Associate Prof. Phoebe Koundouri

1. Information Transmission in Technology Diffusion: Social Learning, Extension Servicesand Spatial Effects in Irrigated Agriculture Associate Prof. Phoebe Koundouri Director of RESEES [Research tEam on Socio-Economic Sustainability] Athens University of Economics & Business Visiting Senior Research Fellow Grantham Research Institute on Climate Change and the Environment London School of Economics & Political Science

2. Co-Authors: • Margarita Genius, Department of Economics, University of Crete. • Celine Nauges, Toulouse School of Economics and the University of Queensland. • Vangelis Tzouvelekas, Department of Economics, University of Greece. We acknowledge the financial support of the European Union FP6 financed project FOODIMA: Food Industry Dynamics and Methodological Advances (Contract No 044283). www.eng.auth.gr/mattas/foodima.htm.

3. Keywords: • agricultural technology adoption and diffusion • information transmission • social learning and social networks • extension services • dynamic maximization behavioral model under uncertainty • risk preferences • factor analytic model • flexible method of moments • duration analysis • olive farms

4. Aim and Contribution • To investigate the role and sources information transmission in promoting irrigation technology adoption and diffusion (TAD) with socioeconomic, production risk , environmental & spatial considerations . • Why irrigation technology?Central to increasing water use efficiency, economizing on scarce inputs, maintaining current levels of farm production, particularly in semi-arid & arid areas [CAP, WFD, etc.] • Structure of Work and Methods: - dynamic maximization behavior model - econometric duration analysis model merged with: - factor analysis for identification of info transmission paths and peers - flexible method of moments for risk attitudes estimation - applied to micro-dataset olive farms in Crete • Findings: both extension services and social network are strong determinants of TAD, while the effectiveness of each type of informational channel is enhanced by the presence of the other. Policy Implications!

5. Background Literature

6. Several empirical studies, in developed & developing countries, on modern irrigation TAD patterns: e.g. Dinar, Campbell and Zilberman AJAE 1992 Dridi and Khanna AJAE 2005 Koundouri, Nauges and Tzouvelekas AJAE 2006, etc. Evidence that: • economic factors: e.g. water and other input prices, cost of irrigation equipment, crop prices • farm organizational and demographic characteristics: e.g. size of farm operation, educational level, experience • risk preferences with regards to production risk • environmental conditions: e.g. soil quality, precipitation, temperature …matter in explaining TAD.

7. …also TAD patterns are conditional on knowledge about new technology: • Besley and Case AER 1993 • Foster and Rosenzweig JPE 1995 • Conley and Udry AER 2010, etc. Sources of Information and Knowledge: • Extension Services (private or public). World Bank: Rivera & Alex 2003; World Bank 2006; Birkhaeuser et al. 1991 Usually ES target specific farmers who are recognized as peers. • Social Learning: Rogers 1995: via peers (homophilic or heterophilic neighbors).

8. Peers: farmers exerting a direct or indirect influenceon the whole population of farmers in their respective areas. Homophilic Heterophilic Farmers may also follow or trust the opinion of those that they perceive as being successful in their farming operation, even though they occasionally share quite different characteristics. Farmers exchange information and learn from individuals with whom they have close social ties and with whom they share common professional or/and personal characteristics (education, age, religious beliefs, farming activities etc.) Measuring the extent of information transmission via different channels& identifying its role in TAD is difficult: • 1. Set of peers difficult to define (Maertens and Barrett AJAE 2013): we need to go beyond the simplistic definition of peers as (physical) neighbors. • 2. Distinguishing learning from other phenomena (interdependent preferences & technologies; related unobserved shocks) that may give rise to similar observed outcomes is problematic (Manski RES 1993).

9. Theoretical model

10. Expected efficiency gains are uncertain at adoption decision time.Uncertainty can be reduced via accumulation of knowledge:- Extension Services, before and after adoption- Social Networks, before and after adoption- Learning-by-doing /using, after adoption • Farmers decision to invest in a new irrigation technology (NIT). • NIT improves irrigation effectiveness (shift in the production technology). At each time period the farmer decides whether to adopt by comparing : • Expected Cost of adoption (decreasing function of time) • Expected benefit of adoption depending on information transmission & accumulation, socio-economic (including risk attitudes) & environmental characteristics

11. Farm's j technology, continuous twice-differentiable concave production function: yj: crop production xjv: vector of variable inputs (labor, pesticides, fertilizers, etc.) xjw: irrigation water xjw< , risk of low (or negative) profit in case of water shortage. Adoption allows hedging against the risk of drought and consequent profit loss. Aj: technology index: irrigation effectiveness index: (water used by crop)/(total water applied in field) Aj⁰ with conventional technology Aj* with new technology farmer produces same y using same xv and lower xw. Aj = Aj* : max irrigation effectiveness is reached Aj* > Aj⁰ : max irrigation effectiveness cannot be reached with Aj⁰ May require time before the new technology is operated at A*. : the expected, at time s, efficiency index for t, under the assumption the new technology is adopted at time τ. c , : fixed cost of NIT known at period t.

12. Modeling the timing of Adoption Fixed time horizon T s τ Info gathering period s+1 τ+1 Info gathering period s + 2 Info gathering period ………………………………… 0 τ (adoption time) t (time) E(Profit) for t until T Decision: Adopt or Not E(Profit) for t until T Decision: Adopt or Not E(Profit) For t until T Decision: Adopt or Not ………………………..

13. Expected Discounted Profit Functions: • expected discounted crop, irrigation water, variable input prices (assumed dynamically constant by farmer). • positively linearly homogeneous and convex in p, wv, and ww • non-decreasing in crop price and irrigation technology index • -non-increasing in variable input and irrigation water prices. No Adoption during t: Adoption at τ: Farmer max over τ her temporal aggregate discounted profit:

14. Farmer’s Trade-off : Benefit: Delaying investment by one year allows the farmer to purchase the modern irrigation technology at a reduced cost. Cost: Delaying adoption by one year results in producing with the conventional less efficient technology and bearing a higher risk of water shortage (thus a loss in expected profit). Note: Farmer considers that technology efficiency index will remain constant after adoption because she does not have enough information to predict the evolution of the technology efficiency after adoption (which is a complex function of learning from others and learning-by-doing). The model could be extended to allow for the farmers anticipating learning after adoption. Such an extension would need to incorporate assumptions about farmer-specific learning curves, which will differ between adopters based on initial adoption time and farmer-specific socio-economic characteristics. Such an extension does not alter the learning processes of our model, neither before, nor after adoption, but it does make the first order conditions less clear.

15. Adoption Decision: Expected Discounted Equipment Cost: • At any point in time, s, farmer j assumes a rate of decrease for the discounted equipment cost: • is a decreasing value of k, and converges to , the asymptotic discounted equipment cost for farmer j at time s, as k→∞. Adoption Equation: The quantity represents approximately the expected excess discounted cost, between choosing to adopt the new technology at time s+1, namely, as soon as possible, and postponing the adoption for a very long period, namely, for a period where the rate of decrease of the equipment cost is practically zero.

16. Heterogeneity of Adoption DecisionDeriving from heterogeneity in E(π),which derives from: • Farm-specific expected cost for technology and farm-specific Water Efficiency Index, depending on farm specific Knowledge Accumulation Via: extension services before and after adoption social learning before and after adoption learning by doing after adoption • Farm Specific Knowledge Accumulation depends on socioeconomic characteristics (age, education, experience) farm location identification & behavior of influential peers • Farm-specific characteristics input & output prices risk preferences spatial location environmental conditions (defining min water crop requirements)

17. Incorporating Risk Attitudes in the Analysis Endogenous Technology Adoption Under Production Risk: Theory and Application to Irrigation TechnologyKoundouri, Nauges, Tzouveleka2, AJAE 2006 Investigate the microeconomic foundations of technological adoption under production risk and heterogeneous risk preferences. Methodology: - Construct theoretical model of adoption by risk-averse agents under production risk - Approximate it with flexible empirical model based on higher-order moments of profit. - Derive risk preference from estimation results. - Use risk preferences to explain adoption through a discrete choice model. Results: - Risk preferences affect the prob. of adoption: evidence that farmers invest in new technologies as a means to hedge against input related production risk. - The option value (value of waiting to gather better information) of adoption, approximated by education, access to information & extension visits, affects the prob. of adoption.

18. Technology Choice Depends: Antle (1983, 1987): max E[U(π)] is equivalent to max a function of moments of the distribution of ε (=exogenous production risk), those moments having X as arguments. Agent's program becomes:

23. Survey Design & Data description

24. Survey carried out in Crete during the 2005-06 cropping period. • Part of EU funded Research Program FOODIMA.5 • Agricultural Census (Greek Statistical Service) used to select a random sample of 265 olive-growers located in the four major districts of Crete. • Farmers were asked to recall: - time of adoption (drip or sprinklers) - variables related to their farming operation on the same year (: production patterns, input use, gross revenues, water use and cost, structural and demographic characteristics). • A pilot survey showed that none of the surveyed farmers had adopted before 1994. • Interviewers asked recall data for the years 1994-2004 (2004 being the last cropping year before the survey was undertaken). • All information was gathered using questionnaire-based interviews undertaken by the extension personnel from the Regional Agricultural Directorate. • Out of the 265 farms in the sample, 172(64.9%) have adopted drip irrigation technology between 1994 and 2004.

26. Table 1: Definitions and Summary of the Variables (cont.)

27. The variable of interest in empirical application is the length of time between the year of drip irrigation technology introduction (1994) and the year of adoption. Mean adoption time is 4.68 years in our sample.

28. Production Risk & Moments of Profit Distribution • In order to capture the impact of this uncertainty on farmers' adoption decision we follow Koundouri, Nauges, and Tzouvelekas (2006) utilizing moments of the profit distribution as determinants of adoption. • Using recall data on: - olive-oil revenues - variable inputs (labor, fertilizers, irrigation water, pesticides) - fixed (land) input • Estimated the linear profit function: The residuals have been used to estimate the kth central moments (k=1,…,4) of farm profit conditional on variable and fixed input use.

29. Measurement of Information Transmission • Each farmer provided information on: - number of extension visits on her farm prior to the year of adoption - age and educational level of her peers (according to farmer) • Data on farm location: -geographical distance between the farmer and extension agencies - geographical distance between farmers and her peers

30. Measurement of Information Transmission • Stock: stock of adopters on the year the farmer adopted • HStock: stock of homophilic adopters (farmers in same age group (6 year range) and with similar education levels (2 year range)). • RStock: stock of homophilic adopters as identified by the farmer (computed as stock of homophilic adopters among those identified by farmer as belonging to his reference group). • Dista : average distance to adopters • HDista: average distance to homophilic adopters • RDista : average distance to homophilic adopters as identified by the farmer • Ext : no. on-farm extension visits until the year of adoption • Hext: no. on-farm extension visits to homophilic farmers • RExt : no. on-farm extension visits to homophilic adopters as identified by the farmer • Distx : distance of the respondent to the nearest extension agency • HDistx : average distance of homophilic farmers to the nearest extension agency • RDistx : average distance of homophilic adopters as identified by farmer, to the nearest extension outlet

32. Econometric model : Duration analysis & factor analysis & Flexible method of moments

33. Duration Model of Adoption and Diffusion Survival models in statistics relate the time that passes before some event occurs to one or more covariates that may be associated with that quantity of time.Duration model formulated in terms of conditional probability of adoption at a particular period, given that adoption has not occurred before and given farmer-specific information channels, socioeconomic (& risk attitudes), environmental &spatial characteristics.

34. Empirical Hazard Function • Assume T follows a Weibull distribution the hazard function is: • α : scale parameter • α> 1: hazard rate increases monotonically with time • α < 1: hazard rate decreases monotonically with time • α = 1: hazard rare is constant • vector zit : variables that determine farmers' optimal choice Some vary only across farmers (e.g. soil quality and altitude) other vary across farms and time (e.g. cost of acquiring the new technology) • β : corresponding unknown parameters

35. Marginal Effects on Hazard Rate and Adoption Time

36. Factor Analysis:Identification Peers and Information Transmission Paths Peers are not JUST physical neighbors! • Statistical method used to describe variability among observed, correlated variables, in terms of a potentially lower number of unobserved variables, called factors. • The observed variables are modeled as linear combinations of the potential factors, plus error terms. • The information gained on interdependencies between observed variables is used later to reduce the set of variables in a dataset. • Estimated Factors will be included in vector zit in our empirical hazard function.

37. SOCIAL NETWORK CHANNEL: Observable Indicators for Latent Variable 1:Total no. of adopters in farmer's reference group • Stock: stock of adopters on the year the farmer adopted • HStock: stock of homophilic adopters (farmers in same age group (6 year range) and with similar education levels (2 year range)). • RStock: stock of homophilic adopters as identified by the farmer (computed as stock of homophilic adopters among those identified by farmer as belonging to his reference group). • SOCIAL NETWORK CHANNEL: Observable Indicators for Latent Variable 2: Distance of farmer to adopters in her reference group • Dista : average distance to adopters • HDista: average distance to homophilic adopters • RDista : average distance to homophilic adopters as identified by the farmer • EXTENSION SERVICES CHANNEL: Observable Indicators for Latent Variable 3:Overallexposure to extension Services • Ext : no. on-farm extension visits until the year of adoption • Hext: no. on-farm extension visits to homophilic farmers • RExt : no. on-farm extension visits to homophilic adopters as identified by the farmer • EXTENSION SERVICES CHANNEL: Observable Indicators for Latent Variable 4:Distance of farmer to Extension Agencies • Distx : distance of the respondent to the nearest extension agency • HDistx : average distance of homophilic farmers to the nearest extension agency • RDistx : average distance of homophilic adopters as identified byfarmer, to the nearest EA

38. Empirical Factor Analytic Model • Aim: To proxy 4 latent variables using the 12 observable indicators: x : the vector of 12 observable indicators : latent components, (4x1) random vector with zero mean and variance-covariance matrix I μ : vector of constants corresponding to the mean of x Γ : (12x4) matrix of constants v: (12x1) random vector with zero mean and variance-covariance matrix Factor analytic model estimated using principal components method with varimax rotation. [From the perspective of individuals varimax seeks a basis that most economically represents each individual: each individual can be well described by a linear combination of only a few basis functions.]

39. Note: Estimation of Factor Analytic Model • Principal component analysis (PCA): Mathematical procedure: uses an orthogonal transformation to convert a set of observations of correlated variables into a set of values of linearly uncorrelated variables called PCs. (no. PCs ≤ no. of original variables. • Varimax Rotation: maximizes the sum of squared correlations between variables and factors. Achieved if: (a) any given variable has a high loading on a single factor but near-zero loadings on the remaining factors (b) any given factor is constituted by only a few variables with very high loadings on this factor, while the remaining variables have near-zero loadings on this factor. • If these conditions hold, the factor loading matrix is said to have simple structure, and varimax rotation brings the loading matrix closer to such simple structure. • From the perspective of individuals varimax seeks a basis that most economically represents each individual: each individual can be well described by a linear combination of only a few basis functions.

40. Estimation of Factor Scores to be incorporated in Hazard Function • All pair-wise correlations between the 12 observed InfoVar are significant at the 0.01 level • All 12 InfVar are used in order to predict each of the four latent variables • Assuming multivariate normality of observable indicators and ξi, we estimate factors scores ξmi, m=1,…,4, for the ith farmer (s = 12 InfVar):

41. Table 3: Estimation Results of the Factor Analytic Model for Informational Variables

42. Estimation of Proportional Hazard Model(:effect of a unit increase in a covariate is multiplicative with respect to the hazard rate) • Using regression calibration we approximate : • By:

43. Empirical results & policy implications

44. Empirical Analysis • Sample of 265 randomly selected olive-growing farms in Crete, Greece. • Estimate higher moments of profit (FMM). • Estimate factor scores (PCA & varimax rotation). • Merge profit moments & factor scores in hazard function and estimate a duration model (right censored ML) • Consistent standard errors via stationary bootstrapping (Politis & Romano 1994) • Empirical Results: • Determination of diffusion curve of modern irrigation technology • Insights on impact on diffusion process & adoption time of: • information and learning channels • risk preferences • other socio-economic factors • environmental and spatial characteristics -ve coefficient implies a negative marginal effect on duration time before adoption: faster adoption.

47. Note: Estimation Robustness Check • Estimation of hazard function including (model A.1) & excluding 4 latent variables (model A.2). • All key explanatory variables in both models are found statistically significant. • Signs of estimated parameters remarkably stable between models • Reduced model underestimates the effects of age and tree density on mean adoption time while it overestimates the effect of education, crop price, and mean profit. • Akaike and the Bayesian information criteria indicate that the full model is more adequate in explaining variability in farmers' adoption times. • Predicted mean adoption times are not statistically different: 5.76 and 5.74 in the full and reduced model.

48. Discussion of Results I : Epidemic Effects Scale parameter of the Weibull hazard function is statistically significant and well above unity in both models. Endogenous learning as a process of self-propagation of information about the new technology that grows with the spread of that technology: • the pressure of social emulation and competition: not highly relevant for farming business • learning process and its transmission through human contact: captured explicitly via the latent information variables • reductions in uncertainty resulting from extensive use of the new technology: learning-by-doing effects

49. Empirical Results II : Extension & Social Learning EXTENSION SERVICES • Exposure to extension services induces faster adoption (-0.306 years). • The bigger the distance from extension outlets the shorter the time before adoption (- 0.0531) Extension agents primarily targeting farmers in remote areas, as these farmers are less likely to visit extension outlets. SOCIAL LEARNING • Larger stock of adopters in the farmer's reference group induces faster adoption (-0.293 years). • Greater distance between adopters increases time before adoption (0.172 years). The impact of social learning is comparable to the impact of information provision by extension personnel (mean marginal effects on adoption times are -0.293 and -0.306 for the stock of adopters and exposure to extension services, respectively).

50. Empirical Result III: Complementarity of Information Channels • Interaction term between the two channels of information transmission is statistically significant and negative: complementarity. • The passage of information cannot be made ONLY by using rules of thumb (manuals and blueprints) mainly utilized by extension personnel, but instead it also requires strong social networks between olive-growers already engaged in learning-by-doing. • The complementarity between the two communication channels in enhancing technology diffusion points to the need of redesigning the extension provision strategy towards internalizing the structure and effects of farmers' social networks.