likely to adopt maize cultivars proven to increase and stabilize maize production (Malacarne2019).

The objective of this study is to assess adoption patterns of different SSI technologies in selected regions of Ethiopia, Ghana and the United Republic of Tanzania (Tanzania). In addition, we use multivariate analysis to examine associations between plot-level SSI adoption and various biophysical and economic variables measured at different levels (plot, household and community) and estimate cluster effects. We define SSI technology as any gravity-based, manual, or motorized irrigation technology and model irrigation adoption using a hierarchical structure that considers plots nested within (farm) households, and households nested within communities. The study countries represent a broad set of agroecological conditions and have acknowledged the contribution of irrigation to their agricultural growth and development agenda. We find significant variation in SSI technologies across countries, highlighting the need for localized solutions to enhance the efficiency of existing systems and address possible constraints in access to complementary inputs that include agricultural labor and inorganic fertilizer. The variation in the role of cluster effects in explaining the variability in SSI adoption suggests the need to identify interventions that address constraints at different scales that matter to the specific context the most.

The rest of the paper is organized as follows. Section 2 discusses the irrigation policy environment in study countries and the relevant literature on determinants of SSI adoption. Section 3 describes the research method and presents descriptive results. Section 4 presents the results from the multivariate regression analysis. Section 5 concludes the paper by identifying potential policy interventions to enhance the use and efficiency of SSI in the target regions.

2. Literature

2.1. Irrigation policy environment

Irrigation has a significant potential to enhance agricultural production, reduce poverty and enhance food security by increasing the quantity and stability of (nutritious) food supply, enhancing incomes and reducing production risks (Burney and Naylor2012b; de Fraiture and Giordano2014; Hirvonen and Hoddinott2014; Ofosu et al.2014; Smith2004). Interest in the role of irrigation to enhance productivity, income, and resilience especially in Africa has strengthened in light of current and expected climatic risks (Williams et al.2018; Woodhouse et al.2017). Relative to large-scale irrigation schemes that require high initial and maintance costs, SSI carry lower costs and are under the control of smallholders. The internal rate of return to SSI in Africa is estimated at 28%, about four times that of large-scale dam-based investments (You et al.2011).

The study countries have acknowledged the importance of irrigation to achieve their development goals through various incentives. While 16%, 25% and 14% of the cultivated area in Ethiopia, Ghana and Tanzania can potentially be irrigated, only 5%, 0.5% and 2%, respectively, of their total cultivated area is equipped for irrigation (FAO2018). This includes area using full control irrigation systems (surface irrigation, sprinkler irrigation or localized irrigation where low pressure water is distributed in a pre-determined pattern through a piped network); floodwater harvesting or spate irrigation; or irrigated lowland areas (cultivated wetland, inland valley bottoms and areas along rivers).

Ghana’s 2007 national water policy aimed to more than triple the total irrigated area by 2020 (Governement of Ghana2007) despite the high cost of setting up new irrigation systems and the low share of irrigation investment of the total agricultural spending (3%) (Namara et al.2014; Regassa et al.2011; World Bank2017). Several large-scale public irrigation schemes have performed sub-optimally due to poor maintenance and a “civil service approach” to their management (Owusu2016). The 2010 Ghana Irrigation Policy focused on the development of small- and micro-scale irrigation schemes and enhancement of the efficiency and productivity of existing schemes, while promoting social inclusion and environmental sustainability.

Tanzania acknowledges the need for improving its irrigation infrastructure to transform and enhance the productivity and competitiveness of the agricultural sector. Since 2006, it has vowed to expand irrigation coverage to one million hectares by 2020 primarily through the development and rehabilitation of formal and mostly large-scale irrigation schemes (United Republic of Tanzania2009), while traditional and SSI systems are often viewed as less desirable or inefficient (Mdee et al.2014; Passarelli et al.2018). Traditional irrigation practices, known locally as vinyungu, are widely used as an alternative to rain-fed agriculture and without outside interventions to grow crops in valley bottoms by harnessing available water from rivers, springs and large river flood plains (Mkavidanda and Kaswamila1994).

In spite of Ethiopia’s huge water endowment, its agriculture is yet to fully benefit from irrigation due to policy, institutional, infrastructure and market constraints. Irrigation investments have historically focused on large-scale irrigation schemes with mixed results (Awulachew et al.2010). In recent years, SSI has become a policy priority to support rural poverty alleviation and growth (MOFED2006), as well as climate change adaptation (GoE2007). Fiscal and non-fiscal legislation have been introduced to promote private sector investment in the form of duty-free importation of equipment (including for irrigation), tax reductions and credit support for suppliers (FAO2015). All duties on agricultural machinery and irrigation equipment have been fully lifted as of April 2019.

2.2. Determinants of SSI adoption

Higher adoption of SSI in SSA is crucial for sustainable agricultural development, improved adaptation to climate variability, and food security. Despite proven effects of irrigation on water use efficiency as well as the amount, diversity and stability of agricultural production, its adoption in the region is very low. Adoption of improved agricultural technologies including SSI is influenced by several socioeconomic and biophysical factors operating at different scales. Extensive research has been done to examine role of liquidity constraints, infrastructure, missing or imperfect input and output markets, as well as informational and institutional barriers (Barrett2008; Dillon and Barrett2017; Duflo et al.2008, 2010; Foster and Rosenzweig2004; Suri2011). Specific to SSI, imperfect or missing markets for irrigation technologies will lead to suboptimal adoption even among farmers with purchasing power. Limited access to modern irrigation equipment and other complementary inputs can force farmers to adapt production decisions by substituting less efficient technologies (e.g., manual irrigation) for more efficient ones (e.g., motorized irrigation).

Plot characteristics, such as ownership status, size and location, can affect adoption through their effects on households’ incentive to invest on the land, the type of investments to be made, and proximity to water sources. For example, farmers who operate bigger plots may have more incentive to invest in capital-intensive SSI technologies compared to those who operate smaller plots, other things remaining the same. Motorized irrigation technologies are generally more capital-intensive than manual technologies, but they reduce drudgery and facilitate irrigation of bigger areas of land (Boutraa et al.2011; Shah et al.2013).

Household-level factors, such as wealth, household head characteristics (e.g., age, gender and education), family size, access to agricultural extension and financial services, can also affect irrigation adoption (Doss2006). Female-headed households often have limited access to productive inputs, information and advisory services, and markets that subsequently affect the rate and intensity of adoption (Peterman et al.2014; Theis et al.2018; Upadhyay2004). Better educated farmers with access to credit and agricultural extension services may be more likely to adopt modern irrigation technologies as are wealthier households (Ndiritu et al.2014; Salazar and Rand2016; Shah et al.2013).

While the gender of the household head is often examined in adoption studies, it is not uncommon for female farmers in male-headed households to manage their own plots (Koppen et al.2012). This highlights the need to control not only for the gender of the head, but the gender of the plot manager. Male and female farmers may opt for different technologies due to differences in access to input markets, agricultural extension services, social norms and preferences among other factors. Adoption of irrigation technologies often has gendered impacts, for example, Ghanaian men are found to take over irrigation using motorized pumps that were given to women (Bryan and Garner2020).

Smallholders often face shortages in agricultural labor, including for SSI. A study in Ethiopia found that close to half of SSI users reported facing labor shortages, twice the level of non-irrigating households (Asayehegn et al.2011). Labor constraints often deter adoption especially among households who are unable to hire or access labor through community-based labor sharing arrangements. At the same time, adoption of labor-intensive agricultural technologies including manual irrigation can worsen labor shortages and affect time spent on non-agricultural activities (Komarek et al.2018; Long et al.2009; Mafongoya et al.2006; Moser and Barret2003).

The acquisition, use, and transfer of land in Africa is regulated by both customary laws and statutory laws enacted at the national level. Even in countries where national land legislations do not officially accept customary laws, local chiefs and ethnic leaders often have considerable de facto power in allocating land (Michalopoulos and Papaioannou2015). Inheritance remains the main means through which land is acquired, with the patrilineal inheritance system being the most common (Linkow2019). Land tenure insecurity not only reduces the incentive to make longer-term investments but limits households’ access to credit given that land is an important source of collateral (Mwamakamba et al.2017). Finally, landscape level factors, such as land characteristics (e.g., soil type), climatic factors (e.g., precipitation), and access to input and output markets, affect adoption through their impact on the suitability and profitability of SSI (Worqlul et al.2017).

Our study contributes to the literature on SSI in several ways. First, we analyze highly comparable cross-country data to examine cross-country commonalities and differences. Second, while the bulk of the empirical work analyzes the determinants of a binary irrigation adoption outcome measured at the household level, we examine the determinants of plot-level adoption of different irrigation technologies-gravity, manual or motorized-to examine opportunities for and constraints to adoption of specific irrigation technologies. This is important since costs, expected returns, production risks, and other factors that affect farmers’ choices vary by the type of irrigation technology. Farmers in Africa often operate more than one plot with varying biophysical (e.g., soil type) and socioeconomic (e.g., the gender of the plot managers) characteristics that can influence the quantity and quality of inputs applied on the plot, making the plot (versus the household) an ideal unit of analysis. Finally, by modelling SSI adoption using multilevel modelling (plots nested in households that are nested in communities), we measure the contribution of group-level errors to the unexplained variation in irrigation adoption.

3. Material and Methods

3.1. Data and summary

Data were collected as part of a project (called ILSSI¹) that aims to promote the use of SSI. Details about the household surveys are provided in the supplemental material. Two rounds of surveys were implemented in each country during 2014-2018 to collect detailed agricultural data, including on the use of SSI. Ethiopia and Tanzania have a bimodal production, with a long rainy season between June and September (Ethiopia) and March and June (Tanzania) and a short rainy season between February and April (Ethiopia) and October and January (Tanzania). Northern Ghana has a single rainy season covering June through October. We use agricultural production and irrigation use data for short rainy seasons. Table 1 summarizes the study sample by country and survey round.

SSI technologies vary by energy or pressure requirement (gravity versus pressurized irrigation), the manner of application or conveyance (surface, subsurface or overhead), and wetted area of the crop root (flood, drip or sprinkler). We find that rivers and groundwater are the main sources of irrigation water in Ethiopia, each accounting for about 44% of the irrigated plots. Rivers are by far the most common water sources in Tanzania (80%) while groundwater and dams account for, respectively, 53% and 24% of the irrigated plots in Ghana (Figure 1). In terms of methods for transporting water, flood irrigation is the most common in Ethiopia and Tanzania accounting for, respectively, 41% and 57% of the irrigated plots. On the other hand, about 85% of the irrigated plots in Ghana use traditional methods, such as buckets, watering hoses, and watering cans. These cross-country differences highlight contextual differences that may promote or hinder irrigation adoption as well as determine the scale of irrigation and water use efficiency. It has been noted that up to 40% of water transported through flood irrigation can be lost due to soil leakage, waste, deep percolation loss, evaporation or runoff potentially making it one of the most inefficient irrigation methods (Butts2019).

Since the drivers of adoption vary by technology and given the small sample for some of the SSI technologies, we group SSI technologies used by study households into three as shown in Figure 2–gravity-based, manual (hand or foot pumps, watering cans, buckets or hoses), and motorized (petrol, diesel, kerosene or solar pumps). Relative to motorized technologies, gravity-based technologies have lower energy input to move water. After the water is delivered to the highest elevation end of a sloped field or the edge of a level field, it will be released and distributed down or across the fields through gravity using different water delivery systems. Most of the plots in Ethiopia are irrigated using manual (43%) or motorized (39%) technologies; Tanzanian irrigators are more likely to use gravity-fed methods (67%) while those in Ghana are more likely (89%) to use manual technologies. Exceptionally high use of manual technologies in Ghana has previously been documented with over 60% of surveyed households reporting buckets as the most common irrigation technology (Namara et al.2014).

We construct several plot, household and community-level variables that may affect irrigation adoption. Plot-level variables include the size, soil quality, and water holding capacity of the plot, plot’s ownership status, the gender of the plot manager, the amount of labor and inorganic fertilizers used per hectare, and the distance between the plot and the homestead. Household-level factors we examine include household size and wealth, characteristics of the household head (age, gender and education), households’ access to agricultural extension and financial services, and their membership in farmers groups.

Some of the cross-country differences in sample characteristics include smaller plots in Ethiopia; higher incidence of joint (by household head and spouse) ownership of plots in Ethiopia (60% versus 40% and 17% in Tanzania and Ghana, respectively); higher intensity of inorganic fertilizer application in Ghana; and better access to agricultural extension agents in Ethiopia (93% versus 58% and 45% in Tanzania and Ghana, respectively). About 90% and 84% of the plots in Ethiopia and Ghana, respectively, were acquired through inheritance or allocation by local government or community chiefs while 37%, 26% and 25% of the plots in Tanzania were acquired through, respectively, inheritance, purchase and rental (supplemental Tables A2–A4).

Across the board, households reported using higher agricultural labor per hectare on irrigated plots relative to non-irrigated plots, with the highest intensity of labor use in Ethiopia associated with manual irrigation (Figure 3, Panel A). Statistical tests (Goldman and Kaplan2018) reject the null that the cumulative distribution functions do not vary by plot’s irrigation status at the 1% level. Labor use per plot is the highest for Ghana, on average, where manual irrigation technologies were used on about 90% of the irrigated plots (Figure 3, Panel C).

3.2. Multivariate analyses

We investigate the determinants of SSI adoption using a hierarchical or multilevel model with random intercept. Generalized linear models assume independent model residuals that fails to hold when data have a hierarchical structure. Omitted cluster-level variables often introduce unobserved inter-cluster heterogeneity and clustering thereby violating the assumption of independent residuals. A classic example is the clustering of students within classrooms, and of classrooms within schools (Goldstein2003). A single-equation framework that disregards such nesting will underestimate standard errors and increase the likelihood of Type I error (Hox2000). Hierarchical modelling allows analysts to simultaneously estimate the effects of group-level variables and group effects. Hierarchical modelling has also been used to study the determinants of the adoption of different types of irrigation technologies (Fan and McCann2020).

Intra-household plot-level irrigation use may be correlated due to factors such as wealth and gender of the household head. For example, irrigating multiple plots may be an optimal strategy for a household that already invested resources to acquire irrigation technologies (e.g., motorized pumps). Groundwater irrigation often requires farmers to dig out open wells, deeper boreholes or ponds and apply water to their fields using a motorized pump, watering cans or buckets. As such, a household’s financial (to purchase motorized pumps) and labor (to manually transport water using traditional methods) assets will likely introduce household-level clustering. Another factor that may cause plot-level correlation in irrigation is the gender of the household head. Female-headed households often adopt irrigation at a much lower rate than male-headed households (van Koppen et al.2012).

Irrigation adoption within a given area (e.g., community) may be correlated due to factors such as amount of rainfall, access to and quality of extension services, as well as irrigation water sources and technologies. For example, in gravity-fed surface irrigation systems, water is conveyed from surface sources (e.g., rivers or reservoirs) and is distributed to individual fields through a network of canals, relying on gravity as the driving force. Canal construction often requires resources that are beyond the means of a single household and hence the type and quality of network of gravity-fed surface irrigation available in an area will likely affect the extent of irrigation adoption in the area. Households in communities with more adopters may also be more likely to adopt as has previously been documented (Salazar and Rand2016).

To guide model selection, we first estimate a three-level, two-level and standard multinomial (for Tanzania and Ethiopia) or binomial (for Ghana) logit model without clustering using the gsem command in Stata software. To recall, the nominal response variable for irrigation adoption has four categories for Tanzania and Ethiopia (no irrigation, gravity, manual and motorized) and two categories for Ghana (no irrigation versus manual irrigation). Multinomial logit models are multi-equation models where a response variable with $m$ outcomes will generate $m-1$ equations, each of which is a binary logistic regression comparing a given outcome with the reference group. Next, we conduct two likelihood-ratio (LR) tests for each country (see Wooldridge (2010) for general discussions). The first LR (LR1) test compares the three-level irrigation model (plots nested in households, and households nested in communities) with the traditional logit model. The second LR (LR2) compares the three-level model with the two-level model (plots nested in households). An LR test compares log likelihoods of two models and tests whether the difference is statistically significant. If the difference is significant, the less restrictive model (i.e., three-level model in our case) is said to fit the data better than the more restrictive model (i.e., two-level or traditional logit model in our case). The LR test statistic has a chi-squared distribution with degrees of freedom (dof) $k$, where $k$ is the difference in the number of dof for the two models being compared.

Chi-squared (${\chi }^2)$ and probability ($p)$ values from LR tests are as follows. For Ethiopia: LR1: ${\chi }^2=607$, $p=0.000$ and LR2: ${\chi }^2=325$, $p=0.000$; for Tanzania: LR1: ${\chi }^2=151$, $p=0.000$ and LR2: ${\chi }^2=70$, $p=0.000$; for Ghana: LR1: ${\chi }^2=6.4$, $p=0.040$ and LR2: ${\chi }^2=4$, $p=0.045$. These tests reject the null hypothesis that a traditional logit (LR1) or two-level logit model (LR2) is nested within a three-level model. Stated differently, accounting for the three-level nesting structure of the data improves the overall fitness of the irrigation adoption model. Within a multilevel framework, the number of clusters is more important than the number of units per cluster to accurately estimate the standard errors of the variances of the latent variables (Sommet and Morselli2011). The number of communities (level-3) in our sample ranges between 12 and 17 depending on the country, while the number of households (level-2) and plots (level-1) is relatively large. Likelihood-ratio tests (see Wooldridge (2010) for general discussions) confirm that three-level multinomial logit model is more appropriate than both the classical multinomial logit model with single intercept and a two-level multinomial logit (with plots nested in households), significant at 1% for Tanzania and Ethiopia and at 5% for Ghana (see supplemental Table A5).

Since almost all irrigated plots in Ghana used manual technologies, we estimated a three-level binomial logit model. For Tanzania and Ethiopia, we estimate a three-level multinomial logit model shown in Eq. (1) where the latent response ($Y_{ijk})$ has four categories-non-irrigated (base category), irrigated with gravity, irrigated using manual technologies and irrigated using motorized technologies. The corresponding multinomial logit link function is shown in Eq. (2). Technology-specific analysis of determinants is important since costs, (expected) returns and production risks vary by irrigation technology that in turn influence farmers’ choices. For example, evidence from nine African countries shows that net value added is significantly higher both per acre and per labor for fields irrigated with motor-based technologies but marginally higher for gravity-based and manually irrigated fields relative to non-irrigated fields (Shah et al.2013).

where $m$ is an index for response category; $j$ and $k$ are indices for, respectively, the second (household) and third (community) levels; $i$ is index for first (plot) level assumed to be nested in $j$ that is in turn nested in $k$; $Y$ is SSI adoption status that takes one of the four outcomes described above (non-irrigated, gravity-based, manual and motorized); $P$ is the probability of the respective outcome, with one category (non-irrigated category) used as the base outcome for which all parameters and the random error are set to zero; $\alpha$ is the intercept term; $B$ is a matrix of fixed-effect coefficients of plot, household, and community level conditioning variables in $X$ assumed to be the same across response categories.

The parameters ${\xi }_j\mathrm{\text{and}}{\psi }_k$ capture unobserved heterogeneity, latent variables or random effects at the household and community levels, respectively, while ${\epsilon }_{ijk}$ is the plot-level model error term. The errors at the second and third levels are assumed to be independent with multivariate normal distribution [${\xi }_{jk}^{\prime }={({\xi }_{jk}^2,\dots ,{\xi }_{jk}^{\mathrm{\text{M}}})}^{\prime }\sim iidN({0,\mathrm{\Sigma }}_{\xi })$; ${\mathrm{\Psi }}_k^{\prime }={({\psi }_k^2,\dots ,{\psi }_k^M)}^{\prime }\sim iidN(0,{\mathrm{\Sigma }}_{\psi })$. Assuming conditional independence of the random errors, the likelihood function is computed by integrating the column vector of random effects for each level ($u)$ as shown in the following equation :

where ${\theta }^{\prime }=({\alpha }^2,\dots ,{\alpha }^M,B^2,\dots ,B^M,{\mathrm{\Sigma }}_{\xi },{\mathrm{\Sigma }}_{\psi },{\mathrm{\Sigma }}_{\epsilon })$, $R^r$ is the set of values in an $r$-dimensional space, $f(\cdot )$ is the conditional probability density function, $\varphi (\cdot )$ is the multivariate normal density, and ${\eta }_u$ and ${\mathrm{\Sigma }}_u$ are the mean and covariance of $u$, respectively. For the three-level model, the conditional density function is given by

where $j^{\prime }$ and $k^{\prime }$ are the number of observations at the second and third level. Given that the integrals in Eq. (2) do not have a closed-form solution, the mean–variance adaptive quadrature method (see Rabe-Hesketh et al. (2005) for details) is employed here. The random effects are not directly estimated as model parameters, but instead are summarized as variance components $\xi$ and $\psi$ that measure the unexplained model variation explained by the group-level errors. The conditional $g$-level intra-cluster correlation (ICC) (${\rho }^g)$ is given by where ${\sigma }_l^2$ is the estimated $l$-level variance and $\gamma$ is the estimated variance of the first (plot) level residual given by ${\pi }^2∕3$ for a logistic distribution (McCullagh and Nelder1989; StataCorp2013). In the case of three-level modelling, only two ICC are estimated for the third or community level (${\rho }^3)$ and the second or household level (${\rho }^2)$. The parameter ${\rho }^2$ measures the correlation between irrigation status of plots $i$ and $i^{\prime }$ nested in the same household while ${\rho }^3$ measures the correlation between irrigation status of plots $i$ and $i^{\prime }$ nested in the same community but different households.

For Ethiopia and Tanzania, we report the exponentiate coefficients-the relative risk ratio ($\mathrm{\text{RRR}})$. The RRR between responses $m$ and $m^{\prime }$ measures how the risk of the outcome falling in $m$ relative to falling in $m^{\prime }$ changes for a unit change in a conditioning variable $(X)$ and is computed as shown in Eq. (6) (omitting subscripts). Considering $m^{\prime }$ as the no-irrigation category, for each SSI type $m$ (gravity, manual or motorized), an $\mathrm{\text{RRR}}>1$ indicates that the likelihood of using $m$ relative to the likelihood of no irrigation on a plot increases as $X$ increases while an $\mathrm{\text{RRR}}<1$ indicates the opposite association between the likelihood of using $m$ relative to $m^{\prime }$ decreases with increase in $X$.

where $exp$ means exponentiation and $Z$ is a matrix of conditioning variables excluding $X$. Given the binary irrigation adoption outcome indicator for Ghana (non-irrigated versus manual irrigation), a three-level binary logit model is estimated and we report exponents of parameter estimates from the binary logit model, known as odds ratios (OR). See Wooldridge (2010) for general discussion about binary logit model and OR. OR measures how the risk of a plot being irrigated manually relative to being non-irrigated changes for a unit change in $X$, with $\mathrm{\text{OR}}>1$ indicating a positive association between $X$ and the likelihood of using manual irrigation relative to no irrigation and $\mathrm{\text{OR}}<1$ indicating a negative association between the same. For completeness, we also present results from traditional multinomial/binomial logistic regression without accounting for data clustering in the supplemental material.

4. Regression Results and Discussion

4.1. Cluster effects on variability in irrigation use

Table 2 presents estimated cluster-level variances from the three-level irrigation model, estimated separately for each country, along with ICC estimates computed using Eq. (5). Plot-level (level-1) variance component for the standard logistic model is given by $3.3(=\pi 2∕3)$; household-level (level-2) ICC captures the correlation in irrigation adoption status of plots within a household; and community-level (level-3) ICC captures the correlation for plots in different households within a community. Community-level effects explain 46% (Tanzania), 35% (Ethiopia) and 23% (Ghana) of the total variation in SSI adoption not explained by the plot-level model while household-level effects explain 54% (Ethiopia), 27% (Ghana) and 22% (Tanzania).

Community-level ICC is the highest in Tanzania (0.46) while household-level ICC is the highest in Ethiopia (0.88). As shown in Figure 2, the share of motorized capital-intensive technologies is the highest in Ethiopia while that of gravity-based technologies is the highest in Tanzania. Gravity-based system operations generally affect the entire community, while the use of groundwater pumps is relatively less dependent on community-level factors.

These results suggest that interventions to promote SSI may be more effective in Tanzania and Ethiopia if they target, respectively, community- and household-level determinants of water availability. Community-level investments in Tanzania in the form of, for example, investments in dams, conveyance and applications systems as well as strengthening the capacity of WUAs may be an effective strategy. Indeed, both the Tanzania National Water Policy (2002) and Water Resource Management Act (2009) acknowledge the importance local institutions, such as WUAs, to better manage water allocation by engaging in maintenance activities, collecting seasonal fees, and determining the water requirement of farmers’ crops (Aarnoudse et al.2018; Woodhouse et al.2017).

The fact that ICC is relatively low in Ghana could partly be due to the dominance of individual and shallow groundwater wells. While household-level factors are relatively more important in explaining adoption patterns in Ethiopia and Ghana, there are major differences in irrigation technologies. The dominance of manual irrigation in Ghana is likely related to the poorer agroecological and socioeconomic conditions, whereas study areas in Ethiopia are in regions with high agricultural potential.

4.2. Determinants of irrigation use in Ethiopia

Regression results for Ethiopia show several variables having significant association with irrigation use (Table 3). At the plot-level, the gender of the plot decision maker, travel time between homestead and the plot, soil type, as well as the intensity of labor and inorganic fertilizer use are all important. About 4% and 33% of plots in our sample were managed solely by women (female-only) and men (male-only), respectively, with 60% managed jointly. The lower incidence of plot ownership and management among women has been documented across Sub-Saharan Africa (Sheahan and Barrett2017). The odds of using manual or motorized technologies on male-only plots are lower as are the odds of using motorized technologies on female-only plots, compared to jointly managed plots. Specifically, the relative risk of using motorized versus not using irrigation would be expected to decrease by a factor of 0.6 for male-only plots and by 0.19 for female-only plots, both relative to jointly managed plots and significant at the 10% level. Comparing the share of female-managed plots of the total number of motorized or manually irrigated plots, the share is lower for motorized plots (4%) versus manually irrigated plots (9.5%). Previous evidence from several African countries also shows a higher likelihood of using manual technologies, such as buckets among women farmers, while men farmers are more likely to use motorized technologies (Njuki et al.2014; Theis et al.2018; van Koppen et al.2012).

Relative to clay soils, sandy soils have lower likelihood of being irrigated using gravity-based technologies but a significantly higher likelihood of being irrigated using motorized technologies. Soil type is one of the factors affecting irrigation strategy, with clay soils having lower water infiltration rate but higher retention capacity while the opposite is true for sandy soils (Goldy2012). The higher infiltration rate makes gravity-based surface irrigation more challenging on sandy soil, while the effectiveness of other methods, such as sprinklers, is unaffected by infiltration rate (Alhammadi and Al-Shrouf2013).

The intensity of agricultural labor and inorganic fertilizer use is significantly higher for all irrigated plots except those that are manually irrigated. Strong complementarity between irrigation and inorganic fertilizer has previously been documented, where use of irrigation or inorganic fertilizers on their own may result in suboptimal outputs unless adequate amounts of both are used (Wichelns2006; Yilma and Berger2006). In areas with unpredictable rains, irrigation can increase the incentive to use inorganic fertilizer since stable supply of soil moisture enhances fertilizer use efficiency.

Higher labor intensity on irrigated plots could be due to increased labor use associated with the extraction, delivery and application of water or due to higher labor use on irrigated plots for other agricultural activities, such as fertilizer application. The use of buckets, watering cans and hoses is widespread in Ethiopia and other countries in the region significantly increasing labor demand including of family members, limiting irrigated area, and rendering SSI an unprofitable business (Ringler2021). Labor demand for fertilizer application can also be quite high depending on the technique for fertilizer application. According to one study in Ethiopia, for example, labor demand for microdosing and top dressing of maize plots is about twice that for broadcasting (Sime and Aune2014).

Most of the household-level socioeconomic variables we controlled for do not have a significant association with the likelihood of irrigation use except lower likelihood of manual irrigation among households with older or more educated heads or those with more household members, higher likelihood of using manual irrigation among households that belong in traditional savings groups (known locally as iqqub), and higher chance of using motorized technologies among households with better educated heads. The association between biophysical suitability for SSI and likelihood of using motorized irrigation is also positive, although the magnitude is relatively weak.

Results from multivariate logit model that does not account for data clustering show that, as expected, a lot more parameters have a statistically significant association with SSI (see supplemental Table A5). These include the size of the plot (positive association), total land area operated by the household (positive) and livestock wealth (positive association). Nonetheless, these estimates are likely prone to Type I error given the discussion in Sec. 3.2 and are presented here by way of checking the sensitivity of parameter estimates from our preferred hierarchical model.

4.3. Determinants of irrigation use in Tanzania

As summarized in Figs. 1 and 2, surface or flood irrigation using rivers is the most common in Tanzania, where surface irrigation is widely practiced using furrows and basins with conveyance by both lined and unlined canals. Given that water flows downhill, application of flood irrigation requires fields that are flat or have been scraped flat. Unlike with manual technologies, flood irrigation may allow farmers to cultivate bigger areas of land.

Results summarized in Table 4 indeed show that bigger or flatter plots are more likely to be flood irrigated. Flat plots also have significantly higher odds of being irrigated using manual technologies. Relative to purchased plots, rented plots are less likely to be manually irrigated but more likely to be irrigated using motorized technologies. This is in line with the descriptive summary where the share of motorized rented-in plots is about three times that of motorized purchased plots (3.6% versus 1.3%). Most Tanzanian farmers cultivate land designated as community land and managed through traditional tenure systems with a relatively more active land rental market from which up to 9%, mostly young, farm households benefit (Abay et al.2021; Ricker-Gilbert and Chamberlin2018). This may be due to farmers with access to groundwater not being able to afford motorized pumps and deciding to rent out their plots instead. The fact that motor pumps can easily be moved around may also make them easier to use by farmers who can afford to rent land.

Most of the household-level control variables do not have a significant association with irrigation adoption, except significantly lower use of motorized irrigation among female-headed households (versus male heads) and higher odds of motorized irrigation among households who interacted with agricultural extension agents. Lower adoption of irrigation among female-headed households has previously been documented (Koppen et al.2012).

These results hold when we estimate a multivariate logit model (see supplemental Table A7). In line with the finding for Ethiopia, bigger plots have higher odds of being cultivated using gravity-based technologies. Unlike the finding for Ethiopia, and relative to jointly managed plots, plots solely managed by women are more likely to be irrigated using motorized or manual technologies. On the other hand, female-headed Tanzanian households (15% of the sample) have significantly lower odds of using motorized and gravity-based irrigation, relative to male-headed households. While female-headed households often are less likely to invest in and adopt technologies, women may invest more in plots that they control (Chamberlin et al.2015; Theis et al.2018; van Koppen et al.2012). Results from the simple multivariate logit regression are reported in supplemental Table A6 for completeness.

4.4. Determinants of irrigation use in Ghana

Table 5 presents regression results for Ghana. Relative to plots allocated to households by local chiefs, those inherited or rented-in appear to have lower odds of being irrigated. Land under customary rules accounts for about 80% of the total land in Ghana with traditional rulers or chiefs having significant authority to govern and manage land (Antwi-Bediako2018). In our sample, about 43% of the plots were allocated to households by local chiefs while 42% were inherited. Further research is needed to better understand the role of different land tenure rights on irrigation adoption in the region.

In line with the findings for Ethiopia and Tanzania, agricultural labor use intensity is positively associated with irrigation. As shown in Figure 1, more than 85% of the irrigated plots used buckets, hoses, and watering cans that require extensive labor just for transporting water. In most cases, family labor is used to operate irrigation systems with little use of hired labor. Irrigating farmers in the region have previously identified high labor requirements and drudgery among the factors that made them abandon irrigation (Namara et al.2014).

The only household-level variable that has a statistically significant and negative association with manual irrigation use in our model is the amount of loan received over the past year. This seemingly counterintuitive result may be explained by the fact that rural financial institutions in Sub-Saharan Africa generally do not offer targeted financial products for irrigation development (Merrey and Lefore2018). Specific to Ghana, it is not uncommon for smallholder farmers to engage in small and informal businesses to supplement their income during the lean season and cope with climatic shocks (Nin-pratt and Mcbride2014), and rural financial institutions often are less reluctant to lend to non-agricultural businesses. Results from the simple binomial logit model are comparable to the results discussed here (see supplemental Table A7).

5. Conclusions and Implications

Sub-Saharan Africa’s (SSA) heavy reliance on rainfed agriculture with the associated climatic risks limit agricultural productivity and the potential to improve rural livelihoods. The challenges of rainfed production are expected to worsen in the coming decades with climate change. Despite the potential for irrigation to transform agriculture in the region, only a small fraction of the cultivated land is currently irrigated. Irrigation investments have historically focused on large-scale dam-based schemes, but there is now strong interest in the role of SSI in bridging the yield gap, promoting resilience to climatic risks, and tackling poverty and food insecurity. SSI technologies range from simple hand watering methods (e.g., hand buckets, hose and watering cans) used for home gardens and nurseries, to manually operated small treadle or motorized (using kerosene, diesel, electric or solar) pumps, and gravity-fed systems. Compared to large-scale schemes, SSI technologies require low capital investments and management skill, while their labor demand depends on the specific technology used to lift, transport and apply water.

We examined the associations between plot-level SSI adoption and various socioeconomic and biophysical factors in selected regions of Ethiopia, Ghana, and Tanzania—countries that have vowed to make irrigation an integral part of their agricultural development agenda. The multivariate analysis employs hierarchical modeling, where we assume (and confirm) data nesting-plots nested within households and households within communities. Descriptive results show significant cross-country differences in the sources and methods of transporting water. Surface irrigation based on river water is the most common in Tanzania; manual irrigation using groundwater is the most common in Ghana; and Ethiopian farmers are more likely to use manual and motorized technologies to transport water from rivers and groundwater.

These differences highlight the need for localized solutions to promote SSI and amplify its economic benefits. For example, surface irrigation is often prone to a high level of inefficiency in conveyance and on-site application and, when that happens, there is a need for interventions that address inefficiencies. Similarly, manual technologies, like watering cans that are the most used technologies in Ghana and other places in the region, can be made more efficient (e.g., by adding a hose on top of the spout to create sprinkler effects) and easier to manage (e.g., by introducing a carry-pole for easier water delivery).

Agricultural labor, especially for manual technologies, and inorganic fertilizer are positively associated with SSI use across study countries. In settings where there are severe labor shortages, enhancing access to labor-saving water lifting, delivery, and distribution technologies will be essential. This can be achieved through motorized and solar-powered pumps that are productivity enhancing for both land and labor. Solar-powered technologies, claimed by some as the most promising solutions to tackling climatic risks and hunger in Africa, may be especially appealing to farmers who live in off-grid areas, that is, most of the farmers in SSA. In this regard, addressing supply chain constraints, including through rental services, can enhance availability and reduce high initial capital outlays of poor farmers. Higher fertilizer use on irrigated plots shows the importance of an integrated approach to maximizing yields and their stability. Our finding of higher odds of motorized irrigation on rented-in plots in Tanzania and lower odds of manual irrigation on inherited and rented-in plots in Ghana calls for additional research to better understand possible impact pathways.

Looking forward, and as Africa’s irrigated area expands, it is imperative for policymakers and water sector stakeholders alike to pay attention to socioeconomic inequalities that may arise, as well as health and environmental externalities. Monitoring changes in the quantity and quality of water resources, raising awareness (through extension agents for example), and strengthening the capacity of community-based institutions become increasingly important for more sustainable and efficient water management.

Before concluding our paper, we highlight some of the limitations of our study by way of identifying areas for future research. We do not have data on whether water delivery techniques are entirely due to farmers’ own efforts and investments or are partially supported by community-level investments in water infrastructure. To the extent there are interhousehold differences in this regard, plot-level modelling of irrigation use may not adequately capture resource implications of SSI adoption. In addition, the irrigation data used in the study are not nationally representative in any of the three countries, and as such results obtained for a given country may not hold in other parts of the country. It is also worth noting that multivariate regression results presented here are associations not causation, as the latter requires a different study design. We do not know, for example, whether a household’s decision to use manual (versus motorized) irrigation is driven by its access to labor (family, communal or hired), or whether its decision to adopt manual technologies has subsequently led to more use of family labor for irrigation at the expense of time spent on other livelihood and household activities, such as childcare. The same can be said about the other control variables in our model (e.g., fertilizer use) that are likely affected by irrigation decisions made prior. Future studies on SSI adoption should explore the role of farmer’s locus of control given that correct perceptions of climatic risks are essential for farmers to adopt appropriate adaptation strategies including SSI. Finally, we acknowledge that the study findings may not be extrapolated to Africa, especially considering the diversity in agroecology and irrigation potential in the region and the purposive selection of study areas.

Acknowledgment

This paper is funded by the Feed-the-Future Innovation Laboratory for Small Scale Irrigation (ILSSI), led by Texas A&M and supported by USAID under Cooperative Agreement No. AID-OAA-A-13-00055, and the CGIAR Research Program on Water, Land and Ecosystems. The authors would like to thank partners who collected household survey data analyzed in this study. These are the Association for Ethiopian Microfinance Institutes in Ethiopia, Sokoine University of Agriculture in Tanzania, and the University of Development Studies in Ghana.

Notes

¹ Additional information about the project can be found here: https://ilssi.tamu.edu/.