I M P A C T A S S E S S M E N T D I S C U S S I O N P A P E R 5 Supported by the CGIAR The impact of participation in Diversity Field Fora on farmer management of millet and sorghum varieties in Mali Melinda Smale,1 Lamissa Diakité,2 Amadou Sidibé,2 Mikkel Grum,3 Hannah Jones,4 Issa Seni Traore5 and Hamidou Guindo6 1 Oxfam America, Washington, DC (during the implementation of this research, Senior Re- search Fellow, International Food Policy Research Institute (IFPRI), Washington, DC) 2 Institut d’Économie Rurale, Bamako, Mali 3 Bioversity International, Maccarese, Rome, Italy 4 International Food Policy Research Institute (IFPRI), Washington, DC 5 Aide au Sahel et à l’Enfance Malienne (ASEM), Hamdallaye, San, Mali 6 Institut d’Économie Rurale, Bamako, Mali Acknowledgements Bioversity International is grateful to the African Association of Agricultural Economists, pub- lisher of the African Journal of Agricultural and Resource Economics, for granting permission to re- produce this paper. This paper was first published in the African Journal of Agricultural and Resource Economics 4(1): 23–47. Abstract Malian farmers have been cultivating millet and sorghum for millennia, but they are slow to adopt and develop modern varieties because it is difficult to observe the difference in yields in their fields, given the challenging local growing conditions. Farmer participatory approaches are therefore recommended. This paper applies an instrumental variables method to survey data from Mali to evaluate the impacts of Diversity Field Fora, a type of farmer field school which aims to boost millet and sorghum yields by showing farmers how to manage diverse varieties. Im- pact indicators are expected and recalled millet and sorghum yields, the total number of unique attributes of millet and sorghum varieties stocked as seed, and the relative deprivation of the household farm with respect to these indicators. The findings suggest the project has had results at one of two sites where it has been implemented with the same local leadership and more inten- sively over a longer time frame. Keywords: farmer field school; sorghum; millet; landraces; participatory crop improvement; Mali IMPACT ASSESSMENT DISCUSSION PAPER  1. Introduction Malian farmers have accumulated knowledge of millet and sorghum management that spans millennia. Pearl millet and sorghum are known to have been domesticated in mul- tiple locations scattered across the Sahel – then savanna and now the border of the Sahara (Harlan, 1992). Archaeologi- cal evidence suggests that economies based on cattle, goats, sorghum and pearl millet were established in this region be- tween 5,000 and 3,000 years ago (Smith, 1998). Sorghum and millet are still the major crops of Mali, grown by subsistence oriented producers in an agricultural sector that is almost entirely rainfed. National average yields for both crops are less than one ton per hectare (Touré et al., 2006). For the Malian agricultural sector in general, the most binding constraint is the infertility and unsuitable struc- ture of the soil, which in turn impedes moisture retention in zones with limited rainfall. The first devastating drought in a series occurred from 1968 to 1973. The 1982–1993 period was persistently dry and marked by another severe drought from 1982 to 1984. Although there was more rainfall during the 1994–2003 decade, conditions remained far drier than they had been from 1930 to 1965 (Anyamba and Tucker, 2005). Low yields are often attributed in part to low rates of adoption of improved seed. The most recent draft Agricul- tural Census reports that the proportion of the area under cereals with improved seed does not exceed 10%. By con- trast, 89% of the area is under industrial crops with im- proved seed (BCRA, 2006). Improved varieties of sorghum have been more widely adopted than improved varieties of millet (BCRA, 2006). Low adoption rates have in turn been blamed on poor performance of the formal seed system. Despite an ongoing process of seed sector reform, liberalization of seed markets for sorghum and millet has not advanced as rapidly as liber- alization of grain markets (Vitale and Bessler, 2006; Diakité et al., 2008). The formal seed sectors for sorghum and millet continue to be largely state run, with some participation by registered farmer cooperatives in multiplying seed. So far, commercialization of farmer-produced seed on more than a pilot scale has posed an insuperable challenge. Surpassing the performance of farmers’ own millet and sorghum landraces is not easy. International and national research centers accelerated breeding efforts from 1973, but new cultivars in the dry savannas made little impact on yields (Sanders et al., 1996). Of the improved varieties that performed well on research stations during that period, few performed better than landraces on farms, for several rea- sons. To start with, the imported breeding material was un- suitable. Initially, an emphasis was placed on material that was successful in India but was not adapted to the high soil temperatures in the Sahel (Matlon, 1987). Local sorghum and millet varieties also have photoperiodicity, which ena- bles plants to adjust the length of the growth cycle to syn- chronize with the length of the rainy season. Unfortunate- ly, early selection programs, combined with the effects of drought, led to the gradual elimination of photoperiodism in favor of a range of varieties with short, fixed cycle lengths (Vaksmann et al., 1996). These shortcomings have since been overcome by inter- national and national breeding programs (Weltzien et al., 2006), but challenges remain. Attaining more than margin- al changes in yield is difficult without hybrids, but while promising materials are in the pipeline, none have yet been released for either sorghum or pearl millet. The tremendous variation in climate, soils and production systems means that the degree of plant stress is not only high but also ex- tremely variable within and among fields in close proximity. It takes time for farmers to recognize whether or not a new variety has advantages, which is one argument for farmer participation in testing and evaluation. In addition, decreas- ing public funds have meant that no breeding is conducted for some agro-ecologies, including that of Douentza, where part of this survey was conducted. Thus there is no consensus about whether it is lack of ef- fective demand or supply that constrains farmer use of certi- fied sorghum and millet seed. Because the private sector has not taken responsibility for seed distribution, and the pub- lic sector has failed to supply improved seed in reasonable quantities, some researchers have also called for strength- ening the informal seed system (De Vries and Toenniessen, 2001), but knowledge about how this may be done is only beginning to emerge. Diakité et al. (2005), Bazile (2006) and Weltzien et al. (2006) recommend greater involvement of farmer and community organizations in testing and evalu- ating improved varieties, coupled with decentralized seed production to reduce the time lag between variety develop- ment and adoption, and to reach more remote areas. The purpose of this paper is to evaluate the impact of one such participatory research effort. Diversity Field Fora, which bear some similarities to farmer field schools, have been launched in Mali on a pilot scale. Given the small scale and brief history of the Diversity Field Fora in Mali, and the statistical approach we use to address possible biases, this analysis focuses on measurable, short-term impacts. The findings shed light on the potential of the approach, but should not be understood as a comprehensive evaluation. The next section summarizes key features of the project and Diversity Field Fora, Section 3 describes the methodology used to evaluate impact, Section 4 interprets the findings, and Section 5 concludes. . Project description Background The project entitled ‘Empowering Sahelian farmers to lever- age their crop diversity assets for enhanced livelihood strate- gies’1 was funded by the International Fund for Agricultural Development (IFAD) from 2005, coordinated by Bioversity International (previously, the International Plant Genetic Re- sources Institute, IPGRI) and implemented by a combination  « Projet de renforcement des capacités des agriculteurs sahéliens pour une meilleure gestion des ressources phylogénétiques en vue d’améliorer leurs conditions de vie » IMPACT ASSESSMENT DISCUSSION PAPER  of local, national and international organizations. This second phase built on an earlier IFAD-financed project, conducted from 1999 to 2002 in Mali and Zimbabwe, which focused on the development of participatory strategies for on-farm (in situ) conservation of millet, sorghum and other crops. In the interim, IFAD also funded some farmer field school activities in the village of Boumboro, which became the locus of the San/Tominian project site during the second phase. The goal of the second phase was to support the livelihood strategies of poor farmers in Mali by strengthening their ca- pacity to manage their plant genetic resources. The findings of the earlier project demonstrated the importance of plant genetic resources in the livelihoods of Sahelian farmers. These farmers must be able to meet their staple food needs directly from their harvests in an exceedingly challenging growing environment – or migrate. Well-adapted, diverse local landraces supply them with a range of consumption attributes that are important for preparing local dishes and other end uses, and enable them to match their varieties to heterogeneous moisture and soil conditions and to smooth labor needs over the growing season. Genetic diversity also helps combat the risk of crop losses from biotic and abiotic stresses – the most pressing of which is the variability of rainfall and dry periods at critical points of plant growth. Diversity Field Fora In the first phase of the project, several activities were identi- fied to support the management of crop genetic resources in farming communities. The most sustained and comprehen- sive of these was the notion of Diversity Field Fora (Champs de Diversité, or DFF), which built on the concept of farmer field schools. Farmer field schools are an adult education method de- veloped and widely promoted in Asia to teach integrated pest management practices to groups of farmers. While there is considerable variation in form and content, the basic approach involves teaching farmers how to solve problems, set priorities and conduct experimental research through fa- cilitated, hands-on sessions in fields allocated by the farm- ing community for study. Reviews of the evidence for the impacts of farmer field schools suggest that they have not translated into changes beyond local communities, that they tend to favor the more privileged farmers in those commu- nities (Davis, 2006), and that they provide an unlikely basis for sustained, group activity (Tripp et al., 2005). Tripp et al. (2005:1718) express concern that the assessment of farmer field schools has been ‘insufficient’, and Van der Berg and Jiggins (2007) explain that the methodology for evaluating their impact is still under development, and is characterized by a tension between statistical rigor, which implies a nar- row focus, and comprehensiveness, which leads to a diver- sity of impact indicators and definitions of impact. In this project, experiments in enhancing knowledge of crop genetic resources were designed and conducted by villagers, with technical support from the project staff, on land distributed for that purpose by villagers. Farmers studied both modern varieties and landraces. As defined in project documents, DFF encompass a combination of edu- cational activities, action research and training. They aim to strengthen the capacity of farmers to understand, analyze and manage their own plant genetic resources, by creating a physical space that facilitates the exchange of ideas among farmers, extension agents and researchers and stimulates farmer experimentation. The choice of project sites was guided by two funda- mental criteria of location: they had to be in the drier savan- nas and in an area served by an IFAD investment program. Other criteria for village selection included considerations of feasibility, such as road access to the villages, willingness of farmers and leaders to participate in the project, social co- hesion and availability of plots for experimental purposes. These criteria generate an obvious, but unavoidable, place- ment bias that limits the extent to which findings from this study can be generalized. The two project sites studied here are separated by at least 400 km along a northeast transect on the main road from the capital city of Bamako into the Sahara. Each is located in a unique agro-climatic zone and cluster of ethno-linguistic groups. The village of Boumboro, Commune of Mandiakuy, Cercle2 of Tominian, Region of Segou, is situated in a semi- arid, tropical climate with annual rainfall levels of 450 to 600 mm, which places it in the Sahelo-Sudanian zone. Vari- ation in vegetative cover is linked to variation in soils, and the landscape is a mosaic of cultivated woodland savanna, heavily populated by shea nut trees (karité). Bambara and Bobo are the major ethnic groups at this site. The village of Petaka, Commune of Petaka, Cercle of Douentza, Region of Mopti, is located in the Sahelian agro- climatic zone, which places it within the 200 and 400 mm isohyets. The zone is composed of a series of rocky plateaus and outcroppings, interspersed with sandy plains, forest cover, cultivated areas and pasture. Villages are located on both the rocky plateaus and the plains. The major ethnic groups in this region are Dogon, Peulh and Sonrhaï. Location is also related to the development of infrastruc- ture and services, which generally declines with distance from Bamako. Boumboro is well served by a nexus of feeder roads linked to the city of San on the main tarmac road. This site is closest to Bamako and has a higher density of large weekly markets and other types of physical infrastructure. The closest town to Petaka is Douentza, which is commer- cially linked to a nexus of smaller-scale weekly markets that are dispersed in villages with more restricted physical in- frastructure. . Methods Evaluating impacts The economics literature about evaluating project impacts emphasizes the importance of establishing the appropriate counterfactual. As it is often described, the problem is es-  In Mali, a cercle is an administrative unit in a region, followed by the cat- egories of commune and village. IMPACT ASSESSMENT DISCUSSION PAPER  sentially one of missing data: differences between those who participated in a program and those who did not are observ- able, but it is not possible to observe the status of project participants in the absence of the project. Consequently, dif- ferences due to the influence of the project are easily con- fused with pre-existing differences between participants and non-participants. The term ‘selection bias’ is often applied to the errors in estimation that result from this dilemma. Several types of selection bias occur. When participation is voluntary, fac- tors that influence the likelihood of participation might also affect the outcome of participation. For example, farmers with more income, assets and access to information may be more likely to decide to participate in a project, and attain higher yields whether they participate or not. In other cases, projects or programs deliberately target certain individuals according to their characteristics, such as income. As noted in this case, projects or programs may also be placed in cer- tain zones or sites because of agro-ecology or infrastructural characteristics. To reduce selection bias, economists have proposed a class of statistical approaches that are commonly referred to as treatment models. Ravallion (1994) categorizes meth- ods for evaluating project impact in terms of five basic ap- proaches. Each approach involves an attempt to construct a treatment and a control group in such a way that they differ only with respect to the program, mimicking an experimen- tal situation. The treatment group represents participants. The control group represents the status of participants in the absence of the program, and is composed in such a way that it matches or represents the test group as closely as possible – with the exception that its members did not participate. Each of the five statistical techniques has advantages and disadvantages. With the randomized approach, indi- viduals in each group are chosen at random. This technique eliminates statistical bias but is not always acceptable from a political or social standpoint. After selection, individuals may opt not to participate in ways that are non-random. The matching approach is employed to compose test and control groups by estimating the propensity of individuals to par- ticipate based on their characteristics. Scores are sensitive to the model used to construct them, and while individu- als may be well matched on observable characteristics, they may still differ in those that have not been recorded, or in unobservable characteristics. Bias caused by differences in unobservable characteristics can be addressed through the double-difference approach, which compares treatment and control groups before and after the treatment. This tech- nique eliminates the effects of unobservable characteristics by comparing net changes in variables between two time periods. Reflexive controls compare the same group before and after the program, but ignore changes induced by fac- tors outside the project, which can lead to false attributions. The instrumental variable approach relies on econometric methods to separate the effects of project participation from those of other factors that influence impact. Identifying val- id instrumental variables is the major challenge associated with this method. Valid instrumental variables are those that determine participation but only influence impact through participation. Only the matching and instrumental variables methods are feasible in the context of this study. An experiment could not be designed given that related project activities in Bou- mboro were already in the process of implementation, from 2002. The data on which the analysis is based were collected as a statistical baseline with a relatively small sample size and single-period survey because of cost constraints, which precluded either the reflexive or double difference approach- es. While feasible, the matching approach is not well-suited to analyzing the impact of this project because specific so- cioeconomic groups were not targeted. In addition, the im- pact indicators we use, which are based on the management and use of crop genetic resources, are complex. Handa and Maluccio (2007) conclude that matching is more promising as an approach for evaluating easily measured outcomes, such as those related to child schooling and health, than it is for more complex outcomes, such as expenditures. Sample design The conceptual approach implies a sample design. The test and the control villages must be as similar as possible with respect to agro-ecology and overarching social, economic and institutional conditions. At each site, both test and control villages are located within the scope of an IFAD project. The same non-governmental organization (NGO) that is active in the test village of a site is also active in the control villages. Ethnic representation is also broadly similar. An additional feature of this project is that one means of addressing its goal was to introduce farmers to new culti- vars (both landraces and improved varieties) and to facilitate their own experimental processes as well as their informal exchange of seed and seed-related information. Customar- ily, farmers at the project sites depend on their own harvests or other nearby farmers for their millet and sorghum seed (Sperling et al., 2006; Diakité et al., 2008). Seed transactions most often follow ties of kin and clan, and are shaped by social and cultural norms. An effective DFF would lead to dissemination of materi- als from the locus villages3 to other villages where farmers have social ties. Of the two locus villages, Boumboro has the longer and more extensive history of project activities. Here, staff reported that surrounding villages have been influenced by the project through farmer visits to diversity fields cultivated by participants and farmer-to-farmer ex- changes of information and seed. In their study of the effects of health programs designed to reduce intestinal worms in school children, Miguel and Kremer (2004) refer to such pat- terns as ‘treatment externalities’. To address these externalities, we included in the treat- ment group nearby villages from which farmers had been  The locus village is the village where the DFF were held. Test villages in- clude these and villages indirectly affected by the DFF. IMPACT ASSESSMENT DISCUSSION PAPER  invited to observe DFF conducted by farmers in Boumboro. Farmers sampled from these villages were not considered to be participants, however. At the Douentza site, all farmers in the treatment group are from the village of Petaka because DFF activities were much more recent and had not yet in- volved other villages. In 2006, in the test villages of Boumboro and Petaka, where DFF had been established, all farmers who were ac- tive participants were interviewed. Farmers in other influ- enced villages near Boumboro were sampled. These and the DFF participants constitute the ‘treatment’ group. Non- participants in both Boumboro and influenced villages were also sampled at random; these constitute the ‘control’ group. Other than the census of participants, farmers were selected at random until we had a sample of 150 per site, roughly split between control and treatment groups. The control groups include at least three villages per site that had not been invited to demonstrations or field days. Characteristics of households, farms, seed management, market participation and social capital were collected dur- ing 2006, along with yield information by variety. Data on variety attributes and additional yield data were elicited in 2007, when household demographic information was also reconfirmed. Among those households remaining in the sample, there were only a few with significant changes in composition (births and deaths). Sample attrition, particularly at the San site, and missing responses on some variables, led to an operational sample size for this analysis of 131 farmers (62 treatment; 69 con- trol) at the San site and 149 (56 treatment; 93 control) at the Douentza site, for a total of 280 farmers. Unfortunately, the San sample originally included some farmers from tempo- rary populations who seek permission from longer estab- lished groups to resettle in the area. However, we posit that the remaining sample is representative of the more perma- nent population. In order to simplify the text below, ‘San site’ or ‘San’ is used to refer to the clusters of test and control villages in the Cercles of San/Tominian and the ‘Douentza site’ or ‘Douentza’ refers to the parallel clusters in the Cercle of Douentza. Impact indicators Since the goal of the project is to strengthen the management of crop genetic resources, we defined impact indicators relat- ed to these resources (Table 1). Stronger management of sor- ghum and millet varieties is expected to enhance yields and contribute to more diverse crop genetic stocks held in farm- ing communities. We use as indicators 1) expected yields of millet and sorghum in the presence and absence of drought, 2) two-year average yields based on farmer recall, and 3) the total count of the unique production and consumption attributes supplied by the millet and sorghum varieties held in stock at the time of the survey. In addition, we construct measures of relative deprivation based on these variables. An indicator of inequality, relative deprivation, compares the standing of participants relative to others. The first two outcome variables are yield measurements. Expected yields are calculated from triangular distributions. The triangular distribution is often used as a subjective de- scription of a population for which there is only limited sample data, and especially in cases where the relationship between variables is known but data is scarce. Elicited in terms of only three parameters (the minimum, maximum and mode), the triangular distribution is the simplest ap- proximation to a normal distribution and has been widely applied in analysis of farmer decision making under risk (Hardaker et al., 2004). Yields are highly variable from year- to-year in this growing environment, and expected yields have the advantage that they include the history of the farm- ers’ experience with variety. Although these are subjective rather than objective measurements of yields (typically un- dertaken by weighing the harvest of the crop cut from sub- plots placed in the field), social scientists often argue that it is the perception of the farmer that drives his or her deci- sions and is thus the relevant point of reference. This project is focused heavily on farmers’ perceptions and knowledge systems. Furthermore, low heritability in this production environment also means that a yield measurement taken in one field or point in time will provide limited informa- tion. The survey team also elicited yields that were based on farmers’ recall of 2006 and 2007 harvests, but these were Table 1. Definition of impact indicators used in this study. Indicator of Impact Definition Expected yields = Where i indexes variety grown, α is the proportion of the crop area planted to the variety, δ is the probability the variety is affected by drought, and a,b, and c are the parameters of the triangular distribution (the minimum, maximum, and mode) variety yields. The superscript 1 refers to a drought year and 0 to a year without drought. See Hardaker et al. (2004). Stock of attributes Count of unique production and consumption attributes of all millet and sorghum varieties held in stock in 2006 Relative deprivation of farm household j with respect to variable Y =AD(Yj) P(Yj) , where AD(Yj) is the mean of the variable Y for all farm households in the site with values higher than household j, and P(Yj) is the proportion represented by those households. Y in this study = expected yield of millet and sorghum, and stock of attributes. See Stark and Taylor (1989) and application by Edmeades et al. (2008). IMPACT ASSESSMENT DISCUSSION PAPER  not thought to be representative. Yields in Douentza, for ex- ample, were extremely low due to locust infestation in some villages. Nonetheless, we have also included these as out- come variables. Each farmer surveyed was asked, for each variety grown, the minimum, maximum and mode production from all plots under that variety, in years with and without drought. These were divided by plot area to obtain minimum, maximum and mode yields. To estimate the probability of drought stress, farmers were asked to report the number of years they had grown each variety and the number of years of drought stress during that period. The minimum, maxi- mum and most frequent (mode) yields were then elicited for years with and without drought stress. Unconditional expected yields were calculated on the basis of the prob- abilities of each type of year and corresponding yield esti- mates. The average unconditional expected yield per farm was calculated by multiplying each variety’s unconditional expected yield by that variety’s proportion of total crop area (see equation in Table 1). The third outcome variable, the count of unique produc- tion and consumption attributes, is a rough indicator of the perceived ‘richness’ of traits and uses embedded in the seed stored on the farm at the time of the survey. Farmers were asked to list all millet and sorghum varieties in stock and, for each variety, to list production characteristics and consump- tion uses. From farmers’ responses, a common ‘spanning set’ of production traits and consumption uses was iden- tified. For millet and then for sorghum, each was counted only once for all varieties stocked by each farm household. This indicator recognizes that subsistence oriented farmers manage a set of varieties to address a complex combination of needs and constraints and that, typically, no single variety meets all their needs (Bellon, 1996). Thus, researchers have hypothesized that farmers with multiple objectives, and particularly those who cannot easily achieve their objectives by trading in markets, will manage more diverse varieties. Other indicators are based on the concept of relative dep- rivation – a measure of inequality that has been applied in studies of migrant laborers (Stark and Taylor, 1989). Recently, this concept was applied in a study of hybrid banana adop- tion by Edmeades et al. (2008). An index was constructed for expected millet yields, expected sorghum yields, and the count of unique variety attributes. The index compares the status of each farm in the sample to all other farms at that project site, and is constructed by weighting the mean of all individuals with higher status by their sample proportion. The larger the number, the greater the relative deprivation of a farm with respect to the characteristic in question. Econometric model Instrumental variables regression can be used to explain variation in impact indicators among household farms while controlling for the effects of underlying observable and unobservable factors. Instrumental variables models are special cases of simultaneous regression models in which the causality of the relationship is recursive but the inter- relationship among the error terms of the two equations is explicit. In this case, participation affects outcome variables but outcome variables do not affect participation. The general form of the instrumental variables model is yi = xliθ1 + xliθ2 + vi (1) zi = xliβ1 + yiβ2 + µi (2) where the dependent variables include y, which meas- ures participation, which is an endogenous regressor, and z, which measures the impacts of participation. The vector x1 represents a set of explanatory variables that influence both participation and impacts, and the vector x2 includes instru- mental variables that explain participation only. The error terms of the equations, v and µ, have means of zero but are correlated. In this analysis, participation is a dichotomous variable (1=participation, 0 else). Angrist (1999) has shown that in models with dummy endogenous variables and nonnega- tive outcome variables, if the focus of the analysis is to esti- mate the causal effects of treatment rather than the magni- tude of structural parameters, a number of simple strategies, such as two-stage least squares, can be applied. Two-stage least squares produces consistent estimators of the ‘local average treatment effect’ that are less sensitive to assump- tions about functional form than probit or logit (Angrist and Krueger, 2001:77). To improve the efficiency of estimation when heteroskedasticity is present, the generalized method of moments is recommended. Standard diagnostic statis- tics include tests of a) the endogeneity of the first-stage de- pendent variable, b) the relevance of the instruments, and c) model identification. Explanatory variables (x1) that are common to both the participation and impact equation represent components of the ‘asset pentagon’, as conceptualized in the project ac- cording to the basic livelihoods framework. Human capital is measured in terms of the ratio of economically active per- sons to the total number of persons in the farm.4 The age limit for the numerator was arbitrarily set at 12 years of age (adulthood). The total value of livestock and material assets inventoried by the patriarch of the farm is the indicator of farm physical capital. Variables for human capital and farm physical capital were both highly correlated with total land area, suggesting that farm production is based on fairly fixed proportions of these three critical inputs (use of pur- chased inputs is negligible). Financial capital is represented by per capita cash income, calculated as the total annual ex- penditures for the farm divided by farm size. In this type of agricultural economy, where savings are in the form of land, livestock or other material assets, eliciting cash expenditure is thought to generate less bias than eliciting income. Enu- merators asked the patriarch of the farm to report expendi- tures by category. The total number of unique associations  All human capital variables were inter-correlated (age and education of the household head, number of adult women, number of adult men, number of household members with any education), and the active ratio was the least correlated with other explanatory variables. IMPACT ASSESSMENT DISCUSSION PAPER  to which any member of the farm belongs is used as an indi- cator of social capital. Other key explanatory factors are those related to treat- ment, site and orientation of production. The site variable represents major differences in agro-ecology, ethnicity and social structure, but also in market infrastructure. The treat- ment variable controls for indirect effects of the project in the locus villages and other test villages.5 Specialization of production in one target crop rather than another is ex- pressed in the share of target crop area allocated to millet. Target crops include millet, sorghum and cowpea. The tar- get crop area was summed over plots where these crops were the principal or sole crop. Very few plots were planted to cowpea as the principal or sole crop. The last variable related to orientation of production is the number of markets in which household members buy or sell millet or sorghum. Since these are not commercial farming operations, no certified seed is sold in markets, and other purchased inputs are rarely used, the count of markets has no direct relationship to expected yields. In a commer- cial farming system, a direct relationship between variety choice (a modern variety, for example) and market involve- ment might be expected. Nevertheless, we hypothesize that market involvement may relate to the desire of household members to procure and test new materials in DFF, either because they sell, or because they are food-deficit, i.e. they produce less than their food requirements, or because there are unobservable factors that cause a farmer to seek infor- mation, whether in village markets or in project activities.  Although it would have been preferable to estimate separate models for the direct and spillover effects of the DFF on farmers, it was not possi- ble for the survey team to identify the individual farmers from other vil- lages who had procured seed when visiting Boumboro. This would have required close monitoring over time by a member of the DFF, which must be built into the original research design. Instrumental variables (x2) are those that affect partici- pation but have no effect on impacts other than through participation, and are uncorrelated with the error terms of the impacts equation. These variables are 1) whether or not the patriarch has a Koranic education (instead of, or in ad- dition to, public school attendance), and 2) the binary vari- able ‘treatment’ which controls for the indirect effects of the project. The variable ‘Koranic education’ is not highly cor- related with any of the outcome variables (correlation coef- ficient range from -0.02 to 0.37). Explanatory variables are defined in Table 2, and means and standard deviations of explanatory variables are shown in Table 3. The null hypothesis that variable means are equal between test and control villages cannot be rejected at the 0.05 level of significance, although p-values are significant at the 0.10 level for total asset value and number of markets at the San site, and millet share and number of markets at the Douentza site. Additional comparisons of sample char- acteristics in the project baseline document6 indicate few sta- tistically significant differences with respect to human, farm physical, financial and social capital between the treatment and control groups, with important differences between sites. Differences in outcome variables by treatment and con- trol group are pronounced at the San site for all outcome variables except average millet and sorghum yields in 2005– 06. At Douentza, statistically significant differences are ap- parent only for expected millet yields and relative depriva- tion with respect to millet yields. The 2005 season was an  Available from the authors. Table 2. Definition of explanatory variables. Explanatory variable Conceptual variable Operational variable Site Agroecology, Ethnicity, Market infrastructure 0=San site 1 =Douentza site Treatment Experimental design 0=Control village 1=Treatment village Active ratio Human capital Ratio of economically active persons (>12 years of age) to total number of persons in the production unit Asset value Farm physical capital Total value of livestock and material assets (FCFA) Millet share Specialization in millet vs other target crops Proportion of total crop area planted to millet Per capita income Financial capital Total expenditures per year in FCFA in production unit divided by total number of persons Association membership Social capital Total number of associations in which a members of production unit participates Markets Market participation Number of markets in which members of production unit sell or purchase millet or sorghum Koranic education Instrument Whether or not the patriarch of the production unit has a Koranic education, either instead of or in addition to a public school education FCFA = Franc Communauté Financière Africaine IMPACT ASSESSMENT DISCUSSION PAPER  outlier due to heavy locust infestation, and these farmers (primarily Dogon) are quintessential millet growers. . Findings Instrumental variables regression The findings of the first-stage regressions are presented in Table 4 for millet growers.7 As expected, given the higher rates of participation at the San site and longer project in- volvement, location at the Douentza site reduces the chances that a farm household will include a DFF participant. Eco- nomically active human capital in the farm household bears no direct relationship to DFF participation. Farm households that are wealthier in terms of farm physical capital are more likely to participate. Greater specialization in millet than in sorghum increases the chances a farmer will participate in a DFF. This may be because millet was more widely grown at both sites, or because fewer improved millet varieties than sorghum varieties have been released at the sites – with the result that farmers are in search of new materials and means of improving their own. Social capital positively influences participation – these farm households have a propensity to engage in local associations. Similarly, participants sell or purchase the target crops in a larger number of markets  Probit and linear probability first-stage regressions are shown in the Ap- pendix. They are similar with respect to statistically significant causal fac- tors (site, total asset value, per capita cash income, markets). Table 3. Comparison of means, explanatory and outcome variables, by site, treatment and control villages. San site Douentza site Mean (control village) Mean (test village) P-value Mean (control village) Mean (test village) P-value Explanatory variables Active ratio 0.596 0.633 0.3896 0.5970 0.620 0.2434 Total asset value 737700 879888 0.0928 485178 558901 0.4588 Millet share 0.529 0.523 0.6785 0.867 0.891 0.0772 Per capita income 33892 37435 0.5098 23814 26130 0.2260 Association membership 1.55 1.64 0.1326 2.26 2.07 0.3669 Markets 1.16 1.27 0.0880 2.33 2.04 0.1081 Koranic education 0.318 0.258 0.4541 1.92 2.44 0.1543 Outcome variables Expected millet yields 331 609 0.0001 107 57.3 0.001 Expected sorghum yields 607 874 0.0141 194 216 0.5081 Mean millet yield 2005–06 908 1114 0.7106 766 799 1.000 Mean sorghum yield 2005–06 771 795 0.9926 241 314 0.589 Stock of attributes 25.8 30.0 0.0001 25.9 25.9 0.1305 Relative deprivation (expected millet yields) 200 364 0.0001 72.81 40.5 0.0001 Relative deprivation (expected sorghum yields) 535 433 0.0147 134 126 0.4995 Relative deprivation (stock of attributes) 18.5 11.4 0.0001 13.6 15.7 0.1665 Note: P-value refers to difference of means t-test for active ratio and Kruskall-Wallis test for other variables. Table 4. First stage regression results, determinants of DFF participation. Coefficient Standard error P value site -0.2126 0.0452 0.0000 active ratio -0.0183 0.0918 0.8420 asset value 0.0000 0.0000 0.0870 millet share 0.2546 0.0909 0.0060 per capita cash income 0.0000 0.0000 0.1090 numassoc 0.1068 0.0242 0.0000 markets 0.0739 0.0177 0.0000 edukoran -0.0035 0.0060 0.5620 treatment 0.2701 0.0357 0.0000 constant -0.1664 0.1105 0.1330 Test statistic P value F(9, 255) 12.06 0.0000 Centered R2 0.30 Uncentered R2 0.37 Root MSE 0.25 Shea partial R2 of excluded instruments 0.18 F(2, 255) of excluded instruments 28.8 0.0000 Anderson canonical correlation coefficient 48.9 0.0000 IMPACT ASSESSMENT DISCUSSION PAPER  than non-participants, other factors being held constant. The first-stage regression is the same regardless of the out- come variable, since DFF included both sorghum and millet related activities. Second-stage regression results for each impact indicator are displayed in Table 5. Taking into account underlying ob- servable and unobservable factors that predict participation in DFF, participation influences all the outcome variables with a high degree of statistical significance except average recalled sorghum yields. Location at the Douentza site offsets estimated yield im- pacts, but augments stock of attributes and further reduces relative deprivation. The pervasive harshness of this envi- ronment, combined with the social structure of the Dogon, is a strong leveler of inequality. The greater degree of self- reliance makes it important for farmers to stock the varie- ties with attributes that span both their consumption and production needs. While being wealthier enhances the likelihood of partici- pation, it has a negative impact on yields per hectare once participation has been considered – perhaps because these factors are associated with a more extensive operation and crop area that are harder to manage as effectively. As ex- pected, specialization in millet positively influences recalled millet yield, and negatively influences recalled sorghum yield and the total stock of attributes. Per capita cash income is statistically significant only in the recalled millet yield re- gression, and social capital is insignificant in explaining out- comes once DFF participation has been taken into account. Market participation is significant only in the expected mil- let yield and deprivation model. In all except the expected millet yield and attribute stock regressions, the null hypothesis of homoskedasticity was rejected and the regression was estimated with GMM (Gen- eralized Method of Moments), so that robust standard er- rors are reported. The Hausman test led to rejection of the hypothesis that DFF participation is exogenous in all cases except the regressions for expected sorghum yields and the relative deprivation with respect to the stock of attributes. The Anderson canonical correlation statistic based on the first-stage regression is highly significant, supporting the relevance of the instruments. The values of the Sargan sta- tistic (in the presence of homoskedasticity) or Hansen J sta- tistic (in the presence of heteroskedasticity) results in failure to reject the null hypothesis for each of the outcome equa- tions. Thus, the instruments in the first-stage regression are uncorrelated with the error term in the second-stage regres- sion. Rejection of the null hypothesis would have indicated either a misspecification of the second-stage regression or an invalid instrument. Three caveats are important to remember in interpreting findings. The first is that while we assert that use of longer- term, subjective yields is more reliable than actual recent yields, it is evident that these variables measure primarily Table 5. Second stage regression, impact of farm household participation in DFFs on crop genetic resource management. Expected millet yield Expected sorghum yield Mean millet yield (2005–06) Mean sorghum yield (2005–06) Stock of attributes Relative deprivation (millet) Relative deprivation (sorghum) Relative deprivation (attributes) cd 364 543 463 321 5.76 -205 -212 -8.01 0.0000 0.0040 0.0040 0.1060 0.0060 0.0000 0.0020 0.0220 site -331 -683 -539 -409 1.61 -347 -462 -5.27 0.0000 0.0000 0.0000 0.0000 0.1780 0.0000 0.0000 0.0010 active ratio -155 -162 -39.0 2.660 0.949 63.9 47.2 -0.154 0.1190 0.3380 0.8680 0.9850 0.6310 0.1090 0.4020 0.9570 asset value -0.0000388 -0.0000558 -0.0004070 -0.0001920 0.000000253 0.0000167 0.0000179 0.000000367 0.1000 0.0800 0.0000 0.0000 0.5580 0.1050 0.1000 0.4510 millet share -55.0 209 659 -682 -8.21 33.5 -74.1 10.2 0.6270 0.2660 0.0040 0.0000 0.0000 0.4650 0.2230 0.0000 per capita cash income 0.000937 -0.001096 -0.004770 -0.001720 0.0000317 -0.000414 0.000197 -0.0000810 0.4410 0.5820 0.0640 0.2570 0.1950 0.3770 0.7510 0.0250 numassoc 14.70 29.5000 5.80 -1.1400 -0.075 -5.47 -7.94 0.321 0.4100 0.3270 0.9210 0.9730 0.8790 0.4620 0.4460 0.6490 markets -36.9 -11.1 72.3 13.6 -0.390 18.2 2.1 0.272 0.0110 0.6210 0.1700 0.6600 0.3780 0.0060 0.8070 0.6490 Hansen J statistic 1.4430 0.032 0.7010 0.594 0.908 0.988 0.154 0.958 or Sargan test P=0.2246 P=0.8573 P=0.4024 P=0.441 P=0.3407 P=0.3203 P=0.6950 P=0.3278 Uncentered R2 0.6548 0.6284 0.6902 0.7081 0.958 0.8511 0.8774 0.7848 Hausman test P= 0.00296 P=0.14226 P=0.05202 P=0.07029 P=0.05583 P=0.00013 P=0.04773 P=0.26991 Note: In all cases except mean yields and stocks of attributes, the null hypothesis of homoskedasticity was rejected and the regression was estimated with GMM. For GMM, standard errors are robust. P-values are reported under the estimated regression coefficient. IMPACT ASSESSMENT DISCUSSION PAPER 10 changes in perceptions. They are not highly correlated with average yields based on recall, which is to be expected given the year-to-year variability in rainfall conditions and the fact that 2005 was an outlier season at the Douentza site. In Bou- mboro, the project has been operational from 1999 – but not so in Petaka – where project impacts are not evident for ei- ther expected and recalled yields once other variables have been taken into account. In addition, it is important to recognize that unobserv- able factors may explain both the decision to participate and yield expectations, which would contribute to an upward bias in estimated coefficients. Finally, the wide variation in average yields among villages, combined with the fact that most long-term participants live in Boumboro, means that the coefficient on DFF impact transmits a strong village-spe- cific effect.8 As recommended by Angrist, it is the signifi- cance of the coefficient (causality) rather than its magnitude that should be emphasized when interpreting the regression results. Post-estimation Figure 1 compares the cumulative density function (cdfs) of predicted values of expected millet yields. Predicted values account for selection bias by incorporating the effects of fac- tors that explain why DFF participants participate. As com- pared to the mean values shown in the regression results, the cdfs portray project impacts over the full range of predicted values. The graphs support the hypothesis that conditions are less risky for farmers in treatment villages than in con- trol villages. For every predicted value of expected millet yields, the probability that a farm household will harvest less is lower in treatment villages than in control villages. Expected millet yields in treatment villages dominate in the first-order stochastic sense: any farm household in a treat- ment village, whether neutral to or averse to risk, is better off.  The descriptive statistics shown in the Appendix suggest that while the only explanatory variable that differs between Boumboro and other treat- ments villages is market participation, most of the outcome variables differ significantly. Similar comparisons were made for predicted values of relative deprivation with respect to expected yields. In this case, households are worse off with higher values, which imply greater relative deprivation. A cdf lying entirely to the left for test as compared to control villages suggests that the probability of being more deprived relative to other house- holds is always less in test villages – an improvement in sta- tus. The cdfs of relative deprivation dominate stochastically in the first-order sense for expected millet yields in treat- ment villages of the San site (Figure 2), but not for expected sorghum yields; at the Douentza site, they dominate for sor- ghum yields in the treatment villages but not for expected millet yields. For both crops at both sites, from the perspec- tive of farmers’ yield expectations, conditions are not always better in treatment villages than in control villages. Figures were not reported for all outcome variables be- cause of space limitations. In the case of sorghum, while the cdfs for expected yields do not cross, they lie tangent to one another at some points. First-order stochastic dominance is also evident for the total count of attributes. For each pre- dicted total count of attributes, the chances are smaller that a farm household in a treatment village will possess fewer attributes. . Conclusions Despite continued progress in breeding improved sorghum and millet varieties, and the gradual process of seed sector reform, adoption rates are often considered to be relatively low in Mali – a nation whose rural population has depended on sorghum and millet for millennia. Some researchers have recommended more involvement of farmers themselves in evaluating and disseminating promising varieties. This pa- per has evaluated a pilot effort to involve farmers through Diversity Field Fora, which build on the concepts of farmer field schools. DFF aim to strengthen farmers’ capacity to manage diverse millet and sorghum varieties, thereby im- proving productivity. Regression results concur with previous assessments of farmer field schools in demonstrating that participants in 0 .2 .4 .6 .8 1 Cu mu lat ive Pr ob ab ilit y -5 00 0 5 00 1 00 0 F itte d va lu es C on tro l T re atm e nt Figure 1. Cumulative density function of predicted values of expected millet yields, in the presence and absence of drought. 0 .2 .4 .6 .8 1 Cu mu lat ive Pr ob ab ility 1 00 2 00 3 00 4 00 F itte d va lu es C on tro l T re atm e nt Figure 2. Cumulative density function of predicted values of relative deprivation with respect to expected millet yields in the San site. IMPACT ASSESSMENT DISCUSSION PAPER 11 DFF are more likely to possess more social and farm physi- cal capital than non-participants. Participation has a posi- tive impact on expected sorghum and millet yields, recalled millet yields and the stock of variety attributes, enhancing participants’ standing relative to other farmers at the project site with respect to these impact indicators. Farmers in treat- ment villages are better off in terms of expected yields and attribute stocks throughout the range of predicted values. A significantly higher probability of participation at the San site, combined with the regression and descriptive results, indicates that the impacts of DFF were principally observed at this site in the village of Boumboro, where local field staff have been continuously engaged for a much longer period of time. However, inequality with respect to millet and sor- ghum crop genetic resources appears to be greater at the San site than at Douentza. Two policy points with respect to DFF emerge from these findings. First, long-term commitment to fostering local leadership and capacity is likely to be a key factor in achiev- ing impacts with this type of extension approach. The local leader at the San site has been trained on site and abroad, and has established his own NGO in surrounding commu- nities. Second, for precisely this reason, it will be difficult to scale up impacts from one village to many without support- ing, coordinated investments by national public institutions and donors. This analysis has focused on statistical methods that re- duce bias and the measurable, immediate impacts of DFF. Although farmer selection bias associated with participa- tion has been taken into account, the findings cannot be generalized to other communities unless these communities conform to the criteria used to select sites. A more compre- hensive evaluation, which would require the application of additional analytical approaches over a longer time period and in multiple locations – is not yet justified by the amount of funds invested in DFF nationwide. As more participatory research efforts are undertaken in Mali, however, these eval- uations will be crucial. References Angrist J. 1999. Estimation of limited-dependent variable models with dummy endogenous regressors: Simple strategies for empirical practice. MIT Economics Depart- ment Working Paper 99-31. Massachusetts Institute of Technology, Cambridge, USA. Angrist JD and Krueger AG. 2001. Instrumental variables and the search for identification: From supply and de- mand to natural experiments. Journal of Economic Per- spectives 15(4): 69–85. Anyamba A and Tucker CJ. 2005. Analysis of Sahelian veg- etation dynamics using NOAA AVHRR NDVI data from 1981–2003. Journal of Arid Environments 63: 596–614. Bazile D. 2006. State-farmer partnerships for seed diversity in Mali. Gatekeeper Series 127. IIED (International Insti- tute for Environment and Development), London. BCRA (Bureau Central du Recensement Agricole). 2006. Recensement General de l’Agriculture 2004. Ministère de l’Agriculture, Ministère de l’Élevage et de la Pêche, Ministère du Plan et de l’Amenagement du Territoire, Bamako, Mali. Bellon M. 1996. The dynamics of crop infra-specific diver- sity: A conceptual framework at the farmer level. Eco- nomic Botany 50(1): 26–39. Davis K. 2006. Commentary. Farmer field schools: A boon or bust for extension in Africa? Journal of International Agricultural and Extension Education 13(1): 91–96. De Vries J and Toenniessen G. 2001. Securing the harvest: Biotechnology, breeding and seed systems for African crops. CAB International, Wallingford, UK. Diakité L, Sidibé A and Diakité S. 2005. Enquête pour l’identification des paysans semenciers dans les zones et villages propices à la production semencière. Rapport fi- nal. Services d’Experts pour les Ressources Naturelles et l’Environnement au Sahel (SERNES), Bamako, Mali. Diakité L, Sidibe A, Smale M and Grum M. 2008. Seed value chains for sorghum and millet in Mali: A state-based sys- tem in transition. IFPRI Discussion Paper 749. IFPRI (In- ternational Food Policy Research Institute), Washington, DC. Edmeades S, Nkuba J and Smale M. 2008. The impact of hy- brid use on the vulnerability and deprivation of small- holder banana growers in Tanzania. Manuscript. Envi- ronment and Production Technology Division, IFPRI (International Food Policy Research Institute), Washing- ton, DC. Handa S and Maluccio JA. 2007. Matching the gold stand- ard: Evidence from a social experiment in Nicaragua. Manuscript. Department of Economics, Middlebury Col- lege, Middlebury, Vermont, USA. Hardaker JB, Huirne RBM, Anderson JR and Lien G. (Eds), 2004. Coping with Risk in Agriculture. CAB Internation- al Publishing, Wallingford, UK. Harlan JR. 1992. Crops and Man. Second edition. American Society of Agronomy and Crop Science Society of Amer- ica, Madison, Wisconsin, USA. Matlon P. 1987. The West African semi-arid tropics. In: Mel- lor JW, Delgado CL and Blackie MJ, editors. Accelerating Food Production in Sub-Saharan Africa. Johns Hopkins University Press, Baltimore, Maryland, USA. Miguel E and Kremer M. 2004. Worms: Identifying impacts on education and health in the present of treatment ex- ternalities. Econometrica 72(1): 159–217. Ravallion M. 1994. Poverty comparisons. Fundamentals of Pure and Applied Economics. Volume 56. Harwood Aca- demic, Chur, Switzerland. Sanders JH, Shapiro BI and Ramaswamy S. 1996. The Eco- nomics of Agricultural Technology in Semiarid Sub-Sa- haran Africa. Johns Hopkins University Press, Baltimore, Maryland, USA. Smith BD. 1998. The Emergence of Agriculture. Freeman, New York, USA. Sperling L, Weltzien E, Sangaré MB, Shines JSc, Salla Boré S, Bamba A, Traoré Keita CO, Ag Hamada M, Ballo M, Sangaré F, Kanouté M, Sanogo B, Guindo H, Konta B, Sa- nogo S, Traoré A, Loeffen M and Dembélé A. 2006. Seed IMPACT ASSESSMENT DISCUSSION PAPER 1 System Security Assessment (SSSA), Douentza, Northern Mali. Final Report (June). Catholic Relief Services, Mali, and Partners, Bamako, Mali. Stark O and Taylor E. 1989. Relative deprivation and inter- national migration. Demography 26: 1–14. Touré A, Sanogo O, Diakité L and Sidibé A, 2006. Program for Africa’s Seed Systems (PASS). Country Report: Mali. Institut d’Économie Rurale, Bamako, Mali. Tripp R, Wijeratne M and Hiroshini Piyadasa V. 2005. What should we expect from farmer field schools? A Sri Lanka case study. World Development 33(10): 1705–1720. Vaksmann M, Traoré SB and Niangado O. 1996. Le photopé- riodisme des sorghos Africains. Agriculture et Dévelop- pement 9: 13–18. Van der Berg H and Jiggins J. 2007. Investing in farmers: The impacts of farmer field schools in relation to integrated pest management. World Development 35(4): 663–686. Vitale J and Bessler B. 2006. On the discovery of millet prices in Mali. Papers in Regional Science 85(1): 139–162. Weltzien E, Christinck A, Touré A, Rattunde F, Diarra M, Sangaré A and Coulibaly M. 2006. Enhancing farmers’ access to sorghum varieties through scaling up participa- tory plant breeding in Mali, West Africa. In: Almekinders C and Hardon J, editors. AgroSpecial 5: Bringing Farm- ers Back into Breeding: Experiences with Participatory Plant Breeding and Challenges for Institutionalisation. Agromisa Foundation, Wageningen, The Netherlands. IMPACT ASSESSMENT DISCUSSION PAPER 1 Appendix: Additional statistics A. Probit and ordinary least squares regressions explaining DFF participation Determinants of DFF participation, probit regression Variable Coef. Std. Err. P>z Site -1.8142 0.4309 0.0000 Active ratio 0.0240 0.7109 0.9730 Total asset value 0.0000 0.0000 0.0820 Millet share 0.7461 0.7198 0.3000 Per capita cash income 0.0000 0.0000 0.1060 Association membership 0.2040 0.1726 0.2370 Markets 0.5497 0.1402 0.0000 Koranic education 0.0439 0.0481 0.3610 Constant -0.6355 0.8339 0.4460 Number of obs 265 LR chi2(8) 37.67 Prob > chi2 0.0000 Pseudo R2 0.2274 Log likelihood -63.970473 Determinants of DFF participation, ordinary least squares regression Coef. Std. Err. P>t Site -0.2320 0.0498 0.0000 Active ratio 0.0387 0.1011 0.7020 Total asset value 0.0000 0.0000 0.0190 Millet share 0.1133 0.0983 0.2500 Per capita cash income 0.0000 0.0000 0.0540 Association membership 0.0368 0.0247 0.1370 Markets 0.0783853 0.0195788 0.0000 Koranic education 0.0040264 0.0065585 0.5400 Constant 0.1531561 0.112754 0.1760 Number of obs 265 F(8, 256) 5.26 Prob > F 0 R-squared 0.1412 Adj R-squared 0.1143 Root MSE 0.2756 B. Comparison of means of outcome and explanatory variables, locus and other treatment villages, San site Boumboro Other treatment P-value Expected millet yields 660 529 0.3318 Expected sorghum yields 1014 673 0.0258 Mean millet yield 2005–06 999 791 0.0263 Mean sorghum yield 2005–06 822 705 0.3598 Stock of attributes 30.74 29.46 0.032 Relative deprivation (expected millet yields) 273 310 0.3308 Relative deprivation (expected sorghum yields) 418 535 0.0258 Relative deprivation (stock of attributes) 9.26 13.20 0.0389 Explanatory variables Active ratio 0.609 0.651 0.3776 Total asset value 1135337 859970 0.4561 Millet share 0.605 0.518 0.1281 Per capita income 36657 38035 0.8703 Association membership 1.89 1.63 0.1418 Markets 2.04 1.04 0.0001 Koranic education 0.556 0.029 0.1203 1 Bioversity International Via dei Tre Denari, 472/a, 00057 Maccarese, Rome, Italy Tel.: (39) 0661181 bioversity@cgiar.org www.bioversityinternational.org © This work by Bioversity International is licensed under a Creative Commons Attribution-NonCommercial-NoDerivs 3.0 Unported License.