Understanding farmers' technology adoption decisions: input complementarity and heterogeneity

Agriculture growth in Africa is often characterized by low aggregate levels of technology adoption. Recent evidence, however, points to co-existence of substantial adoption heterogeneities across farm households and a lack of a suitable mix of inputs for farmers to take advantage of input complementarities, thereby limiting the potential for learning towards the use of an optimal mix of inputs. We use a detailed large longitudinal dataset from Ethiopia to understand the significance of input complementarities, unobserved heterogeneities, and dynamic learning behavior of farmers facing multiple agricultural technologies. We introduce a random coefficients multivariate probit model, which enables us to quantify the complementarities between agricultural inputs, while also controlling for alternative forms of unobserved heterogeneity effects. The empirical analysis reveals that, conditional on various types of unobserved heterogeneity effects, technology adoption exhibits strong complementarity (about 70 percent) between chemical fertilizers and improved seeds, and relatively weaker complementarity (between 6 and 23 percent) between these two inputs and extension services. Stronger complementarities are observed between specific extension services (advice on land preparation) and improved seed and chemical fertilizers, as opposed to simple visits by extension agents, suggesting that additional benefits can be gained if the extension system is backed by “knowledge” inputs and not just focus on “nudging” of farmers to use these inputs. The analysis also uncovers substantial unobserved heterogeneity effects, which induce heterogeneous impacts in the effect of the explanatory variables among farmers with similar observable characteristics. We also show that ignoring these behavioral features bears important implications in quantifying the effect of some policy interventions which are meant to facilitate technology adoption. For instance, ignoring these features leads to significant overestimation of the effectiveness of extension services in facilitating technology adoption. We also document strong learning behavior, a process that involves learning-by-doing as well as learning from extension agents.