Adoption of small-scale irrigation technologies and its impact on land productivity:Evidence from Rwanda

Jules NGANGO,Seungjee HONG

Department of Agricultural Economics,Chungnam National University,Daejeon 34134,South Korea

Abstract In an attempt to identify solutions to the effects of erratic rainfall patterns and droughts that limit agricultural production growth,the Rwandan government has recently increased investments in irrigation development. In this study,we analyze the adoption of small-scale irrigation technologies (SSITs) and its impact on land productivity using cross-sectional data from a sample of 360 farmers in Rwanda. The study uses the propensity score matching technique to address potential self-selection bias. Our results reveal that adoption decisions are significantly influenced by factors such as education,farm size,group membership,gender,extension services,access to credit,access to weather forecast information,risk perceptions,access to a reliable source of water for irrigation,awareness of rainwater harvesting techniques,and awareness of subsidy programs. In addition,the results show that the adoption of SSITs has a significantly positive impact on land productivity. The study concludes with policy implications that highlight the need to promote the adoption of SSITs among farmers as a strategy to improve agricultural productivity and food security in Rwanda.

Keywords:small-scale irrigation,maize production,Rwanda,technology adoption,impact assessment,propensity score matching

1.Introduction

The agricultural sector is Rwanda’s largest economic sector,contributing approximately 31% to the country’s national income and providing employment for nearly 66% of the country’s total population (World Bank 2018). The Rwandan government has identified agricultural development as a key factor that will play a significant role in the achievement of Millennium Development Goal 1,which focuses on the eradication of extreme poverty and hunger (MINAGRI 2018). However,Rwanda’s agricultural production system remains subsistence focused and largely rain-fed where farming is usually limited to the two cropping seasons occurring during the 7-month rainy season. According to World Bank (2021),the distribution of average monthly rainfall in Rwanda during the period 1990-2019 indicate that there is excess rainfall from September until the end of May while the rainfall decreases from early June until late August. In some cases,this rain-fed agriculture is adversely affected by erratic rainfall patterns and droughts,leading to water shortages,crop failure and low crop yields (Nabahungu and Visser 2013). Furthermore,rural households that rely entirely on such subsistence farming systems are likely to face low levels of income,as they do not produce a surplus for the market (Burney and Naylor 2012).

The development of irrigation technologies is frequently cited in the literature as a key strategy among various strategies addressing the challenges of low agricultural productivity in Sub-Saharan Africa (SSA),including Rwanda (Dillon 2011;Youet al.2011;Burney and Naylor 2012). Irrigation water is considered a crucial input variable that can help farmers build resilience to the effects of erratic rainfall and drought. In this way,irrigation helps farmers intensify crop production in both the rainy season and dry season. Therefore,given that Rwanda has plentiful water resources from numerous river basins,lakes,and wetlands (Nabahungu and Visser 2013),irrigation development is given priority in the Strategic Plan for Agricultural Transformation (SPAT),which guides national agricultural policies in Rwanda (Bizoza and Havugimana 2013). Although an estimated irrigation potential of roughly 589 000 ha was identified,the total land area fully developed with irrigation schemes accounts for only 3% of the potential irrigation area (Bizoza 2014). This implies that there is still potential to improve agricultural productivity by expanding irrigated land area. The development of irrigation,as a strategy aimed towards agricultural production growth,rural livelihood improvement,and food security,necessitates enormous investments.

Recently,investments made in irrigation projects in SSA have grown,and the World Bank has doubled the value of loans given for irrigation projects across the region (Youet al.2011). To stimulate agricultural production growth through irrigation development,there is a need to identify incentives and deterrents to the adoption of irrigation. The empirical literature on the adoption of irrigation has identified how various farmers’ socioeconomic,institutional,and environmental factors influence the adoption rate. For instance,Caswell (1991) attributed levels of irrigation adoption to the quality of the soil,input costs,land size,labor wages,weather patterns,and returns in terms of output. Dinaret al.(1992) revealed that the adoption of modern irrigation technologies is driven by the cost of irrigation water,profitability levels,farm management attributes,and environmental characteristics such as the availability of surface water and groundwater. However,apart from the above socioeconomic,farm and environmentally specific variables,Koundouriet al.(2006) noted that the degree of irrigation adoption in agriculture is a function of the likelihood of production risks under such irrigation technologies.

Although these studies do not draw similar conclusions concerning the adoption of irrigation,the explanatory variables used are central to the process of understanding the factors that may influence farmers’ adoption decisions regarding irrigation technologies. Moreover,this stream of literature serves as a useful tool for agricultural extension officers and policymakers in the development and strategic dissemination of irrigation practices.However,there is a limited empirical literature on the welfare impacts of small-scale irrigation technologies (SSITs) in Africa. To date,few studies have analyzed the adoption and profitability of SSITs but no single study has attempted to assess the adoption of SSITs and its impact on farm productivity in Rwanda. Thus,the present article adds value to the emerging body of literature on irrigated agriculture in two ways. First,we provide a clear understanding of the major factors that determine the adoption of SSITs for the purposes of achieving effective and efficient targeting efforts of improving maize yields. Second,we examine the true effects of SSITs on land productivity by controlling for selection biases on yields and adoption decisions. To this end,a propensity score matching (PSM) approach is used to quantify the effects of SSITs by constructing treatment and comparison groups because our data are not generated as a random experiment.

The objectives of this paper are (i) to explore the factors that influence farmers’ decisions to adopt SSITs and their impacts on land productivity in Rwanda;and (ii) to discuss policy implications. In this study,we consider maize yields as a proxy for measuring land productivity.To address our primary objective,we first use the Logit model to estimate the probability of SSIT adoption in relation to various farm and household characteristics.The Logit model uses the maximum likelihood estimation method and assumes a logistic distribution of the error term. Next,we use the PSM approach to assess the impact of SSIT adoption on maize yields.1Yields and land productivity are used interchangeably throughout this paper.Previous studies on impact evaluation (Ali and Abdulai 2010;Becerril and Abdulai 2010;Dillon 2011;Villanoet al.2015) have used PSM as an appropriate model to manage the issue of selection bias when using cross-sectional data.

The rest of this paper is organized as follows. Section 2 describes the current state of small-scale irrigation in Rwanda. The analytical framework used is presented in Section 3. In Section 4,we list the sources of data and descriptive statistics employed. Section 5 presents our analysis of the empirical results. Finally,conclusions and policy recommendations are given in Section 6.

2.Current state of small-scale irrigation in Rwanda

Rwanda has identified a potential irrigation area of approximately 589 000 ha. However,only 18 000 ha of this land is developed with irrigation schemes (Bizoza 2014). A large portion of the irrigated area (approximately 99%) is supplied with surface water while groundwater irrigation supplies roughly 1% (AQUASTAT 2017).Irrigation schemes in Rwanda are categorized into large-and small-scale schemes. Here we describe small-scale irrigation schemes as those privately owned and managed by individuals or groups of farmers. In Africa,small-scale irrigation schemes are more widely adopted than other forms of irrigation due to their low-cost infrastructure investments and operational management (Xieet al.2017). On the other hand,large-scale irrigation schemes are dam-based wetlands financed by public institutions and development partners (Xieet al.2017). The implementation of large-scale irrigation schemes involves high investment costs,and their management is more bureaucratic. Consequently,recently imposed agricultural policy initiatives focus on the development and promotion of SSITs to intensify agricultural production in Rwanda (Kathiresan 2011;Nahayoet al.2017).

SSITs commonly used by farmers in Rwanda include drip and sprinkler irrigation systems with portable diesel pump units,pipes,and rainwater harvesting tanks,among other equipment. To ensure the easy access and affordability of irrigation kits and other farm input (e.g.,inorganic fertilizers,improved seeds,and mechanization),the Rwandan government has implemented a targeted input subsidy program since 2007. The SSIT subsidy program is the most recent agricultural subsidy policy imposed in Rwanda. It was introduced in July 2014 with the aim of promoting private farmer-based investments in irrigation (MINAGRI 2018). The subsidy intervention attempts to address challenges surrounding the high cost of irrigation equipment,which is attributable to Rwanda’s hilly topography. The subsidy scheme covers 50% of the investment cost and the government allows for tax-free importation of irrigation equipment to help make irrigation even more affordable to farmers (MINAGRI 2018).

Farmers’ access to irrigation is also facilitated by private service providers who supply and install irrigation kits in rural areas. Financial institutions also work closely with the Ministry of Agriculture and Animal Resources of Rwanda (MINAGRI) to assist poverty-stricken farmers who cannot afford 50% of the investment cost by offering them a loan or lease of SSIT kits. Although SSITs are privately owned and managed by farmers,the government has employed irrigation engineers in each district to assist farmers with the maintenance of SSIT equipment and water distribution channels (MINAGRI 2018). Eligibility criteria for farmers to obtain the subsidy include:(i) managing an arable land area of 1 to 10 ha;(ii) project profitability (to be determined by a financial evaluation team);(iii) availability of irrigation water;and (iv) use of a low pressure pump capacity not exceeding 5 bars of pressure (MINAGRI 2018).

Currently,the majority of farmers who have adopted SSITs are located in Rwanda’s Eastern Province that is particularly vulnerable to drought as its topography is characterized by lowlands with high temperatures (Twagiramungu 2006). Throughout the country,SSITs are mainly used for the production of fruit,vegetables,and staple crops,such as maize and wheat.

3.Analytical framework

3.1.Modeling the adoption decision and impact problem

The adoption of agricultural technologies in most developing countries tends to be constrained by limited financial resources,market imperfections,poor rural infrastructures,and a lack of information (Zeweldet al.2015). Despite these constraints,Asfawet al.(2012) argued that farmers should adopt a given farm technology only if its use can maximize their utility in terms of net returns. Thus,following Asfawet al.(2012),we apply the random utility framework to model the adoption of SSITs. We assume that farmers are risk neutral and opt for the agricultural technology that maximizes their utility function subject to input costs and other constraints.More formally,let us defineUiAas the utility that a farmeriderives from the adoption of SSITs andUiNas the utility derived from non-adoption. Under such assumptions,the farmer will normally adopt a form of technology if the utility derived from its adoption is greater than the utility of non-adoption,that is,(Abdulai and Huffman 2014). Since these two utility values are unobservable,we can express them as a function of observable factors such as farmer’s characteristics and technology attributes in the latent variable model as follows:

whereDiis a binary latent variable that takes a value of 1 for farmers who adopt SSITs and a value of zero otherwise.Xiis a vector of household characteristics and irrigation technology attributes;αis a vector of parameters to be estimated;andεiis the error term whereεi～N(0,σ).

Theoretically,the adoption of SSITs is expected to have a significant impact on an increase in agricultural production. Hence,in assessing this relationship,we assume that the outcome variable (Yi) (i.e.,land productivity in the present study) is a linear function of the adoption dummy variable (Di) and of other explanatory variables (Xi) as specified below:

Estimation eq.(2) can reflect a direct effect of the adoption of SSITs on the outcome variable,which is land productivity. However,this procedure might produce inconsistent estimates,as it considers that the adoption of SSITs to be barely affected by exogenous factors. Moreover,treatment assignment is not random due to individual self-selection and purposive program placement. Specifically,the issue of selection bias arises when unobserved variables create a correlation between the error term (εi) of the adoption specification and the error term (μi) of the outcome model. In this case,the use of ordinary least squares (OLS) will tend to yield biased estimates (Ali and Abdulai 2010).

To address the issue of selection bias,previous empirical studies have used various econometric models,including the instrumental variable (IV) method,Heckman’s two-step method,difference-in-difference (DID) matching,and propensity score matching (PSM).However,Heckman’s two-step method is limited in its dependence on the restrictive assumption that unobserved variables are normally distributed (Becerril and Abdulai 2010). The IV method is limited by a difficulty with finding at least one variable in its selection model to serve as a proper instrument in the outcome estimation.Further,the IV method depends on using a functional form in the estimation of the outcome equation (Ali and Abdulai 2010). The DID matching method is another estimator that yields reliable,unbiased estimates as it corrects for selection bias. However,the DID method is only applicable in studies that use panel data (Ali and Abdulai 2010),which are not used in the current study.To overcome the shortcomings of the above approaches,we use the PSM method proposed by Rosenbaum and Rubin (1983) to address the issue of selection bias with a cross-sectional dataset. The PSM method relaxes the functional form and distribution assumptions applied in outcome model specifications (Becerril and Abdulai 2010).

3.2.Propensity score matching (PSM) method

To estimate the impact of SSITs,we compare the outcome variables derived when farmers receive and do not receive treatment. The construction of this counterfactual outcome constitutes the main challenge in the assessment of the impact of technology adoption. Following Rosenbaum and Rubin (1983),the average treatment effect (ATE) in a counterfactual framework is estimated as:

Assuming that there is no selection bias,we measure the impact of SSITs on households that have adopted them as the average treatment effect on the treated (ATT) (Gilligan and Hoddinott 2007;Ali and Abdulai 2010;Dillon 2011):

whereXis a vector of household characteristics andE(|X,D=1) is the counterfactual outcome. Since the counterfactual outcome is not observed,the estimation of the ATT in eq.(5) is likely to be biased (Ali and Abdulai 2010).

We use the PSM method to match individuals who adopt SSITs with non-adopters with similar distributions on various observed covariates (Becerril and Abdulai 2010;Villanoet al.2015). The PSM,which is also referred to the probability of adopting SSITs,is based on two assumptions. The first is the conditional independence assumption,which suggests that for a given set of observable covariatesX,the adoption status and outcome variables are independent (Imbens and Wooldridge 2009).2As noted by Rosenbaum and Rubin (1983),the propensity score determined under the assumption of conditional independence is expressed as P(X)=Pr(D=1|X)=E(D|X). That is,the conditional distribution of pre-adoption characteristics (X) is the same for both adopters and non-adopters.The second assumption is the common support condition,which states that individuals with the same covariates have a positive probability of being both adopters and non-adopters such that 0＜Pr(D=1|X)＜1 is the overlapping condition (Becerril and Abdulai 2010).

Propensity scores for the adoption of SSITs are estimated using a logit model. Next,we match the treatment and comparison groups using three matching algorithms commonly employed in the estimation of average effects of a particular program or treatment when using cross-sectional data:nearest neighbor matching (NNM),kernel-based matching (KBM),and radius matching (RM) estimators. The nearest neighbor matching estimator is constructed such that each treated individual is matched with the control individual with the closest propensity score (Becerril and Abdulai 2010). Under the kernel-based matching estimator,each treated individual is matched with a weighted average of all control individuals within the region of common support (Heckmanet al.1997).3These weights are inversely proportional to the distance between the propensity scores of adopters and non-adopters.On the other hand,the radius matching algorithm involves matching a treated observation with the corresponding control observation lying within the specified range of propensity scores (the caliper) (Zeweldet al.2015). After applying the matching procedure,it is necessary to check whether the propensity score estimation is able to balance the distribution of variables across treatment and control groups.

The premise behind the balancing test is to verify whether there are no remaining differences in the covariates of the treatment and comparison groups (Ali and Abdulai 2010). In this context,Sianesi (2004) proposed a method for comparing pseudo-R2values obtained before and after matching to perform this diagnostic statistic. The pseudo-R2shows how well the independent variables influence the probability of participating in the program. After matching,there should be no systematic differences in the distribution of variables of interest across the groups of adopters and non-adopters and consequently,the pseudo-R2should be lower (Becerril and Abdulai 2010). Another balancing test commonly used in the literature involves computing the standardized mean bias to determine whether the observed biases were reduced after matching (Zeweldet al.2015). In addition,the covariate balancing assumption requires that the joint significance of all independent variables be rejected after matching (Ali and Abdulai 2010). However,the elimination of systematic difference,i.e.,an absence of observed bias in the distribution of covariates between treatment and comparison groups,does not ensure the robustness of estimators or an absence of hidden bias. Thus,there is a need to conduct a sensitivity analysis of estimated average treatment effects to manage the issue of hidden bias. For this purpose,we apply the bounding approach and examine whether the effect of unmeasured covariates on the outcome variables is strong enough to undermine the matching process (Zeweldet al.2015).

4.Data and descriptive statistics

4.1.Sampling and data collection

The present study uses cross-sectional data collected from a survey of 360 farmers located in the Bugesera,Kirehe,and Nyagatare Districts of the Eastern Province of Rwanda. The survey was conducted from July to August 2019 using a structured questionnaire. Prior to the formal survey,the questionnaire was pretested with 25 randomly selected farmers to determine the adequacy of our survey instrument. A multistage sampling technique was used to obtain a sample of 360 household farmers. In the first stage,after consulting the Ministry of Agriculture and Animal Resources of Rwanda (MINAGRI),three districts were purposively selected based on the predominance of small-scale irrigation practices and intensive maize production in these districts. Four administrative sectors were randomly selected from each district in the second stage. Then,a random sample of respondents was selected from each sector for personal interviews. Based on the list of farmers obtained from each extension officer at the sector level,a total of 1 197 individual farm household units were counted and recorded in all 12 sectors. Due to limited resources and time,34 respondents were randomly selected in each sector which make up a total sample of 408 farm households. However,after data cleaning,a total sample of 360 farm households remain valid (see Table 1 for a detailed distribution of the sample households by district). The survey solicited information on maize yields,farm-specific characteristics,the socioeconomic characteristics of farmers,and the use of small-scale irrigation practices.

4.2.Descriptive statistics

The summary statistics of variables presented in Table 1 indicate that roughly 41% of agricultural households in the study area have adopted SSITs. The average age of the sampled household heads is approximately 47 years and roughly 69% of the studied households are headed by males. The average level of education in the sample area is around 6 years of primary education. The average household includes approximately 7 members,and the average number of livestock owned by households is approximately 1.37 TLU.4TLU stands for tropical livestock units across various categories of livestock. This value is converted as follows:0.7 for cows;0.45 for heifers;0.1 for goats;0.1 for sheep;0.01 for chicken;and 0.2 for pigs (Source:Zeweld et al.2015).On average,households cultivate maize on a land area of 1.7 ha and produce 1 990 kg ha-1.

The results reported in Table 2 show significant differences between small-scale irrigating and nonirrigating farmers in terms of their farm-level and socioeconomic features. Specifically,the average yield of maize obtained by small-scale irrigators (2 250.29 kg ha-1) is significantly higher than that of nonirrigators (1 809.47 kg ha-1). In addition,irrigating farmers are highly distinguishable with regard to asset ownership,i.e.,they own more land and livestock than their non-irrigating counterparts. Considering household demographics such as age,family size,and education levels,our results indicate that irrigating farmers have higher average figures than their non-irrigating counterparts (see Table 2). We also find a significant difference in the proportion of maleheaded households among small-scale irrigators and non-irrigators. The descriptive statistics given in Table 2 further show that small-scale irrigators enjoy better access to extension services,credit,weather forecast information,and reliable sources of water than non-irrigators.

Table 1 Description of variables and summary statistics

Table 2 Differences in household characteristics by adoption category (sample mean)

It is also notable that small-scale irrigating farmers are more aware of rainwater harvesting techniques and considerably more willing to try new technologies than non-irrigators. However,there appears to be no statistically significant difference between irrigators and non-irrigators in terms of subsidy policy awareness. The results also show that 65% of non-irrigators manage farmland on steep slopes,which is an obstacle to water pumping systems,whereas only 28% of irrigating farmers manage this type of land. Irrigators are also distinct with regard to their membership to farmers’ associations,where a larger number of irrigators belong to farmers’ associations than no irrigators.

Noting the mean differences discussed above,in the next section we use econometric models to provide empirical evidence on adoption decisions and their impacts.

5.Results and discussion

5.1.Determinants of the adoption of SSITs

The results from the maximum likelihood estimation of the logit model for determinants of the adoption of SSITs in Rwanda are reported in Table 3.5We could not include temperature and precipitation variables in our logit model due to the lack of data. In addition,irrigation infrastructures were not included in the logit model because there are no differences in the existing irrigation infrastructures among three districts of the study area.The results indicate that eleven variables are statistically significant in influencing the probability of adopting SSITs. Specifically,high-level education significantly increases the likelihood of adopting SSITs. This finding is consistent with Zeweldet al.(2015) on the adoption of small-scale irrigation in Ethiopia and reflects the fact that education allows farmers to analyze and make appropriate decisions on the adoption of agricultural technologies (Abdulai and Huffman 2014).With respect to gender,male-headed households were found to be less likely to adopt SSITs than those headed by female,corroborating the results of Aliet al.(2016).However,the studies of Mandaet al.(2016) and Nahayoet al.(2017) indicated that female-headed households are less likely to adopt new agricultural technologies due to women’s limited access to and control over resources in SSA. Further,land size has a significantly positive effect on the likelihood of adopting SSITs. A plausible explanation is that farmers who own larger areas of land can easily use this land as collateral to obtain loans for investments in irrigation equipment. Similar results have been given in previous studies on technology adoption (Becerril and Abdulai 2010;Asfawet al.2012;Villanoet al.2015).

As expected,we find that farmer association members are more likely than non-members to adopt SSITs,consistent with previous studies demonstrating the importance of social capital in increasing the rate of farm technology adoption (Abdulai and Huffman 2014;Zeweldet al.2015). Regarding the effect of extension services on adoption decisions,our results indicate that households frequently visited by extension agents are more likely to adopt SSITs than those not receiving such frequent visits. As a plausible explanation,agricultural extension systems enable farmers to obtain information and practical knowledge on farm technologies (Abdulai and Huffman 2014). Similar findings have been given by Asfawet al.(2012),Abdulai and Huffman (2014),and Ali and Abdulai (2010) on technology adoption in Ethiopia,Ghana,and Pakistan,respectively. In addition,access to credit was found to be a significant factor increasing the likelihood of SSIT adoption. Indeed,this finding corroborates the findings of Abdulai and Huffman (2014),who acknowledge the significance of credit in helping rural farmers accumulate enough capital to invest in these technologies.

The results given in Table 3 also show that access to weather forecast information significantly increases the likelihood of adopting SSITs. The availability of weather forecast information enables farmers to make appropriate decisions in their farming operations,including those made on the use of irrigation. This finding supports earlier work on the adoption of farm technologies in Nigeria (Wossenet al.2017). Furthermore,the positive and significant coefficient of risk perceptions implies that farmers who are willing to try using new agricultural technologies exhibit an increased likelihood of adopting SSITs. This result supports the findings of Koundouriet al.(2006) on the adoption of irrigation technologies in Greece,noting that farmers who are not risk-averse are likely to adopt irrigation technologies to manage production risks and uncertainty. On the other hand,Marianoet al.(2012) found that risk-averse and profitoriented farmers tend to be more attracted to high-yielding technologies than others in the Philippines.

Table 3 Estimates of the logit model for determinants of the adoption of SSITs in Rwanda

We also find that access to a reliable source of water for irrigation plays a significantly positive role in farmers’ decision on SSIT adoption. This is an expected result since proximity to a reliable water source (e.g.,dams,rivers,and lakes) clearly facilitates the development of irrigation practices. Similar findings are also provided by Mangoet al.(2018) on the adoption of small-scale irrigation in southern Africa. Likewise,an awareness of rainwater harvesting practices significantly increases the likelihood of adopting SSITs. This is likely due to the fact that rainwater harvesting is considered an alternative source of water for irrigation. This finding corroborates a recent study on the adoption of small-scale irrigation in southern Africa (Mangoet al.2018). The results further show that the likelihood of adopting SSITs is significantly higher for farmers who are aware of subsidy programs on small-scale irrigation than for those not aware of subsidy programs.

5.2.Effects of SSITs on land productivity

As mentioned in the methodology section,we use the PSM approach to estimate the impact of SSITs on land productivity. PSM balances the distribution of independent variables for groups of SSIT adopters and non-adopters. Fig.1 displays the distribution of propensity scores and the region of common support for the two groups. As proposed by Caliendo and Kopeinig (2008),the density distributions of the estimated propensity scores for adopters and nonadopters should satisfy the common support condition. Consequently,27 of the treated observations that are found to be off-support need to be excluded from the analysis to ensure appropriate matching with respect to observable household characteristics for both adopters and nonadopters.

Fig.1 Distribution of estimated propensity scores by treatment and region of common support.“Untreated”refers to the group of non-adopters;“Treated:on support”refers to individual observations for the adoption group that reveal a corresponding match in the non-adoption group;“Treated:off support”refers to individual observations for the adoption group that do not show a corresponding match in the non-adoption group.

Average treatment effects on the treated (ATT),which illustrate the effect of using SSITs on maize yields,are presented in Table 4. ATT values were estimated using three PSM algorithms commonly applied in the empirical literature:NNM,KBM,and RM. The results from all three matching algorithms indicate that on average,the use of small-scale irrigation has a significantly positive effect on maize yields. Specifically,the adoption of SSITs increases average maize yields by roughly 193 kg ha-1when using NNM,by 197 kg ha-1when using KBM,and by 200 kg ha-1when using RM (Table 4). In other words,this implies that on average,farmers who use SSITs have roughly 193-200 kg ha-1higher maize yields than nonadopters. This result is consistent with Dillon (2011) who found the adoption of small-scale irrigation in northern Mali to have a significantly positive impact on maize productivity. Nevertheless,Adeotiet al.(2009) found no significant link between the adoption of agricultural technologies and productivity in Ghana.

As discussed in Section 3,covariate balancing tests and sensitivity analysis were required to assess the quality of matching processes and the robustness of our results. The results shown in Table 5 indicate that a very high degree of matching quality was attained. Table 5 shows a considerable reduction in bias resulting of the matching process. The estimates indicate that the level of bias is substantially reduced from 63% before matching to roughly 9-13% after matching,corresponding to a total reduction in bias of 79-86% (Table 5). In addition,the comparison of pseudo-R2values shown in the second and third columns of Table 5 demonstrates that in all matching algorithms,the pseudo-R2is lower after matching than before matching. This suggests that there are no systematic differences in the distribution of independent variables between irrigators and nonirrigators after matching. Moreover,theP-values of the likelihood ratio tests are very high after matching (Table 5),implying that the joint significance of the explanatory variables is rejected after matching. In general,the results of all covariate balancing tests applied in this study reveal that there is no systematic difference between irrigators and non-irrigators with respect to the distribution of covariates after matching.We can thus confirm that the proposed specifications of the propensity score estimation procedure are successful in terms balancing the characteristics of irrigators and non-irrigators.

Table 4 Propensity score matching (PSM) estimates of the impact of small-scale irrigation on land productivity and results of the sensitivity analysis

Table 5 Covariate balancing tests before and after matching

Hidden biases resulting from unobservable variables is further checked by applying the Rosenbaum bounds sensitivity analysis technique. The sixth column of Table 4 presents critical level of hidden bias (also known as gamma) results for all three matching estimators.The results obtained from KBM and RM indicate that the estimated gamma value is 3.75 at the 5% level of statistical significance. This value suggests that the estimates average treatment effects of adopting SSITs on land productivity are insensitive to unobserved biases that would triple the odds of using SSITs. In conclusion,the findings are robust to hidden biases,satisfying the conditional independence assumption of PSM.

6.Conclusion and policy recommendations

To enhance agricultural production in Rwanda,the Rwandan government has increasingly put efforts into promoting and disseminating agricultural technologies.Therefore,it is very important to assess the effects of using such technologies in improving agricultural productivity. The present paper analyzes factors that influence farmers’ decisions to adopt SSITs and the subsequent impacts of adoption on land productivity (i.e.,maize output per ha) in Rwanda. Our analysis is based on cross-sectional data collected from a sample of 360 farmers. The logit model is used to estimate determinants of the adoption of SSITs and the results show that SSIT adoption is significantly influenced by factors such as education,farm size,group membership,gender,extension services,access to credit,access to weather forecast information,risk perceptions,access to a reliable sources of water for irrigation,awareness of rainwater harvesting techniques,and awareness of subsidy programs. The results for the impact of SSIT adoption on land productivity were obtained through PSM estimation.The findings reveal that the adoption of SSITs significantly increases maize yields.

Such findings imply that government interventions should focus on helping liquidity-constrained farmers obtain easy access to credit. The results also highlight a need to improve farmers’ access to information related to weather forecasts,rainwater harvesting techniques,and irrigation subsidies to promote the adoption of SSITs. Indeed,policy measures such as strengthening effective extension services and farmers’ associations could help farmers overcome information barriers,and hence facilitate the adoption of SSITs. In addition,efforts to improve the education levels of farmers could increase SSIT adoption rates. There is also a need for policies and strategies aimed at helping farmers to access reliable sources of water in proximity to each village. Furthermore,our findings for the impact of SSIT adoption on land productivity highlight the need to increase investment in SSITs,which is a policy intervention that will require public-private partnership. Although SSITs are privately owned by farmers,the government still has a significant role to play in increasing the adoption rates and promoting the sustainability of irrigation development in Rwanda.

It is also imperative to acknowledge the limitations of this study. First,cross-sectional data restricts the extension of our research findings beyond one year and limits the control of selection bias due to problems of unobserved heterogeneity. Thus,future research could consider extending our analysis with panel data to better understand the effects of irrigation over time. In addition,future research should pay a particular attention to the impacts of irrigation on other important outcome variables,such as income,poverty,consumption expenditures,and food security status.

Acknowledgements

This work was supported by the Research Scholarship of Chungnam National University,South Korea.

Declaration of competing interest

The authors declare that they have no conflict of interest.