ANALYZING TECHNOLOGY ADOPTION: CHALLENGES AND LIMITATIONS OF MICRO-STUDIES Cheryl R. Doss Yale University

Cheryl Doss Yale Center for International and Area Studies P.O.Box 208206 New Haven, CT 06520-8206, USA 203-432-9395 (office) 203-432-9886 (fax)

Cheryl.Doss@yale.edu

2 Abstract: Although the questions about the use of improved technologies in developing country agriculture have expanded to include the roles of policies, institutions and infrastructure, most micro-level adoption studies can not address these important policy issues. Drawing on an extensive review of the literature on the adoption of agricultural technologies, this paper suggests alternative approaches for designing technology adoption studies to make them useful for policy makers. It explores the generic limitations of cross-sectional adoption studies carried out in small number of communities and discusses the problems faced in conducting such studies. Recommendations include the use of sampling approaches that allow data from micro-studies to be generalized to higher levels of aggregation, adherence to clearly defined terms that are standardized across studies, and carefully examining the assumptions that often underlie such studies. In addition, the use and interpretation of proxy variables for the policy variables are discussed.

Keywords: agricultural technology, adoption

3 ANALYZING TECHNOLOGY ADOPTION: CHALLENGES AND LIMITATIONS OF MICRO-STUDIES 1. Introduction For most of the world’s poorest countries, and especially those in Africa, agriculture continues to offer the leading source of employment and to contribute large fractions of national income. In many of these countries, however, agricultural productivity is extremely low. Clearly, increasing agricultural productivity is critical to economic growth and development. One important way to increase agricultural productivity is through the introduction of improved agricultural technologies and management systems. National research programs exist in most countries, and working with a network of international centers operating under the auspices of the Consultative Group on International Agricultural Research (CGIAR), these research institutions have worked to develop new agricultural technologies and management practices. A challenge for agricultural researchers, however, is to understand how and when new technologies are used by farmers in developing countries. For this task, agricultural scientists have turned to social scientists, asking for improved understanding of the mechanisms underlying technology adoption. Over the years, researchers have worked to answer changing questions about agricultural technology adoption. Initially, policy makers and researchers sought simple descriptive statistics about the use and diffusion of new seed varieties and associated technologies such as fertilizer and irrigation. Concerns arose later about the impact of technology adoption – on commodity production, on poverty and malnutrition, on farm size and input use in agriculture, on genetic diversity, and on a variety of social issues. Numerous researchers (many associated with the CGIAR centers) have developed innovative methodologies for addressing such concerns, carried

4 out surveys and collected enormous amounts of data to describe and document the adoption of new agricultural technologies. Yet many questions remain. At the simplest level, we still have considerable gaps in our knowledge of which technologies are being used, where, and by whom. Bigger questions have also arisen. Scholars and policy makers are asking about the roles of policy, institutions, and infrastructure in increasing agricultural productivity. Today, studies of agricultural technology adoption are used widely in four areas of agricultural policy: · Assessing the impact of agricultural research1 · Priority setting for research · Evaluating the distributional impacts of new technology · Identifying and reducing constraints to adoption. These questions are complex; they require more complicated research methodologies than did the first-generation studies of diffusion. Simple descriptive statistics do not offer much insight into the process of technology adoption or productivity growth. As a result, much of the published literature on technology adoption in recent years has focused on methodological issues, trying to model the process of technology adoption and to get empirical measures of the importance of different factors. This literature has wrestled with deeply embedded problems of simultaneity and endogeneity – problems which were not important in the first generation of descriptive research. Too often, however, these methodological advances have not been translated into improved policy-oriented studies that can yield information useful to policy-makers. In particular, many studies of technology adoption give disappointingly meager information about

1

For example, see Evenson and Gollin, 2003.

5 the importance of agroecological variables and policy environments – and by ignoring these variables, they may give flawed information about the impact of other policies. This paper shows that some “adoption studies,” by their very design, can give little more than descriptive information. The paper also suggests concrete alternative approaches to survey design that can increase the usefulness of the adoption studies that are now widely undertaken. With no additional resources, these studies could be made to offer far more valuable information. In particular, this paper discusses how to conceptualize and measure some of the variables used in adoption studies and how to interpret the econometric results for policy purposes.

Technology Adoption: Current Trends in the Literature The literature on technology adoption is currently moving in three directions. These include 1) innovative econometric and modeling methodologies to understand adoption decisions, 2) examinations of the process of learning and social networks in adoption decisions and 3) micro-level studies based on local data collection intended to shed light on adoption decisions in particular contexts for policy purposes. Much of the recent literature focuses on new methodologies to deal with issues of endogeneity and simultaneity of decisions. Feder, Just and Zilberman reviewed the literature in 1985. Subsequently, Besley and Case (1993) discussed the different econometric approaches used in these studies. Since then, many economists have tried to resolve these underlying issues. To sample a few of the approaches used since then, Smale et al. (1995) modeled adoption as three simultaneous choices – the choice of whether to adopt the components of the recommended package, the decision of how to allocate different technologies across the land area, and the

6 decision of how much of some inputs, such as fertilizer, to use.2 Staal et. al (2002) used GIS data to examine smallholder dairy farms in Kenya. Holloway and colleagues (2002) used spatial econometric models to understand the adoption of high-yielding rice varieties in the Philippines. Batz et al. (2002) develop an approach to predict the speed and ceiling of technology adoption, using relative investment, relative risk and relative complexity for smallholder dairying in Kenya. Related to these are new methods that simulate the adoption and diffusion of technologies (Berger, 2001; Dimara and Skuras, 2003, Mahmoud, 2004). These new methodological approaches are important developments for our understanding of technology adoption issues. A second strand of literature focuses on the learning and social networks involved in technology adoption. This literature is not necessarily focused on agricultural development, but several papers have used episodes of agricultural technology adoption as a examples of social learning. Conley and Udry (2003) model the adoption of pineapple production practices in Ghana and find that social learning is important in the spread of the new technologies. Foster and Rosezweig (1995) find that own experience and neighbor’s experiences with high yielding varieties in India significantly increased the profitability from these varieties. Considerable more work is needed in this area to understand how the use of technologies spread. Finally, a third strand of literature, aimed primarily at agricultural technology policy, asks about particular technologies and why they are not being adopted in given locations. For example, from 1996-98, the International Center for Wheat and Maize Improvement (CIMMYT) collaborated with national research institutions in East Africa to conduct 22 micro-level studies of technology adoption in Ethiopia, Kenya, Tanzania, and Uganda. These studies looked at the

2

Leathers and Smale (1995) use a Baysian approach to examining the sequential decisions.

7 adoption of improved varieties of wheat and maize, as well as adoption of chemical fertilizers. They provide useful descriptive information on who is using improved seed and fertilizer in some areas of East Africa.3 Many similar studies are referenced throughout this paper (e.g., Ransom, 2003; Hintze, 2003). Increasingly, attention has shifted from the adoption of new crop varieties to the adoption of new management practices. Although some of these studies make methodological contributions, others contribute primarily by providing information on localized situations of interest to policy makers. This paper focuses on conceptual and measurement issues associated with adoption studies. These issues are relevant to all three of the strands of the literature. Using new methodologies or econometric techniques cannot resolve the problems if the data are fundamentally inappropriate to the question being asked, or if the questions are not the appropriate ones. Nor do empirical studies help policy makers if they are based on poorly conceived data.

3.

What Micro Adoption Studies Can Show Many adoption studies are based on an initial desire to gather basic information about the

use of modern varieties and inputs and to identify constraints to technology adoption and input use. Local governments often need this information for policy making. Micro surveys can provide such information, often at lower expense than full-fledged agricultural censuses. In addition to generating descriptive data about technology diffusion, micro studies can provide useful background information about the farmers who are currently using a technology and those who are not. For example, relatively little is known at present about the farmers who
3

The reports on the individual studies are available from CIMMYT and many are referenced in

8 use modern varieties or fertilizers in much of Africa. Most national governments in the region do not systematically collect or report such data, in contrast to some other parts of the world. Without basic descriptive information on who is using the technologies and who is not, it is difficult to know how to formulate policies aimed at improving agricultural productivity. Microlevel studies of technology use can document some of this information. Cross-section analysis at the micro-level can answer important questions about technology use. At the most basic level, we can find out what crops farmers are actually growing in their fields, and how they are growing them. We can also learn about their decisionmaking processes – by asking farmers about what factors were important to their choices of crops and technologies. Cross-section data can also tell us about farmer preferences. We can learn about growing conditions in specific areas and what varietal characteristics are important to farmers. In addition, we can learn about farmers’ perceptions of the constraints that they face. Cross-section analysis can also provide some information on patterns of adoption and disadoption. Information on whether or not farmers have ever used improved technologies can be collected, as well as information on what they are currently using. Obviously, these patterns can be more clearly analyzed if we have panel data on farmers over many years, but even a cross-section survey can show whether specific technologies have been tried and discarded by farmers, whether they are used intermittently, or whether they have never been tried at all. Farmers are usually able to provide information on why they did not adopt a new technology. Sometimes their answers provide important insights into the constraints facing farmers. Other times, multiple constraints are binding, so that removing the listed constraint would not necessarily result in the farmers’ adoption of technology.

this paper. For a synthesis of all of the studies, see Doss, et al. 2003.

9 Information on the profitability of a given technology can sometimes be determined from cross-sectional analysis of micro-level data. Adoption studies typically do not collect data on costs of production, but understanding the conditions under which improved technologies are profitable would add to our understanding of adoption decisions. Thus, we can obtain a description of the current practices of farmers through micro-level studies. Studies of this kind can explain what farmers are currently doing and may be able to explain what factors influence their decisions.

4.

Limitations of Micro Studies – Generic Issues Micro studies can provide important descriptive information on the current use of

agricultural technologies by farmers. However, these studies do not – and cannot – address many other important research and policy questions. This section will argue that although some shortcomings of micro studies can be dealt with through careful survey design, some are intrinsic and reflect the fact that these studies are based on data collected at a single point in time. Simply put, there are some questions that cannot be answered with cross-section micro-level data.

Lack of dynamics One fundamental limitation of micro-level adoption studies is that cross-section data do not permit analysis of the dynamics of technology adoption. These surveys typically collect cross-section data on adopters and non-adopters. Comparisons between the two groups are interesting, but they cannot tell us as much as studies that look at the same farmers before and

10 after they encounter a new technology.4 Similarly, cross-section data cannot tell us much about the impact of a new technology on the well-being of farmers or farm communities – nor on the distributional effects. For example, researchers might like to know the extent to which new technologies have changed the relative and absolute incomes of farmers. But if we only observe adopters and non-adopters, we do not know whether differences in their income or wealth are causes or effects of technology adoption (or both, or neither). To understand the dynamics of adoption decisions, rather than just developing static descriptions of particular areas, it is necessary to develop panel data sets. Technology adoption decisions are inherently dynamic. Farmers do not simply decide whether or not to permanently adopt an improved variety, but instead they make a series of decisions: whether or not to try planting an improved variety, how much land to allocate to the improved variety, whether or not to continue to grow it, whether to try a different improved variety. Decisions about other input use and management techniques are similarly complex. Decisions in one period depend critically on decisions made in previous periods. To understand these decisions, farmers’ decisions need to be followed over a period of time. This is best done with panel data sets of farmers. Ideally, we would start to follow farmers before they adopt improved technologies, but panel data studies may be useful even if they are not strictly “before” and “after” studies. Having more than one observation per farmer allows us to control for heterogeneity across households. Since many farmers have already adopted some form of improved technology, we may need to be satisfied with following farmers over time and observing the changing patterns of

It may be possible to collect some retrospective data from farmers, but any retrospective data should be carefully interpreted from data on current practices, since recall and selection biases may be present.

4

11 use of improved technologies. Panel data would allow us to look at changes in the use of improved materials, both in terms of varietal replacement and the extent of adoption by individual farmers. Panel data would also help us to understand the distributional impacts of new technology. Since many things change within rural communities, both in response to new agricultural technologies and in response to changes in outside forces, panel data are needed to sort out the effects. With panel data, we can begin to answer questions such as whether the benefits of being an early adopter continue once many farmers have adopted the technology. The sample would need to be large enough, and distributed across a wide enough geographical area, to capture variation in policies, institutions, infrastructure, and level of economic development. Because collection of panel data sets requires a major commitment of time and resources, we should not dismiss the need for cross-section analyses of individual sites. Yet, to understand the long-term dynamics of adoption, it is necessary to develop panel data for key locations. Generating the additional information will likely involve considerable expense, but the payoffs could be large in terms of our understanding of technology adoption in Africa and elsewhere.

Lack of variation within samples A recurring problem with micro studies is that there may not be much variation across households in a small survey with respect to variables of interest. For example, in a survey of a few villages in close proximity to each other, it will be difficult to get much information about the impact of credit or labor market failures. Use of credit and hired labor often depend on both the characteristics of the farmer and the characteristics of the village or region. When all of the respondents live in the same area, there may not be much variation among farmers with respect

12 to market access and infrastructure. If this is the case, then these variables should not be included in the econometric analysis. (However, it may still be useful to collect these data, as will be discussed in the following section.) Similarly, agroecological factors often influence the adoption of technology. Typically adoption studies include location variables for the village or region. Studies done in one or two regions will often have relatively little agroecological variation. Alternatively, where there is agroecological variation, appropriate variables will pick up the variation in rainfall, soil quality, and production potential. However, these variables may also pick up variation unrelated to agricultural potential, such as infrastructure and availability of markets for inputs and outputs. Without sufficient variation within the samples, these different relationships cannot be disentangled. It would be useful to have a measure of agricultural potential that shows more variation at the local level – even at the farm level, where possible. It might still be important to include a control for agroecological zone, but then this variable would be interpreted differently.

5. Challenges in the Design and Implementation of Single Studies Careful analysis of existing studies suggests that there are a number of ways in which to improve micro-level analyses. These suggestions are relevant both for the collection of data and for the interpretation of the results. Many adoption studies include explanatory variables without providing any theoretical justification. They frequently do not clearly describe why the particular variables are used or what they are expected to capture. Often the dependent variable is not clearly defined. Careful attention to the variables that are included (and justification for both those that are included and those that are omitted) will make analyses more useful to policy makers and agricultural

13 researchers. In this section, I focus on some widely used variables, some alternative specifications, and the interpretation of the results from econometric estimations. It is rare for social scientists to have variables that exactly measure what we are most interested in – most of the variables that we use are good approximations at best. So it is useful to think about how closely our measures track what we are interested in. Defining and interpreting results obtained from imperfect proxies is key to the usefulness of adoption studies at the micro-level.

Defining an adopter One key issue that the CIMMYT studies bring out is the question of what is meant by an “adopter” of a technology. The definitions of an adopter vary widely across studies, even across the 22 studies that CIMMYT conducted in East Africa examining the adoption of improved varieties of wheat and maize and fertilizer. What exactly is an adopter? This proves to be a complicated question with no obvious “correct” answer when discussing the adoption of improved varieties.5 It may be even more often difficult to define the adoption of management practices such as those now being promoted for environmental benefits and sustainability. In defining “adoption,” the first thing is to consider whether adoption is a discrete state with binary variables (a farmer either is, or is not, an “adopter”) or whether adoption is a continuous measure. The appropriateness of each approach may depend on the particular context. Many studies use a simple dichotomous variable approach.6 For example, a farmer may

5

It is important to define an adopter not only when it is the dependent variable, but also when it is an independent variable, such as when Brush et al. (1992) examine the impact of the adoption of improved varieties of potato on the genetic diversity of potatoes in the Andes. 6 A number of the CIMMYT studies followed this approach, including Salasya, et al. 1998, Hailye et al. 1998, and Beyene et al. 1998.

14 be defined as an adopter if he or she was found to be growing any improved materials. Thus a farmer may be classified as an adopter and still grow some local materials. This approach is most appropriate when farmers typically grow either local varieties or improved varieties exclusively (for example, the adoption of a wholly new crop), or when the management practice is something that cannot be partially implemented. If the interesting aspects of adoption are related to partial adoption (e.g., situations where farmers are increasingly planting more of their land in improved varieties while continuing to grow some local varieties), or to gradual shifts in management practices, then a continuous measure of adoption is more appropriate. Many studies use measures of the proportion of land allocated to new technologies as the measure of adoption.7 Which measure is used may be important. Gebremedhin and Swinton (2002) find that the factors that affect the initial adoption of social conservation technologies are different from those that affect the intensity of their implementation. Often, we seek to understand farmers’ partial adoption of new technologies. In an examination of why farmers in Malawi allocate land to a several different maize varieties, Smale et. al (2001) find that no single explanation – portfolio diversification, safety-first, experimentation or input fixity – explains the land allocation patterns as well as a model that includes all four explanations. Thus, choosing how to model adoption will depend on the type of technology, the local context, and the research questions being examined. Defining “adoption” may be further complicated due to the complexity of defining the technology being adopted. For the adoption of improved seeds, the CIMMYT studies used

Including: Mussei, et. al 2001; Bisanda et al., 1998; Katinila et al., 1998, Nkonya et al., 1998, Kaliba et al., 1998a and 1998b; Mafuru, 1999; Gemeda, et al., 2001; Degu, et. al, 2000; and Kotu, et al. 2000.

7

15 several different definitions. In some cases, farmers were said to be adopters if they were using seeds that had been “recycled” for several generations from hybrid ancestors.8 In others, adoption was identified with following the extension service recommendations of using only new certified seed (Bisanda 1998; Ouma 2002). Since the definition of adoption encompasses a wide range of dissimilar practices, the results from these studies are not comparable. The definitions of adoption of management practices may be even more complex. Studies should explicitly state how they are defining these terms. In cases in which the full range of farmer behavior is not known a priori, it may make sense to ask farmers for detailed information. The researcher can then create an appropriate adoption measure using this detailed data. The researcher might, for example, collapse detailed survey data into an ordered variable (such as whether farmers are using new seed of improved varieties, recycled varieties, or local varieties). This would require the use of multinomial estimation, rather than a simple binary model. We may also want to know whether farmers are growing one improved variety or multiple varieties on their farms. Since many farmers grow more than one variety, measures of the proportion of land planted to improved materials are often used; this type of measure does not easily lend itself to more than one definition of “improved materials.” Collection of detailed data would also allow the creation of measures of adoption that are comparable across studies.

In maize breeding, the genetic integrity of hybrid seeds decays rapidly from one generation to the next. If farmers use hybrid seed in one year and then save the seeds to replant in the following year, this is referred to as “recycling” of hybrids. Within a few years, it yields populations that have little genetic relationship to the original hybrid seed – although there may be some traits that are retained. Recycling of maize seed is common in Africa. For example, many farmers in the Tanzanian CIMMYT studies, (Kaliba et al. 1998b, 1998c; Nkonya1998) had recycled hybrid seeds for twelve years or more. Whether the use of recycled seed should count as “adoption” is a debatable point and perhaps depends on the situation, but it should not be treated as the same as using new hybrid seed.

8

16 Finally, in defining an adopter, we may also be interested in farmers’ histories of technology use. In order to develop such histories, we must ask not only whether a farmer is currently using a particular technology, but they can also whether he or she has ever used it. This helps to distinguish farmers who have never tried a technology from those who have tried it and discarded it. In many studies, both categories are treated as “non-adopters,” which may conceal important differences. Given the complexity of adoption measures, and given the potential value of having compatible measures of technology adoption across studies, it may make sense for one institution or a group of institutions to take a leadership role in conceptualizing the adoption of new crop varieties and management techniques in agriculture. The CGIAR centers, for example, could perhaps develop appropriate adoption categories for different crops that would be available to researchers and that would make sense across a variety of environments and allow comparisons.

Demographic variables A number of explanatory variables similarly require further consideration. One set of commonly used right-hand variables encompasses the demographic characteristics of farmers. “Age” is straightforward, if not always measured accurately. Usually it is expected that older farmers have more experience, but this may be counteracted by younger farmers being more innovative. Lapar (1999) suggests that for technologies requiring long-term investment, age may also indicate the time horizon of the farmer, with younger farmers having a longer frame in which to gain the benefits. “Farming experience” is a bit more difficult to measure, and it is important to define exactly what farming experience entails: Should the survey count all farming experience?

17 Experience farming one’s own plot? Experience farming this particular plot? Experience farming this particular crop? All of these measures have been used, but it is not always clear what they are attempting to capture. Neither age nor experience has much policy relevance, other than for policy-makers to understand the demographic characteristics of farmers using these technologies.9 It is also expected that farmers with higher levels of education will be more likely to use improved technologies. Many studies find this effect. However, the definitions of adoption again may be important. Weir and Knight (2000) find that, in Ethiopia, household-level education affects whether a farmer is an early or late adopter, but is less important in determining whether or not the farmer ever uses fertilizer. To examine the effects of education on adoption of technologies, many studies enter the years of education on the right hand side of the equation.10 This approach assumes that there is a linear relationship and that each year of education has a similar effect on the outcome. Since we don’t expect the relationship to be linear, it may be better to include dummy variables for different levels of education.11 We might, for example, expect a big difference between farmers who are literate and those who are not. The difference between five and six years of education may have much less of an effect on the adoption of improved technologies than the difference between three and four years of education. Depending on the particular context, it may be

If age and/or experience are in fact related to adoption decisions, then they must of course be included as explanatory variables to prevent biasing the estimates of other policy-relevant variables. However, it may also be worth pursuing these variables further; are they in fact capturing cohort-specific changes in other conditions (e.g., education, health) that might have some policy relevance? 10 For example, Boahene, et al. 1999. 11 Alternatively, it is possible to assign weights to different years of education, based on the marginal values obtained from Mincer-type equations; but this is seldom feasible for the kinds of studies under consideration here.
9

18 appropriate to include two or three dummy variables indicating different levels of education. In some contexts, especially where levels of education are very low and adults are more likely to have attended adult education programs than formal schooling, literacy might be a better measure.12 One additional issue to consider is whose education level should be included? Frequently, the head of the household and the farmer are assumed to be the same person, and the education level of this person is included. Yet, they may not always be the same person. In one study that looked at this issue, Asfaw and Admassie (2004) included both the education of the household head and the education of the person with the highest level of education in the household. This assumes that production is enhanced by having someone in the household with more education, even if it is not the head. Foster and Rosenzweig (1996), in a careful study using nationally-representative longitudinal household data from India, find that schooling does have a positive impact on the profitability of technical change, but that the relationships are complex. The returns to schooling are increased by technical change and technical change in turn leads to greater private investment in schooling. Similarly, Wozniak (1984) shows that innovative decision-making is an activity that intensively uses human capital. Thus, the incentives for farmers to attain human capital increase as the technological environment becomes more dynamic. At a macro level, then, the technology environment may affect the level of schooling. The final demographic factor that is important to consider is that of gender. Frequently, studies simply include a dummy variable indicating whether or not the head of the household is

12

For example, Croppenstedt et al. 2003, use dummy variables for literacy, whether the farmer has completed grades 1-3 and whether the farmer has completed grade four or more for rural Ethiopia.

19 man or a woman. This assumes that the head of the household is the farmer and the primary decision-maker. These three roles may not all go together. In Ghana, a study by Doss and Morris (2001) found that the gender of the farmer was not significant in determining the adoption of improved varieties of maize and of fertilizer, but farmers in female-headed households were less likely to use these new technologies. They suggest that male heads of households are better able to access the resources needed to use improved technologies, for themselves and for their wives. Thus, it is useful to collect data both on the gender of the head and the gender of the farmer and to determine which household member is actually making the decisions about the crops and inputs.

Wealth measures Measures of wealth are often used on the right-hand side of adoption models. Wealth is expected to affect technology use for a number of reasons, including that wealthier farmers have greater access to resources and may be more able to assume risk. The challenge here is to find measures of wealth that do not also contain substantial information about other factors related to technology use. For example, land holdings are often used to measure farmers’ wealth, but this measure also picks up information about whether there are economies of scale in production using improved technologies. Land holdings may also reflect the social status and prestige associated with owning land, and the ability of a farmer to obtain credit. Many different measures of landholdings may be used, and there is not one measure that is always appropriate. One issue is whether the researcher is trying to control for wealth or trying to consider the availability of suitable land. If the issue is one of trying to measure wealth, then the value of all land should be used. If the researcher is interested in whether the farmer is

20 constrained by the amount of available land, then they need to consider the availability of suitable land, not just that owned or farmed by the particular farmer. Many of the measures that are used are endogenous to the decision being considered, especially the amount of land planted to the crop in question or the proportion of land devoted to this crop.13 These measures may provide information on the correlations between farmers devoting much land to these crops and growing new varieties, but they cannot tell us anything about the causal relations. Related to the issue of land ownership is that of land tenure. Land that is owned may be a good measure of wealth, although whether the land can be capitalized depends on the context. Farmers may have secure tenure to the land without formally owning it. They may also be renters or sharecroppers. The form of tenure may affect the adoption decisions, not only through the wealth effects, but also through the farmer’s willingness to invest in the long term quality of the land. For example, using US data, one study finds that not only are there differences between renters and owner-operators in the adoption of conservation practices, but also that there are differences between cash-renters and share-renters (Soule, et al., 2000). But a study in Haiti finds no relationship between tenure and adoption of agricultural technology. Instead, they find that peasants are more concerned with political and economic insecurity than insecurity of land tenure (Smucker et al. 2000). Similarly, in Ethiopia, Shiferaw and Holden (1998) use a measure of whether the peasant expects to be able to use the land over his or her lifetime. But this measure was not significant in explaining the adoption of land conservation technologies. Thus,

13

For example, in a US study looking at adoption of GM soybeans or corn, both total farm area and the proportion of soybeans or corn as a proportion of total crop output were included as explanatory variables and both were found to be negatively related to the adoption of GM crops. This approach assumes that farmers first decide how much soybeans or corn to plant and then whether or not to plant GM varieties, rather than the decisions being made simultaneously (Hategekimana and Trant, 2002).

21 there are many questions remaining about these relationships, and it is important to consider the many different effects that land may have on adoption. An alternative measure of wealth – livestock ownership – is complicated by the fact that oxen provide draught power as well as manure. Farmers with more livestock may be wealthier and thus more likely to adopt fertilizer. But simultaneously, they may have more access to draught power and less need for inorganic fertilizer. Pitt and Sumodiningrat (1991) note that the positive relationship that they identify between adoption of high-yielding varieties and the value of livestock holdings may be related to the effect of the diversity of income sources on a household’s willingness to take on a riskier investment. Negatu and Parikh (1998) use a measure of the number of oxen per unit of land. In this case it is clear that they are not including the animals as an explicit measure of wealth, but instead as a measure of the level of animal traction available. In some places, farmers may own nonagricultural assets, which may be good indicators of wealth. These may include ownership of a bicycle, TV, radio or other consumer goods. Indicators of housing type (including whether the roof is thatched or corrugated metal) may be useful measures (e.g. Shiferaw and Holden 1998). Although it may be important to include land and livestock in the regression analyses, caution should be used in interpreting the results as simple measures of the effect of wealth.

Access to credit Researchers are often interested in whether farmers have access to cash or credit, because the lack of such access may constrain farmers from using technologies that require initial investments – whether outlays for seeds and fertilizer at the start of the growing season or large

22 cash expenditures for machinery, or investments in infrastructure in fields. The lack of access to cash or credit is often seen as an indication of market failures that government or NGOs should help to resolve. Many adoption studies include a variable that is meant to be a measure of credit availability. The best measure would be whether there is a source of credit available to the farmer. This would mean a source of credit for which the farmer is eligible, at a reasonable cost, both in terms of time and money. However, such a measure is often not available. Instead, many studies ask whether or not the farmer used credit (Boahene, et al. 1999; Negatu and Parikh 1999) or measure the level of credit use. This measure is problematic. Measures of credit use don’t distinguish between farmers who chose not to use available credit and farmers who did not have access to credit. Economic theory tells us that farmers will borrow only if it is profitable to do so – where profitability depends on the price of credit, and the potential returns of investment. By contrast, lending institutions will extend credit most readily where they think it is profitable to do so. Conceptually, the distinction between supply and demand for credit is important if we are trying to determine whether credit market failures are important constraints to technology adoption, but in practice it is difficult to disentangle them. Some creativity can be used to devise an appropriate credit variable. One innovation was to include a measure of whether the farmer had ever received credit (Salasya, et al. 1998). This measure still is not perfect, but it is a better measure of access than the simpler question of whether the farmer used credit in the current period. This measure only works if there have not been major changes in the credit availability in the area during the period covered and if farmers have not changed location or material circumstances. If a credit facility in the area has closed down recently, however, this may be a poor measure of current credit availability, since farmers

23 who had access to credit in the past may no longer have access. Another measure of credit that has been used is the predicted use of credit or the predicted membership in a credit club (Smale et al. 2001; Zeller et al. 1998). Croppenstedt et al. 2003 use the proportion of farmers in the local area in Ethiopia who bought fertilizer on credit in a given year or paid cash by obtaining a loan from a family or friend. Ownership of land is often thought to be a prerequisite for obtaining credit. For example, in Ethiopia, farmers must have at least 0.5 ha under maize in order to participate in the credit scheme for maize. In Kenya, the Seasonal Credit Scheme requires that farmers have at least 5 acres of land. Thus, farmers with smaller amounts of land will not have access to formal credit through these channels. In some circumstances, it may be possible to assume that if any farmer in a village who meets the land requirements obtained credit, then others with similar or greater landholdings also have access to credit. But it is important to consider both the formal rules and how they are applied in practice. In a study conducted in Malawi, Diagne and Zeller (2001) tried to obtain information about the potential for credit by asking farmers if they could borrow money. This captures farmers’ perceptions of whether or not they have access to credit. This perception may be important in determining technology use. However, even here, the availability of credit may depend on its proposed use. For example, farmers may be able to borrow for fertilizer, but not for consumer purchases. Similarly, if we are concerned about whether the lack of cash is a constraint to the adoption of technology, what we really need to know is whether the farmer has access to a source of cash. Again, there are no obvious measures to use. Income is clearly endogenous to the adoption decision. Yet it is often included as a right-hand variable (e.g. Boahene, et al. 1999).

24 One way that researchers have tried to resolve this is by asking whether the household has any nonfarm income (Herath 2003). It is expected that this is less related to the adoption decision than farm income; but the choice to have a household member engage in wage labor or non-farm income-generating activities may be made simultaneously with the decision of which agricultural technology to use. Thirtle et al. (2003) find nonfarm income positive and significant in explaining adoption of GM cotton in KwaZulu-Natal, South Africa. They attribute this finding to both access to cash and to households with nonfarm income being less likely to be risk averse. Smale et al. (2001), use remittances as a measure of exogenous income. While the income of the remitting household may not be correlated with technology use in the farm household, the decision of how much to remit may be. One other possible measure is data on local labor markets. Information as to the depth of the labor market would suggest whether or not households could work off-farm to earn funds to invest in agriculture. In addition, farmers may be able to obtain cash by drawing on savings or selling assets. We would need to know whether the farmer had savings or whether the farmer had assets and access to a market in which to sell them. One study on Malawi used the value of maize stocks from last year’s harvest as the measure of cash availability (Smale et al. 2001). Access to cash, like access to credit, depends upon the use for which the cash is needed. Access to cash or credit is difficult to measure, so it is important to be careful in interpreting measures that are trying to capture this effect.

Access to information Another variable that we may be interested in is access to information. Farmers must have information about new technologies before they can consider adopting them. Since

25 extension services are one important means for farmers to gain information on new technologies, variables about extension services are often used as a measure of access to information. As with measures of credit market functioning, what is usually measured is whether a farmer used the extension service. Studies often consider the number of extension visits received by the farmer (Boahene, et al. 1999, Ouma et al. 2002, Herath 2003), whether or not the farmer received any extension visits in a particular period (Tiruneh et al., 2001; Ransom, 2003), or whether the farmer attended a demonstration field day. They also may control for whether the farmer was a contact farmer or hosted an extension field (Ensermu, 1998). Studies may also separate out contacts with government extension agents from those provided by private firms (Wozniak, 1987). None of these measures captures whether the information was available to the farmer; instead, they indicate both whether the information as available and whether the farmer took advantage of it. Thus, farmers may have had access to information, but chose not to fully obtain or use it. More rarely, an effort is made to look at the effectiveness of extension, for example, by measuring whether the farmers are aware of the relevant recommendations. This measure actually captures both whether the information resources were available and whether the farmer took effective advantage of them. Thus, it may tell us whether farmers who are aware of the technology and understand it are more likely to use it, but it isn’t a measure of access to information. The simultaneity issues were addressed by Bindlish and Evenson (1997) and they still find that in Burkina Faso, areas with more extension services (measured in staff hours) recorded higher crop yields and that the higher yields were associated with participation in extension contact group activities.

26 Other approaches have tried to include measures of the farmers’ perceptions of the problems and the recommended technologies (e.g. Negatu and Parikh, 1999). For example, Shiferaw and Holden (1998) first model whether or not the farmers recognize land erosion as a problem, before analyzing whether the farmers adopt soil conservation technologies. Most of these measures of extension and information should not be interpreted as measures of access to information. As with credit, they instead measure the equilibrium levels of information, where the supply of information and the demand for it by farmers intersect.

Access to labor markets Other variables of interest relate to the availability of labor. Many researchers suggest that labor market failures discourage farmers from adopting improved varieties and fertilizer. The argument is that where labor markets do not function effectively, households must supply their own labor for farm activities, so they may choose not to adopt varieties that would require more labor at any specific time, such as harvest or weeding, than the household can provide. Just as it is hard to measure access to cash or credit, it is difficult to measure a household’s access to labor. The measure that is often included is household size, either measured as “all household members,” “adult household members,” or “adult equivalents.” However, all of these measures are influenced by decisions about agricultural production. Household size, especially when we consider extended households, may depend, at least in part, on the productive capabilities of household. For example, Chipande (1987 ) finds an inverse relationship between the number of female-headed households and the agricultural potential in a region, suggesting that men remain in households in the rural areas when the returns to agriculture are sufficiently high. If the returns to agriculture are low, some members may migrate to towns in search of employment.

27 Similarly, marriage patterns and the formation of new households depend, in part, on the availability of productive land. It may be useful to have descriptive information on the size of households that do and do not use these technologies, but a causal relationship should not be inferred. Another approach that is used is to include a dummy variable indicating whether or not hired labor was used on the farm (Ouma et al. 2003). Lapar (1999) included a measure of local labor exchange groups to capture the availability of labor. Again, it may be useful to know whether the use of hired labor or labor exchange groups is correlated with the use of improved technologies, but they are clearly endogenous to the decision of which varieties and technologies to use. We might expect that the availability of labor in local markets would affect agricultural technology use. When there are local labor markets, farmers can hire labor as needed. Members of farmers’ households may also sell labor to obtain cash as necessary. The relationship between the local labor market conditions and technology use needs to be explored on a case by case basis. The measures widely used in adoption studies are often not adequate to make policy recommendations about labor markets.

Dealing with endogeneity A final issue that arises in using cross-section data from a single study concerns the interpretation of results. Studies that focus on a cross-section of the population and compare adopters to non-adopters cannot be used to analyze the characteristics of farmers at the time of adoption. For example, simply noting that adopters have larger landholdings than non-adopters does not tell us whether those farmers who initially had larger landholdings were more likely to

28 adopt the improved technologies, or the larger landholdings are a result of adopting the technology. We might expect that both would be true, but additional data would be necessary to draw such conclusions. Similarly, comparing farm size of adopters with that of non-adopters does not tell us much about whether new technologies are biased towards large farms. Large farm sizes may be an effect of adoption, not a cause. Similarly, extension may be correlated with technology use, but again the causal relationships may not be clear. Extension agents may identify farmers who are innovators and spend more time with them. Thus, being a likely adopter may result in more extension visits. These examples suggest that studies that examine current farmer and farm characteristics as explanatory variables for the current use of improved technologies should be careful in the interpretation of the results. The results should be interpreted as a correlation between current technology use and the characteristics. They should not be interpreted as saying that farmers with larger farms or more extension visits are more likely to adopt the technology. And they should not be interpreted as saying anything about the changes in farm size and technology use. In contrast, other variables are less likely to be endogenous to technology decisions. For example, in many places, few adults gain formal access to education after they have begun farming. They may start to attend literacy programs or adult education programs, but the level of education reflected in their formal school attendance will probably reflect that at the time of the technology adoption decision. In such cases, an education variable may provide some information about the impact of farmer education on technology adoption as well as on current use. One thing that clearly does not change based on technology use is the gender of the farmer. Thus, we can assume that any farmer who is now male was a male when he adopted the

29 technology. However, we should remember that the farmer who is now farming the land may not have been the one who made the initial decision to adopt the improved technology. For example, a de facto female household head who was not originally involved in the decision to adopt improved varieties may continue to use practices originally initiated by her husband. In most cases, it should be relatively easy to determine whether this is an issue. Thus, knowledge of the context and a good understanding of causal relationships is important for interpreting the results of adoption studies. Many analyses appear to treat variables as exogenous, when in fact they are clearly endogenous. Where endogeneity is systematic, we need to remember that regression analysis can still provide information about correlation, even if it does not provide evidence of causality. It may be useful to learn that the farmers using a technology are wealthier, use credit, and receive extension visits. However, we should not interpret the results as meaning that farmers who are wealthier, use credit and receive extension visits are necessarily more likely to adopt new technologies.

Representative Samples: For micro studies to have broader usefulness, a key issue is to ensure that the samples are appropriately selected. Many studies are located in specific areas, where there is a particular question about technology use. For example, many of the CIMMYT East Africa studies focused on areas where the adoption levels were known or expected to be high. Although deliberately targeting these areas was useful as a first step to show that some areas were in fact using improved varieties and fertilizer, it did not explain why some areas had adopted and others had not. This type of survey design and sample selection method raises a number of questions. Although it is valuable to know how farmers are adopting new technologies in the main centers

30 of production, surveys of this kind do not generate much information about aggregate impacts. For this, we need samples that can be generalized up to some higher level of aggregation. If new varieties or management practices are encouraging the spread of a crop into new areas, especially into marginal areas, these approaches will not capture it. Nor can we be confident that the studies are providing representative information even at a more micro level. In fact, several studies specifically acknowledge that the study areas were not representative – for example, when the sites were chosen for ease of access (Kotu et al., 2000; Beyene, et al. 1998). It seems likely that these areas have relatively high levels of adoption compared with other areas, but there is no way to know this from the data collected. Ideally, samples should be selected in such a way that generalizations can be made about adoption levels for a country or region – or some other aggregate level, such as an administrative district or an agro-ecological zone. Or they should be selected in such a way that generalizations can be made about groups of farmers – such as large-scale farmers, small-scale farmers, commercial farmers, subsistence farmers, male farmers, or female farmers. This may be done through representative samples. In some instances, it may be useful to oversample some areas to obtain enough data on particular regions or farmer categories to be able to obtain statistically significant results about this group. In these cases, the sampling weights should be made available so that it is possible to generalize to a larger scale. Using a sample that is representative of only maize or wheat growing areas is problematic if we are concerned about national-level policy since these areas are likely to expand or contract in response to new technologies and policies. Representative samples will allow the data to be more readily used for impact assessment. There are three major types of impacts that we might be interested in: productivity;

31 poverty and health; and environmental. Most micro studies do not in themselves collect enough data to do impact assessment adequately. Yet, if the data sets are representative, we may be able to use the data, in conjunction with data from other sources, to perform some kinds of impact assessment.

6.

Problems of Designing Studies to Be Pooled Even with one-time, cross-section studies, with some care it is possible to collect data in

such a way that comparisons are possible across study sites and through time. In order to keep open this possibility, however, it is important to exercise considerable forethought in the design of the surveys. If concepts are defined in similar ways and data are recorded in comparable fashion, the data from disparate micro studies can be combined for various types of metaanalysis. This can be particularly useful for analyzing the “big issues” that cannot be addressed within a single micro study. For example, no single micro study can effectively address the impact of government policies or institutions on technology adoption. But a coordinated set of comparable studies might yield information of this kind. If the studies are not designed to be compatible, however, no amount of ex post analysis will be able to get at the larger questions.

Compatibility of definitions and concepts To keep open the possibility of meta-analysis or synthetic analysis, it is important to pursue some degree of compatibility of definitions and concepts across studies. As noted above, for example, it is important to have comparable definitions of technology adoption – so that across different study sites we can compare outcomes effectively. But this issue is not limited to the definition of adoption.

32 Consider the question of how the agricultural potential of an area affects adoption. The influence of agricultural potential can be assessed to some extent within individual microsurveys if information is collected from individual farmers on agricultural potential. A more revealing analysis could come from cross-study comparisons, since there is more variation. To make cross-study comparisons, however, we need comparable measures of potential, such as rainfall levels and patterns and soil type and fertility. We need quantitative measures that can be directly compared, rather than simply a qualitative judgment that one study site has a higher potential than the other. Phrases such as “high potential area” are often used to describe villages in different studies – but the definitions of “high potential” are not easily comparable. A “high potential” maize area in Tanzania may in fact have lower potential than a “moderate potential” site in Kenya. Next, consider a comparable question about market access. We would expect that areas with higher levels of market access would use more improved technologies, since market access is necessary for purchasing inputs and selling outputs. In order to examine this, we need to collect information on access to markets for inputs and for outputs. Some of the information needs to be at the level of individual farmers – how far do they have to go to the nearest local market. In addition, we need to collect information on the distance to the nearest major market. Distance measures should be in miles (or kilometers), time, and cost. Information on other institutions related to market access might also be useful. In particular, information on credit availability and local labor markets may be needed. For credit, it might be important to know whether there are formal credit facilities and where they are located. In addition, we need to know what the requirements are for farmers to obtain credit. (If the practice differs significantly from the rules, then both should be noted.) In addition, if there

33 are informal sources of credit, including savings and credit associations or moneylenders, this information may also be important. Using this information, it may be possible to gain a sense of whether credit is available to farmers in the area. The extent to which there are functioning local labor markets will affect the ability of farmers to obtain labor and to obtain cash for purchasing other inputs. But these variables must be recorded in some fashion that allows comparison across study sites – and preferably across moments in time. Similarly, to address questions about the intensification of agriculture and the adoption of technology, it is useful to have compatible measures of land use and population distribution across study sites. Regional measures of population density do not necessarily tell us about the pressure on agricultural land, since not all of the land may be suitable for agricultural production. It may also be useful to have farmers’ perceptions on whether there are shortages of land or whether additional land is available to expand agricultural production. Considering the high cost of primary data collection, it is important that survey data be fully exploited, not only by those who conducted the research, but also by other researchers who may have additional questions and techniques. In order for this to be possible, it is important to document and store survey data in ways that will facilitate their use by others. This includes making the questionnaires, codebooks, and data available. Without documentation, including the specific wording of questions, some of the data are difficult to interpret.

7.

Conclusion and Recommendations Many of the issues that motivate adoption studies are “big issues” that cannot easily be

studied with small sample studies carried out in geographically limited areas. Issues such as which types of policy and market environments best support the adoption of improved

34 technologies and increased agricultural productivity are important, but these cannot be addressed with small-sample cross-section micro studies. A number of fairly simple procedural changes can, however, both improve the microlevel analyses and make it possible to compile the micro-level studies into broader syntheses that can address big questions, especially those about infrastructure, institutions, and policy. Researchers designing micro-level studies should be careful to pursue sampling approaches that allow the data to be generalized to higher levels of aggregation. With representative sampling, there is more potential to address the questions of interest to policy makers and agricultural researchers. Organizations carrying out portfolios of adoption studies should develop standard formats and definitions, particularly with respect to two aspects of data collection. First, they could design a standardized community survey that reports appropriate data on infrastructure, institutions, and agroecological conditions for each study site – in a format that would be standardized across communities and countries so that comparisons could be made. Second, they could establish a typology for defining technology adoption that incorporates several levels of adoption (e.g., for seed: local varieties, recycled improved varieties, new seed for improved varieties) and allows distinction between full adopters, partial adopters, and those who have disadopted improved technologies after trying them. This would allow for a more nuanced understanding of technology adoption – and again allow for cross-country comparison. Finally, it is important for all researchers involved in adoption studies to re-think the implicit assumption behind most adoption studies – namely, that the so-called “improved technology” is better than the existing technologies – and the corresponding policy recommendation, that farmers need to be convinced to use these new and better technologies.

35 There is some recognition that farmers face constraints, such as the lack of credit, but implicitly most adoption studies assume that the new technologies are better.14 There are three reasons that farmers do not adopt improved technologies. The first is simply that they are not aware of them – or that they are not aware that the technologies would provide benefits for them. Farmers may also have misconceptions about the costs and benefits of the technologies. The second reason is that the technologies are not available, or not available at the times that they would be needed. The third reason is that the technologies are not profitable, given the complex sets of decisions that farmers are making about how to allocate their land and labor across agricultural and non-agricultural activities. Institutional factors, such as the policy environment, affect the availability of inputs and markets for credit and outputs and thus, the profitability of a technology. Simply noting that a farmer has not adopted a “recommended” technology does not necessarily imply that the farmer would be better off if he or she did so. As researchers, we need to understand better the challenges that farmers are facing. We need to focus on the broader issue of how to increase agricultural production – realizing that new technologies may be a key component. Rather than simply asking whether farmers are using improved technologies, we need to be asking them about their levels of production and finding ways to increase it, through improved technologies, improved infrastructure and institutions, and improved policies. To summarize, micro studies of technology adoption may provide valuable information. There are ways in which micro studies can be improved. Ideally, some can be turned into panel studies – or combined with other micro studies to allow for richer cross-section variation.

14

To understand why farmers do not adopt improved varieties, some work has looked at the varietal characteristics to see if consumer preferences were a limiting factor (Hintze et al., 2003; Smale 1995, 2001).

36 Acknowledgements: An earlier version of this paper was written for International Center for Wheat and Maize Improvement (CIMMYT), based on an analysis of CIMMYT’s adoption studies in East Africa. I gratefully acknowledge the input and suggestions from Wilfred Mwangi, Hugo Verkujl, Hugo de Groote, Michael Morris, Melinda Smale, Mauricio Bellon, Douglas Gollin, and Prabhu Pingali as well as the work of the many researchers at CIMMYT and national research centers who participated in the local studies. The paper was revised while the author was visiting at the Centre for the Study of African Economies, Oxford.

References Asfaw, A. and Admassie, A. 2004. The role of education on the adoption of chemical fertilizer under different socioeconomic environments in Ethiopia. Agricultural Economics 30: 215-228. Batz, F.J., Janssen, W., Peters, K.J., 2003. Predicting technology adoption to improve research priority -setting. Agricultural Economics 28, 151-164. Berger, T. 2001. Agent-based spatial models applied to agriculture: a simulation tool for technology diffusion, resource use changes and policy analysis. Agricultural Economics 25, 245-260. Besley, T., Case, A., 1993. Modeling Technology Adoption in Developing Countries. American Economic Review 83 (2), 396- 402. Beyene, H., H. Verkuijl, and W. Mwangi. 1998. Farmers’ seed sources and management of bread wheat in Wolmera Woreda, Ethiopia. Mexico, D.F.: Institute of Agricultural

37 Research (IAR now EARO), and International Maize and Wheat Improvement Center (CIMMYT). p.30. Bindish, V., Evenson, R.E., 1997. The Impact of T&V Extension in Africa: The Experience of Kenya and Burkina Faso. The Word Bank Research Observer 12 (2), 183-201. Bisanda, S., W. Mwangi, H. Verkuijl, A.J. Moshi, and P. Anadajayasekeram. 1998. Adoption of maize production technologies in Southern Highlands of Tanzania. Mexico, D.F.: International Maize and Wheat Improvement Center (CIMMYT), The United Republic of Tanzania, and The Southern African Centre for Cooperation in Agricultural Research (SACCAR). p. 38. Boahene, K., Snijders, T.A.B., Folmer, H., 1999. An Integrated Socioeconomic Analysis of Innovation Adoption: The Case of Hybrid Cocoa in Ghana. Journal of Policy Modeling 21 (2), 167-184. Brush, S.B., Taylor, J.E., Bellon, M.R., 1992. Technology adoption and biodiversity in Andean potato agriculture. Journal of Development Economics 39, 365-387. Chipande, G.H.R.1987. Innovation adoption among female headed households: The case of Malawi. Development 18:315-27. Conley, T.G., Udry, C.R., 2003. Learning About a New Technology: Pineapple in Ghana. Unpublished Paper. Yale University, Economic Growth Center. Croppenstedt, A., Demeke, M., Meschi, M.M., 2003. Technology Adoption in the Presence of Constraints: the Case of Fertilizer Demand in Ethiopia. Review of Development Economics 7(1) 58-70. Degu, G., W. Mwangi, H. Verkuijl, and A. Wondimu. 2000. An Assessment of the Adoption of Seed and Fertilizer Packages and the Role of Credit in Smallholder Maize Production in

38 Sidama and North Omo Zone, Ethiopia. Mexico, D.F.: International Maize and Wheat Improvement Center (CIMMYT) and Ethiopian Agricultural Research Organisation (EARO). p. 24. Diagne,A., and M. Zeller. 2001. Access to credit and its impact on welfare in Malawi. Research Report 116. Washington, D.C.: International Food Policy Research Institute, xiii, 153 Dimara, E., Skuras, D., 2003. Adoption of agricultural innovations as a two-stage partial observability process. Agricultural Economics 28, 187-196. Doss, C.R. 2003.Understanding Farm Level Technology Adoption: Lessons Learned from CIMMYT’s Microsurveys in Eastern Africa. CIMMYT Economics Working Paper 0307. Mexico, D.F.: CIMMYT. Doss, C.R, W. Mwangi, H. Verkuijl, and H. DeGroote. 2003. Adoption of Maize and Wheat Technologies in Eastern Africa: A Synthesis of the Findings of 22 Case Studies. CIMMYT Economics Working Paper 03-07. Mexico, D.F.: CIMMYT. Doss, C.R. and M. Morris. 2001. How does gender affect the adoption of agricultural innovations? The case of improved maize technology in Ghana. Agricultural Economics, 25(1): 27-39. Ensermu, R., Mwangi, W., Verkuijl, H., Hassena, M., Alemayehu, Z., 1998, Farmers wheat seed sources and seed management in Chilalo Arwaja, Ethiopia. Mexico, D.F.: International Maize and Wheat Improvement Center (CIMMYT) and Ethiopian Agricultural Research Organisation (EARO). Evenson, R.E., Gollin, D., 2003 Assessing the Impact of the Green Revolution, 1960 to 2000. Science 300, 758-762.

39 Feder, G., Just, R.E., Zilberman, D., 1985. Adoption of Agricultural Innovations in Developing Countries: A Survey. Econ. Develop. Cultural Change 33(2), 255-298. Foster, A.D., Rosenzweig, M.R., 1996. Technical Change and Human-Capital Returns and Investments: Evidence from the Green Revolution. The American Economic Review, 86 (4), 931-953. Foster, A.D., Rosenzweig, M.R., 1995. Learning by Doing and Learning from Others: Human Capital and Technical Change in Agriculture. The Journal of Political Economy 103 (6) 1176-1209. Gebremedhin, B. and Swinton, S. 2003. Investment in soil conservation in northern Ethiopia: the role of land tenure security and public programs. Agricultural Economics 29: 69-84. Gemeda, A., G. Aboma, H. Verkuijl, and W. Mwangi. 2001. Farmers’ Maize Seed Systems in Western Oromia, Ethiopia. Mexico, D.F.: International Maize and Wheat Improvement Center (CIMMYT) and Ethiopian Agricultural Research Organisation (EARO). p 32. Hailye, A. H. Verkuijl, W. Mwangi, and A. Yalew. 1998. Farmers’ wheat seed sources and seed management in Enebssie area, Ethiopia. Mexico, D.F.: Institute of Agricultural Research (IAR now EARO), and International Maize and Wheat Improvement Center (CIMMYT). p. 37. Hategekimana, B., Trant, M. 2002. Adoption and Diffusion of New Technology in Agriculture: Genetically Modified Corn and Soybeans. Canadian Journal of Agricultural Economics. 50, 357-371. Herath, P.H.M.U., Hiroyuki Takeya. 2003. Factors determining intercropping by rubber smallholders in Sri Lanka: a logit analysis. Agricultural Economics, 29 (3):159-168.

40 Hintze, L.H., Renkow, M., Sain, G. 2003. Variety characteristics and maize adoption in Honduras. Agricultural Economics 29: 307-317. Holloway, G., Shankar, B. and Rahman, S. 2002. Bayesian spatial probit estimation: A primer and an application to HYV rice adoption. Agricultural Economics 27:383-402. Kaliba, A.R.M., H. Verkuijl, W. Mwangi, A.J.T. Mwilawa, P. Anandajayasekeram, and A.J. Moshi. 1998a. Adoption of maize production technologies in Central Tanzania. Mexico, D.F.: International Maize and Wheat Improvement Center (CIMMYT), The United Republic of Tanzania, and The Southern African Centre for Cooperation in Agricultural Research (SACCAR). p. 29. Kaliba, A.R.M., H. Verkuijl, W. Mwangi, D.A. Byamungu, P. Anadajayasekeram, and A.J. Moshi. 1998b. Adoption of maize production technologies in Western Tanzania. Mexico, D.F.: International Maize and Wheat Improvement Center (CIMMYT), The United Republic of Tanzania, and The Southern African Centre for Cooperation in Agricultural Research (SACCAR). p. 40. Kaliba, A.R.M., H. Verkuijl, W. Mwangi, A. Moshi, A. Chilangane, J.S. Kaswende, and P. Anadajayasekeram. 1998c. Adoption of maize production technologies in Eastern Tanzania. Mexico, D.F.: International Maize and Wheat Improvement Center (CIMMYT), The United Republic of Tanzania, and The Southern African Centre for Cooperation in Agricultural Research (SACCAR). p. 40. Katinila, R., H. Verkuijl, W. Mwangi, P. Anadajayasekeram, and A.J. Moshi. 1998. Adoption of maize production technologies in Southern Tanzania. Mexico, D.F.: International Maize and Wheat Improvement Center (CIMMYT), The United Republic of Tanzania,

41 and The Southern African Centre for Cooperation in Agricultural Research (SACCAR). p. 20. Kotu, B. H., H. Verkuijl, W.Mwangi, and D. Tanner. 2000. Adoption of Improved Wheat Technologies in Adaba and Dodola Woredas of the Bale Highlands, Ethiopia. Mexico, D.F.: International Maize and Wheat Improvement Center (CIMMYT) and Ethiopian Agricultural Research Organisation (EARO). p. 26. Lapar, M.L.A., Pandey, S. 1999. Adoption of soil conservation: the case of the Philippine uplands. Agricultural Economics 21:241-256. Leathers, H.D., Smale, M., 1991. A Bayesian Approach to Explaining Sequential Adoption of Components of a Technological Package. Amer. Journal Agr. Econ. Aug. 1991, 734-742. Mafuru, J., R. Kileo, H. Verkuijl, W. Mwangi, P. Anandajayasekeram, A. J. Moshi. 1999. Adoption of maize production technologies in Lake zone of Tanzania. Mexico. D.F.: International Maize and Wheat Improvement Center (CIMMYT), the United Republic of Tanzania, and the Southern African Center for Cooperation in Agricultural Research (SACCAR). p. 39. Mahmoud, C. and Shively, G. 2004. Agricultural diversification andintegrated pest management in Bangladesh. Agricultural Economics 30:187-194. Mussei, A., J. Mwanga, W. Mwangi, H. Verkuijl, R. Mongi, and A. Elanga. 2001. Adoption of Improved Wheat Technologies by Small-Scale Farmers, Southern Highlands, Tanzania. Mexico, D.F.: International Maize and Wheat Improvement Center (CIMMYT) and the United Republic of Tanzania. p. 20.

42 Negatu, W. and Parikh, A. 1999. The impact of perception and other factors on the adoption of agricultural technology in the Moret and Jiru Woreda (district) of Ethiopia. Agricultural Economics 21: 205-216. Nkonya, E., P. Xavery, H. Akonaay, W. Mwangi, P. Anandajayasekeram, H. Verkuijl, D. Martella, and A.J. Moshi. 1998. Adoption of maize production technologies in Northern Tanzania. Mexico, D.F.: International Maize and Wheat Improvement Center (CIMMYT), The United Republic of Tanzania, and The Southern African Centre for Cooperation in Agricultural Research (SACCAR). p. 42. Ouma, James, Festus Murithi, Wilfred Mwangi, Hugo Verkuijl, Macharia Gethi, Hugo De Groote. 2002. Adoption of seed and fertiliser technologies in Embu District, Kenya. Mexico, D.F.: Kenya Agricultural Research Institute (KARI) and International Maize and Wheat Improvement Center (CIMMYT). Pitt, M.M., Sumodiningrat, G., 1991. Risk, schooling and the choice of seed technology in developing countries: a meta-profit function approach. International Economic Review 32 (2) 457-473. Ransom, J.K, Paudyal, K., and Adhikari, K. 2003. Adoption of improved maize varieties in the hills of Nepal. Agricultural Economics 29:399-305. Salasya, B.D.S., W. Mwangi, H. Verkuijl, M.A. Odendo, and J.O. Odenya. 1998. An assessment of adoption of seed and fertilizer packages and the role of credit in smallholder maize production in Kakamega and Vihiga Districts. Mexico, D.F.: Kenya Agricultural Research Institute (KARI) and International Maize and Wheat Improvement Center (CIMMYT). p. 36.

43 Shiferraw, B. and Holden, S.T. 1998. Resource degradation and adoption of land conservation technologies in the Ethiopian Highlands: A case study in Andi Tid, North Shewa. Agricultural Economics 18:233-247. Smale, M., Heisey, P.W., and Leathers, H.D. 1995. Maize of the Ancestors and Modern Varieties: The Microeconomics of High-Yielding Variety Adoption in Malawi. Econ. Develop. Cultural Change 43(2), 351-368. Smale, M., Just, R.E., Leathers, H.D., 2001. Land Allocation in HYV Adoption Models: An Investigation of Alternative Explanations. Amer. J. Agr. Econ. 76, 535-546. Smale, M., Bellon, M.R., Aguirre Gómez, J.A., 2001. Maize Diversity, Variety Attributes, and Farmers’ Choices in Southeastern Guanajuato, Mexico. Econ. Develop. Cultural Change 50(1), 201-225. Smucker, G.R., White, T.A., Bannister, M., 2000. Land tenure and the adoption of agricultural technology in Haiti. CAPRi Working Paper No. 6, CGIAR System-wide Program on Property Rights and Collective Action. International Food Policy Research Institute, Washington, DC. Soule, M.J., Tegene, A., Wiebe, K.D., 2000. Land Tenure and the Adoption of Conservation Practices. Amer. J. Agr. Econ. 82 (4), 993-1005. Staal, S.J., Baltenweck, I., Waithaka, M.M., deWolff, T., Njoroge, L., 2002. Location and uptake: integrated household and GIS analysis of technology adoption and land use, with application to smallholder dairy farms in Kenya. Agricultural Economics 27, 295-315. Thirtle, C., Beyers, L., Ismael, Y., Piesse, J., 2003. Can GM-Technologies Help the Poor? The Impact of Bt Cotton in Makhathini Flats, KwaZulu-Natal. World Development 31 (4) 717-732.

44 Weir, S., Knight, J., 2000. Adoption and Diffusion of Agricultural Innovations in Ethiopia: The Role of Education. CSAE Working Paper WPS2000-5, Center for the Study of African Economies. Oxford University, UK. Wozniak, G.D., 1984. The Adoption of Interrelated Innovations: A Human Capital Approach. The Review of Economics and Statistics 66(1), 70-79. Zeller, M., Diagne, A., and Mataya, C. 1998. Market access by smallholder farmers in Malawi: implications for technology adoption, agricultural productivity and crop income. Agricultural Economics 19: 219-229.