Network Proximity and Business Practices in African Manufacturing ˚ ns So Marcel Fafchamps and Ma ¨ derbom Patterns of correlation in innovation and contractual practices among manufacturing firms in Ethiopia and Sudan are documented. Network data that indicate whether any two firms in the utilized sample do business with each other, buy inputs from a common supplier, or sell output to a common client are used for the analysis. Only limited support is found for the commonly held idea that firms that are more proximate in a network sense are more likely to adopt similar practices. Indeed, for certain practices, adoption decisions appear to be local strategic substitutes: if one firm in a given location uses a certain practice, nearby firms are less likely to do so. These results suggest that the diffusion of technology and new business practices may play a more limited role in spur- ring growth in Africa’s manufacturing sector than is often assumed in the present policy discussion. JEL codes: O1, D2, D4. Although technological upgrading and institutional innovation are critical for growth, these factors are particularly critical in Africa, where productivity has re- mained low. This fact begs the question of why productivity-enhancing innova- tions have not diffused equally to different countries or regions (Parente and Prescott 1994). Since Griliches (1958), the dominant model of technology adop- tion is one in which information about a more productive technology diffuses through the economy, and the new technology is subsequently adopted by indi- vidual firms. In this model, obstacles to the circulation of information, such as Marcel Fafchamps is a professor at the University of Oxford and a Co-Director of the Centre for the Study for African Economies; his email address is marcel.fafchamps@economics.ox.ac.uk. Ma ˚ ns So¨ derbom (corresponding author) is a professor at the University of Gothenburg; his email address is mans.soderbom@economics.gu.se. The authors thank Matt Jackson, Yann Bramoulle ´ , Alain de Janvry (Editor), three anonymous referees, and participants at seminars at the University of Gothenburg, Trinity College Dublin, the University of Wageningen, the Centre for the Study for African Economies conference Economic Development in Africa, Oxford, 2010, and the Nordic Conference on Development Economics, Copenhagen, 2011, for their very helpful comments. The authors also thank Magdi Amin for his encouragements and support, Gebrehiwot Ageba for his help with the geographical and social network data in Ethiopia, the World Bank for providing the data, and Rowena Chu and Michele Valsecchi for their high-quality research assistance. Ma ˚ ns So¨ derbom thanks Sida for financial support. A supplemental appendix to this article is available at http://wber.oxfordjournals.org. THE WORLD BANK ECONOMIC REVIEW, VOL. 28, NO. 1, pp. 99– 129 doi:10.1093/wber/lht020 Advance Access Publication July 3, 2013 # The Author 2013. Published by Oxford University Press on behalf of the International Bank for Reconstruction and Development / THE WORLD BANK. All rights reserved. For permissions, please e-mail: journals.permissions@oup.com 99 100 THE WORLD BANK ECONOMIC REVIEW social or economic segmentation, delay technology adoption. Delays may also arise because of funding constraints or adoption costs, such as learning by doing, experimentation, and adjustment costs. As a result, pockets of backward technol- ogy may remain. This general view pervades much of the economic discourse on growth and de- velopment. A form of diffusion externality is either built into or hidden in all of the endogenous growth models in which technological innovation fuels growth (Parente and Prescott 1994; Romer 1990; Grossman and Helpman 1991; Aghion and Howitt 1992). The literature on the Industrial Revolution and the rise of the Western world describes how innovations in technology and business practices diffuse to neighboring enterprises, towns, and countries (North 1973; Mokyr 1990). Similarly, the literature on agglomeration effects ascribes a key role to the diffusion of innovative technology and business practices to nearby firms (Jacobs 1969; Fujita, Krugman and Venables 1999; Muendler, Rauch, and Tocoiand 2012). Analogous ideas underlie much of the literature on the productivity bene- fits from FDI and international trade (Casella and Rauch 2002; Tybout 2000). Supplier-client relationships are considered one important channel of diffusion among firms (Jacobs 1969; Rauch and Casella 2003). Another channel is compe- tition between firms in the same market, especially foreign firms (Kraay, Soloaga and Tybout 2002). Another strand of the economics literature has examined the diffusion of inno- vations within countries and regions. A shared assumption that underlies much of this literature is that by interacting, firms learn from each other about techno- logical and institutional innovations that raise productivity. Although there is a rigorous body of research on technology diffusion among farmers (Griliches and Lichtenberger 1984; Young and Burke 2001), much of the existing literature on manufacturing firms in developing countries remains descriptive and relies prin- cipally on case studies (Sutton and Kellow 2010; Sutton and Kpentey 2012; Sonobe and Otsuka 2011). In this paper, we offer statistical evidence on the diffusion of innovations among manufacturing firms in Ethiopia and Sudan.1 Our approach is to examine whether innovative business practices are correlated more strongly between firms that are relatively close in a network or a market sense. We find some evidence for a correlation in business practices, but the evidence is less convincing than one would expect if the diffusion effects were strong. Furthermore, we find evi- dence that along some dimensions ( principally geographical distance), firms are more similar to distant firms than to nearby firms. This observation suggests that some adoption decisions are local strategic substitutes: if some firms adopt a certain practice, the incentive for other firms to adopt it appears to be reduced. This phenomenon is partly confirmed by noting that the practices for which we 1. The Sudanese data used in this paper were collected in 2007. We began work on this paper in 2009, before South Sudan seceded from Sudan in July 2011. Thus, our sample includes firms in what is now South Sudan. Throughout the text, “Sudan” means Sudan prior to the secession. ¨ derbom Fafchamps and So 101 find evidence of strategic substitution, namely, R&D and vocational training to workers, are the very practices that are the most vulnerable to free riding by other firms. Overall, the evidence for diffusion and complementarities is weaker than one might expect given the emphasis in much of the current policy discus- sion on diffusion and agglomeration economies as a source of improved firm per- formance in Africa (Collier 2007; Page 2012). This paper is organized as follows. Section 2 discusses the conceptual frame- work and some key methodological issues, section 3 describes the econometric testing strategy, section 4 provides information about the data, section 5 presents the econometric results, and section 6 concludes. I. CONCEPTUAL FRAMEWORK: DIFFUSION IN NETWORKS Consider two economic agents, i and j, in a network.2 The diffusion of a practice along the network means that i is more likely to adopt the practice if j has adopted it. This statement is equivalent to the assertion that the adoption deci- sions of i and j are strategic complements. To formalize this observation, let gij ¼ f0, 1g denote a network link between two agents i and j and define the network matrix as G ; [ gij ], where gii ¼ 0. Suppose that there are N agents. We follow Liu et al. (2012) and Bramoulle ´ and Kranton (2011) in writing the payoff of agent i as follows: 1 pi ¼ ai yi þ ggi y þ ryi gi y À y2 2 i where yi denotes the action of agent i, y ; [ y1, . . ., yN] is a vector of the actions of all of the agents, gi ¼ [ gi1, . . . giN] is a vector of the neighbors of i, the Greek letters are parameters, and the final term represents the cost of taking action yi, which is assumed to be quadratic for the sake of simplicity. Each agent chooses yi ! 0 to maximize the payoff pi. The first-order condition for an interior solution is yi ¼ ai þ rgi y: ð1Þ The parameters a i, g, and r are now straightforward to interpret: ai is a profit- ability parameter; r indicates whether the actions are strategic complements (r . 0), strategic substitutes (r , 0), or neither complements nor substitutes (r ¼ 0); and g indicates whether there are positive externalities ( g . 0), negative externalities ( g , 0), or neither positive nor negative externalities ( g ¼ 0). Note that it is possible for the externalities to be negative ( g , 0) even if the actions are strategic complements (r . 0) and vice versa. 2. A network consists of links between the nodes in a finite collection of nodes (for example, firms). See Jackson (2009). 102 THE WORLD BANK ECONOMIC REVIEW The equilibria are action vectors y that solve the system of Kuhn-Tucker condi- tions, which combine the first-order conditions (1) with yi ! 0 8 i [ N. The inte- rior solutions y satisfy y ¼ ðI À rGÞÀ1 A where A ; [a1, . . . aN]. When the actions are strategic complements (r . 0) and ai ! 0 for all i, a sufficient condition for an interior equilibrium is that r is smaller than the largest eigenvalue of G.3 If ai 0 for all i [ N, then there exists an equilibrium with y ¼ 0, but there may be other equilibria as well.4 Bramoulle ´ and Kranton (2011) characterize the equilibria that arise in network games with strategic substitutes (r , 0) and show that the equilibrium configuration ultimately depends on the lowest (that is, most negative) eigenval- ue of G. With strategic substitutes, most equilibria have some agents choosing yi ¼ 0, and (some of ) their neighbors choose a strictly positive yi (that is, the actions of neighbors tend to be dissimilar). In contrast, when actions are strategic complements, the actions of neighbors reinforce each other; thus, they tend to be similar (see also Jackson 2009). These observations form the basis of our testing strategy: let y ; y À EðyÞ ¼ ðI À rGÞÀ1 A ~ ~ where A ˜ ¼ A 2 E(A). The covariance matrix y; is of ~   yÞ ¼ E ðI À rGÞÀ1 A Covð~ ~A~ 0 ðI À rGÞÀ1 ð2Þ where the ais that enter matrix A are unobserved   by the researcher. If the ais are ~ ~ independent and identically distributed, E AA ¼ s2 I and the above expression 0 can be simplified as follows:   yÞ ¼ s2 E ðI À rGÞÀ1 ðI À rGÞÀ1 : Covð~ When the matrix G is sparse (that is, when few gij ¼ 1), the ij elements of the matrix E½ðI À rGÞÀ1 ðI À rG0 ÞÀ1 Š that correspond to the existing links ( gij ¼ 1) are approximately proportional to r2. Other elements are functions of higher powers of r and are much smaller than the elements that correspond to the linked pairs ij. In contrast, if r ¼ 0 and the ais are independent and identically distributed, then Cov(~ y) is a diagonal matrix, and Cov(~ yj ) ¼ 0 for i = j. yi, ~ Therefore, it is possible to test whether r = 0 by determining whether the values of y are more or less similar for linked pairs than for unlinked pairs. However, if 3. For this statement to be true, it is sufficient that r is smaller than one over the maximum degree of any agent (Jackson 2009). 4. To illustrate the point, let N ¼ 2, a ¼ 2 1 and r ¼ 2. If y2 ¼ 0, then the y1 ! 0 constraint is binding and y1 ¼ 0. If y2 ¼ 1 then y2 ¼ 2 1 þ 2  1 ¼ 1. Thus, we have two equilibria: ( y1, y2) ¼ (0, 0) and (1,1). ¨ derbom Fafchamps and So 103 E(A˜A˜ 0 ) is not a diagonal matrix (that is, if the ais are correlated), it is possible that Cov(~ yj ) = 0 even when r ¼ 0. This possibility is an important caveat to yi, ~ keep in mind when interpreting our results; similar practices could be due either to strategic complementarity r or to a correlation between the ais (that is, a corre- lation in the profitability of taking action y between linked firms). Manski (1993) calls such correlations contextual effects. By similar reasoning, dissimilar practices may be due to strategic substitution or a negative correlation in the ais. Strategic complementarity may arise for a variety of reasons. For example, the desire to imitate others or to conform to a social norm may be reinforced by peer pressure (Young and Burke 2001) and may result in complementarity. Another possibility is that the adoption of an innovation by others lowers the output price, which forces agent i to adopt the same innovation to remain competitive. In contrast, strategic substitution would arise if agent j takes action yj and the in- centive for individual i to take the same action weakens. For example, the possi- bility of free riding has long been recognized in experimentation; agents may wait for their friends and neighbors to experiment with a new technology before deciding whether to adopt it themselves (see Foster and Rosenzweig [1995] for an application to farming). The training of workers is another possible area where strategic substitution may be important: if firm j decides to train its workers, firm i may decide to try to poach them instead of training its own workers. In addition, a desire to avoid competition may be a driving force of stra- tegic substitution. For example, if firm j decides to design its products for a high-end market, it could be optimal for firm i to tailor its products to a low-end market. Diffusion Dynamics If information diffuses between linked agents, then, in the long run, we expect all of the connected agents to have the same information. The connection may be direct, whereby the agents are linked to each other, or it may be indirect, which implies that the agents are linked through others. This insight was initially for- malized in the context of epidemiologic models on networks (see Jackson [2009] and Vega-Redondo [2006] for excellent summaries of this literature). It follows that when information has had time to percolate through the network, adoption patterns within a giant component depend exclusively on the distribution of the benefits from adoption, namely, the ais, and on the local strategic complements and substitutes r. If agents have dissimilar ais or if r , 0, we expect sporadic adoption of business technology and practices, in which some agents adopt these things but others do not, although they all have the same information. In con- trast, if agents have sufficiently similar ais and r ! 0, we expect all of the agents in the same giant component to adopt similar technology and practices irrespec- tive of whether they are directly linked. However, the latter expectation is not true in the short run. If information circulates slowly, adoption decisions are more likely to be similar among agents who are directly linked. 104 THE WORLD BANK ECONOMIC REVIEW Business Practices Thus far, we have discussed strategic complements and substitutes in general terms. Here, we briefly discuss specific business practices for which we have data, and we speculate about whether they are more likely to be strategic complements or substitutes for manufacturing firms in a developing country. 1) Technology: The adoption of more advanced equipment and machinery is likely to be a strategic complement within a given sector and region. Because regional firms in the same sector compete with each other, they must keep up with each other in terms of productivity. However, some firms may strategi- cally choose to focus on niche products and markets that are poorly served by other firms to avoid competition (Fafchamps 1994). Such behavior may lead to differences rather than similarities in the technology decisions of firms in the same location. 2) Internal organization: Innovations in the internal organization of a firm should follow similar logic. If other firms gain a competitive edge by adopting a better organizational structure, competitors should follow suit. However, this maxim may not apply to firms that eschew competitive pressure by focus- ing on niche markets and products (see the previous item). 3) R&D: If firms compete through innovation, high R&D by some firms will induce others to invest in R&D as well. Therefore, we expect R&D to be a strategic complement unless firms can act as free riders by imitating the inno- vations of other firms or by choosing R&D strategically to avoid competition. 4) Vocational training of workers: If better-trained workers raise productivity, competition between firms will lead them to train workers if new recruits are insufficiently qualified. However, firms may free ride and hire workers who have been trained by other firms instead of providing their own training. Thus, vocational training may be a strategic complement or a substitute. 5) Contractual practices: Because contractual practices involve other firms by definition, strategic complementarities in this area are likely to be stronger. For instance, if one firm imports from abroad or subcontracts part of its produc- tion, other firms may find it easier to import or subcontract in the same way. However, we cannot a priori rule out strategic substitution, such as if firms purchase inputs from the importing firm rather than importing these inputs themselves. 6) Reputational sanctions: Because reputation sanctions contain a strong public-good component, they are very likely to exhibit strategic complementar- ity. Indeed, the threat of exclusion from future trade has the strongest deterrent effect if all of the firms in the industry participate. Hence, the incentive to adopt a reputational sanction is highest when most other firms have already adopted it. The above discussion suggests that different types of proximity may have differ- ent effects. In principle, the strategic complementarities that arise from ¨ derbom Fafchamps and So 105 information exchange apply to all of the practices listed above. If information pertaining to technological, organizational, and contracting innovations circu- lates through supplier-client relationships, we expect such proximity to matter. The strategic complementarities that arise from competition should generate the strongest similarity among firms that share the same market, such as firms in a given sector and location. This observation is most relevant for technology, inter- nal organization, and R&D because other channels of adoption diffusion are ex- pected to be less important in these areas. If upstream and downstream firms face different competitors, which is probable, the strategic complementarities that are driven by competition are expected to be smaller between firms that are located at different levels of the value chain. It follows that if we use geographical prox- imity as a proxy for competition, supplier-client proximity, which identifies dif- ferent points on the value chain, may be associated with less similar practices. Thus far, we have discussed the adoption of practices. It is also possible to in- vestigate payoffs directly, such as by analyzing firm performance and growth. In section 1 in the online appendix, we derive an expression for the covariance in profits across firms.5 We show that if g = 0 (that is, if externalities are present), then positive externalities manifest as proximate firms that have similar performance. In contrast, negative externalities imply dissimilar performance. Furthermore, we show that even in the absence of externalities, firms’ perfor- mance may be similar because of a correlation in firm-specific conditions ai and aj; these correlations are the so-called contextual effects. The potential presence of these contextual effects precludes the interpretation of correlated firm perfor- mance as evidence of externalities. Diffusion across Heterogeneous Firms Firms are heterogeneous, and diffusion patterns across firms are likely to depend upon enterprise characteristics.6 For example, the scope for the diffusion of inno- vations between sectors may be limited if these sectors use technologies that are very different. Similarly, organizational practices that are suitable for large cor- porations may not be useful for microenterprises. In the model, this discrepancy is captured by differences between firms in the profitability parameter ai. If the adoption of new technologies and innovations is dichotomous, the likelihood of adopting can be expressed as l(ai þ rgi y), where l(.) is a logit or probit function. Whereas firms with a low ai are unlikely to adopt irrespective of what neighboring firms do (that is, irrespective of rgi y), firms with a high ai are likely to adopt regardless of what others do. Thus, strate- gic complements and substitutes are very relevant for firms with intermediate values of ai; for these firms, adoption may only be beneficial if neighboring firms 5. The online appendix can be obtained at http://soderbom.net/Fafchamps_Soderbom_Online_ Appendix_2013.pdf and at http://wber.oxfordjournals.org. 6. Heterogeneity across firms has increasingly been recognized in the recent literature; see, for example, Melitz (2003) and Melitz and Ottaviano (2008). 106 THE WORLD BANK ECONOMIC REVIEW adopt (if adoption decisions are strategic complements) or do not adopt (if these decisions are strategic substitutes). It is reasonable to assume that once they have been informed of an innovation, firms with a high ai would adopt it first and other firms would adopt it later owing to rgiy effects. Therefore, we expect to observe the network-driven diffu- sion of innovation only among firms that are somewhat different, but the firms cannot be too different. Hence, the extent to which practices and technologies diffuse may vary, de- pending on the heterogeneity across firms. For instance, if all of the firms in sector A share a high aA for a particular innovation and the firms in sector B have a lower aB but a large r, we expect all of the firms in sector A to adopt new innovations and technologies irrespective of whether they are linked. In contrast, we expect the firms in sector B to be more likely to adopt the same practices if they are linked to the sector A firms. In this example, although the correlation in adoption between firms within the same sector is not affected by network prox- imity, the correlation in adoption between firms in different sectors is stronger if these firms are linked. It is also possible that firms are heterogeneous within sector A; whereas some firms may have a high ai and adopt new innovations and technologies, others may have a lower ai and adopt if and only if they have an adopting neighbor. Similarly, the firms in sector B may all have a low ai and may not adopt regardless of whether they are linked. As these two contrasting exam- ples illustrate, it is not entirely clear a priori what makes firms too similar or too different for network effects to affect diffusion. The economic importance of diffusion across heterogeneous firms is potentially high. For example, if r is small in dissimilar firms, the diffusion of innovations will be more difficult in economies that are populated by very heterogeneous firms (much of sub-Saharan Africa has this characteristic). In such a context, not much should be expected from social networks and their ability to speed the dif- fusion of new ideas. Heterogeneity is also important from a methodological point of view because if we fail to take heterogeneity into account, we will under- estimate the importance of networks for the subset of firms in which diffusion is occurring. As a result, we could erroneously accept the null hypothesis that net- works play no role, and we must remember this point when we interpret our re- gression results. II. TESTI NG STRATEGY In this section, we outline the testing strategy that follows from the above reason- À we observe whether ing. Each enterprise is a node, and Á an enterprise i has adopted a practice yi. The vector gij ¼ g1ij ; g2ij ; : : : ; gMij represents the supplier-client links between two enterprises i and j, and dij represents the geographical distance between them. We want to test whether two enterprises i and j are more likely to have a similar practice y if they are close in a network and geographical sense— that is, whether some or all of the elements of g ij are equal to one or if dij is small. ¨ derbom Fafchamps and So 107 For this purpose, we estimate models of the following form:   yi À yj  ¼ g u þ vdij þ jxi Àxj jb þ uij ð3Þ ij where u ¼ðu1 ; u2 ; : : : ; uM Þ is a vector of coefficients that are associated with network links, v is a coefficient that reflects the relationship between the geograph- ical distance and the outcome similarities, jxi À xj j is a vector of absolute differenc- es in the control variables x that is included to reduce omitted variable bias, b is a vector of parameters, and uij is an error term.7 A negative um in (3) means that y is more similar when firms i and j have a link gmij ¼ 1. For the geographical distance dij, the interpretation of the sign of v is exactly the opposite of the interpretation of the sign of um. Conversely, a positive um or a negative v would mean that linked or nearby firms are more dissimilar. If y is more similar in proximate firms, then this occurrence is consistent with a situation in which adoptions by different firms are strategic complements; in contrast, if y is dissimilar in proximate firms, then adoption by different firms may be a strategic substitute. A positive b means that firms that share a similar x tend to have a more similar y. A negative um does not by itself imply network diffusion because firms i and j may have correlated technology and contractual practices for reasons other than network or geographical À proximity, Á such as because they are subject to similar contextual effects ðcorr ai ; aj . 0Þ that are not adequately controlled for by the quantities jxi Àxj j. If these unobserved contextual effects were more strongly correlated in linked firms, they would bias um below 0. Hence, if we find a signifi- cantly negative estimate of um, the reason may be either diffusion or unobserved contextual effects. However, if um is positive or not significantly different from zero, the net effect of the diffusion and the contextual effects is likely to be posi- tive or zero. There are two possible exceptions to the above scenarios. The first exception is when diffusion is rapid and all firms belong to a single connected network. In this case, our identification strategy will fail because the similarity of firms will depend exclusively on their ais and not on the distance between them. Hence, we will observe a zero u even though diffusion across network links is taking place. The second exception is when the strategic complementarities and substitutes precisely offset each other. Although it is possible, this situation seems unlikely. If unobserved contextual effects can only generate positive correlations in tech- nology and business practices, which is likely, then a nonsignificant u indicates that the network diffusion is zero, and a positive u suggests that the presence of 7. There are two reasons to estimate (3) in terms of its absolute deviation and not in a covariance form as in (2). First, most of the outcome variables we investigate are binary; as a result, the only information they  contain  is whether yi ¼ yj. In this case, (3) boils down to a linear probability model because yi À yj  ¼ 1 if yi = yj and 0 otherwise. Second, in the more general case when y is continuous, model (3) captures the main idea behind (2) but offers the advantage of being more robust with respect to outliers compared to using ðyi À yÞðyj À yÞ as the dependent variable. 108 THE WORLD BANK ECONOMIC REVIEW strategic substitution affects adoption decisions. However, we cannot completely rule out the possibility that a negative correlation between practices could be the result of a negative correlation in the profitability of adoption ai. For instance, if an innovation such as subcontracting is profitable for upstream firms but not for downstream firms, then firms that are linked as suppliers and clients will have negatively correlated practices because suppliers, by definition, are upstream rel- ative to their clients. Equation (3) is a dyadic regression. The dependent and independent variables are defined for every pair of firms i, j in the data, which implies that there are n  (n 2 1) observations that underlie the regression (n denotes the number of firms). Dyadic observations are not independent because the residual uij is correlated with uik. To compute standard errors that are robust with respect to the correla- tion in the error term across firms, we use the bootstrapping procedure that is de- scribed in section 1 in our online appendix. I I I . D ATA To implement our testing strategy, we use detailed firm-level data that were col- lected under the leadership of the World Bank in Ethiopia and Sudan. Virtually the same questionnaire and sampling strategies were used in the two countries. The data on the Ethiopian firms were collected as part of the Ethiopia Investment Climate Survey, which was implemented by the Ethiopian Development Research Institute in mid-2006.8 The survey covered 14 major cities located in seven regions of Ethiopia, and 42 percent of the observations came from Addis Ababa. The survey included firms with at least five permanent employees in four sectors: furniture, wood, and metal; food and beverages; leather and leather products; and textiles and garments. In all, 360 manufactur- ing firms were surveyed. The data from the Sudanese firms were collected as part of the Investment Climate Survey, which was launched in November 2007 and conducted by H&H Consultancy. This company is a Sudanese management con- sulting firm with expertise in conducting complex surveys.9 Thus, the data were collected before South Sudan seceded from Sudan in July 2011; therefore, our sample includes firms in what is now South Sudan (see note 1). The survey covered 432 manufacturing firms in eight states, and most of these firms were private. The capital city of Khartoum accounted for 52 percent of the sample ob- servations. No sector represented more than 20 percent of the sample; hence, the survey was diverse in terms of sector. The largest sectors were food and beverages (18 percent) and fabricated metal products (16 percent). Microenterprises were not covered. After deleting the observations that had too many missing values, 8. See Mengistae and Honorati (2009) for details on the survey methodology. For a thorough report on the survey, see the World Bank (2009). 9. See H&H Consultancy (2008) for details on the survey methodology. ¨ derbom Fafchamps and So 109 we obtained a sample of 304 firms for Ethiopia and 401 firms for Sudan.10 This subset of the observations formed our baseline sample.11 Summary statistics are shown in table 1, and the variables that constitute our control vector are presented first. More mature firms and firms with higher-quality management should be more adept at recognizing the value of new technologies and business practices. Female ownership is included because female-headed busi- nesses have been shown to be less growth oriented (de Mel, McKenzie and Woodruff 2009; Fafchamps 2003). We also include firm size, which is represented by the (log of the) total firm employment. The average log employment is 3.37 in Ethiopia (which corresponds to 29 employees) and 2.91 in Sudan (which corre- sponds to 18 employees). Next, we report information on firm practices. Initially, we focus on the vari- ables for which strategic complementarities across firms are a priori thought to be less strong, such as innovation. We end by considering the variables for which strategic complementarities are likely to be the strongest, such as reputation mechanisms. Within each category, adoption by a given firm may be correlated across individual practices, and this correlation may be positive or negative (if some practices are partial substitutes for each other). In this case, examining each practice separately yields inefficient inferences. To guard against this possibil- ity, we follow the approach suggested by Kling, Liebman, and Katz (2007) and summarize the available information within each category using factor analysis. Thus, we construct an additional dyadic dependent variable from the first principal components. The factor loadings for each category are reported in table 1. The first variable that we consider is a dummy variable that indicates whether a firm introduced a new product in the year preceding the survey. Between one- third and one-half of the surveyed firms responded positively to this question. Approximately one-half of the firms invested in plants and equipment in the pre- vious year in both countries. A nonnegligible proportion of the surveyed firms had spent money on R&D: 13 percent for Ethiopia and 23 percent for Sudan. In addition, we note some usage of information technology (IT), mostly in the form of email. At the time of the surveys, few manufacturing firms in Sudan or Ethiopia had a website. Information on labor management and investment in human capital is pre- sented next. We find a higher ratio of nonproduction workers to total employ- ment in Sudan than in Ethiopia, which suggests that Sudanese firms are less able to manage their workforce with a small number of clerks and managers.12 10. Maps of the survey locations are shown in section 3 in the online appendix. 11. For some of our outcome variables, there are missing values in the baseline sample. Therefore, some of our regressions will be estimated on a smaller sample than the baseline sample. 12. Fafchamps and So ¨ derbom (2006) argue that the ratio of nonproduction workers to total employment indicates the ease with which firms manage their labor force. They show that many African firms have a high ratio of nonproduction workers to total employment despite the relative simplicity of their production processes. T A B L E 1 . Summary Statistics Ethiopia Sudan 110 Std. Std. Obs. Mean dev. Loadings Obs. Mean dev. Loadings THE WORLD BANK ECONOMIC REVIEW 1. Firm characteristics Firm age (years) 304 17.93 16.1 401 15.21 14.1 Education of top manager(a) 303 2.71 1.20 399 2.92 1.25 Experience of top manager (years) 304 14.5 9.77 395 17.2 12.9 Any female owner?(b) 304 0.23 382 0.15 Log(firm employment) 304 3.37 1.66 399 2.61 1.14 2. Innovation and R&D Did the firm introduce a new product last year?(b) 304 0.35 0.70 391 0.48 0.58 Did the firm invest in plants & equipment last year?(b) 304 0.52 0.67 400 0.46 0.70 Does the firm conduct any R&D?(b) 304 0.13 0.72 388 0.23 0.74 IT usage (0 ¼ nothing, 1 ¼ email, 2 ¼ website) 304 0.59 0.76 0.48 401 0.45 0.78 0.74 3. Human capital and labor management Ratio of nonproduction workers to total employment (c) 304 0.27 0.17 0.24 398 0.42 0.30 0.22 Any in-house training of staff last year?(b) 304 0.28 0.83 397 0.27 0.80 Staff sent to formal training course last year?(b) 304 0.28 0.84 398 0.12 0.80 4. Contractual practices Any direct imports of inputs?(b) 304 0.31 0.67 401 0.51 0.74 Do you sell on credit?(b) 304 0.53 0.65 401 0.64 0.73 Does firm subcontract production?(b) 302 0.12 0.33 382 0.09 0.22 5. Reputation mechanism If you have a dispute with a customer, will other customers find out?(d) 304 1.049 0.948 0.47 400 0.808 0.934 0.48 If another firm has a dispute with a customer, will you refuse to deal with that 304 0.457 0.815 0.67 401 0.783 0.954 0.65 customer?(d) If you have a dispute with a customer, will other firms refuse to deal with that 304 0.474 0.717 0.43 401 0.788 0.899 0.63 customer?(d) If you have a dispute with a supplier, will other suppliers find out?(d) 304 0.914 0.926 0.46 401 0.783 0.925 0.69 If you have a dispute with a supplier, will other firms refuse to deal with that 304 0.398 0.682 0.47 401 0.656 0.861 0.64 supplier?(d) (a) 1 ¼ less than secondary, 2 ¼ secondary, 3 ¼ vocational, 4 ¼ university. (b) 0 ¼ no, 1 ¼ yes. (c) Nonproduction workers include professionals, managers, administrators, and sales personnel. (d) 0 ¼ no, 1 ¼ maybe, 2 ¼ yes. Source: Authors’ computations based on data described in the text. ¨ derbom Fafchamps and So 111 In both countries, a substantial minority of firms had provided in-house or exter- nal training to their workers, but the majority had not. The next panel of table 1 covers contractual practices. Firms were asked whether they imported inputs directly from abroad. Although buying directly from abroad requires trust, it is likely to improve the quality of the raw materials that are used in a firm’s production process. We find a difference between the two countries: landlocked Ethiopia lagged behind Sudan. Firms were also asked whether they sold on credit to any of their customers. A majority of manufactur- ing firms sell on credit to at least some of their customers, but a large minority does not. The data also show that subcontracting part of a firm’s production to other firms is rare. Next, we examine the extent to which the surveyed firms rely on reputation to enforce contracts with suppliers and clients. The respondents were asked five closely related questions: (i) If you have a dispute with a customer, will other cus- tomers find out? (ii) If another firm has a dispute with a customer, will you refuse to deal with that customer? (iii) If you have a dispute with a customer, will other firms refuse to deal with that customer? (iv) If you have a dispute with a supplier, will other suppliers find out? (v) If you have a dispute with a supplier, will other firms refuse to deal with that supplier? For each of these questions, we code y ¼ 2 for “yes,” y ¼ 1 for “maybe/do not know” and y ¼ 0 for “no.” Hence, high values correspond to stronger reputation effects. The summary statistics present- ed in table 1 suggest that news about a dispute often travels to customers and suppliers. These statistics also suggest that the reputational sanction imposed on the customers and suppliers that are involved in a dispute is not severe; firms typi- cally continue to deal with customers and suppliers that have been involved in a dispute. Similar results have been reported by Bigsten et al. (2000) and Fafchamps (2004) for African manufacturing. A key module of the survey contains information about the names of the firms’ trading partners and their approximate geographical locations. The re- spondents were asked to name up to three clients and three suppliers.13 Using the information from this module, we construct simple measures of network proxim- ity between the firms in the two samples. Summary statistics for these measures are reported in table 2. We begin by constructing a dyadic dataset of unique firm pairs. For instance, because there are 304 firms in the Ethiopian sample, there exist 304  303/2 ¼ 46, 056 unique enterprise pairs (i, j ) in that sample. For each pair (i, j ), we con- struct dummy variables that capture the different concepts of network proximity. When two firms are close in that network, we consider them to be linked. The most direct network proximity measure that we use is whether i and j buy or sell 13. Because the majority of firms (approximately 70 percent) list three names, there is truncation in the observed network because some existing links are not recorded. This problem may cause a downward bias in the estimated network effects. 112 THE WORLD BANK ECONOMIC REVIEW T A B L E 2 . Dyadic Data Ethiopia Sudan Number of unique enterprise pairs 46,056 80,200 i & j trade with each other (number of pairs) 60 5 i & j have a common supplier (number of pairs) 481 171 i & j have a common client (number of pairs) 273 678 Average distance between i & j (kilometers) 282 421 Minimum distance between i & j (kilometers) 0 0 Maximum distance between i & j (kilometers) 876 1,770 Source: Authors’ computations based on data described in the text. from each other. We are only able to identify a small number of such links in our data: 60 in Ethiopia and 5 in Sudan. The fact that there are so few upstream and downstream links among the sample firms is partly driven by the focus of the surveys on light manufacturing because clients are seldom manufacturers. We also construct dummy variables that indicate whether i and j have a common supplier or a common client. These types of links are more common: there are 481 supplier-based links and 273 client-based links in the Ethiopian data and 171 supplier-based links and 678 client-based links in the Sudanese data. These network proximity variables constitute the core of our vector gij. The last prox- imity dummy is the distance dij, which is defined as the log of the distance between i and j plus one. I V. E M P I R I C A L A N A L Y S I S Our objective is to test whether the outcomes and practices that are related to technology, human capital, contracting, and reputation are more similar among firms that are close to each other, either in a network sense or geographically. To that end, we estimate the parameters of model (3). Our estimation technique is linear regression (ordinary least squares), and standard errors are bootstrapped to make them robust with respect to heteroskedasticity and correlation in error terms across firms. We refer to the presented results as baseline results. Additional results are available in the online appendix. Innovation and R&D We begin by investigating the association between geographical and network proximity and innovation and R&D. We construct dyadic dependent variables from dummy variables that measure whether firms introduced a new product in the previous year, invested in plants and equipment in the previous year, or con- ducted any R&D. A fourth outcome variable is constructed based on a firm-level measure of the extent of IT usage; the value of this variable is zero if IT is not ¨ derbom Fafchamps and So 113 used at all, one if the firm uses email, and two if the enterprise has a business website.14 The dyadic regression results are shown in table 3, columns [1]–[4] for Ethiopia, and columns [6]–[9] for Sudan. In columns [5] and [10], we report results in which we use the first principal component of all four categories to con- struct the dyadic dependent variable. The estimated network proximity coefficients differ in the two countries. For Ethiopia, the dummies that track whether i and j trade with each other, have a common supplier, and have a common client are statistically nonsignificant. For Sudan, we obtain a negative and statistically significant coefficient of trade in the R&D regression (column [8]), and we obtain negative and significant coefficients (at least at the 10 percent level) of having a common supplier in the regressions for investment (column [7]), R&D (column [8]), IT usage (column [9]), and the first principal component (column [10]). Hence, network proximity seems to be associated with a more similar approach to innovation and R&D throughout the firms in Sudan, but not in Ethiopia. Some of these estimated effects are large; for example, the likelihood that firms report the same answer (yes or no) to the ques- tion about whether money was spent on R&D is 34 percentage points higher for firms that trade with each other than for firms that do not trade with each other. However, because of the small number of direct links in the Sudanese data (see table 2), the estimated coefficients of direct trade should be interpreted with caution. Furthermore, we find that Sudanese firms with a common client tend to differ more than other firms with respect to R&D and IT usage. This finding is not consistent with the notion that network proximity tends to result in similar practices regarding innovation. Next, we consider the role of geographical distance between firms. For Ethiopia, the distance coefficient is negative in all five of the specifications shown in table 3, and this coefficient is statistically significant at least at the 10 percent level in four of these specifications. Hence, geographical proximity tends to be as- sociated with greater differences in innovation practices. The results are similar for Sudan: the distance coefficient is negative and highly statistically significant in the models for R&D (column [8]), IT usage (column [9]), and the first princi- pal component (column [10]). These results suggest that for technology, strategic substitution effects dominate strategic complementarities for firms that are located near each other. The control variables in these regressions have explanatory power. The esti- mated coefficients of the same-sector dummy are negative in all of the specifica- tions except [6], and these coefficients are often statistically significant. This result indicates that, as expected, firms in the same sector tend to have similar in- novation practices. Differences in firm size, which are measured as the absolute difference in the log of employment, are positively associated with differences in 14. Here, the three levels of usage are combined. The results for the alternative specifications that model email and website use are shown separately in tables S1.E and S1.S, columns [1] –[2], in the online appendix. The results are similar to those shown in table 3. 114 T A B L E 3 . Correlates of Dyadic Differences: Technology Acquisition THE WORLD BANK ECONOMIC REVIEW Ethiopia Sudan [1] [2] [3] [4] [5] [6] [7] [8] [9] [10] Did the firm Did the firm Did the firm invest in Does the IT usage Did the firm invest in IT usage introduce a plants & firm (0 ¼ nothing, First introduce a plants & Does the firm (0 ¼ nothing, First new product equipment last conduct 1 ¼ email, principal new product equipment conduct any 1 ¼ email, principal last year? year? any R&D? 2 ¼ website) component last year? last year? R&D? 2 ¼ website) component jyi 2 yjj jyi 2 yjj jyi 2 yjj jyi 2 yjj jyi 2 yjj jyi 2 yjj jyi 2 yjj jyi 2 yjj jyi 2 yjj jyi 2 yjj i & j trade 0.0492 2 0.0471 0.0957 2 0.0703 0.0792 0.0147 2 0.205 2 0.339** 0.451 2 0.302 with each (0.126) (0.117) (0.108) (0.188) (0.240) (0.350) (0.326) (0.140) (0.572) (0.237) other i & j have 2 0.0175 2 0.00152 0.0461 2 0.0630 2 0.0202 2 0.0799 2 0.154* 2 0.183*** 2 0.272** 2 0.310** common (0.0400) (0.0456) (0.0408) (0.0660) (0.0684) (0.0740) (0.0791) (0.0648) (0.109) (0.130) supplier i & j have 0.0666 2 0.0259 0.0110 0.0602 2 0.0645 0.0247 0.00539 0.123*** 0.229*** 0.178* common (0.0586) (0.0705) (0.0692) (0.0948) (0.0973) (0.0223) (0.0336) (0.0350) (0.0607) (0.101) client log distance 2 0.00478** 2 0.00129 2 0.0104* 2 0.0173** 2 0.0154* 0.000372 2 0.000603 2 0.0122*** 2 0.0221*** 2 0.0176*** btw i & j (0.00238) (0.00175) (0.00543) (0.00694) (0.00833) (0.00166) (0.00223) (0.00301) (0.00531) (0.00601) i & j belong 2 0.0323* 2 0.0561** 2 0.0218 2 0.0567** 2 0.101*** 0.00460 2 0.00129 2 0.0133 2 0.0955*** 2 0.0299 to same (0.0171) (0.0260) (0.0141) (0.0280) (0.0345) (0.0101) (0.0134) (0.0185) (0.0302) (0.0310) sector Abs diff firm 2 0.000611 2 0.00079*** 2 0.00100 2 0.00187 2 0.00322* 0.000121 0.000259 0.000361 2 0.00318*** 2 0.00102 age (0.000572) (0.000273) (0.00105) (0.00131) (0.00173) (0.000283) (0.000369) (0.00103) (0.00120) (0.00134) Abs diff 0.00863 0.0141 2 0.00995 0.0644** 0.0475** 0.00567 0.0409*** 0.00756 0.0189 0.0664*** managers’ (0.00906) (0.00956) (0.00833) (0.0286) (0.0226) (0.00616) (0.0144) (0.00927) (0.0167) (0.0235) education Abs diff 2 0.000763 0.000110 2 0.000577 2 0.000880 2 0.00312* 2 5.03e 2 06 2 0.000301 2 0.000411 2 2.72e 2 05 2 0.000124 managers’ (0.000897) (0.000386) (0.00132) (0.00151) (0.00169) (0.000287) (0.000265) (0.00116) (0.00218) (0.00240) experience Owners’ 2 0.00147 0.000586 2 0.00558 0.0904** 0.00909 0.00113 0.0354* 0.0978** 0.435*** 0.241*** genders (0.0186) (0.00596) (0.0285) (0.0454) (0.0482) (0.00629) (0.0187) (0.0396) (0.0909) (0.0751) Differ Abs diff log 0.00623 0.0219** 0.0304** 0.189*** 0.121*** 0.0232** 0.0525*** 0.0394** 0.167*** 0.198*** Employment (0.00943) (0.00903) (0.0137) (0.0264) (0.0298) (0.00912) (0.0143) (0.0175) (0.0421) (0.0457) Notes: The table shows ordinary least squares results. A constant is included in all specifications. The numbers in parentheses are bootstrapped standard errors that are robust with respect to heteroskedasticity and cross-observation correlation in the error terms involving the same firms. Statistical significance at the 10 percent, 5 percent, and 1 percent levels is indicated by *,**, and ***, respectively. Source: Authors’ computations based on data described in the text. Fafchamps and So 115 ¨ derbom 116 THE WORLD BANK ECONOMIC REVIEW innovation practices in all of the specifications, suggesting that firms of similar size tend to adopt similar practices. There is also some evidence that managers of the same gender or with similar levels of education select similar innovation prac- tices. The coefficients of differences in managers’ experience or firms’ ages are mostly nonsignificant. Human Capital and Labor Management Table 4 shows the results for our regressions on labor management and invest- ment in human capital. We find no evidence that network proximity is associated with greater similarity in training decisions or labor management in firms. In fact, we obtain a positive and statistically significant coefficient of common clients in specifications [3], [4], and [6], which indicates that firms with a common client tend to have more distinct training policies than firms that do not share a common client. The estimated coefficients related to the distance between firms are negative in all of the specifications except [5], and these coefficients are statistically significant in four of the specifications (columns [2], [3], [4], and [6]). As with the results for innovation, this result implies that firms that are located close to each other tend to differ more with respect to their human capital decisions compared with firms that are far apart. This finding is consistent with strategic substitution. One possibility that is often emphasized in the literature on agglomeration effects (Henderson 1988; Glaeser et al. 1992) is that firms hire workers who have been trained by other firms. Thus, if there are more nearby firms that provide the necessary train- ing, a given firm needs to train its workers to a lesser extent. Alternatively, strategic substitution may be driven by incentives to avoid local competition. For example, if two firms with similar human capital produce similar output, they will compete with each other if they are based in the same local market. By locating themselves in different places, both firms would face less competition and presumably higher profits. Another possibility is that firms that are located in the same place decide to differentiate their output, which may lead to differences in technology and human capital demand. Mechanisms such as these would result in the pattern that we observe in the data of greater differences between firms that are located close to each other than between firms in distant locations. Furthermore, we find that, as expected, firms of similar size and firms in the same sector tend to be more similar with respect to their training decisions than firms of different sizes or in different sectors. The coefficients related to the other control variables, that is, differences in firms’ ages and in managers’ education, experience, and gender, are mostly nonsignificant. When they are significant, their coefficients are usually negative, which suggests that greater differences in these firm-level characteristics are associated with closer similarities in outcomes. Contractual Practices Next, we investigate how the following three measures of contractual practices correlate across firms: whether a firm imports inputs directly, whether it sells on T A B L E 4 . Correlates of Dyadic Differences: Human Capital and Labor Management Ethiopia Sudan [1] [2] [3] [4] [5] [6] [7] [8] Ratio of Staff sent to Ratio of Staff sent to nonproduction Any in-house formal training First nonproduction Any in-house formal training First workers to total training of course last principal workers to total training of staff course last principal employment staff last year? year? component employment last year? year? component jyi 2 yjj jyi 2 yjj jyi 2 yjj j yi 2 yj j jyi 2 yjj j yi 2 yj j j yi 2 yj j jyi 2 yjj i & j trade with 2 0.0218 0.0807 2 0.0451 2 0.0514 2 0.0859 2 0.121 0.368 2 0.200 each other (0.0334) (0.107) (0.133) (0.220) (0.0879) (0.263) (0.340) (0.336) i & j have 0.0136 0.0439 0.0113 0.0857 2 0.000342 0.0313 2 0.0195 0.0743 common (0.0141) (0.0473) (0.0413) (0.0799) (0.0442) (0.0842) (0.0549) (0.147) supplier i & j have 0.0161 0.0396 0.127** 0.215* 2 0.0403 0.0826** 0.0210 0.174 common (0.0237) (0.0742) (0.0627) (0.124) (0.0288) (0.0375) (0.0713) (0.142) client log distance 2 0.00193 2 0.0168*** 2 0.0153*** 2 0.0251*** 0.00320 2 0.00803** 2 0.000397 2 0.00612 btw i & j (0.00159) (0.00299) (0.00291) (0.00824) (0.00195) (0.00315) (0.00434) (0.00847) i & j belong to 2 0.00244 2 0.0173 2 0.0304* 2 0.0480* 0.00212 2 0.0341* 2 0.00869 2 0.0644* same sector (0.00491) (0.0123) (0.0177) (0.0262) (0.00623) (0.0189) (0.0159) (0.0374) Fafchamps and So Abs diff firm 2 0.000149 2 0.000207 0.00180 0.00225 2 0.000522* 2 0.000423 0.000882 0.000211 age (0.000336) (0.000720) (0.00114) (0.00210) (0.000299) (0.000662) (0.000992) (0.00183) Abs diff 2 0.00392* 0.0112 0.00644 0.0328 0.00381 2 0.000336 2 0.0145*** 2 0.0148 managers’ (0.00237) (0.0184) (0.0160) (0.0351) (0.00339) (0.00576) (0.00474) (0.0131) education Abs diff 0.000213 2 0.00167** 2 0.000966 2 0.000834 0.000259 2 0.00155*** 0.000637 2 0.000751 ¨ derbom managers’ (0.000477) (0.000803) (0.000838) (0.00165) (0.000598) (0.000501) (0.00108) (0.00185) experience (Continued ) 117 118 THE WORLD BANK ECONOMIC REVIEW TABLE 4. Continued Ethiopia Sudan [1] [2] [3] [4] [5] [6] [7] [8] Ratio of Staff sent to Ratio of Staff sent to nonproduction Any in-house formal training First nonproduction Any in-house formal training First workers to total training of course last principal workers to total training of staff course last principal employment staff last year? year? component employment last year? year? component jyi 2 yjj jyi 2 yjj jyi 2 yjj j yi 2 yj j jyi 2 yjj j yi 2 yj j j yi 2 yj j jyi 2 yjj Owners’ gender 0.00767 0.0181 2 0.0120 2 0.00539 0.0298* 0.0446 0.0847 0.204* differ (0.0111) (0.0258) (0.0177) (0.0403) (0.0176) (0.0375) (0.0542) (0.114) Abs diff log 0.0110** 0.0876*** 0.0987*** 0.230*** 0.0182*** 0.0567*** 0.0780*** 0.207*** employment (0.00490) (0.0155) (0.0143) (0.0309) (0.00569) (0.0170) (0.0205) (0.0499) Notes: The table shows ordinary least squares results. A constant is included in all specifications. The numbers in parentheses are bootstrapped standard errors that are robust with respect to heteroskedasticity and cross-observation correlation in the error terms involving the same firms. Statistical significance at the 10 percent, 5 percent, and 1 percent levels is indicated by *,**, and ***, respectively. Source: Authors’ computations based on data described in the text. ¨ derbom Fafchamps and So 119 credit, and whether it subcontracts part of its production. The results are shown in table 5. For Sudan, we find a negative and highly significant coefficient on the dummy variable that indicates whether firms i and j trade directly with each other in the models for direct imports, selling on credit, and the first principal component. Thus, Sudanese firms that trade with each other tend to have more similar con- tractual practices. In addition, having a common supplier is associated with a greater similarity in direct imports, although this effect is only statistically signifi- cant at the 10 percent level. In contrast, for Ethiopia, the correlation between network proximity and the similarity in contractual practices is weak and non- significant in all of the specifications except for subcontracting, for which we obtain a positive coefficient related to having a common supplier (column [3]). The estimated distance coefficients vary considerably across regressions. In two regressions, they are positive and significantly different from zero (direct imports and selling on credit in Sudan; columns [5] and [6]), which suggests that firms that are close to each other have more similar contractual practices than other firms. However, in two other regressions, the coefficients are significantly negative (direct imports in Ethiopia and subcontracting in Sudan; columns [1] and [7]). For both countries, distance is statistically nonsignificant in the regres- sions that model the difference in the first principal component. Thus, it is diffi- cult to see a pattern here, which may be because the relative importance of strategic substitution and diffusion varies from one contractual practice to another. Regarding the control variables, the pattern is similar to what we ob- served above; whereas firms of similar size and in the same sector tend to have similar contractual practices, for other controls, the results are mixed. Reputation Mechanisms Here, we examine whether there is evidence that network links facilitate the dif- fusion of information on contractual disputes between suppliers and clients. The theoretical literature has emphasized the role of the diffusion of information on contractual disputes along social networks in the development of modern market institutions (North 1990; Greif 1993). Consequently, we expect to find a strong correlation in answers from firms in the same networks. Using the five questions on the perceived consequences of disputes that are dis- cussed in section 3, we code yi ¼ 2 for “yes,” yi ¼ 1 for “maybe/do not know” and yi ¼ 0 for “no.” Then, we compute jyi À yj j for every pair of firms in the data.15 The regression results are shown in table 6a and do not conform to theo- retical expectations. Except for isolated cases in which a network regressor is sig- nificant (columns [6] and [10], but with opposite signs), the social network variables are not significant. One possible explanation is insufficient power: the 15. Columns [3] –[8] in tables S.1E and S1.S in the online appendix show the results for alternative specifications in which the reputation variables are defined as binary variables in the following way: yes ¼ 1 and maybe or no ¼ 0. The results are similar to those in table 5. 120 T A B L E 5 . Correlates of Dyadic Differences: Contractual Practices Ethiopia Sudan THE WORLD BANK ECONOMIC REVIEW [1] [2] [3] [4] [5] [6] [7] [8] Any direct Does firm First Any direct Do you Does firm imports of Do you sell subcontract principal imports of sell on subcontract First principal inputs? on credit? production? component inputs? credit? production? component jyi 2 yjj j yi 2 y j j j yi 2 yj j jyi 2 yjj j yi 2 yj j j yi 2 yj j j yi 2 yj j j yi 2 yj j i & j trade with 0.0245 2 0.0122 0.110 0.114 2 0.423** 2 0.436** 0.0488 2 0.813** each other (0.117) (0.105) (0.119) (0.252) (0.174) (0.177) (0.331) (0.380) i & j have common 2 0.0186 2 0.00130 0.0755** 2 0.116 2 0.145* 0.0150 0.0479 0.0304 supplier (0.0501) (0.0404) (0.0366) (0.0751) (0.0762) (0.0789) (0.0812) (0.135) i & j have common 0.0838 2 0.0139 0.0765 0.184 2 0.0457 2 0.00759 2 0.0310 2 0.0390 client (0.0572) (0.0524) (0.0702) (0.124) (0.0422) (0.0430) (0.0651) (0.0998) log distance 2 0.0123** 2 0.000254 0.00530 2 0.00906 0.00814*** 0.00915** 2 0.00867** 0.00854 btw i & j (0.00572) (0.00143) (0.00671) (0.00796) (0.00314) (0.00422) (0.00417) (0.00923) i & j belong to 2 0.0397** 2 0.0171 2 0.00192 2 0.0368 2 0.0299* 2 0.00602 2 0.00572 2 0.0459 same sector (0.0160) (0.0155) (0.00908) (0.0286) (0.0157) (0.0137) (0.0131) (0.0279) Abs diff firm age 2 0.000424 0.000480 2 0.00207*** 2 0.00417*** 0.000105 2 8.22e 2 05 2 0.000668 0.000303 (0.000804) (0.000556) (0.000636) (0.00145) (0.000240) (0.000582) (0.000563) (0.00107) Abs diff managers’ 0.0299 0.000668 2 0.0204*** 0.0373 0.0215* 0.00583 0.00164 0.0278 education (0.0199) (0.00500) (0.00748) (0.0279) (0.0112) (0.00657) (0.00477) (0.0173) Abs diff managers’ 2 0.00152* 0.000485 2 0.00178 0.00456** 2 9.80e 2 05 2 0.000212 2 0.000397 2 0.000631 experience (0.000779) (0.000642) (0.00120) (0.00210) (0.000305) (0.000728) (0.000633) (0.00121) Owners’ gender 0.0457 0.00391 0.0235 0.0402 0.00257 2 0.0164 0.0109 2 0.00210 differ (0.0297) (0.00948) (0.0306) (0.0420) (0.00677) (0.0126) (0.0382) (0.0392) Abs diff log 0.131*** 0.00259 0.0150 0.138*** 0.0659*** 0.00494 2 0.00398 0.0977*** employment (0.0150) (0.00473) (0.0132) (0.0312) (0.0145) (0.00781) (0.0123) (0.0264) Notes: The table shows ordinary least squares results. A constant is included in all specifications. The numbers in parentheses are bootstrapped standard errors that are robust with respect to heteroskedasticity and cross-observation correlation in the error terms involving the same firms. Statistical significance at the 10 percent, 5 percent, and 1 percent levels is indicated by *,**, and ***, respectively. Source: Authors’ computations based on data described in the text. T A B L E 6 a . Correlates of Dyadic Differences: Perceived Consequences of Disputes Ethiopia Sudan [1] [2] [3] [4] [5] [6] [7] [8] [9] [10] If you If you have If you If another If you have a If you If you have If you If another have a If you supplier have a firm has a customer have a supplier have a firm has a customer have a dispute, will customer customer dispute, will supplier dispute, will customer customer dispute, will supplier other firms dispute, will dispute, will other firms dispute, will other firms dispute, will dispute, will other firms dispute, will refuse to other you refuse to refuse to deal other refuse to deal other you refuse to refuse to deal other deal with customers deal with with that suppliers with that customers deal with that with that suppliers that find out? customer? customer? find out? supplier? find out? customer? customer? find out? supplier? jyi 2 yjj jyi 2 yjj jyi 2 yjj jyi 2 yjj jyi 2 yjj jyi 2 yjj jyi 2 yjj jyi 2 yjj jyi 2 yjj jyi 2 yjj i & j trade 2 0.187 0.0330 2 0.0502 0.124 0.0158 2 0.704* 2 0.209 0.557 0.280 0.366 with each (0.206) (0.178) (0.166) (0.207) (0.161) (0.373) (0.622) (0.515) (0.614) (0.637) other i & j have 2 0.0159 0.0321 2 0.0611 2 0.0335 2 0.0490 2 0.154 2 0.0828 2 0.0368 2 0.105 2 0.113 common (0.0914) (0.0896) (0.0663) (0.0914) (0.0745) (0.166) (0.164) (0.145) (0.141) (0.164) supplier i & j have 0.0480 0.0293 2 0.0588 0.00750 0.0853 0.0580 0.0328 2 0.104 2 0.105 0.112** common (0.0795) (0.129) (0.0990) (0.0916) (0.111) (0.0557) (0.0669) (0.116) (0.111) (0.0475) client Fafchamps and So log distance 0.0110** 2 0.0198*** 2 0.0171*** 0.0104* 2 0.0268*** 2 0.00683** 2 0.00611** 2 0.000383 2 0.00298 2 0.00474 btw i & j (0.00562) (0.00578) (0.00487) (0.00544) (0.00530) (0.00306) (0.00310) (0.00323) (0.00282) (0.00475) i & j belong 2 0.00832 2 0.0163 0.00167 2 0.0201 2 0.0113 2 0.00888 2 0.0213 2 0.00301 2 0.00127 0.0275* to same (0.0272) (0.0258) (0.0187) (0.0254) (0.0186) (0.0213) (0.0281) (0.0194) (0.0186) (0.0164) Abs diff firm 2 0.000181 2 0.000755 2 0.00166 0.000648 5.85e 2 05 2 0.000854 0.000383 0.000530 2 0.000695 0.00154 age (0.000404) (0.00180) (0.00109) (0.000791) (0.00149) (0.000725) (0.00106) (0.000944) (0.000812) (0.00156) ¨ derbom Abs diff 2 0.00403 2 0.00790 2 0.00932 2 0.00511 2 0.00802 0.000897 0.00325 0.00229 2 0.00144 0.00723 managers’ (0.00757) (0.0117) (0.0110) (0.00884) (0.0112) (0.00726) (0.00754) (0.00795) (0.00705) (0.00943) education (Continued ) 121 122 THE WORLD BANK ECONOMIC REVIEW TABLE 6a. Continued Ethiopia Sudan [1] [2] [3] [4] [5] [6] [7] [8] [9] [10] If you If you have If you If another If you have a If you If you have If you If another have a If you supplier have a firm has a customer have a supplier have a firm has a customer have a dispute, will customer customer dispute, will supplier dispute, will customer customer dispute, will supplier other firms dispute, will dispute, will other firms dispute, will other firms dispute, will dispute, will other firms dispute, will refuse to other you refuse to refuse to deal other refuse to deal other you refuse to refuse to deal other deal with customers deal with with that suppliers with that customers deal with that with that suppliers that find out? customer? customer? find out? supplier? find out? customer? customer? find out? supplier? jyi 2 yjj jyi 2 yjj jyi 2 yjj jyi 2 yjj jyi 2 yjj jyi 2 yjj jyi 2 yjj jyi 2 yjj jyi 2 yjj jyi 2 yjj Abs diff 0.000140 2 0.00130 0.00309 0.00114 0.00286 0.000672 0.000130 2 0.00113 2 0.000772 2 0.00118 managers’ (0.000932) (0.00303) (0.00264) (0.00151) (0.00312) (0.000804) (0.000917) (0.000703) (0.000948) (0.00144) experience Owners’ 2 0.00783 2 0.00611 0.0600 0.0184 0.0541 0.0298 0.00190 2 0.0103 0.0157 0.0261 gender (0.0131) (0.0558) (0.0550) (0.0299) (0.0597) (0.0332) (0.0341) (0.0252) (0.0387) (0.0461) differ Abs diff log 0.00230 2 0.00417 0.00893 0.0151 0.0102 0.0219 2 0.0130 2 0.00426 2 0.0267*** 2 0.0311** employment (0.00661) (0.0180) (0.0150) (0.0132) (0.0190) (0.0157) (0.0113) (0.0114) (0.00875) (0.0133) ¨ derbom Fafchamps and So 123 T A B L E 6 b . Correlates of Dyadic Differences: Perceived Consequences of Disputes, First Principal Component [1] Ethiopia [2] Sudan jyi 2 yjj jyi 2 yjj i & j trade with each other 2 0.0757 0.339 (0.146) (0.782) i & j have common supplier 0.00852 0.0100 (0.0841) (0.151) i & j have common client 0.0573 0.0579 (0.118) (0.0874) log distance btw i & j 2 0.0152** 0.00614 (0.00734) (0.00704) i & j belong to same sector 2 0.0130 0.0163 (0.0258) (0.0188) Abs diff firm age 2 0.000904 2 0.00117 (0.00168) (0.000985) Abs diff managers’ education 2 0.00478 0.00139 (0.0133) (0.00863) Abs diff managers’ experience 0.00290 0.000487 (0.00340) (0.00169) Owners’ gender differ 0.00676 0.0369 (0.0613) (0.0481) Abs diff log employment 0.00317 2 0.0111 (0.0164) (0.0136) Notes: The table shows ordinary least squares results. A constant is included in all specifications. The numbers in parentheses are bootstrapped standard errors that are robust with respect to hetero- skedasticity and cross-observation correlation in the error terms involving the same firms. Statistical significance at the 5 percent level is indicated by**. Source: Authors’ computations based on data described in the text. five categorical reputation variables may contain insufficient information to iden- tify the social network coefficients. Additionally, no network variable is signifi- cant in the principal component regressions shown in table 6b; hence, combining the information contained in all five of them does not lead to better results. Furthermore, the coefficients related to the control variables are nonsignificant in the vast majority of cases. There are two possible interpretations of these findings: either the information about contractual disputes does not diffuse along the type of social networks that we have been able to measure, or this information diffuses so well that social links do not matter. One way to identify which of these two interpretations is more likely is to examine the coefficient of the distance variable. Even though information may diffuse rapidly along social networks within certain areas, the diffusion of information need not happen everywhere because strategic comple- mentarities in diffusion create the possibility of multiple equilibria. If there are multiple equilibria, we expect to find that firms that are distant from each other perceive the consequences of contractual disputes differently. 124 THE WORLD BANK ECONOMIC REVIEW However, we do not find distinct perceptions of contractual disputes in distant areas. For Ethiopia, although the distance coefficient is negative and highly sig- nificant in three of the specifications shown in table 6a, it is positive and signifi- cant in the remaining two specifications. Dyadic differences in the principal component that is based on the five individual variables are negatively and signif- icantly related to distance. For Sudan, the distance coefficient is negative and sig- nificant in two out of five individual regressions, and in the remaining cases, it is not statistically significant. These findings are difficult to reconcile with the idea of widespread diffusion of contractual information among firms in the same lo- cation. If multiple equilibria are present, they seem to coexist within locations; thus, whereas some firms recognize that there are reputational consequences to contractual disputes, others in the same location do not. Firm Performance and Growth Thus far, we have focused on those business practices that may diffuse within networks. We have also investigated whether the above results are mirrored in labor productivity and growth rates, which are our measures of firm perfor- mance. The results for the specifications in which the dependent variable is defined as the absolute difference across firms in these performance indicators are shown in section 4 in the online appendix. For Ethiopia, we find little evi- dence that firms that are closer in the social-network sense have more similar per- formance. For Sudan, we find evidence that firms that share the same supplier have more similar performance than other firms. However, other results related to network links and geographical distance are weak. The overall conclusion from our analysis is that network links and geographical proximity are not strongly associated with any convergence in the performance of firms. Heterogeneous Diffusion and Networks Now, we return to the points raised in section I that are related to the diffusion patterns across heterogeneous firms. We ask whether the reason we find only limited evidence of diffusion is that the firms are too heterogeneous. To investi- gate whether the evidence for diffusion is stronger among pairs of firms in the same sector, we interact our network and distance variables with a dummy that tracks whether firms i and j belong to the same industrial subsector, and we add these new interaction terms to the baseline specification. To minimize the number of explanatory variables, the same industry dummy is interacted with a single network variable anylinkij, which is a dummy variable that is equal to one if there is any link between firms i and j. Such a link would amount to direct trade, a common client or a common supplier. The results for all of the outcome variables are shown in tables S2.E and S2.S in the online appendix. For Ethiopia, the sector-network interaction term is stat- istically nonsignificant in every specification, and the sector-distance interaction term is significant in only one specification (formal training; see table S2.E, column [8]; the term has a positive sign). For Sudan, the sector-network ¨ derbom Fafchamps and So 125 interaction term is statistically nonsignificant in every specification, and the sector-distance interaction term is significant in just one specification (direct imports; see table S2.S, column [10]; the term has a positive sign). These results suggest that sector heterogeneity is not the reason for limited diffusion. They also imply that strategic substitution is equally strong within each sector and across different sectors. We repeat this type of analysis, which focuses on firm-size heterogeneity instead of sector heterogeneity. To that end, we interact anylinkij and the distance variable (dij ) with a dummy that tracks whether firms i and j are of similar size, and we add these new interaction terms to the baseline model.16 The results are shown in tables S3.E and S3.S in the online appendix. For Ethiopia, the size-network inter- action term is statistically nonsignificant in all regressions. However, the size- distance term is negative in the vast majority of cases and is often statistically significant. This result suggests that strategic substitution is stronger across firms of similar size than across firms of differing size, which may be because geographical- ly close firms strategically choose to differentiate themselves from each other to reduce competition. For Sudan, the network-size interaction term is statistically nonsignificant in all specifications, and the size-distance interaction term is signifi- cant in just three specifications (table S3.S, columns [6], 11 and 20). On balance, we find little evidence that size heterogeneity is a likely reason for slow diffusion, and we note that the results for Ethiopia lend further support to the idea that strategic substitution may be important. Market Differentiation within Towns Finally, we investigate how the estimated coefficients of the geographical distance change if we add a dummy variable sametownij to the baseline specification. This dummy variable is equal to one if firms i and j are located in the same town and zero otherwise. We want to establish whether market differentiation within towns causes the result that a shorter geographical distance between firms is asso- ciated with greater differences in business practices. It seems plausible to suppose that strategic substitution is strongest within towns. If markets are localized such that firms in different towns pose no competitive threat to each other irrespective of the distance between these towns, events in town k will not affect the strategic decisions of firms in town l = k. In this case, the relevant geographical circum- stance is whether firms are in the same town; thus, conditional on sametownij, distance does not matter. Thus, by adding sametownij to the set of explanatory variables, we generalize the baseline’s functional form with respect to the effect of distance. The results that are based on this specification for all outcome variables are shown in tables S4.E and S4.S in the online appendix. For Sudan, the coefficients on sametownij are always negative whenever they are significant, which suggests 16. Firms are defined as having a similar size if the absolute log of the difference in employment is less than 0.2. 126 THE WORLD BANK ECONOMIC REVIEW that strategic substitution effects do not primarily operate within towns. For Ethiopia, the picture is more mixed: although we obtain positive and significant coefficients in five of the regressions shown in table S4.E columns [1]–[24], we obtain negative and significant coefficients in three regressions. Overall, the support for the idea that strategic substitution effects operate primarily within towns is quite limited. V. C O N C L U S I O N S In this paper, we have documented empirical patterns of correlation in the adop- tion of innovation and contractual practices among manufacturing firms in Ethiopia and Sudan. Our empirical analysis is based on network data that indi- cate whether any two firms in our sample do business with each other, buy inputs from a common supplier, or sell output to a common client. We also exploit data on firms’ locations to investigate whether firms that are near each other tend to be more similar or dissimilar than firms that are geographically far apart. Our results can be summarized as follows: (i) for Sudan, but not for Ethiopia, there is some evidence that network proximity is associated with similar innova- tion strategies; (ii) for both countries, there is relatively strong evidence that firms that are located close to each other differ more with respect to innovation than firms that are far apart; (iii) there is no evidence that network proximity is associ- ated with greater similarity in training decisions or labor management across firms; (iv) there is some evidence that firms that are located close to each other differ more with respect to training decisions than firms that are geographically far apart; (iv) for Sudan, but not for Ethiopia, there is some evidence that network proximity is associated with similar contractual practices; (vi) differenc- es in contractual practices across firms are only weakly related to geographical proximity; (vii) there is no evidence that network proximity is associated with greater similarity in the perceived consequences of disputes; (viii) there is some evidence that geographical proximity is associated with greater differences in the perceived consequences of disputes; and (ix) except for supplier-based links in Sudan, differences in firm performance are only weakly related to geographical and network proximity. Overall, the strongest results are for innovation. Thus, our results provide limited support for the commonly held idea that firms that are more proximate in a network sense are more likely to adopt similar contractual and technological innovation practices. Furthermore, we find some evidence that for certain practices, adoption decisions are local strategic substi- tutes. Hence, if one firm adopts, other nearby firms are less likely to do so. What should we make of these results? First, we again note that a correlation in practic- es does not imply diffusion because there may be unobserved contextual effects. Second, the evidence presented here does not imply that the diffusion of innova- tion between firms can never be important or even critical for growth. However, diffusion between firms should not be taken for granted. Many of the firms in ¨ derbom Fafchamps and So 127 our sample follow antiquated business practices even when some neighboring firms do not, which is consistent with the observation that firms in developing countries are often more heterogeneous than firms in developed countries (see, for example, Bloom et al. [2012] for evidence that the quality of management practices is more heterogeneous for the firms in Brazil, China, and India than in the U.S.). Third, it is possible that we searched for diffusion in the wrong place (that is, among existing firms) because it is possible that, in general, the diffusion of innovations takes place not because existing firms learn to imitate each other but rather because new firms emerge that adopt innovative practices. This inter- pretation is consistent with the findings reported in the exporting literature: although there is limited evidence that incumbent firms learn from exporting, there is ample evidence that firms that begin exporting are more productive than average, even when they are new entrants (Clerides, Lach and Tybout 1998; Fafchamps, El Hamine and Zeufack 2008). Fourth, we acknowledge that our data have certain limitations. One potentially important limitation is that the survey asked for a maximum of three clients and suppliers, which implies that we do not have complete coverage of all of the network links. In addition, it is likely that our network-link variables are measured with error, which may cause the network effects to be underestimated in our analysis. Despite these caveats, we note that in several ways, the evidence for diffusion and complementarities is much weaker than one might expect given the emphasis in much of the current policy discussion on diffusion and agglomeration economies as sources of im- proved firm performance. REFERENCES Aghion, P., and P. Howitt. 1992. “A Model of Growth through Creative Destruction.” Econometrica 60 (2): 323– 51. Bigsten, A., P. Collier, S. Dercon, M. Fafchamps, B. Gauthier, J.W. Gunning, A. Oduro, R. Oostendorp, C. Pattillo, M. So¨ derbom, F. Teal, and A. Zeufack. 2000. Contract Flexibility and Dispute Resolution in African Manufacturing. Journal of Development Studies 36 (4): 1–37. Bloom, N., B. Eifert, A. Mahajan, D. McKenzie, and J. Roberts. 2012. “Does Management Matter? Evidence from India.” Unpublished manuscript. Stanford University, Stanford, California. ´ , Y., and R. Kranton. 2011. “Strategic Interaction and Networks.” Mimeograph. Duke Bramoulle University, Durham, NC. Casella, A., and J. Rauch. 2002. “Anonymous Market and Group Ties in International Trade.” Journal of International Economics 58 (1): 19 –47. Clerides, S., S. Lach, and J. Tybout. 1998. “Is Learning-by-Exporting Important? Micro-Dynamic Evidence from Colombia, Mexico, and Morocco.” Quarterly Journal of Economics 113 (3): 903–47. Collier, P. 2007. The Bottom Billion. Why the Poorest Countries are Failing and What Can Be Done About It. New York: Oxford University Press. de Mel, S., D. McKenzie, and C. Woodruff. 2009. “Are Women More Credit Constrained? Experimental Evidence on Gender and Microenterprise Returns.” American Economic Journal: Applied Economics 1 (3): 1– 32. 128 THE WORLD BANK ECONOMIC REVIEW Fafchamps, M. 1994. “Industrial Structure and Microenterprises in Africa.” Journal of Developing Areas 29 (1): 1– 30. ——— 2003. “Ethnicity and Networks in African Trade.” Contributions to Economic Analysis and Policy 2 (1): article 14. ——— 2004. Market Institutions in Sub-Saharan Africa: Theory and Evidence. Cambridge, MA: MIT Press. ¨ derbom. 2006. “Wages and Labor Management in African Manufacturing.” Fafchamps, M., and M. So Journal of Human Resources 41 (2): 346–79. Fafchamps, M., S. El Hamine, and A. Zeufack. 2008. “Learning to Export: Evidence from Moroccan Manufacturing.” Journal of African Economies 17 (2): 305–55. Foster, A.D., and M.R. Rosenzweig. 1995. “Learning by Doing and Learning from Others: Human Capital and Technical Change in Agriculture.” Journal of Political Economy 103 (6): 1176– 209. Fujita, M., P. Krugman, and A.J. Venables. 1999. The Spatial Economy: Cities, Regions, and International Trade. Cambridge, MA: MIT Press. Glaeser, E.L., H.D. Kallal, J.A. Sheinkman, and A. Shleifer. 1992. “Growth in Cities.” Journal of Political Economy 100 (6): 1126–52. Greif, A. 1993. “Contract Enforceability and Economic Institutions in Early Trade: The Maghribi Traders’ Coalition.” American Economic Review 83 (3): 525– 48. Griliches, Z. 1958. “Research Costs and Social Returns: Hybrid Corn and Related Innovations.” Journal of Political Economy 66: 419. Griliches, Z., and F. Lichtenberger. 1984. “Intraindustry Technological Flows and Productivity Growth: A Reexamination.” Review of Economics and Statistics 66 (2): 324– 29. Grossman, G.M., and E. Helpman. 1991. “Quality Ladders and Product Cycles.” Quarterly Journal of Economics 106: 557–86. H&H Consultancy. 2008. “Sudan Investment Climate Survey I: Final Survey Report.” Report submitted to the World Bank. Henderson, J.V. 1988. Urban Development: Theory, Fact, and Illusion. New York: Oxford University Press. Jackson, M.O. 2009. Social and Economic Networks. Princeton, NJ: Princeton University Press. Jacobs, J. 1969. The Economy of Cities. New York: Random House. Kling, J.R., J.B. Liebman, and L.F. Katz. 2007. “Experimental Analysis of Neighborhood Effects.” Econometrica 75 (1): 83 –119. Kraay, A., I. Soloaga, and J. Tybout. 2002. “Product Quality, Productive Efficiency, and International Technology Diffusion: Evidence from Plant-Level Panel Data.” Policy Research Working Paper Series 2759. The World Bank, Washington, DC. Liu, X., E. Patacchini, Y. Zenou, and L-F. Lee. 2012. “Criminal Networks: Who is the Key Player?” FEEM Working Paper No. 39. Fondazione Eni Enrico Mattei, Milan, Italy. Manski, C.F. 1993. “Identification of Endogenous Social Effects: The Reflection Problem.” Review of Economic Studies 60: 531–42. Melitz, M.J. 2003. “The Impact of Trade on Intra-Industry Reallocations and Aggregate Industry Productivity.” Econometrica 71 (6): 1695– 725. Melitz, M.J., and G.I.P. Ottaviano. 2008. “Market Size, Trade, and Productivity.” Review of Economic Studies 75 (1): 295 –316. Mengistae, T., and M. Honorati. 2009. “Constraints to Productivity Growth and to Job Creation in the Private Sector.” Chapter 2 in Ethiopia Investment Climate Assessment. Washington, DC: World Bank. Mokyr, J. 1990. The Lever of Riches: Technological Creativity and Economic Progress. New York and London: Oxford University Press. Muendler, M-A., J. Rauch, and O. Tocoiand. 2012. “Employee Spinoff and Other Entrants: Evidence from Brazil.” International Journal of Industrial Organization 30 (5): 447–58. ¨ derbom Fafchamps and So 129 North, D.C. 1973. The Rise of the Western World. Cambridge, UK: Cambridge University Press. ——— 1990. Institutions, Institutional Change, and Economic Performance. Cambridge, UK: Cambridge University Press. Page, J. 2012. “Can Africa Industrialize?” Journal of African Economies 21 (AERC Supplement 2): ii86– 125. Parente, S.L., and E.C. Prescott. 1994. “Barriers to Technology Adoption and Development.” Journal of Political Economy 102 (2): 298–321. Rauch, J., and A. Casella. 2003. “Overcoming Informational Barriers to International Resource Allocation: Prices and Ties.” Economic Journal 113 (484): 21 –42. Romer, P.M. 1990. “Endogenous Technological Change.” Journal of Political Economy 98 (5) pt.2: S71– 102. Sonobe, T., and K. Otsuka. 2011. Cluster-Based Industrial Development: A Comparative Study of Asia and Africa. London: Palgrave McMillan. Sutton, J., and N. Kellow. 2010. An Entreprise Map of Ethiopia. London: International Growth Centre. Sutton, J., and B. Kpentey. 2010. An Entreprise Map of Ghana. London: International Growth Centre. Tybout, J. 2000. “Manufacturing Firms in Developing Countries: How Well Do They Do, and Why?” Journal of Economic Literature 38 (1): 11 –44. World Bank. 2009. Ethiopia - Toward the Competitive Frontier: Strategies for Improving Ethiopia’s Investment Climate. Washington, DC: The World Bank. Vega-Redondo, F. 2006. Complex Social Networks. Cambridge, UK: Cambridge University Press. Young, H.P., and M.A. Burke. 2001. “Competition and Custom in Economic Contracts: A Case Study of Illinois Agriculture.” American Economic Review 91 (3): 559– 73.