Can Agricultural Cooperatives Reduce Poverty? Heterogeneous Impact of Cooperative Membership on Farmers' Welfare in Rwanda
Abstract
We analyze the inclusiveness and effectiveness of agricultural cooperatives in Rwanda. We estimate mean income and poverty effects of cooperative membership using propensity score matching techniques. We analyze heterogeneous treatment effects across farmers by analyzing how estimated treatment effects vary over farm and farmer characteristics and over the estimated propensity score. We find that cooperative membership in general increases income and reduces poverty and that these effects are largest for larger farms and in more remote areas. We find evidence of a negative selection because impact is largest for farmers with the lowest propensity to be a cooperative member.
Improving the productivity, profitability, and sustainability of smallholder agriculture is argued to be the main pathway out of rural poverty in developing countries. Institutional innovations are believed to play a crucial role in this because they can help farmers overcome market failures (Hazell et al. 2010). There is a renewed interest in producer organizations such as cooperatives as an institutional tool to improve market participation of smallholder farmers, increase farm incomes, and reduce rural poverty (Bernard and Spielman 2009; Bernard and Taffesse 2012; Fischer and Qaim 2012a; Fischer and Qaim 2012b; Markelova et al. 2009; Shiferaw et al. 2009). Cooperatives are often associated with collective action and social capital and are therefore often thought to be more poverty reducing than other types of institutional innovations such as contract farming. To have an effect on poverty these emerging institutions need to be both inclusive (i.e., poor farmers need to participate) and effective (i.e., creating an impact on the income and well-being of participating farmers). The effectiveness of agricultural cooperatives in increasing farm revenue and farmers' income has been documented in different contexts. In a context of developing countries with a high incidence of poverty in rural areas, a positive income effect does not necessarily imply a poverty-reducing effect. If the poorest farmers are excluded from cooperative membership, cooperatives might aggravate existing inequalities and fail to contribute to poverty reduction. Although necessary to fully understand poverty effects, the inclusiveness and effectiveness of institutions are rarely assessed together.
The objective of this article is to assess the inclusiveness and effectiveness of agricultural cooperatives in Rwanda and evaluate their impact on poverty. Our focus on Rwanda is particularly relevant because of a rapid spread of agricultural cooperatives in the country and because of a high incidence of poverty in rural areas; the poverty headcount ratio at the national rural poverty line is 49% (World Bank 2014). Agricultural cooperatives are seen as an important institutional vehicle to improve the performance of the smallholder farm sector and to achieve rural poverty reduction in the country (Government of Rwanda 2011). Their numbers have been increasing very rapidly during the past couple of years (Verhofstadt and Maertens 2014), but it remains unclear what the consequences are for rural poverty reduction. A handful of qualitative studies point out that cooperatives in Rwanda are exclusive and aggravate existing inequalities in rural communities (Ansoms 2009; Ansoms 2010; Nabahungu and Visser 2011; Pritchard 2013) but there are no studies that quantify income and poverty effects in a comprehensive way.
We look at membership of smallholder farmers in agricultural cooperatives and analyze the impact of this membership on household income and poverty. We estimate mean income and poverty effects as well as heterogeneous treatment effects across farmers. We use several propensity score (PS) matching (PSM) techniques to estimate the average treatment effect (ATE) of cooperative membership on farm income and the likelihood of being poor. We then analyze how the estimated treatment effect varies over various farm and farmer characteristics and over the estimated PS. We find that cooperative membership in general has a positive impact on farm income and a negative impact on the likelihood of being poor but that the effect varies with farm size, distance to the market, and the availability of labor in the household.
Literature Review
Various empirical studies have verified how inclusive agricultural cooperatives are and investigated which farmers are included in (or excluded from) cooperatives. In general, participation in agricultural cooperatives is found to be closely linked to human and social capital (Hellin, Lundy, and Meijer 2009). For example, some authors find that farmers' level of education, age, farming experience, and access to social networks and information have a positive effect on the likelihood of cooperative membership (Bernard and Spielman 2009; Fischer and Qaim 2012a; Francesconi and Heerinck 2010; Ito, Bao, and Su 2012; Markelova and Mwangi 2010; Matuschke and Qaim 2009; Okello and Swinton 2007; Zheng, Wang, and Awokuse 2012). But physical capital and farmers' asset endowments matter as well. For example, land and livestock holdings are found to have a positive (but sometimes decreasing) effect on the likelihood of farmers to participate in agricultural cooperatives (Bernard and Spielman 2009; Fischer and Qaim 2012a; Ito, Bao, and Su 2012). Some studies conclude that the poorest farmers are excluded (e.g., Francesconi and Heerinck 2010; Ito, Bao, and Su 2012), whereas others point to a middle-class effect with both the poorest and the wealthiest farmers least likely to participate (Bernard and Spielman 2009; Fischer and Qaim 2012a). The prevailing evidence suggests that there are constraints to enter agricultural cooperatives in terms of requirements for human, social, and physical capital and that cooperatives are to some extent exclusive.
Concerning the effectiveness of agricultural cooperatives, there is a growing body of evidence that cooperatives positively affect farm revenues and farmers' income. Empirical studies have indicated positive effects of cooperative membership on producer prices and market participation (e.g., Abebaw and Haile 2013; Bernard, Taffesse, and Gabre-Madhin 2008; Bernard and Taffesse 2012; Fischer and Qaim 2012a; Francesconi and Heerink 2010; Holloway et al. 2000; Ito, Bao, and Su 2012; Shiferaw et al. 2009; Wollni and Zeller 2007). In addition, cooperative membership was found to increase the likelihood of adopting improved technologies such as mineral fertilizer (e.g., Abebaw and Haile 2013; Francesconi and Heerink 2010; Shiferaw et al. 2009). Some studies point to a positive impact on farm incomes and profits (e.g., Fischer and Qaim 2012a; Ito, Bao, and Su 2012; Vandeplas, Minten, and Swinnen 2013). Yet, the effects of cooperative membership on poverty have rarely been analyzed.
The overall poverty impact of cooperatives hinges on both inclusion (i.e., whether the poor participate in cooperatives) and effectiveness (i.e., whether there are income gains and who is gaining most). Therefore it is important to look beyond mean treatment effects. Only a handful of studies have specifically analyzed how the impact of cooperative membership changes with farm and farmer characteristics. Bernard, Taffesse, and Gabre-Madhin (2008) find that cooperative membership leads to a higher degree of commercialization for cereal farmers in Ethiopia, but the effect is larger for the largest farms and even negative for very small farms. Ito, Bao, and Su (2012) find that the impact of cooperative membership on farm income for watermelon farmers in China is twice as large for small farms as it is larger farms. Fischer and Qaim (2012a) find that the effects of participating in banana cooperatives in Kenya on commercialization, technology adoption, and farm income are more pronounced for the smallest farms. Abebaw and Haile (2013) study the impact of cooperative membership on the likelihood that farmers use fertilizer in Ethiopia and find a stronger positive effect for less-educated farmers and in more remote areas.
In this article, we examine heterogeneous treatment effects or how the effect of cooperative membership varies over farmers. An examination of heterogeneity in treatment effects stems from the program evaluation literature. The evaluation of development projects and public health programs (e.g., Basu et al. 2007; Lechner 2002; Millimet and Tchernis 2013) has moved beyond mean impact studies into studies analyzing the distribution of impacts within the treated subjects. Studying heterogeneous treatment effects is important from the perspective of program targeting. If those participants who are most likely to benefit from a treatment are selected or self-select into treatment, expanding the group of treated subjects can reduce the average effectiveness of the treatment. If participants are those that are not most likely to benefit from the treatment, expanding the participant group or retargeting the program can increase the ATE (Xie, Brand, and Jann 2012; Djebbari and Smith 2008). Estimating heterogeneous treatment effects of cooperative membership can reveal whether there is positive or negative selection (Brand and Xie 2010; Xie, Brand, and Jann 2012), which is important to understand how cooperatives can be more effective in reducing poverty.
Background
In Rwanda, the agricultural sector is a key engine for economic development and poverty reduction, contributing 34% to GDP and approximately 90% to employment (Government of Rwanda 2011; World Bank 2014). Rwandan agricultural policies and strategies focus on intensification and increased market orientation of the smallholder agricultural sector, and cooperatives are seen as an important vehicle to achieve this (Government of Rwanda 2011). The number of agricultural cooperatives in the country has expanded very rapidly during the past years, from 645 in 2008 to 2,400 in 2013, and cooperatives are most prevalent in the horticulture, coffee, and maize subsectors (Verhofstadt and Maertens 2014).
There is some variation in how cooperatives are organized and how they function. Roughly three types of cooperatives can be identified: (i) land cooperatives, (ii) land and marketing cooperatives, and (iii) production cooperatives. Land cooperatives are service cooperatives that focus on the joint acquisition of agricultural land. Cooperative members collectively purchase or rent land, either from private landowners or from the state. Members are allocated a specific part of the cooperative land and cultivate the allocated land individually. Apart from the cooperative land, farmers usually also cultivate their own plots, which they own or rent individually. Land and marketing cooperatives are similar, but they additionally focus on joint marketing of agricultural produce. Produce from cooperative plots, and sometimes also produce from farmers' individual plots, is marketed through the cooperative, and postharvest handling and storage is organized collectively. Some cooperatives do not allow side-selling; others do. Sales revenues are mostly distributed to individual cooperative members according to the volumes they supplied. Production cooperatives organize land acquisition, cultivation of crops, and marketing of produce jointly. The cooperative land is cultivated collectively through communal labor, and all produce from this land is sold through the cooperative. Members are either paid a collective share of the sales revenue or the revenues are kept within the cooperative as savings. All types of agricultural cooperatives offer some additional services to their members, especially the provision of agricultural inputs such as improved seeds and fertilizers, and sometimes also the provision of agricultural equipment such as hoes and shovels. The majority of cooperatives also give some form of credit to their members, either in cash or in kind, and organize agricultural trainings.
In this study we focus on land and marketing cooperatives in the Muhanga district in the southern province of Rwanda, where most cooperatives are active in the maize and horticulture sectors. In those cooperatives, land acquisition, input provision, and marketing are done collectively whereas production is done individually. In line with the existing empirical evidence, we expect that membership in these cooperatives improves farmers' income—for example through reduced transaction costs, increased access to inputs, and adoption of better technologies. However, in line with qualitative observations from Rwanda (e.g., Ansoms 2009; Ansoms 2010), we expect poorer farmers to be excluded and gain less.
Data Collection
We use data from an original farm-household survey in the Muhanga district that was implemented in the period February–March 2012. A three-stage stratified random sampling technique was used and resulted in the selection of 401 farm-households. In the first stage and based on government reports and personal communication with local government officials and the local cooperative support organizations, we identified twenty-six active cooperatives in the district and selected sixteen of them. In the second stage, we identified the villages where the cooperatives are active and made a random selection of forty villages (imidugudu) out of sixty-one. In the third stage, we stratified households in these villages according to cooperative membership and randomly selected 263 cooperative member households, belonging to sixteen different cooperatives, and 138 control households.
We used a quantitative structured questionnaire, with specific modules on demographic characteristics, land and nonland asset holdings, agricultural production, off-farm employment, nonlabor income, cooperative membership, savings, and credit. The data allow us to calculate farm income and total household income and derive poverty figures. The household survey data were complemented with data from semi-structured interviews with the cooperatives in the sample on cooperative activities, investments, marketing strategies, and organizational set-up.
For the analysis in this article, we needed to drop twelve observations from the original sample because of missing information, and hence we use a sample of 389 farm-households. In addition, we only consider land and marketing cooperatives, including seven cooperatives and 154 cooperative members. Production cooperatives are present in the region and were included in our original sample, but we do not consider these in our analysis because previous research has indicated that these cooperatives do not function well and fail to create gains for farmers (Verhofstadt and Maertens 2014).
Econometric Approach
PS Matching
We use PSM to analyze the inclusiveness and effectiveness of cooperatives. PSM is a technique that is often used in nonexperimental causal studies (Khandker, Koolwal, and Samad 2010). The method involves constructing a statistical comparison group based on a probability model of membership in a cooperative. Cooperative members are then matched on the basis of this probability, or PS, to nonmembers. The impact of cooperative membership is then calculated as the mean difference in income and poverty across these two groups. This method allows us to analyze the likelihood of cooperative membership, the impact of cooperative membership on farm income and poverty, and the heterogeneity in impact across farmers in a comprehensive way.
(1)| Variable | Description |
|---|---|
| Human capital | |
| Female single headed HH | Dummy for single, female-headed households |
| HH head age (y) | Age of the household head in years |
| Square of HH head age | |
| HH head education (y) | Years of education of the household head |
| HH agricultural workers (No.) | Number of agricultural workers in the household |
| HH children (No.) | Number of children (aged < 18 years) in the household |
| Physical capital | |
| Land owned (ha) | The total area owned by the household, expressed in hectares |
| Square of land owned | |
| TLU | The number of tropical livestock units (TLU) owned by the household |
| Social capital | |
| Siblings close by (No.) | The number of brothers and sisters of the household head and his/her partner living close by |
| Market access | |
| Distance to the market (min) | The mean distance to the market, expressed in minutes of walking distance, of the plots under cultivation |
(2)PSM methods are sensitive to the exact specification and matching method (Imbens 2004; Caliendo and Kopeinig 2008). Therefore, as a robustness check, we use four different matching techniques that are commonly used in PSM: nearest neighbor matching with one neighbor, nearest neighbor matching with three neighbors, kernel matching, and local linear matching. With single-nearest neighbor matching, every treated household is matched to the control household with the closest PS. With three-nearest neighbor matching, every treated household is matched to three households that are closest in PS, and Y(0) is calculated as the average of the three matched controls. Matching is done with replacement to ensure that each treatment unit is matched to the control unit with the closest PS, which reduces bias (Dehejia and Wahba 2002). Kernel matching1 uses a weighted average of all individuals in the control group to construct Y(0), with weights inversely proportional to the PS distance between treated and control units. This method uses more information to construct the counterfactual outcome Y(0), which will result in reduced variance but increased bias in case of poorer matching (Caliendo and Kopeinig 2008). Local linear matching is comparable with kernel matching but uses an additional linear term in the weighing function, which helps to avoid bias when the PS of control observations are distributed asymmetrically around the treated observations.
(3)Assumptions
The reliability of PSM estimators depends on two crucial assumptions. First, the common support or overlap condition requires balancing in the covariable distribution between treated and untreated observations to ensure that treatment observations have comparable control observations nearby in the PS distribution (Caliendo and Kopeinig 2008). The PS distribution shows sufficient overlap between treated and controls (see figure A.1 in the appendix). As proposed by Becker and Ichino (2002), we only use observations in the common support region, where the PS of the control units is not smaller than the minimum PS of the treated units and the PS of the treated units is not larger than the maximum PS of the control units. We investigate the balancing properties of the covariables for the case of kernel matching and find that matching eliminates all imbalances between treated units and controls (see table A.1 in the appendix).
Second, the conditional independence assumption states that given a set of observable covariables, potential outcomes are independent of treatment assignment (Imbens 2004). This implies that selection into treatment is based entirely on observable covariables, which is a strong assumption. We perform a simulation-based sensitivity analysis (Ichino, Mealli, and Nannicini 2008; Ito, Bao, and Su 2012) and find that our results are robust to failure of the conditional independence assumption (see table A.2 in the appendix).
Results and Discussion
Comparison of Cooperative Members and Nonmembers
In table 2 we compare cooperative members with nonmember households in terms of observable characteristics. Cooperative member households have a relatively older household head and more agricultural labor force in the household. In general, 22% of households in the sample are single female–headed, which is in line with the 27.7% single female–headed households in the Muhanga district reported by the recent national Enquête Intégrale sur les Conditions de Vie des ménages (EICV) results (NISR 2012). There are significantly less female-headed households among the cooperative member households. There are no significant differences in the education of the household head, the number of children in the household, the distance to the market, and the number of siblings close by (as a measure of social capital). We find that land and livestock holdings in general are quite small, with on average 0.27 hectares of land and 1.1 tropical livestock units per household. Cooperative members have significantly more livestock than nonmembers, 1.8 units on average compared with 0.8 units. Cooperative members have on average 0.34 hectares of land compared with 0.25 hectares for nonmember households, but this difference is statistically not significant. In addition to the land they own, cooperative members cultivate on average an additional 0.1 hectares of land they access through the cooperative.
| Total sample | Nonmember households | Cooperative member households | |
|---|---|---|---|
| (n = 389) | (n = 235) | (n = 154) | |
| Demographic characteristics | |||
| Female single headed HH (dummy) | 22% | 25% | 11%* |
| HH head age (y) | 45.6 | 44.6 | 49.0* |
| (13.3) | (13.6) | (11.7) | |
| HH education (y) | 4.9 | 4.7 | 5.4 |
| (2.9) | (2.7) | (3.3) | |
| HH size agricultural workers (No.) | 1.9 | 1.8 | 2.4*** |
| (0.98) | (0.88) | (1.2) | |
| HH size children (No.) | 2.5 | 2.6 | 2.4 |
| (1.7) | (1.7) | (1.8) | |
| Asset ownership | |||
| Land individually owned (ha) | 0.27 | 0.25 | 0.34 |
| (0.50) | (0.48) | (0.54) | |
| Livestock (TLU) | 1.1 | 0.8 | 1.8*** |
| (1.1) | (0.9) | (1.5) | |
| Market access | |||
| Distance to the market (min) | 47 | 46 | 49 |
| (33) | (32) | (37) | |
| Social capital | |||
| Siblings living close by (No.) | 2.2 | 2.1 | 2.5 |
| (2.5) | (2.5) | (2.5) | |
| Income and poverty | |||
| Farm income (RWF) | 229,529 | 176,682 | 400,422** |
| (307,653) | (194,161) | (491,565) | |
| Poverty incidence | 49% | 54% | 34%** |
- Source: Calculations based on household survey data collected in 2012.
- Notes: Mean values are shown; for continuous variables, standard deviations are shown in parentheses. Cooperative member households are compared with nonmember households using t test. Annual farm income is calculated as the value of crop and livestock production (including nonmarketed produce valued at market prices) minus variable production costs (including purchased inputs, hired labor, land rent, etc.). Revenue transfers from the cooperatives are also added to the farm income, whereas cooperative contribution cost are subtracted. The poverty line is set at 83,000 RWF per adult equivalent per year, which is the Rwandan national poverty line for extreme poverty derived from the 2011 EICV3 survey. HH, household; TLU, tropical livestock unit (NISR 2012).
- *p ≤ 0.1; **p ≤ 0.05; ***p ≤ 0.01.
Cooperative members and nonmember households differ substantially with respect to income and poverty levels. Annual farm income is significantly larger for cooperative members than for nonmembers: 400,422 RWF (equivalent to about 580 USD) for members compared with 176,682 RWF (equivalent to about 260 USD) for nonmembers. Poverty is widespread in our research area, and almost half of the households in the sample are poor. Whereas 34% of cooperative member households are poor, the incidence of poverty is significantly higher (54%) among nonmember households.
Probability of Cooperative Membership
The results of the probit model estimating the propensity of cooperative membership are given in table 3. Marginal effects are reported. We find that households with a more-educated head and households with more agricultural labor force have a higher probability of being members of a cooperative. The estimated marginal effects indicate that an additional year of education of the household head increases the likelihood of cooperative membership by 2.2 percentage points and an additional agricultural laborer in the household increases the likelihood of cooperative membership by 7.2 percentage points. Further, we find that owning more land decreases the likelihood of being a cooperative member, whereas owning more livestock increases it. An additional hectare of land decreases the likelihood of cooperative membership by 19 percentage points, and an additional unit of livestock increases it by 9.1 percentage points. Distance to the market has a significant negative effect on the probability of cooperative membership, with every hour farther away from the market decreasing the likelihood of cooperative membership by 12 percentage points. Other variables such as the sex and the age of the household head, the number of children in the household, and the social capital indicator do not have a significant impact on the likelihood of cooperative membership.
| Variables | Marginal effects | Standard errors |
|---|---|---|
| Female single headed HH | −0.058 | 0.062 |
| HH head age (y) | 0.007 | 0.013 |
| Square of HH head age | 0.000 | 0.000 |
| HH head education (y) | 0.022*** | 0.007 |
| HH agricultural workers (No.) | 0.072*** | 0.024 |
| HH children (No.) | −0.024 | 0.015 |
| Land owned (ha) | −0.190** | 0.090 |
| Square of land owned | 0.033 | 0.022 |
| TLU | 0.091*** | 0.020 |
| Distance to the market (min) | −0.002** | 0.001 |
| siblings close by (No.) | 0.001 | 0.011 |
| Pseudo-R2 | 0.118 | |
| LR χ2 (11) | 61.8 | |
| Prob > χ2 | 0.000 | |
| Observations | 389 |
- Source: Estimations based on household survey data collected in 2012.
- Notes: Estimated marginal effects are reported. HH, household; TLU, tropical livestock unit.
- *p ≤ 0.1; **p ≤ 0.05; ***p ≤ 0.01.
These results indicate that cooperative membership is biased toward less-remote farm-households and farm-households with more human capital and that, hence, the agricultural cooperatives under study are to some extent exclusive. Remoteness seems to be a constraint for cooperative membership. Farm-households farther away from the market are less likely to be organized in cooperatives, although those households face higher transactions costs, which likely makes cooperative marketing and input acquisition more beneficial. There also seem to be human capital constraints to participate in cooperatives because farm-households with a more-educated household head and more agricultural workers have a higher likelihood of being a cooperative member. Also physical capital matters for cooperative membership, but results are more ambiguous. Livestock ownership positively influences the likelihood of cooperative membership, whereas land ownership has a negative effect. The latter effect might imply that access to cooperative land is a main driving force for land-poor households to participate in cooperatives.
Income and Poverty Effects of Cooperative Membership
The estimated ATEs are presented in table 4. It is important to note that the estimated ATEs are consistent over different matching techniques, which is an indication of the robustness of the PSM estimates. The estimated effect of cooperative membership on farm income is significantly positive and similar in magnitude across the matching methods. We find that cooperative membership increases annual farm income by 40% to 46%. Given that the average annual farm income in the area is 229,529 RWF (table 2), this result comes down to an average income effect of cooperative membership of approximately 105,500 RWF (or approximately 155 USD). The estimated effect of cooperative membership on the likelihood of being poor is significantly negative and similar in magnitude across the matching methods. We find that cooperative membership reduces the likelihood of being poor by 10 to 14 percentage points. Given that the average incidence of poverty in the research area is 49% (table 2), this result comes down to an average poverty reducing effect of cooperative membership of 6.8%. These results indicate that agricultural land and marketing cooperatives in the area are effective at improving farmers' welfare.
| Dependent variables | Single-nearest neighbor matching | Three-nearest neighbor matching | Kernel matching | Local linear matching |
|---|---|---|---|---|
| Log (farm income) | 0.43*** | 0.40*** | 0.40*** | 0.46*** |
| (0.15) | (0.16) | (0.14) | (0.14) | |
| Poverty | −0.10** | −0.14** | −0.12** | −0.13** |
| (0.050) | (0.069) | (0.056) | (0.055) |
- Source: Estimations based on household survey data collected in 2012.
- Note: Standard errors are shown in parentheses.
- *p ≤ 0.1; **p ≤ 0.05; ***p ≤ 0.01.
Our findings about agricultural cooperatives in Rwanda are similar to findings in the literature on contract farming, a widely studied institutional innovation. Although there is a growing amount of recent evidence that contract farming has a positive effect on farm performance and farmers' welfare (e.g., Dedehouanou, Swinnen, and Maertens 2013; Maertens and Swinnen 2009; Minten, Randrianarison, and Swinnen 2009; Rao and Qaim 2011), contract farming is often found to be exclusive and biased toward better-off or middle-class farmers (e.g., Maertens and Swinnen 2009; Neven et al. 2009; Rao, Brümmer, and Qaim 2012).
Heterogeneous Treatment Effects
The results of the graphical and statistical analyses of heterogeneous treatment effects are given in figures 1 and 2. We consecutively discuss the results for heterogeneity over the PS, land ownership, market access, and demographic characteristics.
Heterogeneity over the PS
Figure 1 shows how the estimated income and poverty effects vary over the estimated PS. The results indicate that the ATT on farm income varies significantly with the PS and that the slope is negative. This means that the effect of cooperative membership on farm income is largest for households with the lowest propensity to be a cooperative member and decreases with the propensity of cooperative membership. The slope coefficient is estimated to be 1.7, which means that with every 10 percentage point increase in the likelihood of cooperative membership, the impact of this membership on income reduces by 17%. The income effect of cooperative membership even becomes zero in the upper end of the PS distribution. The ATT on the likelihood to be poor increases with the estimated PS, but the slope coefficient is not significant.
Heterogeneity of treatment effects over the propensity score, land ownership, and market access
Note: Linear and quadratic prediction plots with 95% confidence intervals.
Heterogeneity of treatment effects over demographic characteristics
Note: Linear and quadratic prediction plots with 95% confidence intervals.
These results point to a problem of negative selection. Cooperatives are most effective in increasing the incomes of farmers that are least likely to join cooperatives and are least effective—or even not effective at all—in increasing the incomes of farmers who are most likely to join. In other words, farmers who would gain most from cooperative membership are the least likely to join, probably because they face entry constraints in terms of human and physical capital. The subsequent analyses on impact heterogeneity over different farm and farmer characteristics can shed further light on this finding of negative selection.
Heterogeneity over Land Ownership
Figure 1 shows how the estimated income and poverty effects vary with households' landownership.3 The results indicate that there is an inverse U-shaped relation between land and the estimated ATT on farm income. The impact of cooperative membership on farm income is increasing with landownership up to landholdings of about 0.5 hectare and decreasing thereafter. Because approximately 80% of the sampled households own less than 0.5 hectares of land, the income effect of cooperative membership is largely increasing with households' landholdings but at a decreasing rate. The estimated income effect becomes zero at the lower end of the land distribution, from landholdings of about 0.15 hectares onwards. Further, the results indicate that the estimated ATT on the likelihood to be poor significantly decreases with landholdings. Hence, the poverty-reducing effect of cooperative membership increases with land. Again, we find that the ATT becomes zero for very small landholdings.
These findings imply that the cooperatives under study are most effective in increasing farm income and reducing poverty among smallholders with relatively larger landholdings. For land-poor or near-landless farmers, cooperative membership is not effective for improving welfare. These results contradict findings by Ito, Bao, and Su (2012) and Fischer and Qaim (2012a) who demonstrate that cooperative membership in China and Kenya has a larger impact for smaller farms. Given our finding that land ownership has a negative impact on the propensity to be a cooperative member, the results in figure 1 contribute to explaining the observed negative selection. Although households with the smallest landholdings have the highest propensity to join a cooperative, the impact of cooperative membership on their incomes is very small and not enough to lead to poverty reduction. Other authors also pointed to negative selection related to farm size but in the opposite direction. Ito, Bao, and Su (2012) and Fischer and Qaim (2012a) find that the income gains of cooperative membership are more pronounced for small farmers who have a lower propensity to join cooperatives.
A possible explanation for our findings lies in the fact that landholdings are very small in our study area; on average households own only 0.27 hectares. Cooperative membership, even if it improves the access to land through cooperative land acquisition (on average 0.1 hectares), does not have an impact on farm income for very small farms. These farms use few inputs and commercialize small amounts of produce, such that reductions in transactions costs due to cooperative marketing and input purchase are only marginal.
Heterogeneity over Market Access
Figure 1 shows how the estimated income and poverty effects vary with distance to the market. We find an inverse U-shaped relation between market access and the estimated ATT on farm income. The impact of cooperative membership on farm income is increasing with distance to the market up to a distance of approximately two hours (116 minutes is the turning point) and decreasing thereafter. In addition, the estimated ATT on the likelihood to be poor significantly decreases with distance to the market, meaning that the poverty-reducing effect of cooperative membership increases with distance to the market. Both the income effect and the poverty effect become very small and close to zero for households located very close to the market.
Our findings imply that cooperatives are most effective in more remote areas. Because transaction costs are larger in remote areas, cooperative marketing and input acquisition can create larger gains through reductions in transaction costs. Our findings are in line with those of Abebaw and Haile (2013) who point to an inverse U-shaped relation between distance to the nearest road and the impact of cooperative membership on technology adoption in Ethiopia. Given our finding that distance to the market significantly reduces households' propensity to be a cooperative member, the results in figure 1 contribute to explaining negative selection. Farmers in more remote areas are less likely to join cooperatives although the potential benefits are largest for them.
Heterogeneity over Demographic Characteristics
Figure 2 shows how the estimated income and poverty effects vary with some demographic characteristics of farm-households. The estimated ATT on farm income does not vary with the considered demographic characteristics—age and education of the household head and the available labor force—because none of the effects are significant. In addition, we tested whether the estimated ATTs differ according to the sex of the household head4 and found no significant differences. These findings contradict the findings of Abebaw and Haile (2013) that the impact of cooperative membership decreases with farmers' education but are in line with the findings of Bernard, Taffesse, and Gabre-Madhin (2008). The results imply that cooperatives are as effective in creating welfare gains for less-educated farmers as for more-educated farmers, as effective for younger and less-experienced farmers as for older and more-experienced farmers, and as effective for male- as for female-headed farm-households.
The estimated ATT on the likelihood to be poor significantly increases with the available agricultural labor in the household. Hence, the poverty-reducing effect of cooperative membership is largest for households with few agricultural laborers, likely because these are the poorest households and not because the income gains are largest for these households.
Conclusions and Policy Implications
In this article we analyze the inclusiveness and effectiveness of agricultural cooperatives in Rwanda. We find that cooperatives are to some extent exclusive but that they are effective in improving rural incomes and reducing rural poverty. We link the issues of inclusiveness and effectiveness by analyzing the heterogeneity in impact of cooperative membership. We find evidence of a negative selection because the estimated income effect of cooperative membership is largest for farmers with the lowest propensity to be a cooperative member. This negative selection is partially related to the location of farm-households with respect to markets and to land ownership.
Our findings in general support the emphasis of Rwanda on agricultural cooperatives as an institutional vehicle to boost the smallholder farm sector. Based on our findings on the heterogeneity in impact across farm-households we can formulate some very specific implications toward policies on agricultural cooperatives in Rwanda. First, we find that farmers in more-remote areas have a lower propensity to be a cooperative member but that the impact of cooperative membership on the income and poverty of more-remote farmers is larger. This implies that there is scope for expanding membership of agricultural cooperatives and at the same time improving the effectiveness of cooperatives to raise rural incomes and reduce poverty. This calls for an expansion of input and marketing cooperatives in more remote areas of Rwanda.
Second, we find that households that own more land are less likely to be cooperative members (likely because access to land is a driving force to engage in land cooperatives) but that the impact of this membership on income and poverty is larger for these households. We find that cooperative membership is not effective in reducing poverty among land-poor households because the impact on their incomes is too low. Contrary to previous arguments in the literature that cooperatives need to be more inclusive toward the poorest households, our findings indicate that agricultural cooperatives are not a solution for near-landless or land-poor farm-households. Even if cooperative membership marginally increases access to land, the income effect of cooperative membership for these households is too low to get them out of poverty.
Third, we find that cooperative membership is as effective at improving farm income for more-educated farmers as for less-educated farmers, for female-headed households as for male-headed household, and for households with many workers as for households with few workers. This calls for making cooperatives more inclusive toward less-educated, less-experienced, and female farmers and for removing human capital constraints for entry into cooperatives because this would not decrease the effectiveness of cooperatives to improve rural incomes.
Finally, we need to emphasize that our results are specific for the study area in Rwanda. Nevertheless, our results challenge some prevailing judgments about how inclusive agricultural cooperatives are and should be and about cooperative formation and the impact on rural development and poverty reduction in Rwanda. Our results also highlight the importance of looking beyond average (treatment) effects in studies on the impact of institutional innovation in the agricultural sector and studying impact heterogeneity.
Funding
The authors received funding from VLIR-UOS (Vladoc scholarship program) and the KU Leuven Special Research Fund (OT program).
Acknowledgments
We thank Jean Chrysostome Ngabitsinze from the National University of Rwanda in Butare for facilitating fieldwork. We also thank conference participants in Ghent and Leuven for comments on earlier versions of the article.
Appendix
Robustness Checks for Kernel PS Matching
| Covariables | Sample | Mean treated units | Mean control units | % Bias between treated and controls |
% Reduction in bias | t test Mean (treated) = Mean (controls) |
|---|---|---|---|---|---|---|
| Female single headed HH | Unmatched | 0.195 | 0.244 | −11.8 | −1.12 | |
| Matched | 0.199 | 0.182 | 4.1 | 65.1 | 0.37 | |
| HH head age (y) | Unmatched | 51.4 | 46.1 | 43.3 | 4.13*** | |
| Matched | 51.2 | 50.9 | 2.3 | 94.6 | 0.20 | |
| Square of HH head age | Unmatched | 2,788 | 2,271 | 41.3 | 3.98*** | |
| Matched | 2,775 | 2,743 | 2.6 | 93.7 | 0.21 | |
| HH head education (y) | Unmatched | 5.14 | 4.27 | 25.4 | 2.43** | |
| Matched | 5.06 | 5.08 | −0.6 | 97.6 | −0.05 | |
| HH agricultural workers (No.) | Unmatched | 2.50 | 2.11 | 35.9 | 3.47*** | |
| Matched | 2.47 | 2.53 | −5.4 | 84.9 | −0.44 | |
| HH children (No.) | Unmatched | 2.33 | 2.49 | −9.5 | −0.90 | |
| Matched | 2.34 | 2.42 | −4.7 | 50.4 | −0.41 | |
| Land owned (ha) | Unmatched | 37.4 | 36.2 | 2.1 | 0.20 | |
| Matched | 37.9 | 43.8 | −9.6 | −364.5 | −0.76 | |
| Square of land owned | Unmatched | 4,995 | 5,216 | −0.9 | −0.09 | |
| Matched | 5,081 | 6,909 | −7.8 | −727.4 | −0.63 | |
| TLU | Unmatched | 1.60 | 1.09 | 43.8 | 4.25*** | |
| Matched | 1.56 | 1.57 | −0.8 | 98.1 | −0.07 | |
| Distance to the market (min) | Unmatched | 43.9 | 51.5 | −24.5 | −2.32** | |
| Matched | 44.7 | 45.1 | −1.2 | 95.1 | −0.11 | |
| Siblings close by (No.) | Unmatched | 2.13 | 2.22 | −3.7 | −0.35 | |
| Matched | 2.14 | 2.36 | −9.5 | −160.9 | −0.80 |
- Source: Calculations based on data from own household survey 2012.
- Note: HH, household; TLU, tropical livestock unit.
- *p ≤ 0.1; **p ≤ 0.05; ***p ≤ 0.01.
| Neutral confounder | Confounder calibrated to mimic dummy for female-headed households |
|||||
|---|---|---|---|---|---|---|
| Dependent variables | Estimator effecta |
Outcome effectb |
Selection effectc |
Estimator effecta |
Outcome effectb |
Selection effectc |
| (Log) farm income | 1.6% | 1.02 | 1.06 | 1.6% | 0.545 | 0.78 |
| (Log) total household income | −1.6% | 1.00 | 1.00 | 2.1% | 0.58 | 0.73 |
| (Log) total household income/ADEQ | 0.8% | 1.06 | 1.01 | 0.4% | 1.09 | 0.82 |
| poverty | 1.4% | 1.03 | 1.05 | 1.6% | 1.50 | 0.75 |
| Extreme poverty | 0.0% | 1.02 | 1.05 | −1.0% | 0.86 | 0.74 |
- Source: Calculations based on data from own household survey 2012.
- a The estimator effect indicates to what extent the baseline estimation result would change if we could observe an additional binary confounder.
- b The outcome effect measures the estimated effect of the simulated binary confounder on the dependent variables.
- c The selection effect measures the estimated effect of the simulated binary confounder on the selection into treatment; this is the propensity of being a treated household.
Propensity score distribution
Source: Calculations based on data from own household survey 2012.
