1 ∗ Learning About a New Technology: Pineapple in Ghana Christopher R. Udry Timothy G. Conley Yale University University of Chicago christopher.ud [email protected] [email protected] September, 2008 Abstract This paper investigates the role of social learning in the di ff usion of a new agricultural fi technology in Ghana. We use unique data on farmers’ communication patterns to de ne each individual’s information neighborhood , the set of others from whom he might learn. mers adjust their inputs to align with those of Our empirical strategy is to test whether far their information neighbors who were surprisingly successful in previous periods. We present evidence that farmers adopt surprisingly succes sful information neighbors’ practices, condi- tional on many potentially confounding factors including common growing conditions, credit arrangements, clan membership, and religion. The relationship of these input adjustments to experience further supports their interpretation as resulting from social learning. In ad- ∗ The authors have bene fi ttedfromtheadviceofRichardAkresh,Fed erico Bandi, Alan Bester, Dirk Bergemann, Larry Blume, Adeline Delavande, Steven Durlauf, Ana Fe rnandez, Garth Frazer, Chris Hansen, Lars Hansen, ff Russell, Chris Taber, Otto Toivanen, Ethan Ligon, Charles Manski, Francesca Molinari, Stephen Morris, Je Giorgio Topa, and seminar participants at a number o f seminars. We especially thank Yaw Nyarko, Mark Rosenzweig and anonymous referees for their advice. The data used in this paper were collected by Ernest Appiah, Robert Ernest Afedoe, Patrick Selorm Amihere, Esth er Aku Sarquah, Kwabena Moses Agyapong, Esther Nana Yaa Adofo, Michael Kwame Arhin, Margaret Harriet Yeb oah, Issac Yaw Omane, Peter Ansong-Manu, Ishmaelina Borde-Kou fi e, Owusu Frank Abora, and Rita Allotey under the direction of the second author and of Markus Goldstein, who were visiting scholars at the Institute of Sta tistical, Social and Economic Research at the University of Ghana. This research has received fi nancial support from the NSF (grants SBR-9617694 and SES-9905720), International Food Policy Research Institute, World Bank R esearch Committee, Fulbright Program, Social Science Reseach Council and the Pew Charitable Trust. The author s are of course responsible for any errors in this paper. 1

2 dition, we apply our methods to input choices for another crop with known technology and they correctly indicate an absence of social learning e ects. JEL Codes: O12, D83, Q16. ff 1Introduction The transformation of technology is fundame ntal to the development process. For a new technology to be adopted by an agent, particula rly in agriculture, it must be adapted to the circumstances faced by that agent. Its characteristics usually will not be transparent to the new user (Evenson and Westphal (1995)). Consequently, an investment in learning about the new technology is associated with its adoption. If there are multiple adopters of the new technology in similar circumstances, as is often the case with an innovation in agriculture, then the process social . New users of the technology may learn its of learning about the new technology may be characteristics from each other. The role of social learning in promoting growth and technology di usion has been featured ff 1986); Lucas (1988); Aghion and Howitt (1998); in the endogenous growth literature (Romer ( nowledge spillovers is a central idea in the Acemoglu (2007)). Social learning that generates k large literature on urbanization and growth (e .g. Marshall (1890), Jacobs(1969), Porter(1990), Glaeser et. al. (1992)). These interactions are also an integral part of current practice in agricul- tural research and extension systems in develop ing countries. New technologies are introduced either by farmers’ own experimentation or through formal sector intervention and the process of social learning encourages their di usion (Rogers (1995), Bindlish and Evenson (1997)). There ff is a large body of empirical work looking at country and city-level evidence on the role of knowl- edge spillovers and growth (e.g. Glaeser e t. al. (1992), Rauch (1993), Barro (1994)). This paper is an e ort to contribute to the recent, small but growing literature that uses individual- ff level data to measure the quantitative importance of learning from others. Important examples of this work include Foster and Rosenzweig (1995), Bandiera and Rasul (2006), Bayer, Pinto ff and Pozen (2004), Munshi (2004), Du fl o, Kremer and Robinson (2006) and Kremer and Miguel 1 (2007). 1 In contrast, there is a long tradition of empirical studies by economists of the adoption of new technologies in agriculture. Griliches (1957) is the seminal work. For reviews see Feder et al (1985) and Evenson and Westphal (1995). This important literature, however, does not isolat e the role of learning processes from other determinants 2

3 In this paper we investigate learning about a new agricultural technology by farmers in the Akwapim South district of Ghana. Over the las t decade, an established system of maize and cassava intercropping for sale to urban consumers has begun a transformation into intensive production of pineapple for export to European markets (Obeng (1994)). This transformation of the region’s farming system involves the adopt ion of a set of new technologies, in particular the intensive use of fertilizer and other agricultural chemicals. cult for two major reasons. First, the set of Measuring the extent of social learning is di ffi cult to de fi ne. Second, even with a proper ffi neighbors from whom an individual can learn is di fi nition of this set, distinguishing learning fr om other phenomena that may give rise to similar de f learning, individuals may still act like their observed outcomes is problematic. In the absence o neighbors as a result of interdependent preferences, technologies, or because they are subject to related unobservable shocks. 2 Conse- Direct data on information interconnections is typically unavailable to economists. quently, economic investigations of the process of social learning have typically made assumptions that relate observed relationships between individuals - such as geographic proximity - to un- fl ows of information. This set of assumptions is critical for the measurement of the observed 3 For example, extent of social learning, but can rarel y be tested because of data limitations. Foster and Rosenzweig (1995) provide tabulations indicating that ‘friends and neighbors’ are an important source of information about ferti lizer use, but must use village aggregates as the relevant information set for social learning. of adoption. 2 Exceptions include Woittiez and Kapteyn (1998) an d Kapteyn (2000) who use individuals’ responses to questions about their ‘social environments’ to describe t heir reference groups. Romani (2003) uses information on ethnicity and membership in cooperatives in Côte d’Ivoire to infer the probability of information fl ows. Another exception is Bandiera and Rasul (2002), who have information on the number (though not the identities) of people using a new technology known by particular farme rs. Rauch and Casella (2001) is a very useful collection of papers that use direct information on social interactions more generally. 3 In many investigations of learning in developing country agriculture, the reference group is taken to be all farmers in the village (Foster and Rosenzweig (1995), Besley and Case (1994), Yamauchi (2002)). Munshi and Myaux (1998) take exceptional care in the construction of re ference groups for social learning by using external evidence on communication barriers arising from religion. See Manski (1993) for a concise discussion of the importance of reference group designations in identi fi cation of endogenous social e ff ects. 3

4 We have rich data that allows us to address the concerns of neighbor de fi nition more directly. et al (1957) which related adoption of new Our approach draws on the classic work by Coleman antibiotics to the network of social interconnections between the doctors. We collected detailed et information on who individuals know and talk to about farming. Hence we follow Coleman al by de ning information links between agents using responses to questions about which other fi 4 agents they turn to for information. fi ned, the identi fi cation of learning is still a formidable problem. Once neighborhoods are de revents us from inferring that learning e The classic problem of omitted variables p ects must ff ff be present simply from observations on, say, the di usion process of a new technology. The fact that a farmer is more likely to adopt a new technology soon after his neighbors have done ble that is spatially and serially correlated, so might be a consequence of some unobserved varia rather than learning. We believe that correlated unobservables are a general problem in the literature on agrarian technology, and it is apparent that they are important in the sample region (see sections 3.3 and 5.2). We have collected data to mitigate this problem. Our data contains detailed geographic and soil information as well as information on credit and family relationships, allowing us to contro l for otherwise confounding factors. Our identi fi cation problem can be thought of as a special case of the general problem of identi fi cation in social interactions models studied by Manski (1993, 1997), Brock and Durlauf ffi (1999), Mo tt (2001) and others. This literature is concerned with the problem of inferring whether an individual’s behavior is in fl uenced by the behavior or characteristics of those in his ff ects relies on using the neighborhood or reference group. Our strategy for identifying learning e fi c timing of plantings to identify opportunities for information transmission. The staggered speci plantings in our data naturally provide a seque nce of dates where new bits of information may be revealed to each farmer. By conditioning upon measures of growing conditions, we can isolate instances of new information regarding productivity being revealed to the farmer. We then examine whether this new informa tion regarding productivity is associated with innovations in a farmer’s input use in a manner consistent with a simple set of assumptions about the nature 4 Rogers (1995) and Birkhaeuser et al (1991) provide valuable surveys of re search that describes and charac- terizes the set of neighbors from whom agents learn about new innovations in a wide variety of settings. Van den Bulte and Lilien (2001) argue that the social contagion e ff ects found by Coleman et al vanish once marketing e ff ort is taken into account. 4

5 5 of learning. We model farmers’ learning about the productivity of inputs. The two key farmer-chosen fi inputs, fertilizer and labor, are used in essentially xed, known proportions and farmers need to learn about use of this composite input per pin eapple plant. Each harvest opportunity gives the farmer an observation on output for a given c omposite input, and thus reveals information about the productivity of that input level. We focus on fertilizer usage as a measure of this composite input since it is the most novel dimension of this new technology and because it is better measured than labor. is to estimate how farmers’ input decisions Our primary method to test for social learning respond to the actions and outcomes of other farmers in their information network. We know the inputs used and output harvested by each farmer, and thus can infer aspects of the informa- tion conveyed by each ‘experiment’ with the new technology by each respondent. We use our data on the spatial relationship between farms to condition on spatially-correlated but otherwise fl uence both pro fi ts and optimal input choices. We use our data on unobserved factors that in fl ow between farmers to trace the impact of the information revealed by each exper- information iment on the future input decisions of other farmers who are in the information neighborhood of the cultivator who conducted the experiment. fi nd strong e ff ects of news about input productivity in the information neighborhood of We 6 Speci fi cally, we fi nd for a given farmer: (1) he is a farmer on his innovations in input use. more likely to change his fertilizer use after his i nformation neighbors who use similar amounts fi ts; (2) he increases (decreases) his use of fertilizer of fertilizer achieve lower than expected pro after his information neighbors achieve unexpectedly high pro fi ts when using more (less) fertilizer than he did; (3) his responsiveness to news about t he productivity of fertil izer in his information 5 Du fl o, Kremer and Robinson (2006) use a randomized inter vention in Western Kenya to implement the same strategy. They gather data on social connections betwee n farmers, and then provide information regarding the pro fi tability of fertilizer to a random subset of these farmers. This permits them to identify the importance of learning from the experience of others in their environment. In Western Kenya, however, it turns out that information, either from neighbors or from one’s own experience, plays a very limited role in decisions about fertilizer use. 6 We use the male pronoun to refer to farmers because the large majority of pineapple farmers in our data are men. 5

6 neighborhood is much greater if he has only recently begun cultivating pineapple; and (4) he responds more to news about the productivity of fertilizer on plots cultivated by veteran farmers lusions hold when conditioning on the changes and farmers with wealth similar to his. These conc in fertilizer use of farmers who are physically ne arby and who therefore experience unobserved orrelated with his. In addition, they are robust changes in growing conditions that are highly c to a variety of di erent de fi nitions of information fl ow between farmers, and conditional on the ff fertilizer use of farmers with whom he has fi nancial ties. Finally, we apply our methods to a traditional maize-cassava mixture and they (correctly) indicate no evidence of learning about this established technology. 2 A Learning Model This section describes a simple model of lear ning about a new technology that we use to guide our empirical work. We consider risk-neutral far mers, each with a single plot who are concerned fi ts. At time period t farmer i with maximizing current expected pro chooses a discrete-valued . We mark time with the six-week intervals of our survey rounds. On this time scale, x input i,t pineapple output is realized periods after inputs are applie d via the production function: 5 . ) = w ε ( f (1) x )+ ( y i,t i,t i,t +5 +5 i,t ε is an expectation zero productivity shock that is IID across farmers and time and not +5 i,t is a positive, exogenous w observed by either farmers or the econometrician. The variable i,t and is correlated across x growing conditions variable that in uences the marginal product of fl i,t 7 ff This is motivated by the fact that agricultural production is often a ected farmers and time. by spatially and serially correlated shocks to the marginal product of inputs (examples include are observable to farmers but w variation in soil moisture, weeds, or pests). We assume the i,t 8 Pro fi ts not the econometrician. The price of the input x is a constant which we normalize to 1 . ( ;itisthe = w Farmers do not know the function ( f f x )+ ε x. − ) are therefore π +5 i,t +5 i,t i,t object of learning. The information set availabl e to each farmer is that generated by all current 7 The variable w could include a forward looking component, e.g. a rain forecast. i,t 8 In our study area, fertilizer prices and wages are common across farmers within villages and essentially constant throughout the sample time span. 6

7 and past growing conditions and observation of inputs and pro fi ts for all previous plantings 9 , as well as his own previous plantings. conducted by his information neighbors The farmer’s beliefs are conveniently summarized by his subjective expectations. We use the 0 f x ) and E and ( π ) x ) to refer to farmer i subjective expectations of s time t ( ( f notation i,t +5 i,t i,t +5 pro fi ts, respectively. The farmer’s time t of time t problem is to choose inputs to maximize so that: fi ts. This is nothing more than choosing input level x (subjective) expected pro i,t x ( π ̃ all ( x x, )) = w ̃ f − ( x ) ) − x .(2) ≥ w x f ( ̃ E i,t i,t i,t i,t i,t +5 i,t i,t i,t i,t This simple model is illustrated in Figure 1 for a case where the input can take on three values: zero (Z), low (L), and high (H). This gure plots expected pro fi ts as a function of growing fi As drawn, . ) x ( w. is a line with slope x t for any given input level fi Expected pro conditions f i,t the lines re fl ect a situation where beliefs are such that none of the three input levels is dominated. w the zero input choice is optimal, for intermediate w the low level is optimal, and for For small su ffi w the high level is best. The shaded upper envelope of these lines characterizes ciently large w. optimal input choice as a function of Learning will consist of updating beliefs and subjective ) · ( in response to new pieces of information. As the farmer learns, the slopes of f expectations i,t the lines in Figure 1 change and of course this can in fl uence farmer’s input choices for a given 10 realization of growing conditions. Farmers update their beliefs about in response to observations of inputs, growing conditions f 11 This means that at j is observed by farmer i. of farmer and outputs. Suppose that plot j , and the relevant growing conditions x , the inputs used farmer fi π time observes pro i t t j,t 5 j,t − x − π j,t − 5 j,t ( x . This new information allows farmer and to calculate )+ ε i = f w − j,t 5 − j,t j,t 5 w j,t 5 − this informs his beliefs about the productivity of input level x , leading him to update his j,t − 5 Note . ) prior to t ) : f x ( ( x f to ) t previously held expectation (dated at some time 5 j,t i,t j,t − 5 p − i,t p that farmers know growing conditions, so these expectations re fl ect only uncertainty due to ε w, f. Rather than focus on a speci fi c mechanism or type of learning, we and imperfect knowledge of 9 See Section 3.1 for operational de fi nitions of information neighbors. 10 ,ε tion (1). In a more general model in which y ) Figure 1 is drawn for the production func = F ( w ,x i,t +5 i,t +5 t i,t the curves need not be linear. In this case there could be disjoint regions of w in which a given level of input application is optimal. 11 The analysis applies as well for farmer i learning from his own experience. 7

8 consider the empirical implications of a set of three assumptions about the way in which farmers 12 We state our learn that correspond to farmers’ descriptions of their own learning process. 0 update f ( x ) ≡ f ( x ∆ ) − f s ( x ) in response i assumptions in terms of farmer i,t i,t − 5 j,t j,t − 5 − 5 i,t j,t p , x : ) , w to observing the event ( π j,t j,t 5 5 − j,t − f ∆ A. ) has the same sign as π and it increases without ( x ( − E x ( π ( x )) ) − i,t 5 j,t j,t j,t − 5 j,t − 5 i,t j,t p ) ( x . x )) exceeds E ( π ( π bound as − j,t 5 j,t − 5 i,t j,t j,t p ∆ f B. We assume that ( x ) attenuates as experience at input level x increases. An increase i,t 0 s experience reduces the absolute value of ∆ f in response to a given piece of new ( x ) i in i,t ,x ). ,w π information ( 5 j,t j,t 5 − j,t − C. ∆ f ( x )=0 for all x other than x so that this new information local, Learning is . 5 − i,t j,t 0 s beliefs only about the productivity of input level . Beliefs regarding other x i changes j,t − 5 13 input levels are unchanged. The counterpart of these assumptions in Figure 1 is that only one line will have its slope fi tishigher changed in response to each new piece of information, this slope will rise (fall) if pro (lower) than expected, and for a given piece of new information the change in slope will become i observes at time t +1 less pronounced as farmer experience grows. So in Figure 1, if farmer 0 s using, say, fertilizer level L has achieved a higher pro fi t than expected given j that farmer j growing conditions (at, say, point A), then i L from updates his beliefs about the productivity of L L ) to f ( ( ) . f i,t +1 i,t Assumptions A, B, and C have the following main empirical implications: Implication 1: Farmers tend to adjust input use towards surprisingly successful input levels and higher-than-expected pro ts at the currently utilized input level will make farmers less likely fi 12 There are many speci fi c models of learning consistent with these assumptions. In particular, they are con- sistent with independent Bayesian learning about the elements of the support of x , or with standard models of reinforcement learning (Kaelbling et al 1996; Feltovich 2000). 13 The implications of our model do not hinge upon inputs being discrete. We take them as discrete for ease of exposition. We could allow agents update only within one of a set of input ranges or only within a bandwidth of observed input use via a local-average (kernel) regressi on. This would allow us to obtain identical implications with continuous inputs, at the cost of complicating notation and our de fi nition of local learning. 8

9 to change from that level. When observed pro t is higher than expected, then ∆ f fi ( x ) − 5 i,t j,t x x is the optimal input ( ) is positive, the set of w for which f ∆ is positive. When 5 j,t − 5 − j,t i,t increases, hence this input level becomes more likely to be chosen. This is easy to see in Figure 1, where a higher than expected output for a given input level will raise the slope of the w corresponding line, leaving the other li nes unchanged. This increases the set of for which this given input level is optimal and hence chosen. Implication 2: Farmers tend to adjust input use away from an input level that was less x ( is negative. ) ∆ f fi table than expected. When observed pro pro fi t is lower than expected, 5 − j,t i,t x for which w ( x is the optimal input and so it is ) decreases the set of f ∆ Negative 5 j,t j,t − 5 i,t − w less likely to be chosen. Again in Figure 1, one line will rotate down, decreasing the set of where its corresponding input is optimal, perhaps even rendering an input level dominated. Implication 3: An observation of pro fi tsu ffi ciently above expectations will induce a farmer ffi is su )) ( x x ( − ) ciently large, E ( π π If to switch to that level of input use. j,t j,t 5 5 j,t i,t j,t − − p ( x ts under alternate input levels, fi will be large enough to dominate expected pro ) f ∆ i,t − j,t 5 observes i if t in period x to will switch from i Therefore, farmer w. for any given x i,t j,t − 5 p 0 ciently good. De choice of x ,we x that is su ffi s fi ning ∆ x − as x j an outcome of i,t i,t − i,t j,t 5 p summarize this implication as (3) ) x =1 { π − ( x x ( } ) − E ) x ( π ( ( x >c )) ∆ x − j,t 5 j,t − 5 i,t 5 − j,t j,t j,t 5 j,t i,t i,t i,t − p p where 1 {·} is an indicator function equal to one if its argument is true. The threshold value beliefs prior to s ( x ) at any fertilizer level x depends upon growing conditions w ’ , farmer i c i,t i,t 0 and the characteristics of farmers s , ( · ) i particularly i and j, this information revelation f i,t 1 − experience. Implication 4: The probability of changing input levels in response to a given piece of information is decreasing in a farmer’s experience. This is a direct implication of our assumption that experience attenuates changes in beliefs induced by new information. Implication 5: Correlations in growing conditions across space and time can look like social learning to the econometrician who does not observe w . In this environment, growing conditions are positively spatially and serially correlate d. If there is any positive association in beliefs w i,t across farmers, i.e. similar rankings of the productivity of di ff erent input levels, then positive 9

10 correlations in w will result in positive correlations in optimal input choices. In other words, i,t mean that proximate (close in space and time) w positive spatial and temporal correlations in i,t farmers will have similar values and thus if their beliefs about productivity are similar, they w will tend to choose similar input levels. This is t rue regardless of whether farmers learn about input productivity. Moreover, when plantings by di ff erent farmers are staggered in time, as in our application, can easily lead to and w i, j for spatially close positive dynamic correlations between w i,t k + j,t innovations in choices being positively correlated with lagged neighbors’ choices, even if there is no learning at all. To see this, consider the exa mple in Figure 1 with input levels zero, low, and and there is no learning). Take ) ( . ) ≡ f ( . f high and suppose farmers know the technology (so i,t and who are physical neighbors, who are likely to experience similar w at similar farmers j i times i and j both plant , and who thus are likely to choose similar input levels. Suppose that . Suppose, as is common in our setting, that L = x = x and that they choose t at period i,t j,t j plants another plot in, say, period t +3 , and that farmer i plants another plot in t +8 . farmer ciently ffi >w )su let j experience a positive growing conditions shock ( w At time t +3 , j,t +3 j,t = H . Positive spatial and serial dependence in w is x large to induce him to change to +3 j,t making it more likely that i , also, will experience a su ffi ciently consistent with this shock to j will of course w = H .Highvaluesof x large innovation in so that at t +8 he will change to w +8 i,t tend to result in higher than usual values of yields and pro fi ts. The econometrician observes farmer j choosing H and receiving higher than usual yields/pro fi ts,followedbyhisneighbor i changing inputs to This is of course also a sequence of events that would also be predicted if H. there were no change in gro wing conditions, but farmer i learned about the pro fi tability of input level H from observing farmer j get a good outcome after choosing H at time t +3 . Thus it will be very important for us to disentangle the e ects of learning from the reactions to growing ff conditions. Our model does not capture some aspects of learning about a new technology that are im- portant in other applications. These aspects are largely absent or concerns about them strongly mitigated by the speci fi c circumstances of our context. Fir st, there are multiple inputs to farm- ing pineapple. The main inputs, fertilizer and labor, are used in essentially fi xed proportions so a single, composite input re fl ects reality well. Measurements of fertilizer provide our measure of 10

11 this composite input, since it is better measured than labor inputs. Second, we abstract from strategic behavior on the part of farmers. Strateg ic considerations in information transmission are less salient in this environment than in others in the literature. Our surveyed farmers are operating in a competitive environment in both o utput and input markets. The fertilizer choices of any farmer will have no impact on the prices of pineapple, nor on the costs of farm inputs. Third, we model farmers’ information fl ow from neighbors as being generated just from obser- fi t outcomes. This implicitly limits the extent of communication vation of their inputs and pro between farmers and the ability/willingness of a farmer to model others’ behavior. For example, it rules out learning via conversations about thi rd parties’ activities and farmers making infer- ences about a third party’s outcomes from an inf ormation neighbor’s action. Ruling out such ages we study as it is considered inappropriate aspects of learning is well motivated in the vill gossip for a pair of farmers to discuss other farmer s’ activities. Furthermore, farmers have little to no knowledge of others’ information connections. As a consequence, second-order inference from the input choices of other farmers has little power. Fourth, we model farmers as focused on short-run pro fi ts, so there is no scope for experimentation. Our main implications can be shown to survive the introduction of forward looking farmers who take into account the value of experimentation for future pro fi ts (see Conley and Udry 2005, Appendix 1) so this is not a escriptions of the reasoning behind their input major concern. Furthermore, surveyed farmers’ d minant concerns are returns from current rather choices indicates that their overwhelmingly do than future plantings. Our choice to model learning as local is also speci c to our setting. It contrasts with some fi models of learning in other contexts in which in formation about the production function can be deduced regardless of the portion of the production function that is used (Prescott 1972; Jovanovic and Nyarko 1996; Foster and Rosenzweig 1995). The empirical implications of our local learning model are di erent from those of a model of glob al learning. In the latter class ff of models, there can be no gener al implication that farmers adjust inputs towards (away from) surprisingly successful (unsuccessful) levels o f input use. For us, local learning is motivated by both surveyed farmers’ own descriptions of what they learned from past experiences with using fertilizer and a substantial descriptive li terature from Africa (Richards 1985; Amanor 1994). This evidence strongly suggests that an appropriate model should have the feature that farmers 11

12 must use or observe inputs in a given range in order to learn about the corresponding part of the production function. Basic Empirical Approach dels, using the full sample, pre- fi ve implications by estimating mo We will investigate these 6 =0 , and for the change in inputs itself, dicting the occurrence of a change in input use, ∆ x istically describe the key regressors whose Here, we preview our baseline models and heur x. ∆ construction is detailed in the following Sectio n. We estimate a logistic model of the probability x 6 =0 . The regressors of interest re fl ect whether events observed by the that inputs change, ∆ ts above or below expectations, which farmer in between his planting opportunities had pro fi refer to the farmer’s previous planting t we refer to as good or bad news, respectively. Let p ) total observed events t for the share of time x = ( s opportunity. We use the notation good, x i,t p 0 s information neighborhood (from the beginning of the survey until time t )thatare i in farmer and occur between periods t and t. Analogous notation x good news events at input level i,t p p f good/bad news and alternative inputs. These is used for the shares of other combinations o t regressors are expressed as a share of time total observed events so that they attenuate with does not observe any planting at a given increases in experience. They equal zero when farmer i and t .Weestimate input level between t p ⎤ ⎡ α = 6 good, x ( s α )+ ) x = good,x ( x s 1 i,t 2 i,t p p ⎥ ⎢ ⎥ ⎢ (4) 6 =0 } = Λ x Pr { ∆ ⎥ ⎢ ( bad, x = x α ) )+ s x s ( bad, x 6 = + α i,t 4 3 i,t i,t p p ⎦ ⎣ 0 ∆ α ( growing conditions) +( experience and other controls) α + 6 5 where Λ ( · ) denotes the logistic function. The empirica l counterpart of Implications 1 and 2 is that α are positive. ,α are negative and α ,α 2 1 4 3 We investigate Implications 3 and 4 using a regression model for changes in x with the following form: 0 . v = β + M β + β experience and other controls) ( ∆ growing conditions) +( (5) ∆ x i,t i,t i,t 1 3 2 The regressor of interest M is an empirical analog of the right side of equation (3). M is constructed so that if an inexperienced farmer i observes good news using inputs well above , this index will be positive (negative) and large. If the (below) his previous input level, x i,t p will be near/at M or in the absence of good news, x farmer observes good news close to i,t p 12

13 zero. Motivated by Implication 4, M is constructed so that its absolute magnitude is inversely 0 experience. Conditional on growing conditions, this index should be s proportional to farmer i x ffi cient ∆ a good predictor of variation in induced by observations of good news and its coe should be positive. Implication 2 is not informative about the direction or magnitude of changes in response to bad news; hence measures of bad news events do not appear in (5). tt (2001) describes as Our goal is to identify possible learning interactions via what Mo ffi a type of policy intervention which “changes the fundamentals for a subset of the population in a group in an attempt to in fl uence the outcomes of the others in the group.” Our ‘inter- fi vention’ is the realization of surprising (given growing conditions) pro ts by another farmer in one’s information network, which is re ected in M . Implication 5 focuses our attention on fl ff fl the challenge of disentangling the e uences of ects of social learning from the confounding in to construct high spatial and serial correlation in growing conditions. We do so by using the w i,t a variable that permits us to control for changes in growing conditions (see Section 3.3) . The component of our information measure, M, that identi fi es social learning is that orthog- onal to our measures of changes in growing condi tions and other farmer-level characteristics. fi Our identi cation assumption is that this component is uncorrelated with unobserved determi- nants of changes in input use. Evidence in support of this assumption is provided in section 4 andisthefocusofthespeci fi cation test in Section 5.2. There are at least two important concerns about this empirical strategy. First, individuals clearly choose their information neighbors. This presents the possibility that neighborhood sorting or selection e ff ects could in fl uence our results. As we discuss in Section 5.1, details of our empirical results permit us to rule out the most plausible models of endogenous sorting that might lead to spurious results with respect to learning. Section 5.1 also presents results using an arguably exogenous de fi nition of information neighborhoods based upon a prediction of information links between farmers given the deep er social relationships between individuals. Second, those who farm pineapple are a selected group of farmers. Thus our estimates may not be representative of potential learning e ff ects that non-adopters might face if or of β 1 when they adopt pineapple. If there is heterogeneity in the extent to which farmers learn from others, it is possible that those who adopt fi rst are those most responsive to information from would overstate others. If so, our sample of adopters would be selected to be responsive and β 1 13

14 the importance of social learning in the overall population. As we only have information on learning for adopters, we cannot estimate learning e ects for the non-adopting subpopulation ff and this second concern must remain a caveat. However, in Section 4, we investigate the extent of heterogeneity in ability to learn withi n our sampled subpopulation of adopters. 3Data ction of variables. First, we discuss the basic This section describes our data and the constru then discuss the measures we use to de fi ne information features of our estimation sample. We neighborhoods. We then describe the data on f armers’ inputs and outputs, our methods to con- trol for growing conditions, and our methods for approximating farmers’ subjective expectations and innovations in information. Our main data source is a two-year survey (1996-1998) of 180 households, drawn from a 14 Our study region is in the population of 550 households in three villages in southern Ghana. center of a recently growing area of intensive pineapple cultivation. Two enumerators lived in or near each village and interviewed respondents in 15 rounds at intervals of approximately six weeks. In addition to survey-based information, all plots were mapped using global positioning system equipment. This procedure yields accurate measures of plot size and location, data that is seldom available for developing countries. Our main estimation sample is constructed as fo llows. We begin with information on pineap- ple being grown on plots by 132 farmers. Of these plots, 288 were planted during our survey. 406 Plot input data is missing on 3 of these plots, leaving 285 . 77 of these were planted too late in our survey for fertilizer application to be completed before the end of fi 208 eldwork, leaving 8 87 200 plantings. plantings. We are missing data for some rounds on of of these, leaving fi rst planting in our survey by particular farmers, leaving 113 observed changes in these are the fertilizer use. GIS information is missing on 6 of these plots, leaving information on 107 changes in fertilizer use by farmers. Figure 2 depicts the pineapple plots for which we have GIS 47 14 These three villages are a subset of four total villages in which surveys were administered. Pineapple farming was not present in the fourth village. A detailed description of survey procedures, copies of the survey instruments and the data archive can be found at 14

15 and input/output information. Farmers identi fi cation numbers are the horizontal coordinates the round in which the planting began. The rst fi and the vertical coordinate for each point is planting by each farmer is marked with an open circle; second and later plantings are denoted by closed circles. The 47 farmers who contribut e observations on input changes are arranged at 15 the left to make reading the graph easier. shows the pattern of adoption of pineapple in our sample villages: from less than Figure 3 46% of farmers were cultivating in 1990, pineapple spread very rapidly until more than 10% is the total number of pineapple in 1997. We utilize two measures of experience, Experience i,t i t. We also utilize data pineapple plants planted by farmer from the start of our survey until time fi on years of farming pineapple to de ne a binary indicator, dividing pineapple farmers into two groups: veteran farmers who adopted pineapple before 1994, and novice farmers who adopted in 1994 or after. Table 1 reports summary statistics for our data. We report statistics by novice/veteran status as well. Wealth is de fi ned as the value of the non-land assets held by the farmer at the start of the survey period. Veteran farmers are far wealthier. In addition, pineapple farmers as a class are much wealthier than non-pineapple f armers in the area (not shown in table). Those . million cedis versus who do not cultivate pineapple have an average non-land wealth of only 4 16 Pineapple farmers’ wealth reduces the average of million cedis for pineapple farmers. 2 . 3 the potential importance of credit constraints for fertilizer decisions (mean fertilizer use on an average-sized farm is . 08 million cedis, a small fraction of average pineapple farmer non-land wealth). The clan indicator variables denote membership in a particular , or matrilineal abusua clan. The church indicator denotes membership in the most popular church. Members of that church and of one of the matrilineages are ov errepresented among experienced pineapple farmers. We collected information on the soil chemistry (pH and organic matter content) of approximately 80% of the plots. Approximately one-third of farmers report that they have received advice from an agricultural extension agent (from the Ghanaian Ministry of Food and Agriculture) in the past; we do not know when such advice was received. 15 There are 81 farmers who cultivate plots for which we have GIS and input/output information. 11 live in village 1, 32 in village 2 and 38 in village 3. Of the 47 farmers who provide data on input changes, 8 live in village 1, 14 in village 2, and 25 in village 3. 16 Cedis are small units. The exchange rate during the sample period ranged from 1700-2300 cedis/US dollar. 15

16 3.1 Communication and Knowledge One of our main innovations is that we are able to use the survey data to de ne information fi neighborhoods. We base our measure of information availability on direct data about conversa- tions between individuals. Each respondent was questioned about a ran dom sample (without replacement) of seven e village. The samples of individuals produced other individuals from our own sample in the sam responses to the question: “Have you ever gone to ___ for advice about your farm?”. In this case, we say an information link exists between farmers i j if either i responded ‘yes’ to and j this question about j responded ‘yes’ to this question about i . We use responses to this or if question as our benchmark de nition of information neighbors because during the fi eld research fi ently related to the learning process under study. it appeared reliably-answered and it is transpar 17 Not counting farmers Farmers are of course included in their own information neighborhoods. themselves, the median number of information neighbors is 2 . There are several systematic pattern s in information links. Here we brie y summarize esti- fl mates of a model predicting our benchmark neighborhood connections given exogenous farmer characteristics that is fully reported in Appendix 1. Spatial proximity is correlated with the presence of information links but it is not their sole determinant. Information links occur over long as well as short distances. These longer-distance information links are essential to our abil- ity to distinguish the impact of information from that of spatially correlated shocks in growing conditions. Cross-gender links are rare and links are positively correlated with common clan membership and similarity in age. Individuals with di erent levels of wealth are more likely to ff be linked, re fl ecting the strong vertical patron-client ties that exist in these villages. There is no evidence that religion in fl uences information links. Pineapple farmers — especially veteran pineapple farmers — are more likely to be in each others’ information neighborhood than would 2 provides a summary of our baseline information link distribution be expected by chance. Table by experience. Over 20% of veteran pineapple farmers (within each village) have approached each other for advice about farming, while only 6% of non-pineapple farmers are in each others’ 17 There are relatively few plantings with timing that allows individuals to learn from their past plantings so we do not distinguish between such ‘learning by doing’ events and observations in our benchmark speci fi cation. We do however, examine ‘learning by doing’ events seperately in Section 5. 16

17 information neighborhood. A similar pattern is observed using our other information metrics. It may be the case that these information connections were important determinants of the adop- ces of new adoption during our sample period to tion process; however, we have too few instan , we discuss the possibility that farmers vary in their address this question formally. In section 5 ar that the pineapple farmers who comprise our ability to learn from others, and in particul sample are selected along that dimension. s of our main results to varying de fi nitions of the infor- In section 5, we check the robustnes fl mation neighborhood by using three alternative measures of information ow. Two of these measures are based on lists of interactions betwe en respondents during the course of the survey ring labor, exchanging gifts, etc.). Our third (discussing farming, buying or selling goods, hi de fi d on predicted links between individuals using nition of the information neighborhood is base the model discussed in Appendix 1 . Information neighborhoods based on predicted neighbors are less subject to concerns that observed links are endogenously formed in anticipation of obtaining input advice. All of these measures are fully described in Appendix 1 . Finally, we note the surprising fact that pineapple exporters, though they might have an incentive to provide input information, do not a ppear to be an information source. There is no er application from the exporters to whom they evidence that farmers receive advice on fertiliz sell their harvest (nor is there any contract farming). Farmers were asked a series of open-ended questions about sources of information regarding farming, including fertilizer application. In 18 no instance did any farmer mention pineapple exporters as a source of information or advice. 3.2 Inputs and Outputs We focus on farmers’ decisions about the intensity of input use in pineapple production. The two key inputs are fertilizer and labor , which are used in essentially fi xed proportions but with varying intensities per pineapple plant. There is agronomic evidence that pineapple yields are very glove (1972)). In informal interviews, individuals responsive to fertilizer (Abutiate (1991); Purse in the sampled villages expressed substantial uncertainty and con fl icting views regarding the 18 This pattern has changed since the survey. Suri (2008) describes the recent emer gence of contract farming among pineapple farmers in this area. 17

18 19 optimal intensity of fertilizer and associated labor use. on fertilizer usage since it is better measured Our empirical measure of input intensity is based than labor. During the period from six weeks to six months after planting, pineapples are extremely sensitive to nutrient availability (Bartholomew and Kadzimin (1977);Soler (1992)). Our input measure is the amount of fertilizer per plant applied during this period. This corresponds to one to four survey rounds after planting. We adopt the convention of indexing plantings by the round when input application i s complete, four rounds after the pineapple was , is the per-plant amount of fertilizer applied i planted. Thus our input measure for plot x , i,t t − 3 through over the reference period: We observe many plantings of pineapple during each t. of our survey rounds because pineapple production in southern Ghana is not strongly seasonal. thin plots. The plots in our sample are close to Per-plant fertilizer application is uniform wi the minimal viable scale for export farmers. The median plot size in our data is approximately .5 hectares. The novice pinea pplefarmerswhoexhibitthemostevidenceoflearninghavea median plot size of .25 hectares; exporters are reluctant to harvest and export crops from plots 5 plantings in our data were as small as any smaller (only 125 ha.). Plots have to be harvested . on a single day for e ffi cient export of the fresh fruit by air to Europe. It is essential that the fruits mature simultaneously, which requires common treatment across plants within the plot. As a consequence, there is no scope for experimentation with di ff erent levels of fertilizer within plots. Pineapple takes approximately 5 of our surve y rounds to mature after the application of fi i with fertilizer application completed at time t, denoted fertilizer is completed. Pro t for plot ( ) , is obtained by deducting from harvest revenue the value of all inputs, including x π i,t +5 i,t family labor valued at the relevant gender-speci fi c wage. In addition to plantings with harvest revenue data, we have some plantings that wer e unharvested at the end of the survey period but far enough along to enable an accurate forecast of their value at harvest. Our measure ( x ) for these plantings (about a third of the total observed) is constructed using the π of i,t +5 i,t 19 While there are o ffi cial recommendations on fertlizer use from the extension service of the Ghanaian Ministry of Agriculture (which we interpret as the yield-maximizing level of fertilizer use on test plots), these far exceed the levels of application in the sampled villages. The rec ommendation is 400 kg of fertilizer per hectare which is more than 10 times the mean fertilizer use observed in our sample. Only 4 of 208 plantings we observed exceeded the recommended level of fertilizer application. 18

19 farmer’s forecast of revenue at harvest. Respo ndents had no trouble providing such forecasts, perhaps because pre-harvest crop sales are routine for traditional crops like cassava. 3.3 Growing Conditions We expect shocks to growing conditions to be positively correlated across both space and time. Our concern about spatial correlation is motivated by the observation in these villages that growing conditions vary spatially on the scale of hundreds of meters. Soil types and topo- graphical features are highly correlated across neighboring plots, but vary over the village as a whole. Therefore, common village-level weather shocks can have varying impacts across the village. Moreover, rainfall realizations can be di ff erent on opposite sides of a single village. Finally, weeds spread in a broadly continuous ma nner across space, and soil moisture and pest and disease environments are often much more similar on nearby plots than on more distant plots within villages. Positive serial correlation in growing conditions is also to be expected across the six-week long periods. At this time scale, there is substantial correlation in soil moisture, weeds, and pest conditions on a given plot over multiple periods. Thus we anticipate ons for physically close plots planted at di substantial correlation in growing conditi erent but ff near points in time, due to the overlap in much of the environmental conditions they experience. As detailed in Section 2, it is essential for us to control for changes in growing conditions in our empirical work. i between his current and Our control for the change in growing conditions faced by farmer ff previous plantings is based upon the di erence between the observed input choices of other farmers with plots that are close to i ’s current planting and the earlier input choice of farmer i . The assumption underpinning this is that strong positive spatial and temporal correlations in growing conditions will induce farmers to make highly correlated input choices. Therefore, input choices for plots proximate to plot i t will contain a strong signal regarding the at time close as the (plant-weighted) average of fi growing conditions at this place and time. We de x ne i,t 20 Our regressor for the change in ffi i and time t. fertilizer input on plots su ciently close to plot 20 ‘Su ffi ciently close’ means within 1 km and at time t to t − 3 . The median number of physically proximate plots is 12, the maximum is 25. See Figure 4 for a scale map of plot centers within one of the villages. Our results are not very sensitive to the radii chosen here, qualitat ively identical results obtain with a range of 500 to 1500 19

20 0 growing conditions for plot since i i s previous planting at time t at time is then de fi ned as: t p close Γ = − x x (6) i,t i,t p i,t Γ i ∆ x should be a good predictor of changes in input choice by farmer ≡ x − x , that are i,t i,t i,t i,t p due to growing conditions. This regressor can be interpreted as a gap between a target input use of those close to ( i, t ) and the farmer’s previous input choice. We use this construction rather than a measure of changes in geographic neighbors’ input choices because of the unbalanced nature of our panel. Many farmers have single plantings and those with multiple plantings are armers remain useful in constructing our target irregularly staggered in time. Single-planting f ̃ to re fl ect the level of absolute discrepancies between x Γ term. We also employ a regressor i,t i,t p ̃ . and inputs used on plots close to ( Γ i, t is a plant-weighted average (across plots close to i, t ) ) i,t . erence between inputs used on the plot and x of the absolute di ff i,t p with the neighborhood de fi nition based Γ In addition, we construct a regressor analogous to i,t fi nancial rather than geographic neighborhoods. Two farmers belong to each others’ fi nan- on cial neighborhood if they lend to, borrow from, or exchange gifts or hold assets in common with each other at any point during the two year surve y period. This regressor is motivated by alter- estimates might be caused by omitted fi cant β nate explanations that would suggest that signi 1 variable bias because information neighbors share common access to credit arrangements. 3.4 Innovations to Information In this Section we discuss our measures of farmers’ subjective expectations and our operational de fi nition of innovations in information regard ing input productivity. A key object in this mea- ’s subjective expectation, at the time of his previous i π ( x which is farmer )) , ( E sure is i,t j,t j,t +5 p tability for his information neighbor farmer ,ofpro fi j , who used input planting opportunity t p 0 0 inputs at t. Farmer i and growing s information includes knowledge of farmer j x s level x j,t j,t relying , )) . We use observed inputs and pro fi ts to approximate E x ( ( π w conditions j,t i,t j,t j,t +5 p 21 without observing it. ff w on physical and temporal proximity to e ectively condition upon j,t meters and one to four time periods. See Appendix 2. 21 We note that while this method allows us to overcome the absence of data on w, it has the disadvantage of forcing our estimate of expected pro fi tability to be based on information relatively near the time t to t +5 interval. 20

21 ˆ Our approximation, ( π E ts of all plots whose observed input ( x fi )) , is the median pro j,t i,t j,t +5 p and whose time and location are close to plot j t or and dated period choices are close to x j,t 22 . Armed with w tions are approximately the same as shortly before so that growing condi j,t ts, we construct indicator variables for whether pro fi ts these approximations to expected pro fi exceeded or were below ’s expectations, which we refer to as good or bad news events, i π +5 j,t respectively. receives between his planting opportunities Our characterization of the information that i depends on the number of plantings that he observes and their precise timing relative to his ’ for ‘current’). In c (‘ and later at time t i t have plantings at time planting times. Let farmer c p gets information from only a single planting, say by i the modal case (half of all observations) and t . , as depicted in Figure 5 ’s planting are revealed between t farmer j, and the pro fi ts of j p c 0 s input use will be i Recall from equation (3) that the change in farmer £ ¤ − x ) x = ( 1 { π − x ( x ( ) − E (3’) } . π ) x ( x ( )) >c x j,t j,t +5 j,t j,t i,t i,t j,t +5 i,t j,t i,t i,t p p c c p 0 fi t exceeds expectations (it is good news) s pro j That is, the input change will be non-zero if ( x ) . This threshold is implicitly and does so by an amount that also exceeds the threshold c i,t j,t c 0 s beliefs about j i i and a function of the characteristics of farmers including of course farmer , which are not observed w all input levels and his experience, and of the growing conditions i,t c by us. It is necessary for pro fi ts to exceed expectations for the bracketed term in (3’) to equal one ffi cient. Good news about an input level may not exceed the threshold but of course this is not su 0 experience as it attenuates x ) . In particular, this threshold will increase with farmer i s ( c i,t j,t c updates in expectations. We construct a rough empirical analogue to the right-hand side of (3’) 0 s experience, Experience as follows: , i using our good news indicator and a term in farmer i,t c 1 ˆ ) ≡ { x − x [1 (7) π ]( } ( x 0 ) − . E > )) ( π x ( M j,t j,t +5 j,t i,t j,t +5 i,t i,t j,t c p p Experience i,t c Farmers in our model and, we presume, in real life have relatively more past information. However, with our data it appears impossible to incorporate more past information in these estimated expectations. 22 fi ne ‘su ffi ciently close’. For the large majority it means within 1 km of We use a variable bandwith to de are on relatively isolated plots and for these we expand the j and from time t to t − 3 . About 1/4 of x location j,t 3 kilometers. In both cases our baseline de fi nition of proximate inputs are those within geographic neigborhood to coarse categories of x =0 and x> 0 . The robustness of results to these assumptions is examined in Appendix 2. 21

22 If there is learning of the type we have proposed, should be positively correlated with M i,t c j, t results in good changes in inputs conditional on changes in growing conditions. If planting 0 x which is much higher than farmer ), previously used level ( i s news about an input level x i,t j,t p x that is much is large and positive; if instead it contains good news about a level M j,t i,t c , it is large and negative; and it will be near zero if good news concerns input lower than x i,t p will of course be nonzero for some M or zero in the absence of good news). ( x levels near i,t i,t p c ciently large to induce a ffi observations when in fact the innovation in information is not su should be a good predictor of both the change in fertilizer use by farmer . Nevertheless, M i i,t c as our “index of good news input M direction and magnitude of changes in inputs. We refer to it levels.” nition are used for the modal case. Slight variants of this de fi Equation (7) de fi nes M i,t c i observes multiple harvests and/or there is di ff erent timing. When the farmer observes when M uses a plant-weighted t , our baseline de fi nition of and t more than one harvest between i,t c p c 23 When a planting . t and observed between times i farmer j, t average of (7) across all plots t p c ts as is the case for farmer k in Figure 5 ,pro fi , i observes is harvested shortly after t that c 0 i ,but are not fully observed when i has an excellent s input decisions are fi nalized at t π c +5 j,t . Therefore, we are motivated to make use of these signal of the value of impending harvest fl is formed using (7) and discounted to re ect the M observations. Our baseline measure i,t c reduced information compared to fully-revealed h arvests. The tedious details of these weighting schemes are described in Appendix 1, and the ro bustness of our results to changes in weighting is examined in Appendix 2. 4 Empirical Results We begin by presenting some simple cross tabulat ions to illustrate our basic empirical strat- egy. Table provides a cross-tabulation of changes in fertilizer use and the two measures of 3 information fl ow with the most robust estimated e ff ects. The fi rst row in each cell provides a simple count of the number of transitions in each category. The most populous category is one maintaining zero fertilizer use, while transiti ons from positive use to zero use are more common 23 It turns out in our data there are no multiple plantings with good news observations disagreeing about the direction of movement. Thus we never have to average Ms of opposite sign. 22

23 than transitions in the opposite direction. The s econd row in each cell is the within-cell average , our index of good news input levels. A farmer whose previous input level was above that of M it whereas those who M, of his ‘good news’ neighbors input le vel will have a negative value for M positive. observe good news at a higher input level than they had previously used will have Cell-average M is strongly predictive of changes in fer tilizer inputs: it is strongly negative for positive for those who increased fertilizer use those farmers who reduced fertilizer use to zero, from zero, and near zero for those who did not change categories. The third row in the table provides a measure of the amount of bad news information about previously-chosen input levels, the average (within-cell) share of new information that falls in this category. This bad news elated with switch regarding past choices is clearly corr ing input categories. are consistent with the implications of our model of 3 The correlations evident in Table social learning. However, as we have noted above (Implication 5) they could also be generated by spatially- and serially-correlated growing co nditions. Therefore, we move beyond cross tab- ulations and estimate regress ions predicting the occurrence of a change in input use and the cations which include regressors controlling for this confounding source fi change in inputs; speci of variation in input changes. fi cations that we use to examine the relationships We now recall the base regression speci between changes in fertilizer use and information shocks that were introduced in Section 2. Al- ternative speci fi cations and a discussion of robustness are considered in Section 5 and Appendix i 2 and his plot that we use for conditioning be contained in a . We let the characteristics of These characteristics include the farmer’s wealth, soil characteristics, and indicators . z vector i,t for religion, clan, village, round of the planting and the novice farmer indicator. We fi rst estimate a logistic model of the probability of a change in x : ⎤ ⎡ α )+ good,x ( s x α = ) x 6 s ( good, x = i,t i,t 2 1 p p ⎥ ⎢ ⎥ ⎢ (4’) Λ } =0 6 = { Pr x ∆ ⎥ ⎢ ( bad, x = x α ) )+ s x s ( bad, x 6 = + α i,t 4 i,t 3 i,t p p ⎦ ⎣ 0 ̃ + z α α + Γ 6 5 i,t i,t fi ect the share of new information to the farmer in good/bad news and fl The rst four terms re both input categories, as described in Section 3.4. Recall that ( good,x = x s is the share of ) i,p h inputs equal to the farmer’s previous choice, plants on plots associated with good news and wit are negative and and α fi ned likewise. Our model implies α with the other three s terms de 4 1 23

24 ̃ α α are positive. Γ , is our control for growing conditions de fi nedinSection3.3. When and i,t 3 2 0 w s is very di ff erent from the growing conditions i the absolute discrepancy between farmer i,t p ̃ will tend to be large. Therefore it should be correlated with Γ growing conditions nearby ( ) i, t , i,t is anticipated to be positive. α changes induced by changes in growing conditions and 5 Our baseline regression predicting changes in fertilizer use is: 0 β . M (5’) + = v Γ + + z β β x ∆ i,t i,t i,t i,t 2 1 3 i,t Γ induced by correlated growing fi ned in (6) is our crucial control for movements in x de i,t i,t includes wealth, soil characteristics, and indicators for religion, clan, conditions. As in (4’), z i,t village, round of the planting and novice indicator. In addition, it includes the fi nancial analog with the neighborhood de fi nition based on fi nancial rather than geographic neighbor- to Γ i,t is permitted to be conditionally he teroskedastic and spatially hoods. Finally, the error term v i,t correlated across plots as a general function of their physical distance using the spatial GMM approach of Conley (1999). 2 that our key identi fi cation assumption is that conditional on our mea- Recall from section ), our measures of information shocks are uncorrelated Γ sure of changes in growing conditions ( it with unobserved determinants of changes in input use. We cannot test this assumption di- rectly. However, it is true that determinants of input changes are uncorrelated with observed this component of our information measures. Conditional on changes in growing conditions, our measures of information shocks are uncorrelated with of the characteristics of farmers (ex- any perience, wealth, matrilineage, religion, land characteristics) that might in fl uence input choice. Thus we argue it is plausible that conditional on our measure of changes in growing conditions (and our other covariates), surprisingly high pro ts achieved by a given farmer’s information fi neighbors in fl uence his decision to change his input leve l only through that information link: a is evidence of social learning. cant, positive coe ffi cient β signi fi 1 Logit Results 24 4 presents the coe ffi cient and spatial standard error estimates from equation (4’). Table We show results for three descriptors of a cha nge in inputs which vary with respect to their 24 The standard errors in all our speci fi cations use limiting results for cross section estimation with spatial de- pendence characterized by physical distance between the cen troids of each farmer’s set of plots. Serial dependence is allowed for only by use of time (round) dummies. Speci fi cally, spatial standard errors are calculated using 24

25 sensitivity to small changes in input use. The dependent variable in column is an indicator equal to one if the farmer changed A t ,or . We his fertilizer use from zero in his previous planting to a positive value at vice versa fl see that the direction of the in uence of our information and experience variables upon the predicted probability of changing is as impli ed by our model. Observations of bad news at the farmer’s previous input level strongly increas e the predicted probability that he will change ffi cient is positive, signi fi cantly di ff input levels. The estimated coe erent from zero and large. A one standard deviation increase . 12) in a farmer’s observation of bad news at that farmer’s ( th an increase in the probability of changing of previous level of fertilizer use is associated wi percentage points, calculated at the median probability of changing fertilizer use (which is 59 . Similarly, observations of bad news at alternat ive levels of fertilizer use strongly decrease 13%) . 12 )intheshare the predicted probability of changing. A one standard deviation increase ( of bad news at alternative fertilizer levels is associated with a reduction in the probability of changing fertilizer use of 8 percentage points (o ff of the same base probability of 13% ). The point estimate of the e ff ect of good news at alternative levels of fertilizer use on the probability of ff changing is positive, and that of the e ect of good news at the previous level of use is negative, as anticipated, but these coe ffi cients are relatively small (compared to those for bad news) and these estimates’ precision is too low to statistically distinguish them from zero. As would be expected in virtually any model of learning, novice farmers are much more likely to change input levels. In addition, the estimated probability of changing fertilizer levels is signi cantly and very strongly fi 0 s lagged inputs from his geographic increasing in the average ab solute deviation of farmer i neighbors, providing evidence of the importance of positively serially- and spatially-correlated the estimator in Conley (1999) with a weighting function that is the product of one kernel in each dimension tartsatoneanddecreaseslinearlyuntilitiszeroata (North-South, East-West). In each dimension, the kernel s distance of 1.5 km and remains at zero for larger distances. This estimator is analogous to a Bartlett (1950) or Newey-West (1987) time series covariance estimator and al lows general correlation patterns for distances shorter than the cuto ff . Note that plantings by the same farmer are allowe d to be arbitrarilly correlated as they are all distance zero from each other. The inferences reported below are robust to cuto ff distances between 1 km and 2 km. This is largely due to the fact that there is little spatial correlation in our regre ssion errors because we are in tentionally conditioning upon indices involving geographic neighbors’ actions th at provide a good signal of spatially correlated growing conditions. 25

26 unobserved shocks to the productivity of fertilizer. In column , the dependent variable is an indicator that the absolute value of the change B 1 cedi per plant Mean fertilizer use is cedis per plant, and the in fertilizer use is larger than 2 percentile of changes in fertilizer use for those farmers whose fertilizer use changed is cedi th 1 25 per plant. Hence, this column focuses on relatively large changes in fertilizer use. The results are qualitatively similar to those in column A , with the main exception that the positive point ff ect of good news at alternative levels of input use is now negative and very estimate of the e imprecisely measured. C In column iable equal to one whenever the change the dependent variable is a indicator var in input use is non-zero. Obviously, the median p robability of a non-zero change is higher in 53% The pattern of statistical signi fi cance of the estimates remains similar this case: it is now . fi to those of the previous speci cations. As in the other columns, bad news at the farmer’s lagged fertilizer use, bad news at alternative leve ls of fertilizer use, the absolute deviation of the farmer’s fertilizer use from that of his geogr aphic neighbors, and his experience are all quantitatively important determinants of the likelihood of changing fertilizer use. Finally, all three speci fi cations include an indicator for whether the respondent has ever received advice from the local extension agent. We do not know when any such conversation occurred. In column , the estimate indicates that those who have received advice from an C fi extension agent are signi cantly less likely to adjust their fertilizer use. Regression Results ffi cient on Table 5, column A presents the results of estimating equation (5’). The coe the index of good news input levels in the farmer’ s information neighborhood is positive and statistically signi fi cant, as implied by our model. A one standard-deviation increase in M (about 4 ) is associated with an increase in fertilizer use of approximately 4 cedis per plant, which is greater than the median level of fertilizer use p er plant of those farmers who use fertilizer. Round indicator variables are included, but not reported. None of them is individually signi fi cant, nor are they jointly signi fi cantly di ff erent from zero. There is no evidence that changes in input use are signi fi cantly related to inputs used by fi nancial neighbors. In this as in the following columns, changes in fertilizer use are strongly in the direction of the use by one’s geographic neighbors. 26

27 In column B, we examine relationship between experience and a farmer’s responsiveness to information on the pro tability of fertilizer. The coe ffi cients on M for veteran and novice fi ff erent from each other at the level. There is no evidence that farmers are statistically di 3% ws about alternative levels of fertilizer use. veteran pineapple farmers respond at all to good ne For novice farmers, in contrast, a one standard-deviation increase in is associated with an M increase in fertilizer use of approximately 4 cedis per plant. We raised the possibility in section 3.2 that farmers might be heterogeneous in their ability to learn from others, and in particular that lower ability farmers adopt pineapple more s lowly(ornotatall)andreactlesstoinformation from their neighbors. If this is the case, then our use of a sample of current pineapple farmers ral to information from neighbors. The results overstates the responsiveness of farmers in gene in Column B provide some evidence on the importance of this kind of selection. If there is selection such that later adopters of pineapple are less responsive to news, then this selection is su ffi ciently weak that recent adopters are still very responsive to new information. fl uence of the source of infor- Columns C-F present the results of an investigation of the in 0 s reactions. In alternate speci fi cations, we use variants of M de fi ned on i mation on farmer 0 0 information neighborhoods based on s i s information neighbors’ expe- partitions of farmer i rience, farm size, relative wealth and relative soil type. Our novice/veteran indicators are as fi ne large farms as those with plantings of at least 60,000 total pineap- described above. We de 25 Finally, we de fi ne a ple plants over our sample period (27 % of farmers have large farms). classi fi cation of wealth as rich/poor with rich as tho se whose non-land wealth at the start of the survey is greater than the mean non-land wealth (30% of farmers are rich by this de nition). fi 0 information neigh- s nes M separately for novice and veteran farmers in i C fi Column de borhood. The coe ffi M using veteran farmers’ results is large and signi fi cant and that cient on M comprised of novice farmers’ information is not. Column D presents a par- corresponding to 0 0 s i s wealth category (both rich or both information neighbor is in tition depending on whether i M fi poor). Wealth-partitioned cant predictor for same category neigh- is an important and signi ff bors but not for di M partition coe ffi cients in erent category neighbors. For each of the pairs of columns C and D , their di ff erence is statistically signi fi cant with a p − value under 2%. Column E presents analogous estimates with a partition of M depending upon the size of the farms in 25 Median and mean numbers of plants planted by farmers in our sample are 22,000 and 41,000, respectively. 27

28 0 i information neighborhood. Both coe ffi cient estimates are large, positive, and statistically s fi cant. Point estimates suggest that the responsiveness of input use to news from large signi farmers may be stronger than it is to similar news from small farmers. However, these estimates ff presents estimates with M are not statistically di erent from each other. Finally, Column F 0 s neighbor has the same soil type as i (sandy or fi ned using a partition based on whether de i fi cant evidence that news from others with the same soil clay). These estimates provide no signi type matters more to a farmer. In summary, novice farmers appear to be the ones reacting to good news and they tend to react to information revealed by neighbors who are veterans and 26 who have similar wealth. 5 Robustness Checks and Extensions 5.1 Learning-by-Doing, Alternative Information Neighborhoods, Endogenous Sorting There are 19 cultivators in our data who have multiple plantings su ffi ciently far apart in time for the fruit on the earlier planting to be growing bef ore fertilizer is applied on the later planting. These farmers present the opportunity to identify le arning-by-doing alongside the social learning that is the key focus of the paper. Column A of Table 6 presents regression results using with a partition of M using only an individual’s past history and u sing only other farmers’ information. In column A, we see that there is no important or statistically signi fi cant di ff erence between the impact of good news on one’s own plot and that of good news on a neighbor’s plot. In columns − D , we examine whether our fi nding that M predicts innovations in fertilizer B fi nitions of each of these fi is robust to changes in the de nition of an information link. Full de 0 s information is considered to be in j .Incolumn i 1 B alternatives are provided in Appendix , neighborhood if j is named by i when asked a series of open-ended questions about who taught them to farm and from whom they have received farming advice (or .Incolumn vice versa) C ,weusethebroaderde fi nition of an information link if either i or j is listed anywhere in the other’s roster of interactions with other sample members. In column D we de fi ne informa- 26 Of course an important caveat to this summary is that our small sample size constrains us to examine these partitions one variable at a time rather than jointly. 28

29 tion neighborhoods based on the predicted probab ilities for going to another farmer for advice A 1 (corresponding to the estimates in Table ). nition of the information neighborhood, the coe M cient on fi Regardless of the precise de ffi cant and large for novice farmers (the standard deviation of M is approxi- is statistically signi fi 3 . 5 mately 1 for the “Predicted Advice” for the two roster of contacts neighborhoods, and about 0 s i is a novice, good news experiments in i neighborhood). In each case we nd that when fi i changing his fertilizer use in the direction informationneighborhoodtendtobefollowedby of those experiments, conditional on our growing conditions control, village and round e ff ects, 0 wealth, clan and church membership. In contr s ast, there is evidence of responsiveness i and to information by veteran pineapple farmers only for one metric: predicted ask-for-advice. The robustness of our main results to the use o f predicted neighborhoods provides some assurance that they are not driven by sorting/selection e ff ects. It is also reassuring that, across fi nitions, our strong partial correlations of M alternate de with changes in fertilizer are driven as much by farmers moving down in response to good news at lower levels of fertilizer use as by upward movement in response to good news at higher fertilizer levels. Endogeneity due to positive, associative sorting along a characteristi c correlated with fertilizer use (e.g. unobserved wealth) could produce a tendency for good news to be associated with either high or low fertilizer use but we were unable to think of a scenario w here it would induce both. For example, assume high fertilizer levels are more productive than low levels, and assume a positive correlation between unobserved wealth and the amount of fertilizer used. In such a scenario, positive sorting would produce high fertilizer farmers with neighbors prone to good news events at high levels of fertilizer but not farmers lowering their fertilizer being prone to receipt of good news from their low-fertilizer neighbors. We are also con fi dent our results are not driven by reverse-causality sorting. For example, suppose farmers fi rst decided whether to increase or decrease their fertilizer use and then sought out the friendship/advice of surprisingly succes sful farmers who tended to dogmatically use their chosen amount. Such a scenario is unlikely in our setting since our data on information connec- tions were collected at the beginning of a 2 year survey. It is implausible that these farmers premeditated to the extent that they planned fertilizer choices one to two years in advance and chose their contacts accordingly. Moreover, the results using the predicted information 29

30 neighborhoods could not be generated by this type of scenario. 5.2 Learning about Optimal Labor Use in Pineapple and in Established Crops There should be no learning about optimal inputs i n the maize-cassava cultivation that occurs in our study villages. A standard maize-cassava intercrop pattern has been the foundation of the ivation in the 1930s. The characteristics of the economy here since the local decline of cocoa cult n to farmers in these villages. This provides a maize-cassava production function are well-know ‘placebo’ environment in which to test our meth odology. The only non-seed variable input into maize-cassava production is labor; no chemic al inputs are used on any maize-cassava plot in these villages (Goldstein and Udry, forthcoming). In this section we estimate a model of learning about optimal labor use in maize-cassava cultivation. To verify that our results regarding learning about labor use is not an artifact of a peculiar aspect of our data on labor inputs, we also estimate a model of learning about optimal labor use on pineapple farms. Since labor and fertilizer are used in approximately xed fi proportions in pineapple cultivation, we will see the same patterns associated with learning regarding labor that we found with fertilizer. We estimate a regression of changes in labor inputs for pineapple plots and for maize-cassava plots with a speci cation analogous to (5’): fi labor labor 0 labor = δ + M (8) δ + δ . Γ z u + ∆ x 2 3 1 i,t i,t i,t i,t i,t labor labor x Where is the labor input per plant for pineapples and per hectare for maize-cassava. M i,t 27 labor Γ are constructed exactly as above for this labor input. and We expect a positive δ for 1 i,t pineapple plots if pineapple farmers are learnin g from their neighbors about the productivity of for maize-cassava cannot be attributed to learning, because this technology δ labor. A nonzero 1 is well-established. We estimate (8) for the same sample of pineapple plots examined above. Labor inputs are measured over the crucial period early in the life cycle of the pineapple during which fertilizer 27 Labor inputs include both the value of hired labor an d that provided by the farmer’s household. The labor input range was divided into two categories (above and below median) for determining whether pro fi ts were unusually high given inputs. 30

31 inputs also occur. All farmers change their labor inputs across plantings, so there is no need to labor is de ned using the benchmark (asked for advice) fi estimate an analog of the logit (4’). M of Table presents the results of estimating (8). We A information neighborhood. Column 7 0 s lagged labor use from the lagged labor used by his i condition on the average deviation of labor , and its analog for his fi nancial neighbors. geographic neighbors, Γ 0 information neighborhood tend to be s i For pineapple farmers, good news experiments in i changing his labor use in the direction of those experiments’ labor, conditional followed by fi ot characteristics, village and round nancial neighbors’ lagged labor use, pl on geographic and 0 ffi wealth, clan and church membership. The coe cient is also large: a one standard s i ff e ects, and labor deviation increase in M (which is ) is associated with an increase in labor use of approx- 348 imately 682 cedis per plant, which is 37 per cent of the median labor use per plant on pineapple plots (which is 1 , 845 ). Pineapple farmers are learning about the productivity of inputs in the cultivation of pineapple from the experiences of their information neighbors. The data on labor 28 show the same pattern we saw with fertilizer. In contrast, in column B , we see that there is no evidence that maize/cassava farmers 29 The adjust labor inputs to information from the cultivation of their information neighbors. labor is virtually zero (at the point estimate, a one standard M ffi coe cient of the learning index labor 34 is associated with an increase in labor use of thousand cedis M deviation increase in per hectare, while the mean labor use is 466 thousand cedis, and its standard deviation is 567 labor fi cant predictive power for innovations in labor use in maize- has no signi M thousand). cassava production, just as we expect given the familiarity of this farming system in the study region. fi nd that there is a strong geographic correlation For both pineapple and maize-cassava, we 28 Note again that because labor and fertilizer move together in approximately xed proportions (subject to the fi additional measurement error in labor), so this is not independent evidence of learning about inputs on pineapple farms. 29 There are two di ff erences in speci fi cation between the pineapple and maize-cassava regressions: fi rst, in contrast to pineapple, the maize-cassava intercrop system is seasonal. Hence we compare inputs across successive seasons and replace the round indicators in A with season indictors in B . Second, the maize-cassava mixture is grown in all four of our survey villages, while pineapple is g rown in only three villages, hence there is an additional village indicator for maize-cassava. 31

32 in innovations in labor use. There are important spatially- and serially-correlated shocks to the productivity of inputs. This underscores the value of direct data on communication for de ning fi information neighborhoods. In the more typical case in which we had data only on geographic proximity, it would be tempting to rely on this to p roxy for information links. The consequences labor but with of this are presented in column C. We construct a new variable, analogous to M information links replaced with an indicator of geographic proximity. We see in column C n the direction of successful ‘experiments’ in that maize-cassava farmers adjust labor inputs i their geographic neighborhood. The coe ffi cient is large (a one standard deviation increase in the index of experiments in the geographic neighborhood is associated with an increased labor input of 153 466 thousand cedis/ha.) and thousand cedis/ha., compared to mean labor input of fi cant. This result has nothing to do with learning; it is induced entirely by the statistically signi onditions. However, without our direct data on communication strong correlations in growing c we might incorrectly infer the existence of social learning about labor productivity in this well- established farming system. 6Conclusion This paper presents evidence that social learning is important in the di ff usion of knowledge regarding pineapple cultivation in Ghana. We take advantage of data that combines agronomic and conventional economic information with details regarding relationships between farmers to ff address the challenge of identifying learning e ects in an economy undergoing rapid technological change. We fi nd that farmers are more likely to change input levels upon the receipt of bad news about the pro tability of their previous level of input use, and less likely to change when they fi observe bad news about the pro fi tability of alternative levels of inputs. Farmers tend to increase (decrease) input use when an information ne ighbor achieves higher than expected pro fi ts when using more (less) inputs than they previously used. This holds when controlling for correlations fi nancial neighborhoods, and in growing conditions, for common credit shocks using a notion of across several information metr ics. Support for the interpretation of our results as indicating learning is provided by the fact that it is novice farmers who are most responsive to news in their information neighborhoods. Additional support is provided by our fi nding no evidence of 32

33 learning when our methodology is applied to a known maize-cassava technology. Further evidence of learning is provided by changes in pro ts that correspond to input fi changes that appear to be mistakes and those that appear to be correct, subject to a conjecture regarding the optimal level of input use. Learning implies that farmers respond to both signal and noise. So, particularly in the early stages of learning, we expect to see novice farmers make mistakes by switching to what is truly a s uboptimal input level after seeing it perform surprisingly well in a small number of experiments. We have a strong belief that optimal fertilizer use for pineapple in Ghana is much higher than the l evels we observe in our sample; certainly it is 30 Thus we believe that many of the movements from positive to zero fertilizer greater than zero. use are mistakes. About a quarter of the farmers moved towards our conjectured optimal input s of fertilizer use. On average, these farmers levels in response to good news about high level fi t in our sample with average pro fi have the highest growth of pro 122 cedis per tgrowthof plant. We also observe approximately the same number of farmers who make (we conjecture) mistakes by reducing their level of fertilizer i nput in response to good news about low levels of fertilizer. They achieve substantially lower average pro fi tgrowthat 62 cedis per plant. Though such mistakes provide evidence in favor of lear ning, they inherently undermine the ability of our fi data to provide us with good estimates of the value of learning via estimating pro ts associated with optimal inputs. The span of our data is simp ly too short for most farmers to have learned optimal input choices. We have presented evidence that social learning plays a role in the cultivation decisions of these farmers. Information, the refore, has value in these villages, as do the network connections through which that information fl ows. This raises the possibility that farmers consider the consequences for the availability of informat ion when forming the connections that underlie their information neighborhoods. If so, measurement of the extent of social learning is not su cient for adequate evaluation of policy regarding the di ff usion of technology. It is necessary, ffi in addition, to understand the endogenous proces s of information network formation, making this a very important topic for future research. For example, consider the impact of a subsidy o ff ered to one farmer in a village that induces him to use an optimal large amount of fertilizer 30 This belief is based mainly on follow-up visits to these villages several years subsequent to the survey period and the observation that virtually all pineapple farmers now use positive amounts of fertilizer. 33

34 and (with high probability) get high pro ts. The speed with which this information spreads, fi and hence the value of the subsidy, depends upon the choices of the subsidized farmer and others in the village to make and maintain inf ormation linkages. These choices may depend upon the value of the information to each farmer and upon the costs of information links, which may depend upon a rich array of characteristics of the farmers and the social structure of the village. In some contexts, di ff ering religions may be an e ff ective barrier to communication. In others, gender, wealth or family ties may be the most salient determinants of the shape of the information network. 34

35 References [1] Abutiate, W. S. (1991) Guide to the Commercial Production of Pineapple, Papaya and Mango in Ghana. [2] Acemoglu, D. (2007) . Manuscript, MIT Introduction to Modern Economic Growth [3] Amanor, K. (1994) The New Frontier . London, Zed Books. [4] Aghion, P. and P. Howitt (1998) Endogenous Growth Theory . Cambridge, MIT Press. Econo- [5] Bala, V. and S. Goyal (2000) “A Non-Cooperative Model of Network Formation.” . metrica [6] Bandiera, O. and Rasul I (2006). “Social Networks and Technology Adoption in Northern Mozambique” The Economic Journal 116(514): 869-902 [7] Barro, Robert (1991) “Economic Growth in a Cross Section of Countries.” Quarterly Jour- nal of Economics .May. [8] Bartlett, M.S. (1950) “Periodogram Analysis and Continuous Spectra.” Biometrika , 37, 1-16. [9] Bartholomew, D. and S. Kadz Ecophysiology of Tropical Crops imin (1977). “Pineapple.” in by P. d. T. Alvim and T. T. Kozlowski. New York, Academic Press: 113-156. [10]Bayer,P.,R.Pinto and D. Pozen (2004). “Building Criminal Capital Behind Bars: Peer ff E ff ects in Juvenile Corrections.” Working Paper. [11] Berry, D.A. and B. Fristedt (1985) Bandit Problems .NewYork,Chapman&Hall. [12] Besley, T. and A. Case (1994). “Di usion as a Learning Process: Evidence from HYV ff Cotton.” RePEc Working Paper. [13] Bolton, P. and M. Dewatripont (1994). “The Firm as a Communication Network.” Quarterly Journal of Economics 109(4): 809-840. [14] Bindlish, V. and R. E. Evenson (1997) “The Impact of T&V Extension in Africa: The Experience of Kenya and Burkina Faso.” World Bank Research Observer 12(2): 183- 201. 35

36 [15] Birkhaeuser, D., R. E. Evenson, et al. (1991) “The Economic Impact of Agricultural Ex- tension: A Review.” 39(3): 607-650. Economic Development and Cultural Change Handbook of Econometrics, [16] Brock, W. and S. Durlauf (1999) “Interactions-Based Models,” J. Heckman and E. Leamer eds. [17] Carter, M (1997) “Environment, Technology and the Social Articulation of Risk in West Africa.” Economic Developmen t and Cultural Change . 45(3): 557-590. Rational Herds: Economic Models of Social Learning . Manuscript: [18] Chamley, C (2003) DELTA, Ecole normale supérieure. [19] Conley, T. (1999) “GMM Estimation with Cross Sectional Dependence” Journal of Econo- Vol 92(1) p.1-45. metrics [20] Coleman, J. E. Katz and H. Menzel (1957) “The Di usionofanInnovationAmongPhysi- ff cians,” Sociometry 20: 253-270. [21] Coleman, J. (1966). Medical Innovation: A Di ff usion Study . New York, Bobbs-Merrill. [22] Du fl o, E., M. Kremer and J. Robinson. (2006) “Understanding Technology Adoption: Fer- tilizer in Western Kenya, Evidence from Field Experiments.” Manuscript, MIT [23] Economides, N. (1996). “The Economics of Networks.” International Journal of Industrial Organization 14(6): 673-99. [24] Evenson, R. and L. Westphal (1995). Techno logical Change and Technology Strategy. Hand- book of Development Economics. J. Behrman and T. N. Srinivasan. Amsterdam, North- Holland. 3A: 2209-2300. [25] Falk, A. and M. Kosfeld (2003). “It’s all a bout Connections: Evidence on Network Forma- tion” IEW Working Paper 146. [26] Fafchamps, M. (1999). Ethnicity and Credit in African Manufacturing. Manuscript: Oxford. [27] Fafchamps, M. and B. Minten (1999). “Relationships and traders in Madagascar.” Journal of Development Studies 35(6): 1-35. [28] Feder, G., R. Just, et al. (1985). “Adoption of Agricultural Innovations in Developing Countries: A Survey.” Economic Development and Cultural Change 33(2): 255-298. 36

37 [29] Feltovich, N. (2000) “Reinforcement-based v s. Belief-based Learning Models in Experimen- Econometrica. 68(3): 605-641. tal Asymmetric-information Games.” [30] Foster, A. and M. Rosenzweig (1995) “Lear ning by Doing and Learning from Others: Hu- man Capital and Technical Change in Agriculture” Journal of Political Economy , 103(6): 1176-1209. [31] Glaeser, Edward, Hedi Kallal, Jose Scheinkman, and Andrei Shleifer (1992). “Growth in Cities.” Journal of Political Economy 100 (6): 1126-1152. fi [32] Goldstein, M. and C. Udry (forthcoming). “The Pro ts of Power: Land Rights and Agri- cultural Investment in Ghana.” . Journal of Political Economy American Journal of Sociology 78(6): [33] Granovetter, M. (1973). “Strength of Weak Ties.” 1360-1380. ns As Social Constructions - a Framework For [34] Granovetter, M. (1992). “Economic Institutio Analysis.” Acta Sociologica 35(1): 3-11. [35] Griliches, Z. (1957). “Hybrid Corn: An Exploration in the Economics of Technological Change.” Econometrica 25(4): 501-522. [36] Jackson, M. O. and A. Wolinsky (1996) “A Strategic Model of Social and Economic Net- works.” 71(1): 44-74. Journal of Economic Theory The Economy of Cities [37] Jacobs, Jane. (1969). . New York: Vintage. [38] Jovanovic, B. and Y. Nyarko (1996). “Learning by Doing and the Choice of Technology.” 64(6): 1299-1310. Econometrica [39] Kaelbling, L.P., M.L. Littman, and A.W. Moore (1996). “Reinforcement Learning: A Sur- Journal of Arti fi cial Intellegence Research 4: 237-285. vey.” [40] Kapteyn, A. (2000) “Saving and Reference Groups” Tilburg University CentER Working Paper. Journal of Monetary [41] Lucas, R. E. Jr. (1988) “On the Mechanics of Economic Development.” Economics. 22: 3-42. [42] Manski, C. F. (1993) “Identi fi cation of Social Endogenous Social E ff ects: The Re fl ection Problem” Review of Economic Studies 60: 531-542. 37

38 [43] Manski, C. F. (1997) Identi cation Problems in the Social Sciences. Harvard University fi Press, Cambridge. [44] Marshall, A. (1890) . London: Macmillan. Principles of Economics [45] Mo ffi tt, R. (1999) “Policy Interventions, Low-Level Equilibria, and Social Interactions” In Social Dynamics edited by Steven N Durlauf and H. Peyton Young. Washington, D.C.: Brookings Institution Press. [46] Montgomery, James (1996) “The Structure of Social Exchange Networks: A Game- Sociological Methodology 26:193-225. Theoretic Reformulation of Blau’s Model.” ff usion [47] Munshi, K. (2004). “Social Learning in a Heterogeneous Population: Technology Di in the Indian Green Revolution” . 73: 185-215. Journal of Development Economics [48] Munshi, K. and J. Myaux (2006) “Social Norms and the Fertility Transition." Journal of Development Economics , 80(1): 1-38. [49] Newey, W.K. and West, K.D. (1987). “A Simple, Positive Semi-de fi nite, Heteroskedasticity and Autocorrelation Consistent Covariance Matrix.” Econometrica 55:703-708. ff [50] Obeng, I. S. (1994). E ects of Domestic Policies on Production and Export of Non- Traditional Agricultural Commodities: A Case Study of Fresh Pineapples in Ghana, M.Phil Thesis, Department of Agricultural Economy and Farm Management, Univer- sity of Ghana, Legon. [51] Porter, Michael E. (1990) . . New York: Free Press. The Competitive Advantage of Nations [52] Prescott, E. (1972) “The Multi-Period Control Problem Under Uncertainty.” Econometrica 40(6): 1043-1058. [53] Purseglove, J.W. (1972) Tropical Crops: Monocotyledons I . London: Longman. [54] Rauch, James E. (1993) “Productivity Gains from Geographic Concentration of Human Journal of Urban Economics 34:380-400. Capital: Evidence from Cities.” [55] Rauch, James E. and Alessandra Casella, eds. (2001) Networks and Markets. New York: Russell Sage Foundation. [56] Richards, P. (1985) Indigenous Agricultural Revolution: Ecology and Food Production in West Africa. London: Hutchinson. 38

39 [57] Rogers, E. (1995) Di usion of Innovations , 4th Edition. New York: Simon & Schuster. ff [58] Romani, Mattia. (2003) “Love Thy Neighbor? Evidence from Ethnic Discrimination in Information Sharing within Villages in Côte d’Ivoire” Journal of African Economies 12(4): 533-563. [59] Romer, Paul (1986) “Increasing Returns and Long-Run Growth” Journal of Political Econ- (5): 1002-1037. omy.94 Pineapple: Quality Criteria . CIRAD-COLEACP-OCAB, Paris. [60] Soler, Alain (1992) [61] Suri, Tavneet (2007) “Selection and Comparative Advantage in Technology Adoption.” Manuscript, MIT Sloan. [62] Suri, Tavneet (2008) “Technological Change and the Evolution of Markets: The Case of Pineapple in Ghana.” Manuscript, MIT Sloan. [63] Van den Bulte, Christophe and Gary Lilien. (2001) “Medical Innovation Revisited: Social Contagion versus Marketing E ort.” American Journal of Sociology , 106(5), 1409-1435. ff [64] Wilson, R. (1975) “Informational Economies of Scale.” Bell Journal of Economics : 184-195. [65] Woittiez, I. and Kapteyn A. (1998) “Social interactions and habit formation in a model of female labour supply” J. Public Economics (70)2 p. 185-205. [66] Yamauchi, Futoshi (2007) “Social Learning, Neighborhood E ff ect, and Investment in Human Capital: Evidence from Green-Revolution India." Journal of Development Economics , 83: 37-62. 39

40 Appendix 1: Construction of Alternate Information Neighborhoods and the Good/Bad News Indices In this Appendix we provide details concerning the information neighborhoods mentioned in section 3.1 and the construction of our index of good news in those cases in which farmer i 0 plot is fully revealed s might be deciding on his fertilizer input before the outcome on farmer j by his harvest. Two of our alternate indicators for the existenc e of an information link between two farmers ff erent are based on a listing of all the individuals named by each respondent in a number of di contexts. This data includes people named in response to questions designed to record all fi ‘signi cant’ conversations about farming between individuals, and people who were hired by, borrowed from, lent or sold output to, or exchanged gifts, transacted land or jointly held assets 31 fi fi rst de We construct two metrics from this information, ning an with the respondent. information link to exist between two farmers if either reports learning about farming from the other. Because important information might be transmitted during quite casual conversation, we also de fi ne a broader information neighborhood which de fi nes a link to exist if either farmer lists the other anywhere in his or her roster of contacts. Both our baseline “ask for advice” metric and t hese “roster of contacts”-based measures have potential drawbacks. The “ask for advice” measure is based on a random sample of other farmers, and so yields estimates of the informa tion neighborhood of a farmer that are smaller than his actual information neighborhood. The roster of contacts measures include some pairs who probably do not discuss farming activities , and depends upon the respondents’ subjective understanding of ‘signi fi cant conversations about farming.’ In addition, there is some concern that observed information links might be endogenously formed in anticipation of receiving input advice (see Bala and Goyal 2000 or Falk and Kosfeld 2003). Therefore, we also construct a predicted information neighborhood based on estimates reported in Table A1 of the probability of a link (based on the question “Have you ever gone to ___ for advice about your farm?” ) given pair characteristics. There is evidence of spati al correlation in link patterns as the marginal 31 Signi fi cant conversations include, for example, discussions o f techniques for using agricultural chemicals, seeds, dealing with agricultural problems, or crop choice. 40

41 e ff ect of proximity is to increase link probability, but distance is not the sole determinant of mation links if they are of the same gender, the links. Individuals are more likely to have infor same clan, and similar ages. Individuals with di ff erent levels of wealth are more likely to be linked, re fl ecting the strong vertical patron-client ties that exist in these villages. There is no 32 We construct the predicted probability evidence that religion in uences information links. fl 0 information neighborhood from the parameter estimates in Table A1, and s i is in j that farmer predicted 0 j, t that farmer i is constructed as a weighted average of (7) across all plots s village M i,t c and t , with weights equal to this predicted probability (times the number of t between times p c 0 plot). s j plants on Turn now to the timing of information revelation. In about half of our observations, farmer 0 input decisions arenotfullyobservedwhenfarmer i s π i observes farmer k , whose pro fi ts +5 k,t may have an excellent signal i as illustrated in Figure 5. In this case, farmer , nalized at fi are t c t +5 harvestattime t +4 andislikelytohavesomesignalabouta t about a harvest as early +5 as +1 . Therefore, we are motivated to make use of these observations. Our simplest treatment t as in (7) separately for groups of observations with the same lag between M is to construct i,t c and 1 ≤ k for groups of . Letting k = t +5 − t ) refer to this lag, we construct M k ( t and +5 t c i,t c c , 3 , or 4 and examine the robustness of our conclusions to using such a set of Ms. These k =2 of Table A 2 and are discussed in Appendix 2 . However, to A results are presented in column regressor to be a cation by de fi ning our M conserve degrees of freedom in our baseline speci fi i,t c (1) getting a weight of one, (2) aweight M M linear combination of elements of this set with i,t i,t c c 33 . (4) with a weight of . 25 . 75 , and so on down to M of i,t c 32 For the sake of discussion of the quantiative importance of the determinants of link probability, take as ff erence, age di ff erence, and distance (2.9, 10.9, and 1.25 a base pair one with the mean values of wealth di ff erent clans, religions, and where neither party holds an o respectively) with the same gender and soil but di ce. ffi The point estimate of the link probability for this base pair is 22%. This point estimate would drop to 14% if one of the parties held some o ffi ce and increase to 31% if instead they were from the same clan. A reduction in estimated probabilities to around 15% would accompany a n approximate doubling of the base pair’s distance or age di ff erence, individually. Likewise, an approximate doubling of the wealth di ff erence would result in an increase to 31%. If the pair is not of the same gender, the predicted probability of one asking the other for advice drops dramatically to 5%. 33 When we have both multiple harvests observed by farmer i and some of these harvests are at times after t c we use a weighted average for baseline M that has this same pattern of weights based upon lag between t +5 i,t 41

42 Appendix 2: Robustness to Assumptions on Construction of M it InAppendixTable2,weexaminesomeoftheassumptionswehavemadeaboutthetiming of learning, the categories of fertilizer used in cons tructing our proxy for subjective expectations, tioning on soil characteristics. In each case the size of the geographic neighborhood, and condi ∆ x on cation change for our regression of interacted with fi we look at the impact of the speci M our experience indicator. ows from In column A, we examine the assumptions regarding the timing of information fl exible speci neighbors’ pineapple plots. We adopt a more cation which permits the respon- fl fi i siveness of ’s fertilizer use to vary depending upon the lag between his planting and the planting of his information neighbor’s plot. There is virtually no e ff ect of the success of plots planted t − 1 in round i ’s round t planting. As the lag increases, so does the estimated e ff ect of a on successful experiment, until for plots planted at least rounds previously the magnitude of the 4 ffi 1 . 15 . coe cient reaches In columns B and C, we examine the impact of averaging across multiple good news sources in 30% . For those of our observations in which more than one plot enters into the calculation M i,t of M in our standard speci fi cation, we select only the one, largest plot (that is, the plot with the most pineapple plants) that provides information to the farmer. Thus in column B, M i,t is de fi ned as in (7) with x ned as the level of fertilizer use on the largest successful plot de fi j,t 0 s information neighborhood during the relevant time period. In column C we take the in i more draconian step of dropping those observations for which more than one plot enters into the calculation of in our standard speci fi cation. In neither case do the results change in any M substantive way. In column D, we modify our categorization of fertilizer use. Expectations about pro fi tability fi ned over the two coarse categories of x =0 and of fertilizer use had been de 0 .For x> 40% of the plantings in our sample, there is at least one planting in the farmer’s information fi tability of x neighborhood that provides information about the pro ;for 32% of the plantings =0 there is at least one planting in the farmer’s information neighborhood that provides information about x> 0 .Wenowde fi ne expectations over three categories of input intensity: x =0 , 0 < . 5 ,x 5

43 seen in column D, this change in speci fi ect on the results. This cation has no qualitative e ff 85 th percentile of fertilizer use. For larger x ,the less than the x conclusion is robust for any h h cient on M is not signi fi cant at conventional precision of the estimates falls enough that the coe ffi ne expectations over more than three meaningful fi levels. It does not appear to be feasible to de categories given the size of our dataset. fi nition of the geographic neighborhood so that only plots within In column E, we alter the de fi nd that novice farmers 500 meters fall within a plot’s geographic neighborhood. Again, we nputs associated with good news experiments by change their fertilizer use in the direction of i their information neighbors, but that experi enced farmers do not. Very similar results are fi ned as within 1500 meters. obtained when geographic neighborhoods are de haracteristics in the conditioning set. We In column F we include information on soil c lose some observations by doing so, because soil testing was not completed on all plots, but ffi M is positive, large and statistically once again the core result is unchanged: the coe cient on cant for novice farmers but not for experienced farmers. signi fi r results are an artifact of mean reversion In column G we examine the possibility that ou in fertilizer use. Lagged own fertilizer use a ppears both in the dependent variable and in M i,t raising the possibility that mean reversion in fertilizer use, perhaps due to large measurement . In the absence of any cient on cant coe ffi fi M error, might lead us to nd a spuriously signi fi i,t across good news observations would be an estimate of the ff ects, the average x learning e − 1 k,t given π was above its expectation. So M could be interpreted as a x conditional mean of i,t k,t k,t is above or below its unconditional expectation which x noisy, biased estimate of whether i,t 1 − , we think A priori solely due to mean reversion. might be positively correlated with ∆ x i,t M . The in (5’) in addition to Γ this is an unlikely source of spurious results as we include i,t i,t sample size within geographic neighborhoods is considerably larger than that in information neighborhoods. Despite a higher spatial correla tion within geographic neighborhoods, averages within this larger neighborhood will have a smaller variance than averages within information should be a much less neighborhoods. Therefore, if mean reversion were driving correlations, Γ it M is conditioned upon, Γ is above its long run mean. Once x noisy measure of whether 1 i,t i,t i,t − 34 ff er little or no predictive power for ∆ x However, resulting from mean reversion. should o i,t 34 A special case of mean-reverting x would result if our data on inputs were dominated by large amounts of i,t 43

44 to be con fi dent our results are not an artifact of mean reversion, in column G we add the change by a magnitude to the regression. The coe ffi cients on M x lagged fertilizer level 1 i,t − i,t comparable to some of our other alternative speci fi cations and for novice farmers it remains a changes the most dramatically; this is unsurprising ffi cient on Γ signi fi cant predictor. The coe i,t lated and so provide an alternate control for since the lagged input levels are spatially corre spatially correlated growing conditions. o conditioning on lagged own pro ts. Thisiscon fi rmed in These results should be robust t fi cients change little and that lagged pro fi t realization column H, where we show that the key coe ffi fl uence on innovations in fertilizer use. This is consistent with our results on the credit has no in neighborhood. We have no evidence that variations in the opportunity cost of capital are playing important roles in fertilizer choices. in a way that In addition, we show in column Γ I that the results are robust to constructing i,t is constructed. In column I , the Avg. Dev. of Previous is strictly analogous to the way M i,t close ̈ , = ̈ x where x − Γ Use from Geographic Neighbor’s Use variable is constructed as it i,t previous it close and is the (plant-weighted) average of fertilizer input on plots su ffi ciently close to plot i ̈ x it time that had surprisingly high pro fi ts, as de fi ned in section 3.4. t to include fi Γ Column J demonstrates that results are robust to changing the de nition of i,t t+1, t+2 and t+3 in addition to t-3 through t. We input choices of geographic neighbors at time as only t-3 through t comparisons because this corresponds to fi keep our base speci cation of Γ i,t the period during which i is also applying fertil izer and thus provides the closest match to the unobserved Γ i,t Our logit results for responses to bad news at previous input choice are robust to analogs of classical measurement error. However, we think this case is unlikely to have occured as the fi eld research was speci fi cally designed to collect accurate data on farming inputs (including the number of plants planted) and fi cing sample size in exchange for frequent and thorough visits to respondents. output by sacri We also examined the special case of measurement error by performing Monte Carlo experiments (available fi cial measurement error to our fertilizer data. The mean of per-plant fertilizer use is 4 , upon request) adding arti and its standard deviation is 7 ; to our data we added a mean zero normal draws with standard deviation 1 to 7 cant (truncated so that measured fertilizer u ffi cient on M becomes insigni fi se is never negative). The estimated coe at conventional levels when the standard deviation of the added noise is 4 while Γ remains a signi fi cant predictor. 44

45 35 the robustness checks in Table 7. Indices of bad news at farmers’ previous input levels remain more bad news about previously used =0; 6 ∆ cant predictors of whether fi statistically signi x i,t levels increases predicted changes. 35 Columns B through F have analogs for our logistic regr essions and results (omited here to save space) are available upon request. The robustness check in Column G has no analog as the concern of potentially spurious results due the construction of M does not apply to our logits. 45

46 E (x, w) π Figure 1: Updating Productivity Knowledge A wf wf wf wf i,t+1 i,t i,t ( H i,t (Z) -Z ( ) - ( L L ) - ) - H L w L

47 subsequent plantings Farmers Figure 2: initial planting Distribution of Observations Across Farmers and Survey Rounds 5 3 7 9 11 Round of Planting

48 Figure 3: Sample Proportion of Farmers Cultivating Pineapple By Year (retrospective data from authors' survey) .5 .4 .3 .2 Proportion of Farmers Cultivating Pineapple .1 1990 1998 1996 1992 1994 'veteran farmer' 'novice farmer'

49 Figure 4: Roster Connections and Average Pineapple Plot Coordinates Village 3 1 Km

50 Farmer j Farmer k Time Farmer i Planting 0 1 Figure 5: Timing of Cultivat Fertilizer 2 Planting 3 t previous 4 Fertilizer 5 6 Fruit growing Planting 7 tt + 5 8 Harvest 9 Fruit growing Fertilizer ion and Information Revelation 10 Planting 11 Harvest 12 Fertilizer 13 Fruit growing 14 t current 15 Harvest 16 17 Fruit growing 18 19 Harvest 20 21 22

51 Table 1: Descriptive Statistics, Estimation Sample : equal means t (H Estimation 0 by experience) Sample Novice Veteran 1.392 3.767 Fertilizer Use (cedis per sucker) -1.91 1.938 (5.62) (9.66) (3.57) Δ Change in Fertilizer Use ( ) x -0.315 -1.398 0.009 0.61 i,t (4.71) (10.34) (20.02) 0 ≠ 0.500 0.494 -0.05 Indicator of Change in Fertilizer 0.496 (0.50) (0.51) (0.50) 0.889 2.763 3.07 2.331 Wealth (million cedis) (0.71) (2.84) (3.08) Clan 1 Indicator 0.327 0.154 0.379 2.18 0.451 0.538 0.425 -1.01 Clan 2 Indicator 0.487 Church 1 Indicator 3.07 0.231 0.563 6.175 -1.73 5.952 pH 5.876 (0.77) (0.74) (0.61) 2.927 3.007 1.19 Soil Organic Matter (%) 2.694 (1.11) (0.79) (1.20) 0.327 0.269 0.345 0.71 Contact with Extension Agent Indicator -2.58 -1.505 2.870 -0.500 Avg. Dev. of Lagged Use From ) Geographic Neighbors' Use ( Γ (14.18) (7.79) (4.02) i,t Indicies of Good News Input Levels: 0.055 -1.421 0.495 2.03 M Ask advice (2.11) (4.28) (7.80) M Talk frequently 0.163 -1.210 0.573 1.82 (2.33) (8.18) (4.43) -1.582 M Roster of contacts, farm info only -0.167 0.255 2.49 (1.54) (6.34) (3.37) 0.320 M Roster of contacts, full list -0.118 2.59 -1.584 (6.34) (1.53) (3.38) 0.126 M Predicted ask for advice 0.019 -0.340 2.21 (0.47) (1.79) (0.96) -2.33 0.076 0.187 0.117 Share of Good News at Lagged Fertilizer ) Use s(good, x=x (0.30) (0.17) (0.21) i,previous 0.022 -2.28 0.037 0.084 Share of Bad News at Lagged Fertilizer ) Use s(bad, x=x (0.12) (0.19) (0.09) i,previous -0.23 0.039 0.046 0.040 Share of Good News at Alternative ≠ Fertilizer Use s(good, x ) x (0.11) (0.12) (0.17) i,previous Share of Bad News at Alternative -1.45 0.025 0.017 0.049 x ) ≠ Fertilizer Use s(bad, x (0.07) (0.16) (0.10) i,previous 0 1.000 Novice Farmer Indicator 0.230 25 107 number of observations 82 15 number of farmers 32 47 Unless otherwise indicated, cells contain means of the indicated variable for the indicated sample. Standard deviations in parentheses.

52 Table 2: Information Connections by Cohort of Pineapple Adoption Proportion of pairs of individuals in each other's information neighborhood Novice Veteran Not Farming Pineapple Pineapple Pineapple Farmer Farmer Neighorhood Metric Not Farming Pineapple 0.06 0.05 0.07 Response to "Have you Novice Pineapple Farmer 0.09 0.13 0.05 ever gone to ____ for 0.13 0.21 Veteran Pineapple Farmer 0.07 advice about your farm?"

53 Table 3: Transitions in Fertilizer Use and Reciept of Information Current Fertilizer Use Zero Positive Count 55 12 Zero Avg M 0.44 0.86 i,t Previous vg s(bad, x=x ) A 0.090 0.015 i,previous Fertilizer Use 17 29 Count Positive Avg M -0.03 -3.30 i,t ) Avg s(bad, x=x 0.125 0.007 i,previous Count: Number of transitions in each category Avg M our index of good news input levels : within-cell average of M i,t i,t, Avg s(bad, x=x ): within-cell average of the share of plots obsered by i that had bad news i,previous about the input level he used in his previous planting

54 0.03 0.42 0.30 0.42 0.41 0.04 0.35 -0.04 -0.18 -0.34 -0.16 effect at Marginal median prob ve you ever n as a function C respectively. A Std. 1.15 2.59 0.89 1.06 1.04 1.12 0.65 2.94 0.20 0.12 5.84 inuous variables, Error rease in the variable. change in fertilizer 1.33 2.81 1.48 2.86 0.36 0.06 -0.70 -0.75 -1.54 -5.40 15.82 Dept Var: Indicator for non-zero 0.53 0.51 0.56 0.62 0.10 0.41 -0.03 -0.02 -0.11 -0.22 -0.14 effect at Marginal median prob Coefficient Std. 1.45 2.05 0.82 1.40 1.21 0.90 0.80 2.08 0.16 0.08 5.12 Error > 1 Cedi/Plant 2.70 2.47 2.36 0.47 0.16 -0.77 -1.11 -0.58 -1.42 -6.45 14.35 Dept. Var: Indicator for |Change| 0.14 0.86 0.64 0.71 0.39 0.85 0.10 0.59 -0.03 -0.06 -0.08 effect at Marginal median prob Coefficient ABC Std. 3.53 5.15 1.10 2.34 1.21 2.36 1.52 3.21 0.31 0.14 8.08 Error 9.48 3.54 6.51 2.00 3.09 0.77 0.27 -1.42 -0.68 -8.02 between Zero and Positive 23.62 Dept Var: Indicator for Change Coefficient ] i,t Γ ) ) ) ) i,previous i,previous i,previous i,previous x ≠ x ≠ s(good, x = x Good News at Previous Input Use Bad News at Lagged Fertilizer Use s(bad, x = x s(good, x Good News at Alternative Fertilizer Use Table 4: Determinants of Changing Input Use Average Absolute Deviation from Geographic Neighbors' Fertilizer Use [ Bad News at Alternative Fertilizer Use s(bad, x Talks with Extension Agent of physical distance, see footnote 24 for details. Sample Size = 107. Pseudo R-squareds .40, .26, and .31 for columns A, B and Novice Farmer full set of village and round dummies were included but not reported. Information neighborhoods defined using responses to: Ha Marginal Effects: Calculated at the median probability of a change for each column (.13, .32, and .53, respectively). For cont gone to farmer ____ for advice about your farm? Clan 1 this is the change in probability associated with a 1 std. deviation increase in the variable; for dummy variables a 1 unit inc Church 1 Wealth (Million Cedis) Clan 2 Logit MLE point estimates, spatial GMM (Conley 1999) standard errors in parentheses allow for heteroskedasticity and correlatio

55 Table 5: Predicting Innovations in Input Use, Differential Effects by Source of Information e Dependent Variable: Innovation in Per Plant Fertilizer Us A F BCDE Index of Good News 0.99 Input Levels (M ) [.16] i,t M * Novice Farmer 1.09 i,t [0.22] * Veteran Farmer M 0.10 i,t [0.32] -0.13 Index of Good News Input Levels by Novice Farmers [0.37] 1.02 Index of Good News Input Levels by Veteran Farmers [0.17] 1.03 Index of Good News Input Levels by Farmers with Same Wealth [0.18] Index of Good News Input Levels by -0.41 Farmers with Different Wealth [0.32] Index of Good News Input Levels 1.10 on Big Farms [0.14] 0.89 Index of Good News Input Levels [0.18] on Small Farms Index of Good News Input Levels 1.04 Farmers with Same Soil [0.16] 0.91 Index of Good News Input Levels Farmers with Different Soil [0.19] 4.01 4.19 4.12 Novice Farmer 4.22 4.20 [2.65] [2.66] [2.62] [2.77] [2.65] Avg. Dev. of Geog. Nbrs 0.58 0.58 0.58 0.55 0.54 0.59 ] From Prev. Use [ Γ [0.06] [0.06] [0.06] [0.06] [0.06] [0.08] i,t vg. Dev. of Financial A 0.53 0.45 0.40 0.43 0.22 0.24 Nbrs From Prev. Use [0.59] [0.55] [0.61] [0.60] [0.58] [0.58] -7.92 -7.62 -8.24 -7.81 -7.88 Village 1 -8.09 [1.36] [1.16] [1.31] [1.43] [1.43] [1.31] -2.15 -0.61 -1.78 Village 2 -2.17 -1.82 -1.83 [2.02] [2.07] [2.03] [2.02] [1.56] [2.11] 0.41 0.45 0.29 Wealth (Million Cedis) 0.13 0.36 0.29 [0.17] [0.20] [0.20] [0.25] [0.17] [0.17] -2.42 -2.62 -2.53 -2.55 -2.62 Clan 1 -2.68 [1.12] [1.09] [1.29] [1.15] [1.21] [1.11] -0.40 -0.11 -0.15 -0.31 -0.29 Clan 2 -0.11 [1.32] [1.44] [1.30] [1.32] [1.32] [1.30] 0.67 -0.60 0.87 0.88 0.26 Church 1 0.76 [1.06] [1.11] [1.29] [1.12] [1.12] [1.12] 0.70 0.71 0.71 0.71 0.71 0.73 R-squared OLS point estimates, spatial GMM (Conley 1999) standard errors in brackets allow for heteroskedasticity and correlation as a function of physical distance, see footnote 24 for details. Sample Size = 107. A full set of round dummies included but not reported. Information neighborhoods defined using responses to: Have you ever gone to farmer ____ for advice about your farm?

56 Table 6: Learning-by-Doing and Alternate Definitions of the Information Network Dependent Variable: Innovation in Per Plant Fertilizer Use AC D B Roster of Learning by Roster of Contacts: Predicted Doing and Contacts: Full Information Neighborhood Metric Farm Info Advice Learning by Set of Only Others Contacts 1.46 M from own plots only [0.56] 1.22 M from information neighbors [0.31] 1.49 M * Novice Farmer 6.34 1.50 [0.28] [0.28] [1.14] 0.15 4.52 0.19 M * Veteran Farmer [1.80] [0.22] [0.21] Novice Farmer 4.66 4.65 4.01 [2.84] [2.84] [2.77] Average Deviation of Lagged Use From Geographic Neighbors' Use ) Γ ( 0.49 0.49 0.49 0.33 i,t [0.09] [0.09] [0.12] [0.10] 0.72 0.73 R-squared 0.73 0.72 OLS point estimates, spatial GMM (Conley 1999) standard errors in brackets allow for heteroskedasticity and correlation as a function of physical distance, see footnote 24 for details. Sample Size = 107. All of the variables in Table 5 were included but coefficients are not reported. Alternative information neighborhoods are as defined in Section 3.1 and Appendix 1.

57 Table 7: Predicting Innovations in Labor for Pineapple and Maize-Cassava Plots Dependent Variable: First Difference in Labor Inputs for Pineapple and Maize-Cassav a C AB Maize-Cassava Maize-Cassava Pineapple Crop (labor cost in 1000 (labor cost in cedis (labor cost in 1000 per plant) cedis per hectare) cedis per hectare) labor 0.13 1.96 Index of Good News Input Levels: M [0.14] [.86] Index of Good News Input Levels in the 0.32 Geographic Neighborhood [0.12] Average Deviation of Lagged Use From 0.49 0.74 labor Γ Geographic Neighbors' Use [ ] [0.20] [0.08] Average Deviation of Lagged Use From 0.01 0.52 Financial Neighbors' Use [0.07] [0.29] Village 1 162.12 -89.77 -79.29 [301.85] [87.40] [77.69] Village 2 -303.66 -432.70 -246.19 [109.97] [131.54] [289.83] -193.52 -198.11 Village 3 [82.96] [73.17] 166.42 30.36 2.44 Wealth (Million Cedis) [91.55] [35.13] [32.90] 43.62 Clan 1 -159.86 -378.15 [299.52] [247.73] [265.36] -67.78 463.35 -82.42 Clan 2 [239.96] [78.77] [65.29] Church 1 -552.22 -62.06 -48.52 [91.96] [88.30] [253.94] 346 89 346 Sample size R-squared 0.55 0.42 0.24 OLS point estimates, spatial GMM (Conley 1999) standard errors in brackets allow for heteroskedasticity and correlation as a function of physical distance, see footnote 24 for details. Round/season dummies included but not reported. Information neighborhood from: Have you ever gone to farmer ____ for advice about your farm?

58 Appendix Table 1: Determinants of Information Links Coefficient Standard Error -0.55 Either Party Holds Traditional Office 0.26 0.04 0.33 Same Religion Same Clan 0.24 0.43 Same Gender 1.73 0.78 Same Soil Type -0.23 0.27 Absolute Age Difference (years) 0.02 -0.04 Absolute Wealth Difference (million cedis) 0.15 0.03 Distance Between Plot Centers (kilometers) -0.46 0.16 Constant -2.14 0.84 Logit MLE Estimates, Sample Size = 490, Pseudo R-squared =.12. Dependent variable is one if either party answered yes to the question: Have you ever gone to _____ for advice about your farm?

59 Appendix Table 2: Robustness to Changes in Specification Dependent Variable: Innovation in Per Plant Fertilizer Use DE GHI J AF BC Only Drop Obs Ferilizer Largest Geographic Geographic includes Г i,t Soil Flexible with Lagged Categories Good- Neighbor- Neighbor Lagged t+1, t+2, t+3 Charac- Lags in Multiple Fertilizer Zero, News hood within defined as Profits in addition to teristics Learning Good- Use Medium Plot 500m is M t-3 through t News Plots and High Enters M 1.85 1.03 0.98 1.48 1.30 1.11 0.76 0.89 0.34 M * Novice Farmer [0.19] [0.30] [0.22] [0.29] [0.31] [0.20] [0.27] [0.13] [0.21] 1.32 -0.24 0.08 -0.28 -0.08 -0.49 -0.41 0.04 0.52 M * Veteran [0.35] [1.29] [0.35] [0.69] [0.46] [0.52] [0.23] [0.45] [0.31] Farmer Novice Farmer 3.88 3.27 3.96 3.84 4.01 2.87 3.95 5.94 4.05 [2.13] [2.70] [2.71] [2.71] [2.90] [2.62] [2.72] [2.71] [2.71] 4)* M (k = 0.02 [0.69] M (k = 3)* 0.64 [0.14] M (k = 2)* 0.92 [0.25] M (k ≤ 1)* 1.15 [0.71] Lagged Own Fertilizer Use -0.84 [0.22] Avg. Dev. of Lagged Use From Geo Neighbors' 0.45 0.09 0.68 0.50 1.58 0.54 0.50 0.10 0.58 0.46 ) Use ( Γ [0.08] [0.06] [0.12] [0.17] [0.11] [0.25] [0.10] [0.16] [0.08] [0.10] i,t Lagged Own Profits 0.01 [0.01] Soil Organic Matter 0.14 [0.67] Soil pH 4.09 [2.31] Soil Type = Loam 1.40 [1.14] Soil Type = Sandy -5.78 [2.72] Sample size 107 84 107 107 107 107 107 107 107 89 R-squared 0.69 0.73 0.75 0.68 0.68 0.80 0.72 0.72 0.75 0.75 OLS point estimates, spatial GMM (Conley 1999) standard. errors in brackets allow for heteroskedasticity and correlation as a function of physical distance, see footnote 24. All of the covariates listed in Table 5 are included in the regressions, but n ot reported. Alternative specifications are as defined in Appendix 2. . When k=3, only * k parameterizes the number of lags. When k=4, only information from t-1 plantings is included in M t information from t-2 is included; when k=1, information from t-4 and earlier is included.

Related documents