WPS3665 Credit Constraints as a Barrier to Technology Adoption by the Poor Lessons from South-Indian Small-Scale Fishery Xavier Giné World Bank xgine@worldbank.org Stefan Klonner Cornell University stefan@klonner.de World Bank Policy Research Working Paper 3665, July 2005 The Policy Research Working Paper Series disseminates the findings of work in progress to encourage the exchange of ideas about development issues. An objective of the series is to get the findings out quickly, even if the presentations are less than fully polished. The papers carry the names of the authors and should be cited accordingly. The findings, interpretations, and conclusions expressed in this paper are entirely those of the authors. They do not necessarily represent the view of the World Bank, its Executive Directors, or the countries they represent. Policy Research Working Papers are available online at http://econ.worldbank.org. 1 Introduction It is generally recognized that the adoption of a new technology plays a fundamental role in the development process. However, the benefits from the introduction of the technology may be unevenly distributed among the population, especially if the markets do not function properly. While the micro literature on technology adoption and diffusion focuses on "who" adopts and "when" adoption happens, the macro literature has focused on the overall impact of globalization on inequality. In this paper, we bring these two strands of the literature together by studying the diffusion of plastic reinforced fibre (FRP) boats in a fishing village in Tamil Nadu and by analyzing the dynamics of income inequality during this process.1 On a micro level, this paper is interesting because unlike most of the technology adoption literature, we focus on a capital intensive technology. Indeed, in the agricultural context, many of the technologies studied, like the adoption of HYV seeds, the switch from food to cash crops, or the use of chemical fertilizers are not capital intensive. As a consequence, the literature has focused on explaining the impediments that hinder the gains from socially valuable technologies. These impediments usually take the form of information constraints and externalities that lead to a lower private than social value of technology adoption. In our study village, a fisherman who wishes to switch to an FRP boat has to incur an investment that he cannot typically self-finance. When capital markets are imperfect, lenders do not advance the full cost of the boat, so that prospective fibre boat owners are forced to save to make up for the remaining amount. As a consequence, for the same ability level, poor entrepreneurs adopt later than rich entrepreneurs as it takes them longer to accumulate the necessary capital. This results in a socially inefficient timing of adoption. Although a credit constraints model delivers a positive relationship between wealth and the propensity to adopt, there are many competing models of risk aversion or learning 1The economic importance of this particular new technology for South Indian village economies is documented by an article in the leading national newspaper, The Hindu (2001). 2 from others that would also be consistent with this fact and yet point to other explanations that can potentially lead to contradictory policy advice. Therefore, in this paper we try to disentangle the precise way in which asset wealth matters for adoption. The data, which were collected by the authors in 2002 and 2004, cover all 65 boat- owning fishermen of a fishing village where the first fibre boats appeared in 2000. They include a census as well as time series of individual fish sales. These unique data allow us to discard the view that delayed adoption by poor households is due to the smaller scale of their entreprise as each individual operates the same fibre boat. In addition, aversion to income fluctuations is not likely to be important as there is evidence of informal insurance arrangements among boat owners (Platteau, 1985). Since we do find that wealthier entrepreneurs adopt earlier after controlling for individual ability, we conclude that credit constraints faced by relatively poor boatowners are important. To a lesser extent, we find evidence of aversion towards uncertainty about the gains from the new technology. This paper is related to the work of Platteau (1984) and that carefully document the introduction of mechanized boats in Kerala. They study an example of how globalization drastically changed the rural landscape. In the words of Platteau, "by revolutionizing the traditional techniques and methods of catching, preserving and distributing fish, the project succeeded beyond expectations in achieving a `revolution' [...] that changed life in many ways." This process exemplifies what Nissanke and Thorbecke (2004) have in mind when they point out that the relationship between globalization and poverty is complex and may even be non-linear as globalization affects poverty through many channels. In contrast, we can assert that the only major change from 2000 to 2004 in the village we study was the introduction of FRPs. This enables us to focus on the effects of a single channel, namely, technology diffusion. On a more macro level, the paper is interesting because we explore the dynamics of income inequality during the process of technology diffusion. Contrary to some empirical literature on the effects of globalization, we find that inequality follows Kuznets' well- known inverted U-shaped curve.2 Initially, the technological innovation widens the gap 2These studies reject the Kuznets hypothesis, but they focus on the impact of initial inequality on 3 between the rich and the poor, but after the entire community has completed the techno- logical shift, inequality drops to a lower level than before, which implies that in the long run the innovation studied here has benefited the poor more than proportionally. In addition, we provide some simulations to investigate how different counterfactual distributions of initial wealth across the sample affect adoption timings. Here we find that a redistributive policy favoring the poor results in accelerated economic growth and a shorter duration of sharpened inequality, although the quantitative impact of such a policy is small. When we simulate the adoption process for a sample of only rich households, in contrast, the process of adoption is completed ten times as fast as observed in the actual data, implying that rich communities can enjoy the benefit from technological innovation and thus grow considerably faster than poor ones. Our findings are not limited to the specific context of fishing villages but provide insights for any capital-intensive technological innovation among poor small-scale en- trepreneurs who face tight financing constraints. Examples include urban small-scale manufacturers and service providers, such as garment shop owners and rickshaw drivers. The rest of this paper is organized as follows: in the next section we provide an overview of the recent existing literature on technology adoption in primary sectors of low income countries. Section 3 sketches a theoretical framework that illustrates how wealth disparities translate into different timing of technology adoption. Section 4 describes the data and institutional background. Section 5 develops the empirical methodology and presents results. Section 6 simulates the adoption process for alternative distributions of initial wealth. The final section evaluates the findings and draws conclusions. 2 Related Literature Much of the literature that studies technology adoption in developing countries concludes that its pace has been rather slow. Feder et al. (1985), in their excellent review of the early literature point to factors such as credit constraints, aversion to risk and limited growth, and find that it may be detrimental. Thus, while the Kuznets hypothesis may not hold in situations with high initial inequality, it may still me appropriate in contexts of low initial inequality like ours. 4 access to information, to explain why adoption has not been faster. Most of the work they survey uses static models to explain adoption, and so the dynamic properties of adoption are left to heuristic or comparative-static arguments at best. In particular, the role of savings, which may be crucial in contexts where credit or insurance markets are imperfect, especially if the technology is indivisible, does not receive much attention. The literature distinguishes between divisible technologies, such as high yield varieties (HYV) or new variable inputs, and indivisible technologies, such as tractors or the one we study here, FRPs. If the technology is divisible, one can study the intensity of adoption of a given farmer as well as the aggregate intensity in a region. When the technology is indivisible, the decision at the individual level is necessarily a dichotomous variable and only the aggregate intensity is still continuous. In the case of technologies that are not capital intensive, like the adoption of high yield variety (HYV) seeds, lack of credit is not seen as a major constraint and so the literature focuses on how individuals learn from others about the profitability of the new technology. Still, because the costs of gathering information may be fixed, some of these papers also establish interconnections between initial wealth and the pace of individual technology adoption. Most of the more recent literature is concerned with the interaction between learning about a new technology and its diffusion. The first of these contributions is Feder and O'Mara (1982), who show that aggregate adoption at each point in time can follow a sigmoid curve. They consider a scale-neutral risky innovation with risk-neutral farmers holding prior believes about the mean yield of the new technology. Besley and Case (1994) proceed in a similar fashion in their study of the diffusion of a new cotton variety in one of the south-Indian ICRISAT villages. In their model, planting the new variety not only affects current profits but it also generates public information on the profitability of the new versus the old variety. Therefore, there is individual as well as social learning from planting the new crop. They find that adoption occurs with delay because farmers underestimate initially the technology's profitability and because they fail to internalize the positive informational externality created by other farmers when planting the new crop. Among other findings, they conclude that wealthier farmers tend to innovate first because the informational externality is largest to them. Poor farmers 5 adopt later as they benefit from the positive informational externality generated by rich farmers. Foster and Rosenzweig (1995) take for granted that HYV of wheat and rice that became available during the Indian Green Revolution in the mid 1960's yield higher profits than traditional varieties. In their model, however, the profitability of HYV's is dictated by a target input model, whose optimal level has to be learned. The issue is, again, individual versus social learning in that each "trial" with the new variety generates additional information on the optimal level and this information is conveyed not only to the farmer himself, but also to the entire village (at least to some extent). In contrast to Besley and Case (1994), however, planting the new crop comes at the cost of choosing an input level that is far from the optimal, especially in earlier periods when there is little knowledge about the optimal level. Farmers find themselves playing a dynamic public good game, where each farmer has an incentive to wait because information is generated costlessly by another farmer experimenting with the new crop. As a consequence, those farmers who expect the greatest benefits from experimentation adopt first. As in Besley and Case (1994), those are the relatively wealthy farmers because they operate several plots, each of which benefits from the additional information in future cropping periods. Interestingly, their results imply that poor farmers in a community of relatively poor farmers adopt earlier than poor farmers with wealthy neighbors. Bandiera and Rasul (2004) test for non-monotonicity of information spillovers among Mozambiquean farmers to whom a new sunflower variety was made available in 2000. They find an inverted U-shaped relationship between the amount of available information to a farmer and the probability that he adopts, suggesting that social effects on the individual adoption decision are positive when there are few adopters in the individual's information network, and negative when there are many. Unfortunately, they did not collect information on whether the cultivation decisions were individual or collective. If the latter happens, then the network would also play a role in adoption. Differences in asset wealth, however, are not found to impact the adoption decision, but this is not surprising given that the NGO that provided the new variety in their context covered all switching costs. 6 Munshi's (2003) study of adoption of rice and wheat HYV's during the Indian Green Revolution focuses on the effect of the sensitivity of farm-specific growing conditions on the extent of social learning. He finds that for rice HYV's, which are more sensitive to unobserved farm characteristics than the wheat HYV's, individual adoption decisions are less responsive to neighbors' experience. His analysis, however, does not take into account the effect of famers' wealth on their adoption decisions.3 To summarize, all of these papers conclude that there is either a positive or no rela- tionship between individual wealth and the decision to adopt a new technology. Wealth, however, is typically correlated with, or even undistinguishable from other important indi- vidual characteristics, such as farm size, education, access to credit, the capacity to bear risk, availability of other inputs, and access to information. Thus, a positive relation- ship between wealth and early adoption can be due to alternative factors, which are not disentangled by the existing empirical analysis. Policy recommendations, however, may well depend on the nature of the channel through which wealth affects adoption. In the papers focusing on learning, for example, it is generally argued that poor farmers adopt later because their valuation for information generated by initial "trials" with the new technology is lower. Thus, an information campaign about the benefits would result in more adoption. In general, however, it is not clarified, whether alternative channels might also play a role. Other potential candidates are differential risk aversion (see Binswanger et al., 1980), access to capital, or availability of labor. For example, if the technological innovation is labor intensive and wealthier households have better access to the labor market, a wealthier household may adopt earlier just because of labor market conditions. In the present study, we therefore make an attempt to thoroughly identify the channel through which wealth affects adoption decisions. 3Another important paper on the nature of social learning that is not concerned with technology adoption, however, is Conley and Udry (2003, 2004). These authors are concerned with optimal choice of target inputs among Ghanian pineapple cultivators who had only recently adopted the new crop. By identifying information networks, they disentangle the process of information dispersion and show how farmers treat information from an information neighbor differentially, depending on the difference between their own and their neighbor's experience as well as recent success with the crop. 7 3 Context The village of study is located in the southern part of the coast of the gulf of Bengal, close to the pilgrim center of Tiruchendur. With a population of 1,500, there are 75 boats operated by about 250 men. The village has neither a harbor nor a jetty, a fact that restricts operations to beach-landing boats only. All year-round operating vessels have a crew of two to four men and are operated by local households. All of these households belong to the exclusively catholic fishing community of the village, which used to belong to a particular fishermen's caste within the Hindu caste system before collectively converting about 400 years ago. On a typical day, boats leave the shore around 1am and land at the village's market place on the beach between 7 and 11 in the morning. There, local fish auctioneers market the catches to a group of buyers, which comprise local traders as well as agents of nation- wide operating fish-processing companies. During the monsoon months, mechanized vollam-boats with a crew of five from other villages land on the village's beach and market their catches there. The local fishing techniques, catamaran and FRP fishing, continue during that period. According to local fishermen, fish is plentiful enough that no competition with the migrating mechanized boats arises. Instead, it is held that the local economy benefits from the demand generated by the migrant crew members and the increased marketing activity in the village. The catamaran is the traditional fishing technology in southern India. It is a raft-like vessel made of two Alphesia logs tied together with two crossbeams at the two ends. While it was originally powered by either a sail or manpower, all of these boats in our study village were equipped with a 8 or 9 horse power motor by the year 2000. The beach-landing, fibre boat is, in contrast, a recent technology. The fibre-reinforced plastic used in these crafts is a composite material comprising a polymer matrix reinforced with glass fibres. This manufacturing technique has been commonly used in aerospace, automotive and marine industries throughout the western hemisphere since the 1950s. It required an intervention of the Indian government, however, to make this technology available to the small-scale fishing sector of Tamil Nadu. In 1995, the Department of 8 Science and Technology of the Indian central government in New Delhi initiated a special project, under which the Centre for Science and Technology and Socio-economic Devel- opment in Chennai designed a vessel with active participation by fishermen in more than 20 villages along the 1000-km coastline of Tamil Nadu over the course of four years. The form of globalization we see here is thus one where the southern country's government becomes active by itself to make a technology which has long been known in the North available to its population. According to the project, the boat was designed to be cost and fuel-effective, versatile, comfortable, and durable to stand the constant exposure to saltwater (Hindu, 2001). In 2000, the boat was made available to fishermen throughout the Tamil Nadu coast. In the context of our study village, it was the opening of a subsidiary of the domestic FRP manufacturer in nearby Tiruchendur in early 2000, which made FRPs readily available. All the FRPs operated in our study village come from this manufacturer and roughly share the same characteristics. Only two boats measure 21× 7 feet, while the rest measures 18×7 feet. Fishermen alleged that FRPs can cope with rough surf and are, at the same time, more comfortable, faster and more economical than catamarans. All this indicates that the government project has been successful in achieving its goals. Moreover, the FRP can be powered by the same 8 or 9 hp outboard engine, which was already in common use on catamarans at the time the FRP became available. In most cases, catamaran owners that shifted to FRPs continued to use the outboard engine of the catamaran. With the same number of crew, an FRP's landings are about 50% bigger than those of a catamaran. Given the yields of fibre-boat fishing, every owner of a catamaran in the village we interviewed assured that he wanted to switch to a fibre boat as soon as possible. It has to be mentioned, however, that fishing on an FRP requires a different set of skills than those needed to operate a catamaran. For that reason it is common practice among the buyers of fibre boats in the village to hire migrant laborer-fishermen from Kerala as crew members who have previously gathered experience with the new technology. It is instructive to compare the process of modernization in our village of study in Tamil Nadu with the three-decade long process in Kerala carefully studied by Platteau (1984) and others (e.g. Kurien, 1984). In Kerala, the process was triggered by the Indo- 9 Norwegian Project (INP), a large comprehensive development project in the 50s. It called for "the mechanization of the fishing boats, provision for repair facilities, introduction of new types of fishing gear, improvement of processing methods, building of ice plants and supplying of insulated vans and motor crafts for transport of fresh fish" (Sandven 1959). The most successful vessel that was introduced in Kerala under the program was a fully mechanized 32-ft trawler with a powerful 84-90 hp inboard engine. A new trawler cost around Rs. 125,000 in 1978 prices (which equals about Rs. 600,000 in 2004) and had a crew of 15-20 members. In our context, the FRP costs between Rs. 70,000 and Rs. 80,000 in 2004 and a new outboard motor is Rs. 50,000 to 60,000. In summary, while in Kerala the process was induced by a large-scale international development project, in our village it was triggered by a smaller-scale domestic Indian development project. Turning to the issue of vessel financing, in both cases, credit constraints are a key determinant of adoption. In Kerala, external finance was dominated by commercial banks, providing loans of up to Rs. 90,000 in the 1970s. The trawler itself was not accepted as collateral because the risk of depreciation due to an accident was deemed too high. Instead, real estate was often used, explaining why the majority of trawler owners had never been regular fishermen but wealthier villagers. Typically, they were engaged in occupations related to fishing trade, such as fish merchant or exporter. They hired a captain and a crew to operate the vessel and provided incentives by entitling each of them to a share of the sales. Auctioneers only played a marginal role for boat financing in Kerala because the capital requirements were too large. In our study village, in contrast, it is always the fisherman himself who owns the craft while the bulk of external finance for the purchase of FRPs comes from auctioneers who advance loans to boatowners in exchange for the right to market their catches. As for the trawlers in Kerala, FRPs are not accepted as collateral by banks. There is, moreover, evidence that external finance does typically not cover the amount needed for the technology switch. In our sample, boatowners finance about 35 percent of the cost of the FRP from own resources. Despite the credit constraints being a major obstacle to adoption, we do not observe a fibre boat rental market developed, although it is common for trawlers in Kerala. Pre- sumably, if a rental market for fibre boats existed, credit constraints would be less of an 10 issue as relatively poor but talented fishermen could bypass their lack of funds. According to qualitative interviews conducted in the village, it requires great diligence and attention not to damage an FRP and associated gear, such as nets, during operations. In contrast, a hired crew only seeks to maximize catches and according to respondents it cannot be held liable for any damage to the gear or boat. Thus, all respondents agreed that only boat-owning fishermen can make the operation of an FRP economically viable by opti- mally resolving the trade-off between maximizing daily catches and harming the gear. Apparently this trade-off is less severe for substantially larger vessel, like the trawlers in Kerala. We now turn to the marketing of daily fish catches. For catamarans, there is interlink- ing of credit and marketing in our village and it takes the same form as in Platteau's study villages. The auctioneer gives a loan for the purchase of the gear, which at the time of our 2004 interview was about Rs. 15,000 and 25,000. In return, the boatowner has to sell all his daily catches through that auctioneer, who keeps 5 percent of the value of the sales. The boatowner does not repay the principal. As a consequence, the commission comprises a compensation for the marketing services as well as an implicit interest payment on the amount owed. It thus comes as no surprise that more succesful boatowners are granted a larger loan. When a boatowner switches auctioneers, the new auctioneer settles the debt with the previous one. Switching of auctioneers does occur occasionally, in both our and Platteau's data. The superiority of this interlinked share arrangement over separate debt and marketing contracts is a result of, first, limited liability of the fisherman and, second, costless monitoring of the fisherman's day-to-day success by the auctioneer. While marketing and financing became de-linked when trawlers were introduced in Kerala, the contract is still observed in Tamil Nadu with a slight modification. In this new contract, in addition to a commission of 7 percent, the auctioneer keeps another 10 percent of daily sales, which he deducts from the principal owed by the boatowner. Unlike a catamaran owner whose level of debt remains constant, an FRP owner asks his auctioneer for additional funds from time to time. If these additional loans are granted, they bare no interest and are added to the fisherman's outstanding balance. The emergence of this feature of debt reduction and repeated renegotiation can be explained by the following two 11 reasons. First, fibre boat fishing consumes more working capital, such as nets. To cover these costs, the owner of an FRP has to incur expenses between Rs. 5,000 and 20,000 from time to time. Second, since the FRP is a new technology, each individual's ability to operate it is not precisely known initially. Since the auctioneer's cash-flow directly depends on the fisherman's day-to-day success, however, the debt reduction component allows the auctioneer to drive down the debt level of an ex-post unsuccessful fisherman to a level at which the auctioneer's opportunity cost of capital does not exceed his commission income. This latter aspect is the subject of a companion paper (Gin´e and Klonner, 2005). We finally discuss the structure of labor contracts. On catamarans, in Platteau's as well as our study village, typically at least two members (two brothers or father and son) of the family which owns the vessel sail on the boat. The rest of the crew consists of laborer- fishermen who earn a daily minimum wage and a percentage of the value of catches. To ensure daily availability of non-family labor, owners of catamarans often tie laborers by advancing interest-free credit. While Platteau finds that technological change resulted in a de-linking of not only marketing, but also labor and credit contracts, the labor-cum- credit arrangement still prevails among FRP-owning households in our sample. At the same time, FRP fishing is carried out as a family business in the same way as catamaran fishing. 4 A Simple Model of Credit Constraints to Adoption of a New Technology In this section we provide a very simple model of credit constraints that delivers a positive relationship between wealth and the propensity to adopt the new technology. Given the discussion in Section 3, we assume that agents only have access to a savings technology to accumulate assets. Agents can produce with a traditional technology (catamaran) that yields yC or invest in a more profitable technology (fibre boat) that produces yF in expectation. The fibre boat can be purchased at cost K. Since there is no possibility of borrowing, the investment K must come from own resources. In line with Section 3 we 12 may think of K as the cost of the boat net of the loan from the auctioneer. Likewise, yF is income net of debt repayment and commissions charged by the auctioneer. Agents accumulate assets in the following manner: at = yt - ct + (1 + r)at, +1 where r is the interest rate on the savings and at is the level of assets or liquid wealth in period t. We assume that agents start in the first period with an endowment of assets a0. We further assume that the fibre boat is productive enough so that yF > yC + rK. (1) To keep things simple we assume that agents are risk neutral, live infinite periods and discount the future at the rate 1 . Each period, a household has to decide whether to 1+r purchase the fibre boat and how much to save for the following period. More formally, a household's task is to choose the vector of next period's assets {at } and the adoption +1 date t to 1 t max ct {at },t +1 1 + r t=0 s.t at = yt - ct + (1 + r)at - i{t = t}K, +1 at 0, a0 given, +1 where the function i{·} is an indicator function. With the assumed risk neutrality, the agents' discount rate, and the equation in 1, we obtain that it is always profitable to save until the new technology can be purchased.4 Since the marginal utility is linear, agents do not resent zero consumption. The agent will then optimally decide not to consume and save the entire income until enough assets 4Adoption is preferred to non-adoption if 1 t yF > yC + ra0. 1 + r Substituting for t and simplifying we obtain that yF > yC + rK. 13 have been accumulated to purchase the fibre boat. We then have that ct = 0,t < t, at = K, ct = yF,t t and ln rK+yC ra0+yC t = . ln(1 + r) By differentiating the optimal adoption time t with respect to the different parameters of interest, it is easy to see that the higher the initial level of assets a0, the higher the income from the catamaran yC, and the higher the interest rate r, the lower will be the adoption time t. If agents are risk averse instead of risk neutral, then t also depends positively on yF, so that the expected change in income resulting from the technology shift matters. 5 Estimation In this section, we seek to empirically identify the determinants of the timing of technology adoption. As developed in the previous section, a risk-neutral fisherman seeks to adopt the new technology as quickly as possible when he expects the technology switch to increase his income. An important explanatory variable for the adoption decision is therefore the expected change in income resulting from the technology shift. If expectations are unbiased, the ex-post change in observed income for fisherman i can be interpreted as a (most likely noisy) realization of i's expectations. We will therefore first derive the income change for each fisherman who adopted a fibre boat before the interview date and use these results in the subsequent analysis of the timing of adoption. 5.1 Estimating the Income Change from Adoption The goal of this section is to provide estimates of the average income that a fishing household earns with the old and new technology. With the share system that exists in the village for the compensation of both laborers and the capital obtained from an auctioneer, household income is expected to be roughly proportional to monthly fish sales generated by that household. Since both catch quantities as well as daily fish prices are 14 subject to substantial fluctuations, however, the following analysis aims at netting out the individual-specific component in how successfully each technology is operated by a given household. Moreover, we have to allow for the possibility of both individual and social learning when the new technology is used. Learning by doing implies that individual's catches trend upwards after adoption as the individual learns how to use the new technology more efficiently over time. Social learning (or learning from others), on the other hand, implies that an individual can use the expertise other individuals have acquired with the new technology to become more efficient himself. Quite generally, the latter implies that the "learning curve" of an individual, that is his success as a function of time since adoption, depends on the amount of information available at the time he adopts. More specifically, we expect the learning curve of later adopters to be flatter since they start out with relatively more information at the time of adoption. With monthly sales data from 43 fishermen who switched to a fibre boat before the date of the interview, a test for individual as well as social learning is thus facilitated by the regression specification log(ysit) = µsi + t + i{t ti} 1i + 2i + 1tii + 2tii + usit, 2 2 (2) ysit denotes monthly sales (in Rupees) of fisherman i in month t who currently operates technology s, where s = C for catamaran and s = F for a fibre boat. Also consistent with the notation in the previous section, ti denotes the time of adoption by individual i, and i denotes time since adoption, so that t = ti + i. Variable µsi is an individual-specific, technology-dependent intercept term or fixed effect, while t is a month-specific dummy that picks up aggregate fishing conditions and shocks. Finally, usit is an i.i.d. error term with E[usit] = 0. This parametrization assumes that shocks affect sales generated through the old and new technology identically in a proportional sense. This is strictly true as far as price fluctuations (per kg of fish) are concerned as the price indices faced by catamaran and fibre boat fishermen are the same. Whether it is also an appropriate assumption for weather shocks remains an open question. It is to be expected, however, that at least the 15 sign of the shock works in the same way for both types of technology. While specification (2) does not allow for learning by fishermen who are operating the old technology, which has been used over several decades, the term 1i + 2i allows for 2 learning by doing. In particular, we would expect 1 to be larger and 2 smaller than zero, which, at least in the increasing section of the inverted U-shape curve, would indicate positive, but decreasing marginal returns to learning by doing. The term 1tii + 2tii 2 captures the possibility of learning from others by allowing for a different shape of the learning curve for later adopters. Here time since adoption is interacted with a proxy for the amount of information available at the time of adoption by individual i, namely the time that had passed between the first adoption in the village and the adoption of the individual in question. In particular, we would expect the individual learning curve for a later adopter to be flatter since he starts out with more information in hand than any adopter before him. A test for the hypothesis of no social learning can thus be implemented by testing the composite hypothesis HS : 1 = 2 = 0. Analogously, a test for the hypothesis of no individual learning (learning by doing) is given by HL : 1 = 2 = 0. The results of the estimation of Equation 2 together with F-test statistics for HS and HL are shown in Table 1. According to these results, the null hypotheses of no social and no individual learning are rejected, at least at the 10% level. According to the point estimates of 1 and 2, the first adopters in the village experience an increase in sales for roughly the first ten months with the new technology.5 The estimate of 1 on the other hand implies that the individual learning curve starts out flat for a fisherman who adopts a fibre boat 12 months after the first adoption in the village (the absolute value of 1 equals roughly one twelfth of 1). Since the results for the parametrization of the specification in (2) imply "negative 5This is obtained by calculating the maximum of the parabola implied by 1 and 2,1 . 22 16 learning" for part of the sample, we use the insights from the previous estimation for deriving a more restrictive econometric specification, in which there is (positive) individual learning before some cutoff date and none of it afterwards. More specifically, we estimate log(ysit) = µsi + t + 1i{t ti}Di() + usit, (3) where i Di() = if t < t0 + max(0,t0 + - ti) if t t0 + . Here t0 denotes the month of the first adoption in the village while is a cutoff month (counted from the time of the first adoption in the village), after which no increase in individual sales occurs. The shape of the Di function can be explained simply: for fisher- men who adopted no later than months after the first adoption in the village, Di equals a straight line with slope one before date t0 + . From t0 + onwards, it remains at the level attained in that period. Unfortunately, does not lend itself to estimation by OLS methods. Instead, as sug- gested by the literature on structural breaks in econometric models, we estimate Equation 3 for all possible values of and pick the value that yields the highest R2 fit statistic. Following this procedure, we arrive at a value of = 5, which implies that the process of learning about the new technology within the village comes to an end after roughly half a year. This is not surprising given that, in contrast to the duration of an agricultural cultivation cycle, fishing is a daily, and thus a high-frequency activity. Moreover, a value of = 5 can be reconciled with the estimates of Equation 2. Notice that the quadratic function used there is downward sloping for high values of i and thus leads to an upward biased estimate of the duration of learning if the learning curve is in fact flat for high values of i. The results for Equation 3 with = 5 are displayed in Table 2. The estimate of 1 is positive and significantly so, suggesting an initial 11% monthly increase in sales for early adopters. The results for the individual-specific fixed effects, µsi, are graphically depicted in Figure 1. For fixed i, each of the 25 data points has abscissa equal to µCi and ordinate 17 µFi. Notice that, for those fishermen who adopted before t0 + 5, 1Di(5) has been added to µFi. The diagram thus gives the long-run expected gains from technology adoption, which will also be used throughout the rest of this paper. The straight line depicts the 45 line. According to these results, three fishermen suffered a loss in sales of more than 1%, 2 experienced virtually no change (less than 1% change), while 20 enjoyed increases in average sales between 3.5 and 158%. The average change equals 40.2% with a standard deviation of 46.8%. 5.2 Determinants of the Timing of Adoption Since adoption in the context of this study can be interpreted as a one-time transition from one state, catamaran fishing, to another state, fibre boat fishing, the timing of the individual adoption decision is most suitably modelled using methods from the statistical analysis of survival data. In typical applications of these methods, the outcome variable is the time of a "failure", e.g. of a light bulb, conditional on certain characteristics of the item considered (e.g. the manufacturer of the bulb). In our application the "failure" event is the adoption of the new technology. In cases where the technology is divisible, like the adoption of new seeds in agriculture, the farmer with several plots can choose on how many of them to try the new technology. In contrast, a fishing boat is by nature an indivisible productive asset for a household. Moreover, switching technologies is expensive, while in agriculture, a farmer can return to the old crop variety in subsequent growing cycles without incurring a cost from switching back. To summarize, in the context of adoption of new crop varieties in agriculture, the adoption decision is typically both divisible and reversible, while in the present setup, neither of these two properties hold. For the estimation, we adopt the common proportional hazard assumption. According to it, the hazard , that is the probability that i adopts within the next period given that he has not adopted yet, can be factored into a baseline hazard function, which is the same for all individuals in the population, and a function of individual characteristics, xi. 18 Specifically, it is assumed that i(t) = 0(t)exp(xi), where is a vector of parameters. From this structure of individual hazard, the likelihood of each observed adoption time can be derived as a function of the adoption time ti, xi and . An expression for the likelihood can be obtained regardless of whether or not adoption occurred before the date of the interview. When the latter is true, the observation is known as "censored". Using Cox's (1975) semiparametric method of partial likelihood, maximum likelihood estimates of can be obtained numerically without making any functional form assumptions about the shape of 0(t). An individual with characteristics xi has a hazard higher than the sample average if she is more likely to adopt earlier than the average of the sample because she faces a higher probability of switching at any time t after date zero, conditional on not having switched already before t . The sign of the relationship between an explanatory variable, xik say, and the outcome variable ti thus goes the opposite way from an OLS model in which adoption time is regressed on xi: in the proportional hazard model, a positive value of k implies that an individual with a higher value of xik faces a higher probability of making the transition at any given point in time, and thus reduces the expected value of his adoption time, ti. In the OLS model, in contrast, a positive value of k implies that an individual with a higher value of xik adopts later in expectation. From the model of the previous section, one key explanatory variable of interest is the income gain that an individual expects from the transition. Recall that, in our simple model, an individual starts saving to finance the new technology as quickly as possible only if the expected net gain from adoption is positive. Unfortunately, the researcher does not observe individual expected net gain but only a measure of realized net gain, which can be retrieved from µFi and µCi. We interpret realized net gain as a proxy for expected net gain. More specifically, when individual expectations are unbiased, realized net gain equals expected net gain plus a random error term which has expectation zero. Define yi = exp(µF) - exp(µC) 19 as the proxy for expected net gain in absolute terms. When yi is included as a regressor in the vector xi, one potentially faces the problem of a contaminated regressor. The applicable explanatory variable is expected net gain while the variable used is a noisy realization of it. We are thus facing a problem analogous to the one of errors in variables in a linear regression model. The extent of the estimation bias induced by this problem depends of course on how accurate individual expectations are. If individuals can perfectly predict the actual income change, the use of yi as explanatory variable is valid. The wider realized gains are distributed around expected gains, the more severe the bias introduced by using yi. There is yet another reason for why yi is only a noisy measure of the expected income change. Individual shocks may not be i.i.d. in each month, but rather be correlated. For example, if a fisherman falls unexpectedly sick for an extended period of time right after purchasing a fibre boat and this reduces his ability to go fishing, yi underestimates his expected gains. For both of these reasons, we will experiment with two specifications in the empirical analysis. One where yi is included in the xi vector without modification and one where, in the spirit of the two stage least squares model, yi is first regressed on a vector of instruments and its predicted values yi say, are used instead. Indirect evidence for the "noisiness" of yi is provided by the fact that our estimates of yi are negative for one fifth of those households for which both catamaran and fibre sales data are available. For these households, individual rationality seems to be violated as they adopt although they expect smaller profits from the new technology. To give an idea of how income change and adoption time are empirically related, Figure 2 plots ti over yi. If yi is an accurate measure of expected gains, unconstrained economic efficiency dictates that all households which realize a positive income change adopt immediately while those with a negative yi never adopt. When funds available to the fishing village are limited, constrained economic efficiency dictates that households which realize a positive income change adopt in decreasing order of yi. While there is some negative correlation between ti and yi (the correlation coefficient equals -0.04), this relationship is apparently rather weak. This finding may provide a lower bound on the 20 cost of external funds as boatowners do not find it profitable to borrow from elsewhere to obtain the higher income earlier. Alternatively, it may be seen as support of households behaving as if they were risk neutral, because as explained in Section 4, adoption will occur as long as the fibre boat is sufficiently productive. Another set of key explanatory variables refers to the capital market conditions a household faces. Here we consider two categories, income and asset variables. Within the first one, yCi = exp(µCi), average sales generated with the old technology, proxies a household's income stream before adoption. If the technology switch requires own funds that are not present when the new technology becomes available, a household with higher yCi will be able to accumulate the required own funds faster. A significant negative relationship between yCi and ti can thus be taken as evidence for a credit constraint faced by an income-poor household. Another income variable that will be used is the number of household members who earns an income. The second one, the value of the house at the time when the new technology became available, is an important component of the assets a household can collateralize to obtain credit. A significant negative relationship between a0 and ti can thus be taken as evidence i for a credit constraint faced by an asset-poor household. Other variables that will be initially included are household size, literacy, age, both linear and squared, and years as boatowner as a measure of experience. Some descriptive statistics relating to these variables are set out in Table 3. The sample used in this analysis is restricted by the availability of monthly catches using the traditional technology. When we visited the village in 2002, only two auctioneers had records that went back to 1999. Table 4 gives the results of the estimation of the determinants of adoption timing. Notice that Cox's method of partial likelihood does not identify an intercept term. Column 1 gives coefficient estimates together with asymptotic p-values for the full set of regressors, including yi not instrumented. For the three censored observations in the sample used for this estimation we had to impute values of yi. These are obtained by regressing yi of the available 23 uncensored observations for which we have both yCi and yFi = exp(µFi) on house value, yCi, age, age squared, literacy and number of crew members who belong to the extended family, and using the estimated coefficients to generate predicted values 21 of yi for the three households in the sample that had not adopted before the date of the interview. At conventional significance levels, only the value of the fisherman's house is a signifi- cant determinant of the timing of fibre boat adoption. The positive sign of the coefficient means that a wealthier (in terms of assets) household is more likely to adopt the new technology earlier. Of the two variables that proxy for the income status of the house- hold, yCi is significant at the 12% level while the number of family members who earn an income is insignificant. The same applies for household size and age. A Wald chi-square test of the hypothesis that both age coefficients are equal to zero fails to reject with a p-value of 0.58. Interestingly, the sign of the coefficient on literacy implies that a more literate head of household is likely to adopt later, which is in contrast with the findings of other literature on the subject (e.g. Bandiera and Rasul, 2004), which finds a positive relationship between literacy and the pace of adoption. Younger (and relatively poorer) boatowners tend to be more literate than older (but relatively richer). Thus, although work ability may be important, formal education is not. Column 2 gives coefficient estimates for a specification that uses predicted values of yi, yi, for the entire sample. As elaborated above, the concern addressed with this methodology is that there are reasons to believe that yi is a noisy realization of the income change expected by an individual. The problem, however, is to find good instru- ments for yi that do not affect the timing of adoption directly. The best one we could find in our data is the number of crew members employed by the head of household who belong to the extended family. It is, however, still a rather weak instrument. The only two noticeable changes with this estimation procedure are, first, that yCi is now substan- tially less significant and, second, that our measure of experience, years as boatowner, becomes more significant. Finally, the Wald chi-square test of the hypothesis that both age coefficients are equal to zero fails to reject with a p-value of 0.92. Guided by the findings of specifications 1 and 2 and in regard of the fact that the sample underlying this estimation is small, we also estimate a restricted version where the four least significant explanatory variables are omitted. According to column 3 of Table 4, both asset and income poverty significantly delay adoption. Households with a 22 greater realized income gain are likely to adopt earlier, but this relationship is significant only at a level of 0.16. Literacy continues to negatively affect the timing of adoption while, as before, more experience comes together with earlier adoption. Column 4, where the income change is instrumented, confirms these findings. As in the full specification, instrumenting mainly affects the coefficient on yCi, which ceases to be significant at conventional levels with this specification. To summarize columns 1 through 4, we find compelling evidence that asset poverty delays adoption and mixed evidence that income poverty does so as well. On the other hand, households that can expect a larger income change from adoption are not more likely to adopt earlier. 5.3 How Wealth Affects Technology Adoption We now discuss in some detail how asset wealth affects the timing of adoption. We start by considering the arguments of Besley and Case (1994) and Foster and Rosenzweig (1995) that asset wealth accelerates adoption because land-rich households enjoy higher intertemporal benefits from experimentation due to their larger scale of operation. In our sample, in contrast, each household operates exactly one boat before and after the switching of technologies, so that we can safely discard the scale argument. Another channel we can confidently rule out is that wealthy households adopt earlier because of better access to the labor market. In the setup studied here, the same amount of labor is employed to operate the old and the new technology. Each household in our sample which adopts the new technology has operated the old technology before and thus already secured the amount of labor needed for the new technology. What about better access of wealthier households to the new technology? Each house- hold in the sample obtained its FRP from the nearby branch of a domestic FRP manufac- turer. That branch is less than 4 kilometers away from the village and no transaction costs for transportation are incurred from the purchase. Moreover, according to villagers, there has never been a supply constraint ever since the new technology has become available in 2000. It can thus be ruled that wealth works through overcoming a supply constraint or having enhanced access to the new technology. We next examine the relationship between initial wealth and risk-bearing attitudes. It 23 is commonly believed that preferences for risk bearing crucially depend on a household's wealth. In particular, under the plausible assumption of decreasing absolute risk aversion (DARA), households above a certain wealth level choose to incur a given lottery with positive expected payoff while households with wealth below that level choose to stay away from it, although they could accumulate assets to later choose the lottery. Apparently, adoption of an FRP entails two forms of risk. First, the amount of fish catches fluctuates from day to day depending on weather and maritime conditions as well as individual luck. The question, however, is whether these fluctuations are more severe with an FRP than with a catamaran. To obtain an answer, we run the regression log(ysit) = µsi + t + usit separately for s = C and s = F. The resulting root mean squared errors are 0.66 and 0.50, respectively. Thus, controlling for scale by considering the natural logarithm of sales, operating an FRP entails a smaller month-to-month risk than a catamaran. While it may be argued that daily catches may exhibit different volatility patterns across technologies than monthly ones, it is not likely that those are particularly relevant as informal insurance arrangements seem to be prevalent in these villages. In this connection, boatowners report that they can easily obtain a short-term consumption loan from their auctioneer to compensate for a series of bad catches. Second, as argued in the previous subsection, uncertainty about how the gains from using the new technology is likely to be important. This together with the DARA as- sumption can explain later adoption by poorer but ex-post equally successful households. This explanation competes with the remaining one of credit constraints. While we cannot provide a definite answer in favor of either one of the two, we can provide some evidence in favor of the latter. Our survey asked each boatowner the following question: "Why did you wait (are you waiting) to switch to an FRP? [give the most important reason]". By far, the two most frequent answers were, first, "It required a lot of capital", and second, "I was uncertain about the benefits". Table 5 gives some statistics relating to the characteristics of the respondents by their answer to this question. The pattern we 24 find is as follows. First, the capital requirement is mentioned more than 50% more often than benefit uncertainty. Second, wealth among those who cite benefit uncertainty as the main reason is on average more than 25% higher than among those who mention the capital requirement first. This clearly suggests that the capital constraint is more severe for poorer entrepreneurs, in fact to such an extent that it dominates the concern about benefit uncertainty, even though that latter concern is also of greater importance to poorer decision makers when DARA is postulated. While the difference in asset wealth across answers is on the order of 30%, this difference is not statistically significant. In that light, we do not have statistically significant, albeit economically important, evidence for the assertion that a lack of wealth affects the timing of adoption mainly through limited access to capital, and not through aversion to benefit uncertainty or any of the other channels discussed here. 6 Simulation The results suggest that asset poverty delays adoption of the capital-intensive technology. To be more precise, between two households that expect the same increase in average income from adoption, the one with a more valuable house is more likely to adopt first. In this section, we address the policy-relevant question of how alternative distributions of wealth, as measured by house value, can change the diffusion of the technology. We focus on the relationship between the wealth distribution, which will be affected by the different economic policies considered, and the outcome variables mean income (within the sample) and income inequality. To conduct simulations, we first need to specify a baseline hazard function, 0(t). We make the assumption of a constant baseline hazard, 0(t) , given the small sample we have. Moreover, we consider a situation in which each household 25 adopts exactly at the expected value of its adoption time, ti = E[ti|xi], which is of course a function of . With a constant baseline hazard, we obtain ti = e- /. xi Finally, the parameter is calibrated as follows. In our sample, three households have not adopted before the interview date. We thus choose such that the date of the last adoption recorded before the date of the interview matches the fourth to last adoption date in the data simulated with the actual values of xi. Figure 3 plots actual and simulated mean income. Notice that actual mean income uses all ysi for fixed t, that is yFi (yCi) enters the average when household i has (not) adopted before date t. More formally, actual mean income is computed as 1 n n (i{t < ti}yCi + i{t ti}yFi), i=1 where i{·} denotes again the indicator function. The formula for predicted mean income is given by the same expression, except that ti is replaced by ti. The predicted data is generated from the specification of Column 3 in Table 4. Without reproducing the results separately, we note that the shape of the predicted graph remains qualitatively unchanged when the instrumented version, Column 4 in Table 4, is used instead. According to the solid line in Figure 3, there are three obvious "waves" of adoption: at the beginning, then just before one year later, and finally a little more than two years later. Notice that the solid line ends at the 36th month, the last date for which we have data. Our simulation model appears to capture satisfactorily the main features of the data, though the predicted path is smoother than the stair-shaped pattern in the actual data. According to the simulation, the last household in the sample adopts 54 months after the technology has become available. At that time, predicted average income has increased by about 39%. 26 Figure 4 depicts the Gini index of estimated actual incomes and the Gini as predicted by the simulation model. Notice that inequality during the adoption process exhibits the familiar inverted U shape. This reflects, first, that on average adopters experience a substantial increase in income and, second, that it is not the initially income-poor who adopt first because in that case adoption would narrow the income gap between the initially income-rich and poor. In the data, we see an increase of the Gini from 0.34 to 0.38 during the first wave of adoptions. The second wave of adoptions a year later leaves inequality virtually unchanged, while the third wave results in a drop of the Gini of about 20% to a level of 0.31, which is substantially lower than the value that prevailed before the new technology was known. All in all, while the village experiences a substantial increase in inequality over a course of two years, the availability of the new technology can hardly be criticized for its long-term impact on the village economy since, at the same time, average income increases and inequality decreases in a substantial way. Again the prediction satisfactorily captures the main features of the data. It correctly predicts the jump in inequality induced by the first wave of adoptions. The consequences of the second and third wave, however, are less clearly distinguishable in the simulated data, because, according to the dotted line, inequality gradually decreases from the eleventh month onward. The last predicted adoption in the 54th month leads the village to a Gini of 0.285, which is sixteen percent lower than the one at date zero, where all households operate the old technology. We now turn to the simulated policy counterfactuals. We first investigate the con- sequences of redistributive policies. Toward this, we assume that each household in the sample holds just the mean level of wealth observed in the data, i.e. owns a house worth Rs. 75,380. In such a scenario, the credit constraint is loosened for households whose wealth is below average and tightened for the rest. If the relationship between wealth that can be collateralized and the extent to which a household is credit-constrained is concave, we expect adoption to occur more promptly on average with such a policy in place. The results for mean income and the Gini are plotted in Figures 5 and 6, respectively. Ac- cording to Figure 5, equal redistribution does in fact result in a quicker adoption process. According to the simulation, the last adoption occurs a year earlier, in the 42nd instead 27 of the 54th month, than with the actual wealth distribution. The effect on sales over the course of the adoption process, on the other hand, is rather small. With an equal asset distribution, simulated sales never exceed predicted actual ones by more than 7 percent. Moreover, when we focus on differences between simulated and predicted actual sales of more than 3%, simulated sales never lead predicted actual ones by more than five months. According to Figure 6, a similar picture emerges for the dynamics of inequality. While the inverted U contracts by about 20% toward the origin, the change in the general pattern of inequality as measured by the Gini can hardly be judged economically significant. A second set of simulations investigates two extreme scenarios. The first one assumes that each household in the sample holds only the smallest observed wealth, that is each house is assumed to be worth Rs. 20,000. The second one, in contrast, assumes that each household in the sample holds the highest observed wealth, that is each house is assumed to be worth Rs. 500,000. The results for this set of simulations together with the predicted actual values are set out in Figures 7 and 8. We thus consider situations in which all households are either tightly credit-constrained or virtually do not face a credit constraint at all. The mean income and inequality paths for the first simulation very closely follow the respective paths generated from the actual asset data, which suggests that the observed income pattern accompanying the introduction of the new technology closely resembles a situation in which all households are substantially credit-constrained. The results for the second simulation, where the credit constraint is released for the entire sample, are more striking. The dotted lines in Figures 7 and 8 suggest that with uniform asset wealth the adoption process is completed in just five months. As a con- sequence, the village enjoys a substantially higher mean income for about two years by which adoptions in the simulated data lead predicted actual ones. This result suggests that a community in which households face virtually no credit constraints is able to move up the technology ladder much faster than the one investigated by this study. Similarly, only a minor spike remains of the observed pronounced inverted U shape of inequality. 28 7 Conclusions This paper studies the diffusion of a capital intensive, indivisible technology and analyzes the dynamics of income inequality during this process. We find evidence that credit constraints are the key factor that explains the pattern of delayed adoption observed in the data. During the diffusion process, inequality follows Kuznets' well-known inverted U-shaped curve. Consistent with the model and results, a redistributive policy favoring the poor results in accelerated economic growth and a shorter duration of sharpened inequality, although the quantitative impact of such a policy is small. One advantage of this paper over other studies about the impacts of globalization, is that context is well understood. Thus, the specific channel in which wealth matters for adoption can be disentangled. We conclude, like Platteau (1984), that overall our village in Tamil Nadu experienced a success story of globalization: according to the simulated data, the diffusion for the entire sample is completed in less than five years and it results in substantial income gains for everyone. References [1] Bandiera, O. and I. Rasul, 2004. Social Networks and Technology Adoption in Northern Mozambique, unpublished manuscript, London School of Economics. [2] Bell, C. and T. N. Srinivasan, 1989. Some aspects of linked product and credit con- tracts among risk-neutral agents. In P. Bardhan, editor, The Economic Theory of Agrarian Institutions. Oxford: Clarendon Press. [3] Besley, T. and A. Case, 1993. Modeling Technology Adoption in Developing Coun- tries, American Economic Review 83, 396-402. [4] Besley, T. and A. Case, 1994. Diffusion as a Learning Process: Evidence from HYV Cotton, Princeton University Department of Economics Working Paper 5/94. 29 [5] Binswanger, H., J. Dayantha, T. Balaranaia and D. Sillers, 1980. The Impacts of Risk Aversion on Agricultural Decisions in Semi-Arid India, working paper, World Bank, Washington DC. [6] Conley, T. G. and C. Udry, 2003. Social learning through networks: The adoption of new agricultural technologies in Ghana, American Journal of Agricultural Economics 83, 668-673. [7] Conley, T. G. and C. Udry, 2004. Learning about a new technology: Pineapple in Ghana, working paper, Yale University Economic Growth Center. [8] Cox, D. R., 1975. Partial Likelihood, Biometrika 62, 269 - 276. [9] Feder, G., R. E. Just and D. Zilberman, 1985. Adoption of agricultural innovations in developing countries: a survey, Economic Development and Cultural Change 33, 255-297. [10] Feder, G. and G. T. O'Mara, 1982. On information and innovation diffusion: a Bayesian approach, American Journal of Agricultural Economics 64, 145-147. [11] Foster, A. and M. Rosenzweig, 1995. Learning by doing and learning from others, Journal of Political Economy 103, 1176-1209. [12] Gin´e, X. and S. Klonner, 2004. Dynamic Lending with Limited Commitment and Uncertain Borrower Types: Theory and Evidence, working paper, Cornell University Department of Economics. [13] The Hindu, 2001. Boon for small-scale fishermen, The Hindu, March 22, 2001. [14] Kurien, J., 1994. Kerala's marine fisheries development experience. In B. A. Prakash, editor, Kerala's economy: Performance, problems, prospects, 195-214. New Delhi: Sage. [15] Munshi, K., 2003. Social Learning in a Heterogeneous Population. Technology Dif- fustion in the Indian Green Revolution, unpublished manuscript, Brown University. 30 [16] Platteau, J.P., 1984. The Drive Towards Mechanization of a Small-Scale Fisheries in Kerala: A Study of the Transformation Process of Traditional Village Societies, Development and Change 15, 65-103. [17] Sandven, P., 1959. The Indo-Norwegian Project in Kerala, Norwegian Foundation for Assistance to Underdeveloped Countries, Oslo. 31 Table 1. Estimation Results for eq. 1. Variable Parameter Standard t Value Pr > |t| Estimate Error i 0.03852 0.01842 2.09 0.0366 i 2 -0.00187 0.00109 -1.72 0.0862 t*i i -0.00305 0.00131 -2.33 0.0199 t*i i2 0.00000511 0.00003560 0.14 0.8859 F Value Pr > F Test of HS 4.34 0.0132 Test of HL 2.48 0.0842 R-Square 0.6943 No. of obs. 1471 Coefficients for 60 monthly dummies and 30 individual-specific fixed effects for catumeran-operating fishermen as well as 42 individual-specific fixed effects for fibre boat-operating fishermen not reproduced Table 2. Estimation results for eq. 2. Variable Parameter Standard t Value Pr > |t| Estimate Error Di(5) 0.11175 0.03868 2.89 0.0039 R-Square 0.6924 No. of obs. 1471 Coefficients for 60 monthly dummies and 30 individual-specific fixed effects for katumeran-operating fishermen as well as 42 individual-specific fixed effects for fibre boat-operating fishermen not reproduced Table 3. Descriptive Statistics for the Core Sample Variable Mean Std Dev Minimum Maximum Value of House (in thousand Rs.) 75.38 97.74 20.00 500.00 Family members with Income 2.00 1.01 1.00 5.00 Average Sales before Adoption 22052.45 15860.84 5497.34 76017.63 Change in Sales from Adoption 8419.69 10550.01 -8750.07 48339.28 Household Size 6.42 3.03 3.00 17.00 Literacy of Household Head* 0.38 0.49 0 1.00 Age of Household Head 38.46 12.12 21.00 65.00 Years as Boatowner 10.57 5.06 3.00 20.00 Adoption in month** 36.91 8.88 25.00 51.00 * equals one if he reports that he can read or write, and zero otherwise. ** for those households that had adopted before the interview, which took place in the 62nd month. Table 4. Adoption timing Estimation. Dependent variable: month of adoption (1)* (2) (3) (4) Value of House 0.00525 0.10697 0.00557 0.00528 (0.070) (0.064) (0.022) (0.026) Family members -0.17937 0.00521 with Income (0.777) (0.860) Average Income 0.0000429 0.06286 0.0000414 0.0000287 before Adoption (0.122) (0.439) (0.057) (0.158) Income Change 0.0000506 0.0000187 0.0000346 -0.0000074 from Adoption (0.151) (0.786) (0.161) (0.912) Household Size 0.03444 -0.0000289 (0.868) (0.760) Literacy of -0.86901 -0.72087 -0.73734 -0.79233 Household Head (0.156) (0.246) (0.173) (0.138) Age of Household -0.22363 -0.02748 Head (0.311) (0.919) Age Squared 0.00273 0.0004796 (0.298) (0.878) Years as 0.10930 0.12437 0.11890 0.14000 Boatowner (0.222) (0.149) (0.105) (0.065) Log-Likelihood -47.9 -48.9 -48.5 -49.3 Income Change No Yes** No Yes** Instrumented Number of Obs. 26 26 26 26 No. of Obs. 3 3 3 3 Censored * Asymptotic p-value in paretheses ** Instruments: Age, age squared, years as boatowner, number of crew members who belong to the extended family Table 5. Self-reported reason for delay of adoption within the core sample. Answer N* Mean Std Dev Minimum Maximum Capital Requirement 13 69.2 63.7 0 250 Benefit Uncertainty 9 95.5 152.7 20 500 * 4 respondents cited other reasons Figure 1. Individual average profitability with fibre boat over individual average profitability with katumeran for 25 households for which sales data is available for both katumeran and fibre boat fishing. MU1 12 11 10 9 8 7 8 9 10 11 12 MU0 Figure 2. Adoption date over realized absolute income change for 25 households for which sales data is available for both katumeran and fibre boat fishing. corr_dat e_f b1_num 60 50 40 30 20 -10000 0 10000 20000 30000 40000 50000 ydi f Figure 3. Mean income after the new technology became available, actual (dotted line) and predicted by the model (dashed line) meanreal 31000 30000 29000 28000 27000 26000 25000 24000 23000 22000 0 10 20 30 40 50 60 MONTH Figure 4. Income Gini after the new technology became available, actual (dotted line) and predicted by the model (dashed line) gi ni real 0. 40 0. 39 0. 38 0. 37 0. 36 0. 35 0. 34 0. 33 0. 32 0. 31 0. 30 0. 29 0. 28 0 10 20 30 40 50 60 MONTH Figure 5. Predicted actual (solid) and simulated (dotted) mean income. Simulation assumes perfectly equal distribution of wealth (measured by house value) over the sample meanpred 31000 30000 29000 28000 27000 26000 25000 24000 23000 22000 0 10 20 30 40 50 60 MONTH Figure 6. Predicted actual (solid) and simulated (dotted) Gini. Simulation assumes perfectly equal distribution of wealth (measured by house value) over the sample gi ni pred 0. 40 0. 39 0. 38 0. 37 0. 36 0. 35 0. 34 0. 33 0. 32 0. 31 0. 30 0. 29 0. 28 0 10 20 30 40 50 60 MONTH Figure 7. Predicted actual (solid) and simulated mean income. Simulation 1 (dashed) assumes the lowest observed wealth (house value equal to 20) for the entire sample, simulation 2 (dotted) assumes the highest observed wealth (house value equal to 500) for the entire sample. meanpred 31000 30000 29000 28000 27000 26000 25000 24000 23000 22000 0 10 20 30 40 50 60 MONTH Figure 8. Predicted actual (solid) and simulated income Gini. Simulation 1 (dashed) assumes the lowest observed wealth (house value equal to 20) for the entire sample, simulation 2 (dotted) assumes the highest observed wealth (house value equal to 500) for the entire sample. gi ni pred 0. 40 0. 39 0. 38 0. 37 0. 36 0. 35 0. 34 0. 33 0. 32 0. 31 0. 30 0. 29 0. 28 0 10 20 30 40 50 60 MONTH