Experimental Design
Our experimental design will combine c-RCT and qualitative approaches in selected communities in Adamawa and Kwara states.
Random Assignment:
A c-RCT is a field experiment in which clusters of farmers (communities) rather than farmers are randomly allocated to intervention groups. A key property of c-RCT is that inferences are intended to apply at the farmer (individual) level, while randomization is at the cluster or group level. Thus, the unit of randomization (community) is different from the unit of analysis (farmer). The intervention will follow a clustered randomized approach. The randomization takes place at the community level while selection of eligible households in each community is performed by community-based organizations operating in each community. LGAs and Communities will be randomly assigned into three groups/arms, based on the type of treatment and benefit they receive. The c-RCT will have two treatment arms, T1and T2 and a control group C:
• T1: will receive PBR cowpea (Sampea 20-T),
• T2: will receive PBR cowpea (Sampea 20-T) plus inputs.
• C: will receive conventional cowpea seed
Sampling
We will use a multistage sampling procedure to select the states, LGAs, communities and households included in the survey. In the first stage, we purposively select Adamawa and Kwara state in North-East of Nigeria. Both states are among the top cowpea-producing states in Nigeria and are states that also have a relatively high share of female-headed households engaged in cowpea production. No bias will be introduced by targeting these states with high proportion of females, because we expect randomization to even out any differences between treated and control groups. In other words, similar proportions of female-headed households in both treated and control group. We also consider the overall security situation in the states compared with some of the other major cowpea producing states. Adamawa and Kwara provide an interesting case study to analyse the impact of PBR cowpea on cowpea yield, productivity, and costs of cowpea production. Nevertheless, it is important to note that although the proposed study sites in Adamawa and Kwara are currently safe for field work, the security situation has deteriorated in many parts of Nigeria recently. If this trend continues our ability to conduct field work could be subject to some level of uncertainty.
The second stage involves the purposive selection of four LGAs from 25 LGAs of Adamawa state and four LGAs from 16 LGAs of Kwara state. Those eight LGAs from both states are chosen because there is no expected PBR cowpea penetration in these areas by the time of the intervention. The four selected LGAs in each state will be similar in terms of several important contextual factors (size, socioeconomic and agroclimatic conditions, and road and market access). From the eight LGAs, we will select 240 communities that are also similar in terms of the contextual factors, with 80 communities for the control group and the remaining 160 communities for the treatment groups. Within each community, we then list all households, from which we will randomly select 5 cowpea farmers to participate in the study, including from the control communities, for the baseline and end line surveys.
The second stage randomization will include selecting farmers randomly for PBR cowpea interventions within each of the 160 treatment communities from our study sites in Adamawa and Kwara states, where cowpea is an important crop. There are two cowpea cropping seasons: wet season and dry season. For this analysis, we will use the wet crop season. It is the main cowpea-growing season and runs from approximately July to October in the Northern Guinea Savanna agro-ecological zone, where Adamawa and Kwara are predominantly located. The intervention will take place before the beginning of the wet season of 2023.
Treatment Effects
We will evaluate the effectiveness of PBR cowpea adoption by comparing average household-level yield between T1 and C. We will also evaluate the incremental effect of interventions by comparing the average household-level yield between T1 and T2. This will give us the average treatment effect for incremental interventions. Randomization solves the problem of selection bias because households in treatment groups and control group are drawn randomly from the same underlying population; therefore, the average characteristics of these groups do not systematically vary, and any differences observed in the outcomes of interest can therefore be attributed to the interventions. To further minimize the possibility that our study design would be compromised (e.g., units selected for treatment may not, in fact, receive the treatment, or may not receive it in the fashion that is intended by the intervention), we will assign a field team (consisting of research assistants and extension agents) that will regularly visit study sites, engage with the local implementing staff, monitor progress, and report back to the evaluation team.
Data Collection
To exploit power gains from repeated observations, we will collect three rounds of data at baseline, midline and endline (e.g., McKenzie, 2012). We will collect baseline survey prior to the experiment’s rollout to investigate balance. Given that agricultural production depends on many exogenous and largely unobservable factors, such as weather, disease, pathogen and pest pressure and others, the baseline survey enables us to average out the noise in measuring yields, productivity, and costs of agricultural production, generating estimates less vulnerable to bias that might arise due to unusual conditions for all subjects during the experimental period (McKenzie, 2012; Rosenzweig and Udry, 2020). The baseline data will improve the power to estimate treatment effects, such as for examining treatment heterogeneity (McKenzie, 2012).
The choice of variables for the baseline survey is based on those used by similar studies in their orthogonality tests. In particular, we will look at variables used in studies that investigate the adoption of yield-improving technologies and practices using RCTs (Duflo, et al., 2011; Karlan et al., 2014; Ashraf, et al., 2009; Bulte et al., 2014). We will collect household information which consist of 12 modules. We will collect household characteristics such as household size, age, education level of household head, and agricultural assets, as well as more specific information related to cowpea farming, such as yield and productivity change in the last wet season, and distance to the nearest agro-input shop and road, access to extension agent and service about cowpea production and varieties, and households input use. The baseline data will improve the power to estimate treatment effects, such as for examining treatment heterogeneity (McKenzie, 2012). We will conduct the midline survey in 2024 one year after experiment’s rollout and the above interventions. An endline survey will re-survey the same households one year later after the midline to measure impacts of the long-term impact of the of PBR cowpea adoption and the change in the adoption behaviour of farmers. Given that our sample households are PBR beneficiaries or potential beneficiaries, we anticipate that attrition rate will be low or at least not systematic. In the endline survey, we will also have a module intended to collect data on unintended (positive and negative) consequences of the adoption of PBR cowpea (e.g., a decrease in number of insects that affect cowpea in the community).
Spillover effects
Finally, we will use statistical analysis to account for other important but less critical limitations of the experimental approach. One such example is related to spillover effects whereby untreated areas may have profited also from the intervention. The existence of spillovers can lead to underestimation of the treatment effect. Randomizing at the community level, rather than at the farmer level, allows the evaluation to account for spillover effects. To further account the spillover effects on our analysis, we will also expand the sample (200) to cover non-treatment households in treatment communities. This will provide useful information on channels and speed of spillover and diffusion of PBR variety. We will compare PBR cowpea adoption among control households (control households within a treatment one communities) versus control households (control households in a control communities).