MMSS SENIOR THESIS

The Role of Culture in Agricultural Technology Diffusion in Ghana

Keyoung Lee Advisor: Professor Lori Beaman, Department of Economics 6/6/2011

Acknowledgements I thank Professor Lori Beaman, without whose help this thesis will have been simply inexistent. Her patience, humor, and guidance have given me life and motivation to work on this thesis even during the times I wanted to give up. I extend another special acknowledgement to Jihae John Hwang, with whom I lived my academic career at Northwestern. I am glad that our paths more than just crossed. Of course, I thank all my wonderful friends and family. Without them I would have been inexistent.

2

Abstract This paper examines whether cultural traits affect the diffusion of agricultural technologies in developing countries using data collected in Ghana. Various mechanisms through which cultural traits may affect the social learning process are studied, including network formation and indirect peer effects. I test whether religion and clan membership affect the likelihood of adopting pineapple farming, which is a relatively new technology in the study region. I find that while indirect peer effects might take place within clan membership, it does not occur in church membership. The effect exists even after controlling for information networks and learning methods, but does not persist with fixed effects regressions.

3

Introduction There is a curious phenomenon in developing countries that farming technologies do not diffuse as fast as we often hope. Many new agricultural technologies such as high-yield varieties, which seem to have clear benefits, take a slow adoption process. It has been the focus of some development research to better understand the mechanics behind the adoption process of new farming technologies. Once we better understand the aspects that hasten and factors that slow the adoption process, we can create and seek environments where adoption will occur more quickly. To better understand the adoption mechanics, researchers have asked how social learning and networks play a role in agricultural technology adoption processes in developing countries. Development economists have found that farmers indeed learn from other farmers (Conley and Udry 2010, Munshi 2003, Besley and Case 1994), and that who they talk to determine who they learn from (Romani 2004, Munshi and Myaux 2006), and famers sometimes merely imitate other farmers without having complete information on the new technology (Maertens 2010). Social learning is perhaps the most studied mechanism behind the adoption process as it is an interesting economic question that comes at the heart of rational agents. Although social learning seems to be the most plausible explanation for the diffusion of new technologies, evidence on the extent of its effect is mixed. While most studies find that social learning indeed takes place, some studies show that the effect is minimal and that social learning does not completely drive the adoption process of new agricultural technologies in developing countries. Therefore, it is important to examine other ways technologies spread, which motivates this paper to study the effect of culture, and especially

4

peer effects within cultural groups. In other words, does belonging to a more ―progressive‖ cultural group help farmers adopt more easily? Also, economists have conducted research on how social networks are formed in developing countries, as networks are important in determining how farmers learn from each other. Studies have shown that not only geographical proximity but also cultural similarities are important in determining a farmer’s network (Romani 2003). Thus, culture could be an important factor in the agricultural technological adoption process in developing countries via its role in network formation. Cultural traits may enter into the agricultural technological adoption process in several ways. One way may be through network formation. If we believe that people with similar cultural backgrounds join similar networks, we would expect cultural variables to affect network formation, thus affecting the diffusion of knowledge. The other way would be through indirect effects such as imitation, peer effects, and norm-based diffusion. Imitation describes the process in which farmers adopt simply because others are adopting, not because they are making a cost-benefit decision. Peer effects describe how having farmers that are well behaved (in this case adopting) can influence other farmers within a group. Norm-based diffusion describes the process when a new farming technology is considered as breaking the norm, making adoption desirable only after a threshold number of other farmers adopt it, therefore making the new technology into the new norm. Cultural proximity may also affect access to credit and other agricultural inputs, thus playing a role in the accessibility of new technologies. Despite these reasonable ways cultural factors may affect agricultural technology diffusion, the direct link between cultural traits and agricultural technology adoption has not been explicitly studied. In this paper, I examine the role of two cultural variables—church membership and clan membership—in the adoption of pineapple farming in southern Ghana. 5

I will examine whether belonging to the same church or clan with adopters will enhance the probability of adopting for an individual farmer, controlling for explicit communication networks between farmers. I find that farmers belonging to clans with higher proportion of adopters are more likely to adopt pineapples, although the effect of belonging to religious groups with higher proportion of adopters is not statistically significant. The effect of belonging to clans with higher adoption rates persists even after controlling for information networks and learning methods, but fades with fixed effects.

Literature Review There are three major classes of literature that deal with agricultural technology diffusion: innovation diffusion, network theory, and social learning. All three have contributed to the field of agricultural innovation diffusion among the developing poor. Mathematical modeling using differential equations and dynamic modeling have been used to describe the process of innovation diffusion (Vijay and Peterson 1985). A simple diffusion model using differential equations, however, fails to capture the slow nature of the diffusion process that we see in real life (Young 2009). In addition, simple methods assume that the population is homogenous, which is a strong assumption to make that applies only to very specific populations. Peyton Young (2009) explores different mathematical models of diffusion within a heterogeneous population. He lays out three major models of innovation diffusion: contagion, social influence and social learning. Contagion happens much like how an epidemic spreads—adoption occurs on contact with an adopter. In essence, this is similar to imitation. Social influence is the process in which agents adopt as enough people in his group adopts, which is similar to norm-based diffusion. Lastly, social learning describes the process

6

whereby agents adopt after seeing empirical evidence from other adopters on the technology’s advantages. Young’s major concern is finding the model that best describes the reality that most innovations diffuse quite slowly among populations. All three models are able to generate an adoption curve that is slow at first but turns exponential past a certain point. However, Young argues that social learning is the one that makes the most sense at least in economics: only social learning provides a decision-theoretic modeling of the agents’ behavior. Such diffusion models have been studied across many different fields. Obviously, how innovations spread and how agents adopt new technologies are interesting questions outside developing countries. Ryan and Gross (1943) studies the adoption of hybrid corn in two Iowa communities. Griliches (1957), as a follow-up, looks at hybrid corn adoption in the broader community across Midwestern and Southern farmers. These two studies show that the adoption curve of corn did indeed exhibit a slow initial startup but an explosive growth past a certain point. Still, these results seem incomplete as they only provide a rough sketch of the adoption process. These models all lay out an adoption process that implies that in developing countries, a new agricultural technology will be adopted exponentially past a certain point in time. However, it is unclear when and how these points are determined—in other words, we are still uncertain what exactly drives the adoption process. To explore this question, it is important to understand the various mechanisms through which technological adoption occurs. Is it indeed the case that adoption occurs through social learning? Is it the case that only one of the models that Young lays out is correct? Social learning models are a start, but there is a need to expand the models to incorporate certain characteristics of a population, such as cultural heterogeneity or prevalence of gossip, that my speed up or slow down the adoption process. 7

To answer these questions, we first examine the literature on social learning. Many studies that focus on agricultural technology adoption have focused on the role of social learning, although the topic of social learning in developing countries has been relatively new. Most of these papers attempt to use creative datasets to examine if indeed social learning plays a role in how farmers learn to use new technologies. An important paper in the literature of social learning in developing countries is Conley and Udry’s work on pineapple technology in Ghana (2010). They use data collected in Ghana over a two-year period to determine if social learning exists among farmers and through what mechanism it occurs. One of the important advancements the paper makes is in collecting explicit data on a farmer’s information neighborhood—which is usually conjectured using creative techniques, not measured—as well as credit neighborhood. In doing so, they construct a variable accounting for explicit networks and use it in their model to find that farmers learn about pineapple technology through social learning. However, Conley and Udry arrive at the result only after controlling for religion and clan. They leave open the option that cultural variables may affect the learning process. And indeed, after replicating their results, I find that in some of their specifications their learning estimates lose statistical significance without clan and culture variables. Duflo, Kremer and Robinson (2006) also study the determinants of adoption, but they use field experiments in Western Kenya. They break down the process of social learning by laying out three key ways of learning that may occur in agricultural technology adoption: learning by doing (learning form self), learning by watching (learning through education), and learning from others (social learning). They find a surprising result that although these learning processes do happen, the effects are very small. Meanwhile, other factors such as the inability to save may contribute more to the slowness of the adoption process. This result drives us to study roles of learning mechanisms other than social learning. 8

In addition, most papers that deal with agricultural technology adoption in developing countries seek to understand just if social learning happens, not how it happens. The mechanism through which learning occurs is still little understood among development economists. For example, what characteristics in populations foster learning? Kaivan Munshi (2003) studies the role of heterogeneity in farming conditions on social learning in India. He examines high yield variety adoption process in rice growing areas and wheat growing areas, where rice growing regions are more heterogeneous in growing conditions, such as plot characteristics and crop varieties. He finds the very intuitive result that within a heterogeneous population, learning becomes more difficult as what may work for your neighbor may not directly translate to success for you. Thus, farmers are unwilling to take successes of others as positive signals or something to be imitated in more heterogeneous populations than in homogeneous populations. This result shows that cultural heterogeneity may also have an effect on the adoption process—through differences in work ethics or other attitudes. Another way of studying how characteristics of a population affect the learning process is to study how networks are formed. As mentioned, the most common way of thinking about cultural or other traits affecting the adoption process is to examine how they affect the formation of networks. It is intuitive to think that farmers will talk to other farmers who are similar to them. The mechanism of network formation is important because many papers in the learning literature present results that are sensitive to how networks are defined. Romani (2004) discusses evidence of ethnic factors affecting network formation in Côte d’Ivoire. He finds that indeed what he defines as ―social proximity‖—a factor driven by similarities in ethnic and cultural backgrounds—determines who farmers talk to. Therefore, cultural variation can also affect the adoption process by affecting the network structure:

9

people may be more likely to talk to each other if they go to the same church or belong to the same clan. However, while Romani’s result does confirm the role of cultural traits in network formation, it does not in the context of innovation diffusion. Munshi and Myaux (2005) discuss a form of innovation diffusion within culturally heterogeneous population: fertility transition. With an introduction of new contraceptive technologies in Bangladesh, there has been a trend of dropping fertility rates. Munshi and Myaux show that within the same religion network, people are more likely to adopt new contraceptive technologies when more people in the network are known to have adopted them. It is a norm-based approach to adoption of technology. Thus there is evidence that having more adopters in a culture group can speed up the adoption process for people in the group. Maertens (2010) uses a more comprehensive data to separate social learning, social pressures (norm-based explanation) and simple imitation of adopting new breed of hybrid cotton in India. With the data, she is able to identify specific networks between farmers, as well as the perceived progressiveness of certain farmers who adopted early. She finds that there is indeed social learning—it is important for farmers to know the profit of adoption before they consider adopting. However, she also finds surprisingly strong evidence for simple imitation, where farmers merely copy progressive farmers’ decision to adopt without observing the profit of adopting. In addition, she finds that social pressures may influence farmers to not adopt: as more people in a farmer’s network thinks the new technology is harmful (even when it actually is not), the farmer is pressured not to adopt it. Studying agricultural technology diffusion in developing countries is very sensitive to methodology. Doss (2005) lays out the problems and advantages of using microstudies to analyze technology adoption in developing countries. She highlights the general limitations of such microstudies: lack of dynamics, lack of variation within samples, and sensitivity to 10

the definition of certain variables. How the data that we use is limited will be presented in the Data section. A key paper on the methodology of examining social learning processes is by Manski (1993) on the ―reflection problem.‖ The influential paper posits that outcomes for individuals in the same network may be from common but unobserved shocks rather than from social learning. In essence, it is hard to econometrically separate out endogenous effects within population (social learning) with endogenous selection of one’s networks. Thus, determining the causal effect of network-wide variation becomes nearly impossible. Papers within the social learning framework attempt to deal with the reflection problem using different techniques: Conley and Udry (2010) use directly gathered data on farmers’ networks and spatial GMM standard errors (allowing for heteroskedasticity as a function of physical distance), Foster and Rsenzweig (1995) focus on profitability instead of pure adoption, and Duflo, Kremer and Robinson (2006) try to use experimental data to eliminate common unobservable shocks. Three major conclusions arise from examining these papers. First, although there is evidence for social learning, it is unclear whether social learning plays a major role in the adoption process of technologies in developing countries. It is especially alarming that Duflo, Kremer, and Robinson (2006) which has arguably the cleanest and most comprehensive data, find minimal (but still statistically significant) effects of social learning. Second, networks and network formation influence the social learning process and also are influenced by proximity in cultural and social traits. Lastly, there seems to be some evidence that cultural variables affect the adoption process either through their influence on network formation or through their influence on heterogeneity of a population. However, how these cultural variables affect the adoption process is not well studied, apart from a norm-based explanation studied by Munshi and Myaux. 11

Overall, the role of cultural variables as determinants of agricultural technology adoption has not been studied exclusively by development economists. Therefore, this paper seeks to study the effect of culture on the adoption process of new agricultural technologies while being subject to all the difficulties and limitations of network analysis and the social learning literature. It also seeks to break down the different effects cultural traits can have in the adoption process.

Data The data we use comes from Christopher Udry’s two-year survey (1996-1998) of 180 households drawn from 550 households in southern Ghana, a region that has recently begun adopting intensive pineapple cultivation.1 This dataset was used in Conley and Udry (2010) in determining the role of social learning with fertilizer decisions of pineapple farmers. In this data, I consider pineapple as a new agricultural technology, and examine the role of cultural variables—such as differing church membership and clan membership within villages—in the adoption process. One of the most valuable aspects of this data is that it contains explicit network data. Information network data is constructed using a survey of individual farmers asking them to list other farmers they talk to in general and talk to about farming. However, this information is hard to use as many farmers talk to others who may not have been included in the original sample. This means that we have no other data on those who the farmers talk to. Therefore, we use the alternate network data. This was collected by asking farmers the question ―Have you ever gone to ____ for advice about your farm?‖ where the blank would be filled in from a random sample of seven other farmers within their own village. This ensures that we have 1

The data, as well as the surveys used and detailed description of the survey procedures, are publicly available on Udry's webpage: http://www.econ.yale.edu/~cru2/ghanadata.html.

12

additional data on those who we deem to have an information link with the farmer. We say that there is an information link between two farmers if the either responded ―yes‖ to the question. According to Conley and Udry (2010), this data was reliable compared to the first method they used to collect network data. As mentioned above, pineapple in this region of Ghana is a relatively new technology. Appendix Figure 1 is a replication of Figure 3 in Conley and Udry (2010) that plots the proportion of farmers cultivating pineapple constructed retrospectively using the authors’ survey question ―how many years have you been cultivating pineapples for?‖ We can see that about 46% of the farmers were farming pineapple by the end of the survey in 1997, but only about 20% of the farmers were farming pineapple by 1993, which is three years before the survey, and less than 10% of the farmers had pineapple plots in 1990. The 1993 mark is important as we will use an additional measure of adoption by defining the adoption variable as 1 if the farmer is planting pineapple at the time of the survey or has planted pineapple in the last three years. At that mark, about 20% of the farmers were cultivating pineapples. There are two relevant cultural variables in the data: religion and clan. The religion variable describes the individual’s membership in a particular church denomination within a village. The clan variable denotes membership in an Abusua, a matrilineal clan. Matrilineal clan is considered as a cultural variable as each Abusua has specific cultures and customs in Ghana (Berry 1995). One thing to note is that while an individual can choose their church membership, it is impossible for him to choose his clan membership, which is determined by blood. It should also be noted that most of the literature that deals with the role of culture in social learning deals with how cultural variables affect network formation. It is therefore expected that these two culture variables will heavily affect the farmers’ information

13

networks: that when two farmers are in the same cultural group, the likelihood that they are in the same information network will be higher. Conley and Udry examine the question in their paper. Appendix Table 1 is a replication of Conley and Udry’s Appendix Table 1 with the sample that we will be estimating. The table examines determinants of information links, examining which characteristics between two farmers influence their network formation. It is interesting to see that religion is not a strong factor in the formation of a farmer’s information network while clan membership is. Other factors of interest that drive information links are gender, geographical proximity, age, and traditional office. We see that both parties holding traditional office lowers the chance of two being in the same information neighborhood. Being the same gender raises the likelihood for two farmers to be in the same neighborhood considerably. Quite intuitively, being in a similar age group and geographical proximity also raise the likelihood of two farmers talking to each other about farming. Surprisingly, an increase in the wealth difference increases the likelihood of network formation. However, as Conley and Udry describe, the strong vertical patron-client ties between the wealthy and the less wealthy might be driving this effect. The data also contains soil type and fertility data that covers organic material and pH levels of individuals’ plots. These two aspects will allow us to control for more specific individual characteristics than with other datasets. In addition, there is data on the value of all crops that the farmers in the sample harvested, including pineapples. We can use this data to construct profits for pineapple farmers and construct average profits of farmers within each culture group. One word of caution is that there are no farmers that planted pineapples during the two-year period of the survey in Village 4. Although there are pineapples farmers in the 14

village who have farmed pineapples for a long time, this is alarming. Another is that we are only using the cross-sectional variation in the data. This makes us subject to one of the limitations laid out by Doss (2006): using cross-sectional data to study adoption processes is limited due to lack of dynamics. Therefore, it is hard to determine the before-after-effects of adopting the new technology, as well as the specific characteristics of farmers at the time of the adoption. Because technology decisions are inherently dynamic, this poses a problem in the study of adoption. We now describe our sample selection. We first start out with 441 individuals with family and cultural background information. Out of the 441 individuals, 397 individuals have data on years they have been farming pineapple, dropping 74 individuals. We additionally drop 74 individuals for whom religion or clan data are missing, and 46 individuals for whom plot data (area and soil data) are missing. That leaves us with 277 individuals. However, village-religion and village-church groups with less than 4 individuals are dropped, dropping at total of 46 individuals, leaving us with 231 total farmers in the sample. It is also important to define relevant variables. Following Doss (2006), we define adoption variable as a binary variable since pineapples are a new crop, not a variety of local crop such as high-yield varieties. We define plant as a binary variable with a value of 1 for those who have ever planted pineapple, using the data of years planting pineapples. Within the two years of the survey, 8 farmers started planting pineapples for the first time; they were added to the total number of farmers who have ever planted pineapples. It is important to note that the construction of the adoption variable includes those who have been planting pineapples for various years. We here make the crucial assumption that networks and in particular cultural aspects such as clan and church memberships have not varied much over the years. However, we relax this assumption later by using a more contemporaneous

15

definition of adoption, albeit with some caveats. We will see if the contemporaneous definition of adoption affects the estimates. Lastly, we present summary statistics in Appendix Table 2. We see that around 52.4% of farmers in our sample at the time of the survey have ever farmed pineapple. However, a lot of these farmers who have been farming pineapple for a time were not planting pineapples at the time of the survey or during the three-year period prior to the survey. Only 31.6% of all farmers were farming pineapples at the time of the survey or three years prior to the survey. We see that the average percentage of a farmer’s village-religion group planting pineapples is 46.96%, and for village-clan it is 48.7%. On average, individuals have been farming pineapples for 2.6 years and have an average plot of 5.6 acres. We also see that farmers talk to only a small proportion of other farmers in their same religion or clan groups about farming.2

Empirical Strategy In estimating the role of cultural aspects on diffusion, we want to see if there exists some mechanism within culture groups that enhances a farmer’s likelihood to adopt. In the most basic specification, we simply regress the adoption decision on a culture-group specific variable. The most basic framework that we use is the following: (E1) where X denotes a vector of controls. Adopt is a binary variable with 1 meaning that the farmer has ever planted pineapple and 0 otherwise. AP is the percent of farmers that adopted pineapples within a group, with rel subscript for variation in the village-religion level and the clan subscript for variation in the village-clan level. The i subscript denotes the individual 2

We here note that the currency of Ghana (the cedi) have been reintroduced to equal 10,000 old cedis in July 2007. Thus, at the time of the survey, a million cedis are now 100 cedis.

16

farmer; the G subscript denotes the village-religion group and g denotes the village-clan group to which the individual belongs; the -i subscript denotes the fact that the variables are constructed excluding oneself—therefore, the subscript denotes his peers. The coefficients of interest, then, are β1 and β2. A simple interpretation of the coefficient says that a 10% increase in the proportion of adopters in the group is associated with an increase in the likelihood to plant by

%.

We must be careful in interpreting the coefficients as there is a possibility of endogeneity in the model. Our main variables (AP variables) might be absorbing the effect of differences in attitudes and behavior that are common within one culture that are affecting the likelihood to adopt. For example, a particular group may have a particular characteristic within itself that raises all farmers’ likelihood to adopt. To adopt a new technology one must have a sufficient capital base. Therefore, wealthier farmers will adopt pineapples more readily. It is possible that farmers in a certain group are wealthier on average than farmers in another group. Then, there is an endogeneity problem, where unobservable characteristics that are correlated with the independent variables are driving our coefficient of interest (Manski 1993). Of course, as long as these characteristics are observed, they can be included in the regression. We have included the wealth variable to control for the specific instance of wealth. However, other characteristics, such as receptiveness to new technologies, are not observable and will pose problems. The direction of the bias is ambiguous, however, as we cannot conclusively say what kinds of attitudes will lead to higher adoption rates. In addition, it is possible that similar people might be more inclined to choose to be in same cultural groups. For matrilineal clan, determined by bloodline, this may not be a problem. However, it may be possible that church membership may be determined by similarities in characteristics. Moreover, even observed characteristics suffer from measurement error. 17

It is possible to partially address these problems by using fixed-effects regressions. If we use fixed effects within groups, then we can minimize omitted variable bias by eliminating all aspects that are the same between individuals within the same group. Thus, fixed effects will minimize confounding effects that can significantly skew the estimate. Moreover, we can use fixed effects to eliminate of group-wide shocks that might drive the estimates in a certain direction. It is especially important to utilize fixed effects within villages, as it is possible that a certain village may have village-wide shocks in growing conditions that need to be accounted for. However, there are problems with using fixed effects regressions. First, we need to be careful of what aspects of similarities within groups we are eliminating. It is true that by using fixed effects we may be able to find evidence of peer effects—that more adopters within a group leads to a higher likelihood to adopt. However, when we are examining culture, it is possible that certain confounding effects are of interest to us. For example, being in the same church probably fosters similar attitudes by listening to the same sermons and receiving similar pastoral advice. If we find that across different groups, farmers indeed have differing attitudes, this might be a mechanism through which culture affects the spread of innovations. Thus, it may be necessary to consider these effects that may be taken away using fixed effects. Of course, the problem with not using fixed effects is that although we may estimate the role of attitudes in cultural factors, other confounding effects are still in play. Also, it must be noted that after using fixed effects, there may not be sufficient variation in the adoption rate. For example, if there is a particular religion that is represented only in one village and that religion has a particularly high adoption rate, using village fixed effects might eliminate effects driven by that particular religion. Therefore, using religion or clan fixed effects on top of village fixed effects can sap out almost any variation left to be

18

estimated. Although this is not a problem, it will be hard to obtain reasonably large estimates due to the lack of variation. For these reasons, it is important to interpret regressions without fixed effects and with fixed effects carefully. We cannot conclude that the fixed effects model is superior in estimating what our model attempts to find, mostly because cultural traits are immeasurable using simple metrics. One solution would be to find a proxy variable for some of the attitude differences we might expect. Moreover, fixed effects cannot eliminate all the omitted variable bias in the model. Using the village fixed effects allows us to eliminate all characteristics that are identical within villages. However, any unobservable characteristics that are not the same will not be eliminated, and we do not anticipate many of the characteristics will be identical. Especially with the lack of historical data, we cannot account for changing observable characteristics over time. For clan fixed effects, we need to assume that the relationship between unobservable characteristics and clan are same across all villages. This assumption may not be valid, in which case the clan fixed effects will not be able to remove the unobservable characteristics we desire to remove. There is also a technical issue with calculating standard errors. Using a simple OLS model poses a problem in the standard errors due to the way the percentage variable is constructed. The percentage variables for village-religion and village-clan groups are constructed so that every farmer belonging to a certain village-religion or village-clan group has virtually the same value. This violates the assumption of independence across observations, leading to miscalculated standard errors. We can try to correct this by using robust clustered standard errors. Using clustering, variation is allowed at the group level, not at the individual level. The problem with this approach is that robustness holds with clustering only asymptotically. At the minimum, we would need at least 30 clusters to ensure 19

limited robustness (Bertrand, Duflo, and Mullainathan 2003) However, our data does not have the number of clusters that conventionally yield robust standard errors and this seems to be driving the standard errors in the wrong direction, making our estimates falsely significant. To account for the small number of clusters, it is possible to use block bootstrapping to get robust standard errors. Block bootstrapping standard errors are obtained by re-sampling the estimation sample on the group level multiple times and averaging the standard errors calculated from re-sampling. We use the customary number of replications—1,000—to resample. According to Bertrand, Duflo, and Mullainathan (2003), block bootstrapping performs better with more number of clusters and thus our small number of clusters may still pose a problem. Still, it is an alternative to using simple clustering standard errors. Finally, estimating β1 and β2 represent the effect of having more pineapple farmers who have adopted in a farmer’s group. Therefore, it is closest to estimating peer effects or imitation effects. However, there are other effects such as learning and network formation that affect the learning process that are not included in the model. In the specification presented in (E1), those processes are omitted in the regression. These omitted variables are most likely correlated with the peer effects, as each effect drives each other. We have seen that in the case of clan variable, being in the same clan greatly increases the likelihood for two farmers to be in the same information neighborhood. Thus, omitting network formation effects will bias the coefficients of interest in (E1).

Accounting for the Learning Mechanism Therefore we turn to examining the full mechanism behind cultural effects on diffusion. The simple specification used in (E1) cannot separate out different effects of culture on the likelihood to adopt. In essence, in the estimated effect, we might be including different factors of learning such as social learning. Therefore, there is a need to include other 20

variables. In a more detailed specification, we include two additional variables of interest— the learning variable and the network variable—which will be explained below. Before going further, it is important to think about the mechanism through which technological diffusion process might occur. Just as in Hanushek et al. (2002), we need to first define how culture might affect diffusion before we proceed to estimating culture’s overall effect. As we have seen in the literature that exists in this field, the primary methods of diffusion seem to be social learning, network interactions, and other indirect effects such as simple imitation or social pressures (or further, norm-based explanation, depending on the nature of the interactions). We can therefore assume that cultural effects happen in the way that includes all of these factors. In other words, Cultural effects = social learning + direct interaction effects + indirect effects. Therefore, it is desirable to separate these effects out when running the regression. In the simplest specification, this is not possible, and thus the estimate is likely to be overestimated if we want to measure just the indirect effects (which includes imitation) of having more adopters in a farmer’s culture group inducing him to adopt pineapples himself. Thus, we introduce another specification that breaks down the raw cultural effects into three effects that we defined above: (E2)

where Learn is a variable that contributes to social learning, such as observable profits of farming pineapples, and Network is a variable that allows us to control for direct information links between farmers. There are two challenges to this framework. First, the variable Learn cannot be directly included as social learning cannot be directly measured. Second, it is unclear how to construct the network variable. Although we have data on network interactions, it is hard to 21

construct an accurate measure of the extent of interactions between two farmers within his culture group. We will try to address the first problem by using different proxies for Learning and the second by constructing different measures of network interactions. One of the specifications is as follows:

where

is the average pineapple farming profit of the farmer’s culture group during

the survey, talk is a binary variable that equals 1 when a farmer talks to at least one other farmer within his culture group: religion or clan. The interpretation for coefficients β1 and β2 are the same as in the first specification (E1). β3 is the effect of the rise in mean profit within the farmer’s culture group and β4 and β5 are the effect of an increased proportion of talking to other farmers in religion and clan groups, respectively. In this framework, average profit is an imperfect proxy for the existence of social learning. While we do believe that most of social learning comes in the context of Bayesian updating of profitability (Besley and Case 1994, Conley and Udry 2010), average profits alone cannot measure the whole extent of social learning. Similarly, the talk variables constructed will not be a perfect measure of network interactions. This stems first from the data, that not all respondents were asked about farmers in their own clan and own religion, and second from the fact that merely including a simple variable about talking to other farmers cannot generate the full network effect. Another problem is that we only have current data on characteristics while we are using the historical data on the adoption decision. This was briefly discussed in the Data section referencing Doss (2006). Our regressions are in essence examining what current

22

characteristics, such as networks, church membership, and other control variables, have on the historical adoption of all farmers, since we have defined our adoption variable to be equal to 1 when the years of farming pineapple of a particular farmer has been greater than 0. In the sample, there are many farmers who have been planting pineapples for more than 6 years. It is probably the case that the decision to adopt must have been driven by some historical effects, not only current effects. As we are constructing the adoption percentage variables with the adoption variable, these historical effects will also be correlated with our variables of interest, causing omitted variable bias. The fact that the data is not a panel dataset makes it impossible to remove historical data from the error term using fixed effects (as they do in Hanushek et al. 2002). We will therefore have to try different specifications to see how the change in proxy variables and the dependent variable affect our estimates. To address this problem, we can redefine Adopt to only include those who are farming pineapple now, those who farmed pineapple at least once in the past three years, or those who are farming pineapple now and those who farmed pineapple at least once in the past three years. However, farming pineapples now or having farmed pineapple in the last three years is not equivalent to adoption at that particular time—there are indeed cases in the data where one farmed for a long time but has not farmed for the past three years. It would be possible to consider the famers who have been farming but not farming now as those who dis-adopted. But this is a rather strong assumption to make. There are farmers who have not been farming for the past three years who are farming now, but have farmed for longer than three years. This fact shows that just because a farmer is not farming now does not necessarily mean he has dis-adopted. Still, we can use these different measures to try to conduct robustness tests of seeing if different time-period pineapple farming decision variables affect the strength of the estimates.

23

Analysis We consider the percentages of pineapple farmers in each village-religion and village-clan groups. In a sense, we are considering a peer effects or imitations approach to the diffusion of pineapples: the hypothesis is that the higher the proportion of pineapple farmers in each group excluding oneself, the higher the probability of other farmers in the group adopting. A simple regression model with just the percentage variables is presented in Table 1. The regression results show that there is indeed an effect of having a higher proportion of farmers in a group on adoption rates. The coefficient on the religion percentage variable indicates that a 10% increase in the proportion of adopters in your group will raise your likelihood to adopt pineapples by 6.15%. The effect with clan is higher than the effect with religion. A 10% increase in the proportion of adopters in your clan group raises your likelihood to adopt by 6.8%. Both of these estimates are statistically significant at the 1% level, with block bootstrap standard errors. The group at which the block bootstrapping was performed was village-religion in the first column regression and village-clan in the second column regression. Obviously, although we can use these regression results to see if there indeed is some correlation between our dependent and independent variables, we cannot use the results as final. We want to estimate equation (E1) in the Empirical Strategy section by adding control variables. We run two specifications, both of which the results are presented in Table 2. Both specifications include the individual farmer’s wealth, average plot area (as he may own more than one plot), financial assets, farming stock, and gender with 1 being female and 0 being male. The difference between the two specifications is that the second specification adds the soil characteristic variable of a farmer’s average plot (pH and soil organic matter), and an

24

indicator variable for the farmer holding traditional office. We again use block bootstrapping: but only for columns 1 and 3 on the village-clan level as village-clan seems to have consistently more significant results than village-religion. Including control variables decreases the magnitude and the significance of the religion percentage variable for both specifications. A 10% increase correlates to about 0.50.9% rise in the likelihood to plant. On the other hand, clan percentage variable stays significant, with the magnitude of the estimate similar to the result in Table 1—a 10% increase correlates to about 6.5% rise in the likelihood to plant. We also see that in the more specified model, wealth, financial assets, gender, and soil organic matter do play a role in the adoption of pineapples. It is reasonable that as you are more wealthy, you are more likely to adopt new technologies because you have the capital base to do so. A one million cedis (about $60) increase in wealth will lead to about 10.7% increase in the likelihood to adopt, which is fairly low. Curiously, financial assets and soil organic matter both have negative and significant coefficients. We can perhaps think that having a financial asset ties you down to your financial investment, therefore making it harder to invest in new agricultural technologies. Also, having a higher percentage of soil organic matter, which makes the soil more fertile, may induce farmers to plant less cash crops (which pineapples are) and more food and staple crops. Another variable that has a negative impact on adoption is gender. As is the case in many developing countries, women are restricted from many resources, including land rights in the study region (Goldstein and Udry 2008). This will prevent them from adopting new technologies. Gender has a strong role in adoption, as being a woman decreases your likelihood to adopt by about 33%. Overall, the estimate on clan percentage is strong, but this estimate may be driven more by the fact that clan membership affects network connections much more than church 25

membership. As we see in Appendix Table 1, while belonging to the same church group does not have a significant effect on the likelihood for two farmers to talk about farming, belonging to the same clan does. As we do not explicitly include the network variable in this model, it is possible that the clan percentage variable is soaking up the effect of network formation. We address this problem by estimating (E2) using different proxy variables for Learn and Network.

Controlling for Other Mechanisms: Learning and Network The regression results from estimating (E2) are presented in Table 3. We proxy network by inserting the variable signifying how much a farmer talks to another farmer about farming within his own group, using the data on networks described in the Data section. This variable is constructed in terms of proportion: we first take how many people of the same group the respondent was asked about, and then consider out of those how many farmers the respondent talk to about farming. We proxy for learning in two different ways. The first is to use simple averages of group profits, which are the results in column (1) and (2). If we believe that people casually observe others’ profits and that there is some mechanism of indirect spread of information within groups, these variables will proxy for learning. The second is to interact these averages with whether farmers talk to pineapple farmers in their same group. This seems more reasonable as we would expect only those who talk to each other will receive direct signals on the profitability of pineapples. We use both regular and block bootstrap standard errors to preserve standard errors for non-grouped terms. From the results in Table 3, we see that when we merely include simple averages, the clan percentage estimate still is significant, although with decreased effects: about 5.2% per 10% rise in proportion of pineapple farmers in your village-clan group. We also see the religion profit average is statistically significant even at the 5% level when block 26

bootstrapping is used, while clan profit average is not significant. The estimates on the profit averages suggest that a rise in one million cedis of average religion group profit raises the likelihood to adopt by about 6.8%. The regression using the second set of learning proxy variables returns a different result. We discussed that interacting these average profits with the fact that the farmer actually talks to their peers in the group about farming may be a better proxy in the Empirical Strategy section. We see that with this specification, although the learning variables themselves return no statistically significant result, the clan percentage variable maintains its significance even with the block bootstrap standard errors: a rise of 10% in clan percentage variable raises the likelihood to plant by 6.75%, significant at the 1% level. We see that wealth, gender and soil organic matter variables maintain their significance and signs. In fact, the estimates are nearly identical to the estimates presented in Table 2. However, we see that the proxy variable for networks do not have statistically significant results across both specifications. Moreover, with a different specification of learning proxy variables, the signs on the network variables flip (although with no statistical significance), providing us with an unreliable measurement. Of course, this does not mean that learning or direct communication effect in general does not occur. As we have seen in Conley and Udry (2010), there is evidence of social learning among pineapple farmers. These results are merely a reflection of adding the learning and network variables in as a control, not as a strategy to estimate the existence of social learning or effects of direct communication links.

Fixed Effects Fixed effects regressions are implemented to tackle the issue of unobservables within farmer groups. The results are presented in Table 4. Columns (1) and (2) are same 27

specification as (E2) with village fixed effects, the former with regular standard errors and the latter with block bootstrap standard errors. Columns (3) and (4) are with the clan fixedeffects, (3) with regular standard errors and (4) with block bootstrap standard errors. We first examine village fixed effects. When we use village fixed effects, we see that the sign of the religion percentage variable flips, but is not statistically significant with block bootstrap standard errors. We also see that the magnitude of the clan percentage variable falls to about 2.4%. Moreover, the clan percentage variable loses statistical significance with the fixed effects regression. We still see that gender and wealth have the same direction and significance as the non-fixed effects regressions. With the clan fixed effects, the magnitude of the clan percentage decreases even more, to about 1.7%, still with statistically insignificant results. In addition, the learning variables and network variables also lose significance, except average religion group profit. These fixed effects regression results may show that most of the effect of the clan percentage variable came from village-clan specific shocks or attitudes. This means that in terms of pure imitation or peer effect (indirect effects), culture groups may not be significant in spreading agricultural technologies. However, as stated in the Empirical Strategy section, when we use fixed effects we are eliminating culture-specific attitudes from our estimates and reducing variation overall. Although we cannot make a strong conclusion, the results of the fixed effects regressions may suggest that cultural differences in ―unobservables‖ may indeed drive differences in adoption decision among individuals.

Alternate Dependent Variable Lastly, as we have discussed in the Empirical Strategy section, the definition of the dependent variable for adoption may be problematic due to its historical nature. We have classified people as having adopted if they have only been farming pineapples for 1 year or if 28

they have been farming pineapples for 25 years. In reality, in order to more closely examine the process of adoption, we need contemporary data at the time of each farmer’s adoption. Since we do not have that data, we resort to using a different definition of the dependent variable: a binary variable equal to 1 if the farmer is either farming pineapple now or have farmed at least for one year in the last 3 years. The regression results using this definition of the adopt variable is presented in Table 5. Column (1) is OLS regression with regular standard errors, (2) is OLS with block bootstrap standard errors, (3) and (4) are village and clan fixed effects, respectively, with block bootstrap standard errors. What we want to see from these results is whether the estimates on our variables of interest change significantly. The results indicate that statistical significance does not depend on the definition of our dependent variable. Yet, the magnitudes of the estimates seem to change significantly. The OLS estimate on clan effect goes from 0.52 to 0.28 with the alternate dependent variable, while the village fixed effects estimate goes from 0.24 to 0.03 and the clan fixed effects estimate goes from 0.17 to 0.034. This is an imperfect ―robustness test‖ to see whether contemporary or historical characteristics may affect our estimates. The result seems to indicate that changing the dependent variable to current farming greatly changes the result of our estimates. Therefore, this is an indication that the lack of historical data and cumulative effects in the adoption decision process may be a big problem our data cannot address.

Conclusion This paper examined the role of culture and its effect on agricultural technology diffusion by looking at two aspects of Ghanaian culture: church membership and matrilineal clan membership. We wanted to examine whether cultural variables may affect the adoption

29

of new agricultural technologies and if so, by which mechanism the technology diffusion process occurs in the context of culture. Through our literature review, we saw that there is little work done on the direct link between culture and the adoption process. Through our analysis, we found that having more adopters in your cultural group raises your likelihood to adopt pineapple. This effect persisted with more specified models controlling for other learning mechanisms such as learning and network information flows. We examined different mechanisms of innovation diffusion, breaking down the cultural effects into three different categories: indirect effects such as imitation, social learning from profits, and social learning from networks. We utilized the advantage of the Udry dataset of having explicit network data to (albeit imperfectly) separate out these effects. We found that social learning and networks may not have a significant effect in the adoption of new technologies in the context of culture, while indirect effects including imitation seem to have a much more significant effect. This finding comes in line with Maertens (2010) and Munshi and Myaux (2005). However, our model estimates seem to depend greatly on the kinds of proxies we used for learning and networking variables. This suggests direction for further research using a more comprehensive dataset that includes cultural aspects and learning aspects in regions proven to exhibit the social learning process. This would take creative ways to measure learning and networks so that we may quantitatively include these aspects in the estimation instead of using imperfect proxy variables. Moreover, the effects of having more adopters in your cultural group seem to subside with the fixed effects specification. This may suggest that, although a pure imitation effect within cultural groups may not be strong, cultural differences in attitudes and unobservable characteristics may play a role in the adoption process. Although we cannot make strong conclusions from this finding, we leave the possibility open. Further research will need to 30

quantify the unobservable characteristics of different cultural groups, perhaps by using innovative measurement techniques or lab and field experiments. This will allow researchers to directly include differing attitudes of cultural groups into the regression, thereby avoiding omitted variable bias. To examine this question further, it would be necessary to examine locations with more cultural heterogeneity, such as the region in Côte d’Ivoire that Romani (2004) studies. It would also be necessary to collect longitudinal survey data gathering historical information on the adoption process. It is nearly impossible to fully study adoption if we merely have cross-sectional data, as illustrated by Doss (2006). It is necessary to see when, how, and why one decided to adopt and the characteristics at the moment of adoption—as well as the before-and-after observations of adoption that can give us insight into the social learning process. With such data, we can avoid ignoring historical characteristics of adoption that cross-sectional data cannot examine. Through this, if we can truly find the mechanism behind agricultural technology diffusion in developing countries, we can find effective ways to spread new and efficient technologies to the poorest parts of the world.

31

References Bertrand, Marianne, Esther Duflo and Sendhil Mullainathan, 2004. ―How Much Should We Trust Differences-in-Differences Estimates?‖ The Quarterly Journal of Economics, 119(1): 249-275. Berry, LaVerle (ed.). 1995. Ghana: a Country Study. Washington, D.C.: Federal Research Division, Library of Congress. Besley, Timothy J., and Anne Case. 1994. ―Diffusion as a Learning Process: Evidence from HYV Cotton.‖ Princeton Woodrow Wilson School Development Studies Working Paper 174. Conley, Timothy G., and Christopher R. Udry. 2010. ―Learning about a New Technology: Pineapple in Ghana." American Economic Review, 100(1): 35–69. Doss, Cheryl Ro. 2005. ―Analyzing technology adoption using microstudies: limitations, challenges, and opportunities for improvement. Agricultural Economics, 34: 207-19. Duflo, Esther, Michael Kremer, and Jonathan Robinson. 2006. ―Understanding Technology Adoption: Fertilizer in Western Kenya, Evidence from Field Experiments.‖ Unpublished. Foster, Andrew D., and Mark R. Rosenzweig. 1995. ―Learning by Doing and Learning from Others: Human Capital and Technical Change in Agriculture.‖ Journal of Political Economy, 103(6): 1176-1209. Goldstein, Markus and Christopher Udry. 2008. “The Profits of Power: Land Rights and Agricultural Investment in Ghana.” The Journal of Political Economy. 116(6): 9811022 Griliches, Zvi. 1957. ―Hybrid Corn: An Exploration of the Economics of Technological Change.‖ Econometrica, 25: 501-22.

32

Hanushek, Eric

A., John F. Kain, Jacob M. Markman and Steven G. Rivkin. 2003. ―Does

Peer Ability Affect Student Achievement?‖ Journal of Applied Econometrics, 18(5): 527-544. Maertens, Annemie. 2010. ―Who Cares What Others Think (or Do)? Social Learning, Social Pressures and Imitation in Cotton Farming in India.‖ Unpublished. Mahajan, Vijay, and Robert A. Peterson. 1985. Models for Innovation Diffusion. Sage University Paper Series on Quantitative Applications in the Social Sciences, 07-048. Newbury Park, CA: Sage. Manski, Charles F. 1993. ―Identification of Endogenous Social Effects: The Reflection Problem.‖ Review of Economic Studies, 60(3): 531-42. Munshi, Kaivan. 2003. ―Social learning in a heterogeneous population: technology diffusion in the Indian Green Revolution.‖ Journal of Development Economics, 73(1): 185-213. Munshi, Kaivan, and Jacques Myaux. 2006. ―Social Norms and the Fertility Transition.‖ Journal of Development Economics, 80(1): 1-38 Romani, Mattia. 2004. ―Love Thy Neighbor? Evidence from Ethnic Discrimination in Information Sharing within Villages in Côte d’Ivoire.‖ Journal of African Economies, 12(4): 533-63. Ryan, Bryce, and Neal C. Gross. 1943. ―The Diffusion of Hybrid Corn in Two Iowa Communities.‖ Rural Sociology, 8(1): 15-24. Young, Peyton. 2009. "Innovation Diffusion in Heterogeneous Populations: Contagion, Social Influence, and Social Learning," American Economic Review, 99(5): 1899-1924.

33

Regression Tables

Table 1: Simple Regression Models (Unspecified)

Religion percentage

(1)

(2)

plant

plant

.615*** (.172)

Clan percentage

.679*** (.129)

Constant

.235**

.193**

(.0981)

(.0755)

Observations

231

231

R-squared

0.053

0.078

p < 0.1 p < 0.05 p < 0.01. Block Bootstrap standard errors with 1000 replications in parentheses.

Table 2: Simple Regression Models (Specified) (1)

(2)

(3)

plant

plant

plant

.0877

.05

.05

(.282)

(.211)

(.324)

.724***

.654***

.654***

(.197)

(.192)

(.21)

.108**

.107**

.107**

(.0514)

(.0425)

(.0469)

-.00509

-.00476

-.00476

(.0148)

(.00908)

(.0141)

-.153

-.149**

-.149

(.181)

(.0615)

(.228)

-.0017

-.0277

-.0277

(.437)

(.223)

(.461)

-.334***

-.33***

-.33***

(.0656)

(.0657)

(.0589)

.0338

.0338

(.0581)

(.0732)

-.0798*

-.0798**

(.0409)

(.0346)

.0476

.0476

(.0856)

(.0872)

.246

.329

.329

(.171)

(.406)

(.562)

Observations

231

231

231

R-squared

0.244

0.258

0.258

Religion percentage Clan percentage Wealth (million cedis) Mean plot area Financial asset (million cedis) Farming stock (million cedis) Gender (=1 if female) pH Soil organic matter (%) 1 if resp holds trad. office Constant

p < 0.1 p < 0.05 p < 0.01. Block Bootstrap standard errors with 1000 replications in parentheses for column (1) and (3). Regular standard errors in column (2).

Table 3: Network and Learning Variables Added (1)

(2)

(3)

(4)

plant

plant

plant

plant

-.36

-.36

-.0473

-.0473

(.257)

(.378)

(.221)

(.347)

.522**

.522**

.675***

.675***

(.202)

(.244)

(.194)

(.232)

Average Religion Profit * Talk to:

-.0282

-.0282

Pineapple farmers (million cedis)

(.0502)

(.0722)

Average Clan Profit * Talk to: Pineapple farmers

.0577

.0577

(million cedis)

(.0429)

(.48)

Religion percentage Clan percentage Average religion group profit

.0677*** .0677**

(million cedis)

(.0239)

(.029)

Average clan group profit (million cedis)

-.00863

-.00863

(.0129)

(.0192)

Talk to: Religion

.291

.291

.326

.326

(.187)

(.285)

(.233)

(.355)

-.101

-.101

-.283

-.283

(.179)

(.259)

(.228)

(.442)

.107**

.107*

.105**

.105*

(.0427)

(.0578)

(.0434)

(.0553)

-.00622

-.00622

-.00383

-.00383

(.009)

(.0149)

(.00913) (.0147)

-.15**

-.15

-.143**

-.143

(.0618)

(.184)

(.0625)

(.16)

-.06

-.06

-.0368

-.0368

(.222)

(.431)

(.226)

(.424)

-.311***

-.311***

-.31***

-.31***

(.0671)

(.0659)

(.0682)

(.0702)

-.00493

-.00493

.0322

.0322

(.06)

(.0786)

(.0595)

(.0722)

-.0694*

-.0694*

-.0765*

-.0765*

(.0413)

(.0358)

(.0422)

(.0419)

.0948

.0948

.0625

.0625

(.0867)

(.0967)

(.0869)

(.0877)

Observations

231

231

231

231

R-squared

0.289

0.289

0.269

0.269

Talk to: Clan Wealth (million cedis) Mean plot area Financial asset (million cedis) Farming stock (million cedis) Gender (=1 if female) pH Soil organic matter (%) 1 if resp holds trad. office

p < 0.1 p < 0.05 p < 0.01. Block Bootstrap standard errors with 1000 replications in parentheses for column (2) and (4) and regular standard errors in column (1) and (3). Constant omitted from table.

Table 4: Fixed Effects Regression (1) Village-FE

(2) Village-FE

(3) Clan-FE

(4) Clan-FE

plant

plant

plant

plant

-.62**

-.62

-.133

-.133

(.292)

(.377)

(.268)

(.34)

.24

.24

.17

.17

(.24)

(.756)

(.244)

(.809)

Average religion group profit

.035

.035

.0691***

.0691**

(million cedis)

(.0259)

(.0329)

(.0245)

(.0291)

Average clan group profit

-.0151

-.0151

.0097

.0097

(million cedis)

(.0134)

(.111)

(.0145)

(.0785)

Talk to: Religion

.241

.241

.248

.248

(.185)

(.254)

(.186)

(.257)

-.00444

-.00444

-.106

-.106

(.185)

(.253)

(.187)

(.291)

.114***

.114*

.114***

.114*

(.0423)

(.061)

(.0426)

(.0595)

-.00472

-.00472

-.00178

-.00178

(.00898)

(.0144)

(.00892)

(.0155)

-.159**

-.159

-.154**

-.154**

(.0618)

(.206)

(.0616)

(.0773)

-.0438

-.0438

-.113

-.113

(.222)

(.451)

(.221)

(.345)

-.327***

-.327***

-.344***

-.344***

(.0644)

(.0786)

(.0643)

(.061)

-.061

-.061

-.0403

-.0403

(.0755)

(.0895)

(.0615)

(.0998)

-.0535

-.0535

-.071

-.071*

(.0472)

(.0332)

(.0437)

(.039)

1.3**

1.3

.713*

.713

(.56)

(.798)

(.43)

(.818)

Observations

231

231

231

231

R-squared

0.309

0.309

0.312

0.312

Religion percentage Clan percentage

Talk to: Clan Wealth (million cedis) Mean plot area Financial asset (million cedis) Farming stock (million cedis) Gender (=1 if female) pH Soil organic matter (%) Constant

p < 0.1 p < 0.05 p < 0.01. Block Bootstrap standard errors with 1000 replications in parentheses for column (2) and (4). Regular standard errors in column (1) and (3) Fixed effects dummies are not reported in the table.

Table 5: Accuracy of Dependent Variable? Dependent Variable: =1 if is farming pineapple now or have farmed in the last 3 years (1) OLS

(2) OLS

(3) Village-FE

(4) Clan-FE

pine

pine

pine

pine

.107

.107

-.0951

.145

(.216)

(.364)

(.38)

(.365)

.282*

.282

.0342

.0808

(.17)

(.178)

(.379)

(.524)

Average religion group profit

.0469**

.0469*

.0197

.0413

(million cedis)

(.02)

(.0282)

(.0288)

(.0293)

Average clan group profit

.0127

.0127

.00299

.0275

(million cedis)

(.0109)

(.0174)

(.0472)

(.0908)

Talk to: Religion

-.171

-.171

-.204

-.19

(.157)

(.157)

(.154)

(.163)

.0433

.0433

.163

.0765

(.151)

(.175)

(.141)

(.172)

.0277

.0277

.0331

.0241

(.0359)

(.0527)

(.0466)

(.0495)

-.0157**

-.0157**

-.0142*

-.0138*

(.00756)

(.00737)

(.00779)

(.00755)

-.0388

-.0388

-.0485

-.0322

(.0519)

(.19)

(.118)

(.17)

.0438

.0438

.0484

.0311

(.187)

(.24)

(.182)

(.179)

-.427***

-.427***

-.417***

-.431***

(.0563)

(.0908)

(.1)

(.0893)

-.00596

-.00596

-.0843

-.037

(.0504)

(.0653)

(.0795)

(.0718)

-.0921***

-.0921***

-.0672*

-.0962***

(.0347)

(.0316)

(.0377)

(.0351)

.0987

.0987

.0911

.0923

(.0728)

(.1)

(.1)

(.094)

.559

.559

1.32*

.696

(.355)

(.497)

(.685)

(.64)

Observations

231

231

231

231

R-squared

0.421

0.421

0.448

0.441

Religion percentage Clan percentage

Talk to: Clan Wealth (million cedis) Mean plot area Financial asset (million cedis) Farming stock (million cedis) Gender (=1 if female) pH Soil organic matter (%) 1 if resp holds trad. office Constant

p < 0.1 p < 0.05 p < 0.01. Block Bootstrap standard errors with 1000 replications in parentheses for column (2), (3) and (4). Regular standard errors in column (1).

Appendix

.3 .2 .1

proportion

.4

.5

Appendix Figure 1: Proportion of Farmers Cultivating Pineapple (Retrospective from Udry’s Data)

1990

1992

1994 three years past

1996 survey time

1998

Appendix Table 1: Determinants of Information Links everask Either Party Holds Traditional Office

-.698*** (.245)

Same Religion

.0801 (.313)

Same Clan

.504** (.235)

Same Gender

2.22*** (.75)

Same Soil Type

-.278 (.253)

Absolute Age Difference (years)

-.0423** (.0166)

Absolute Wealth Difference (million cedis)

.123*** (.034)

Distance between Plot Centers (kilometers)

-.00042*** (.000151)

Constant

-2.43*** (.795)

N

525

pseudo R-sq

0.118

Logit MLE Estimates. Dependent variable is one if either party answered yes to the question: Have you ever gone to _____ for advice about your farm? Table reconstructed using Udry’s method, but on different sample of interest (all farmers, instead of just between pineapple farmers).

Appendix Table 2: Summary

Count

Mean

Statistics

Standard

Minimum Maximum

Deviation

Plant - Have ever planted pineapple

231

.5238095 .5005173

0

1

Farm pineapple now or in last 3 yrs

231

.3160173 .4659293

0

1

Religion percentage

231

.4696094 .1872613

0

1

Clan percentage

231

.4867095 .2063048

0

1

Years farming pineapples

231

2.597403 3.971397

0

25

Gender (=1 if female)

231

.4632035 .499727

0

1

Financial asset (million cedis)

231

.1611169 .9346574

0

10

Farming stock (million cedis)

231

.0645733 .1627326

0

1.544

Wealth (million cedis)

231

.8559151 1.488389

.015

12.287

Mean plot area

231

5.618286 3.474355

1

21

Average religion group profit (million

231

2.737027 2.050292

.08

9.199908

231

2.656578 2.62794

.038

10.00393

pH

231

6.347246 .5799017

4.85

7.725

Soil organic matter (%)

231

3.2035

1.353953

4.97

Talk to: Religion

231

.0785714 .213746

0

1

Talk to: Clan

231

.0979489 .2224749

0

1

cedis) Average clan group profit (million cedis) .8054459