Model estimation framework

Our main interest lies in understanding the contribution of extension-acquired knowledge to the adoption of CA practices. We used mediation analysis to determine whether and to what extent the adoption effect of participating in project activities was mediated through the resulting higher-quality knowledge of CA. Treatment assignment was defined by three variables: (i) whether the farmer visited a demonstration site on an individual basis (hereafter, demonstration); (ii) whether the farmer attended structured class/field training (hereafter, training) and (iii) whether the farmer attended organized communal farmer field days. As has often been recognized in the literature, the farmers who attended various extension activities in our sample constituted a somehat self-selectng sample. This is because it is likely that those who took the initiative to come to any of the three sessions have characteristics that increased their attendance to these extension sites and events. Therefore, these groups could be a skewed social or economic class in ways we did not or could not observe. Hence, in estimating the mediation effect of interest, we must take into account the possible selection bias in the data. As we explain later, we account for selection bias by using propensity score matching (PSM). This means that our comparison groups have similar propensity scores (are statistically similar) First, we briefly expound on the mediation phenomenon that we used.

A mediating factor helps to explain how a treatment variable works (Vlaeyen and Morley, Reference Vlaeyen and Morley2005). This is equivalent to confirming the hypothesized working mechanism underlying an intervention. Using the random utility framework, consider the i ^th farm adopter (i = 1… N) facing the decision of whether to adopt the available CA practice on plot P (P = 1, …., P). Let U ₀ represent the benefits to the farmer from traditional management practices (MP), and let U_k represent the benefit of adopting the k ^th MP, where k denotes the choice of a CA practice. The farmer decides to adopt the k ^th MP on plot P if Y_ipk = U_k–U ₀ >0. The perceived net benefit of adopting any MP is influenced by how well the farmer understands and implements the various aspects of the MP in question. This is important because implementing the MP accurately as per the recommendation ensures cost-effectiveness, and proven cost-effectiveness would sustain adoption. For the successful implementation of MP, farmers need to internalize the principles of MP in the process of evaluating and trying it on their farms. Therefore, the channel we test is whether farmers' knowledge of MP directly affects the decision to adopt CA practices. This provides the basis for the choice of the outcome and mediating variables. The intervention may translate into better knowledge about MPs in general, thereby increasing its adoption.

Let Y_i represent an outcome variable (adoption of an MP), T_i is the treatment, X_i denotes control variables and M_i is the mediator. Mediation analysis can be conducted by estimating the equations:

(1)$$M_i = {\rm \alpha }_1 + {\rm \beta }_1T_i + {\rm \gamma }_1X_i + {\rm \epsilon }_{1i}$$

(2)$$Y_i = {\rm \alpha }_2 + {\rm \beta }_2T_i + {\rm \gamma }_2X_i + {\rm \sigma }_2M_i + {\rm \epsilon }_{1i}.$$

For the empirical estimation of Equations (1) and (2) to represent a causal mediation analysis, the following conditions should be satisfied: (i) predictor T (treatment) should be causally related to outcome Y before mediator M is added to the model, (ii) T should be causally related to M and (iii) M should be causally related to Y when controlling for T.

The first two conditions assume the exogeneity of the treatment T to Y and M (random assignment of the treatment). In observational studies, treatments are not randomized, which makes it difficult to satisfy the exogeneity assumption. Conditional on using PSM or the Heckman selection model, we can mimic the random treatment assignment, which implies that the average treatment effect is identified. However, calculating the average treatment effect occurring through a mediation variable M_i (T_i), which is also affected by the treatment, requires additional assumptions. The third condition is the exogeneity assumption of the mediator, which states that once we condition on T_i, the mediator should be exogenous. This assumption implies that the effects of unmeasured covariates that confound the relationship between the mediator and the dependent variable have been removed.

To understand why an intervention affects an outcome, we turn to the causal mechanism framework. We follow the reasoning on causal mechanisms by Imai et al. (Reference Imai, Keele, Tingley and Yamamoto2011). The causal effect of treatment might flow through another intermediate variable on the causal pathway from treatment to outcome. The causal pathway is an indirect or mediated effect, which tells us how the effect of the treatment depends on a particular pathway. To define the key quantities of interest in our causal mediation analysis, we rely on the potential outcome framework for causal inference (Rubin, Reference Rubin1974). Assuming all three assumptions stated above are satisfied, the estimates from Equations (1) and (2) can be used to estimate the causal true parameters. Therefore, β₁ denotes the effect of the treatment on the mediator, while σ₂ is the effect of the mediator on the outcome. Thus, the average causal mediation effect (ACME) is computed as σ₂β₁. For ACME to hold, the sequential ignorability assumption implies that the correlation between the two error terms ε_1i and ε_2i is zero. The above equations are estimated assuming this zero correlation between the error terms. Note that β₂ represents the average direct effect (ADE). The average treatment effect (ATE), which represents the effect of the treatment on the outcome variable, can be decomposed into the sum of ACME and ADE (i.e., ATE = ACME + ADE). Given the importance of testing the validity of the sequential ignorability assumption, we use the sensitivity analysis developed by Imai et al. (Reference Imai, Keele, Tingley and Yamamoto2011) to determine if our assumption of sequential ignorability is defensible.

Dependent (outcome) variables

This study focused on explaining the adoption of two elements of CA, namely, the area under conservation tillage (previously denoted as CT) and the area under soil cover (previously denoted as SC). These constitute the basic outcome variables. CT was defined as the practice of opening only seeding holes to plant the seed without tilling the plot. In the use of CT, weeding was mechanically accomplished by shallow weeding (chopping off the weeds at above-ground level or by very shallow digging) or spraying with a herbicide. The use of SC was defined as the application of crop residues (mainly those of maize) as mulch for weed suppression and moisture conservation by forming a protective cover on the surface. The two practices were selected to represent the two most critical aspects of CA, without which CA systems would not succeed. These two elements are also the most novel (especially CT) and require the most improvement in know-how. Although crop rotation and intercropping are part of CA, we did not include them in the analysis because they have been part of the farming system and are widely adopted before CA was promoted in these communities. When combined with crop rotation and intercropping the three (CT, CA and intercropping/rotations) will constitute CA. Therefore, the adoption of CT and SC provides an appropriate case for how adoption is mediated by improved know-how arising from information exposure to the innovation.

During the household interviews, each respondent was asked to provide information on the use of CA and other practices on all their maize sub-plots. A sub-plot was defined as a contiguous patch of land on which maize was grown as the primary crop (solely maize, in rotation or intercropped with any legume) and using a particular MP. For each sub-plot, we obtained information regarding the practices used to grow maize, including whether the sub-plot was managed using any CA component practices: whether CT or normal tillage was used, in addition to whether SC (mulching) was used and other practices. The respondent was asked to estimate the size of each sub-plot and what portion was managed under any of the CA practices, and these areas were standardized to hectare equivalents.

Treatment variables

The treatment variables considered were exposure to the project extension (scaling) activities. Recall that the project implemented a mix of approaches to promote CA practices among the communities around the demonstration sites. The three main extension methods were: individual (informal) visits to CA demonstration sites, organized formal farmer field days and in-class and field-based training sessions on CA and related aspects. The project field teams organized demonstrations in high-visibility areas to showcase CA farming techniques. These demonstration sites were locations along major village roads, near public facilities such as markets and religious compounds, and government offices. While the demonstrations were meant to ‘speak for themselves’, the project teams also organized field days for the community to interact with the extension and project staff. The final aspect was the hands-on training sessions on CA practices. The data (presented later) show variation in the number of times any individual participated in the different activities. Project teams implemented the activities in collaboration with local extension staff, farmer community groups and other nonprofit community development organizations. This multi-stakeholder approach was intended to help mobilize and capitalize on the convening capacity of grassroots organizations. Over a period spanning approximately 7–8 growing seasons, project teams implemented these three methods to deliver the information by asking farmers to observe the CA demonstration by visiting the demonstration sites, attending training and field days and trying the new methods on their farms.

Mediation variable

The basic notion behind the mediation variable is that exposure to demonstrations, field days and training will promote the adoption of CT, SC and related components by improving farmers' technical understanding of CA. A question was asked to let all respondents describe what they understood as CA. The interviewer took note of whether the respondent described something but mentioned none of the three principles (minimum/zero tillage, mulching and crop diversification by rotations or intercropping), or the respondent acknowledged a complete lack of awareness. Therefore, each farmer's response was categorized as (a) no information or knowledge (as above), (b) they described CA but fell short of listing all three principles, mentioning a maximum of two of the three principles (this category was described as having partial or incomplete knowledge of CA), and (c) they mentioned all three elements in their description; these farmers were categorized as having accurate or complete knowledge of CA. In constructing the mediation indicator variable, we categorized farmers into two: those who had accurate (complete) knowledge of CA and those who had only partial or no knowledge.

Other covariates

Finally, other covariates were used in the regression models. These control variables included the total number of members in a household and demographics of the household head and spouse (sex, age, years of education and years of education). Institutional (community) variables included the distance of the residence from the nearest major trading center and distance to the nearest farm input sales point in kilometers. As described above, farm-level covariates included the total area of all plots cultivated under maize and other associated crops (ha.), number of livestock owned, distance from residence to the plot and the plot size on which CA was implemented.