Impact of Aquaculture Training in Ghana: Cluster Randomized Trial

The intervention is the provision of information to farmers. This is combined practical face-to-face training + access to mobile app. Information to be provided in mainly on good aquaculture practices.

2020-03-30

2021-05-28

Productivity (harvest in kg per m2)

Adoption of good aquaculture practices (dummy, count, index/score)

Dummy, count Index or score based on pre-determined list of good aquaculture practices, including record-keeping, hygiene and sanitation

To minimize risk of contamination, the unit of randomization will be clusters (that is, districts).
Due to small sample size and few geographical locations, we plan to use to pair-matching method first and then randomization (that is, assigning randomly 1 to the treatment group and the other to the control group). Within the resulting pairs, the intervention is randomized, and the outcomes measured at the end of follow-up survey. We consider this design to be adaptive, because the construction of the matched pairs depends on the baseline covariates of all candidate units.
We produced 19 district pairs based on cluster size (number of farmers in each district) and region. We used mahascores and mahaselectunique commands in Stata to select the optimal pairing based on the covariates. In each pair, we randomly assigned 1 in the Treatment and the other in the Control group using a coin toss by the research team.

Given random assignment to the treatment, intention-to-treat effects are estimated by ordinary least squares, where the variable of interest is the indicator variable equal to one if the district was assigned to the treatment group (receiving the aquaculture training). The outcome can then be written as:
Y = a + b*T + e
Where Y is the outcome indicator (see above), T is the assignment to the Treatment (provision of information), and b measures the average effect of the intervention (provision of the information).
We checked the balance or equality of means between the treatment and control using various individual and location-level covariates. We used gender, age and education level of owner and manager, region, asset quintile, land area owned, size of biggest pond, total pond size, proportion of income coming from aquaculture and access to extension services. We also checked for the equality of means in intermediate and primary outcome indicators: ratings of the facility, record-keeping, survival rate, stocking, feed use, and productivity. In almost all of them, we could not reject equality of means between the treatment and control. This means the we achieved balance of baseline characteristics between treatment and control. The only covariate that is not balanced is the size of the biggest pond, in which treatment districts have bigger ponds on average across the largest ponds of the sample farmers. We will control for this in the regression analysis (ex-post).
Moreover, while we tried to minimize the risk on contamination and information spillover between treatment and control districts, it is still possible to occur. We will control for the distance between the farmer to the nearest border of the nearest treatment district to capture spillover. In the follow-up survey, we will also ask for detailed interactions of the farmer with other farmers to give some sense of information sharing and spillover especially between treatment and control districts that are near to each other. If any, the effect (b) can be interpreted as the lower bound of the true effect.


We test the null hypothesis (b=0). If rejected, we conclude that the intervention has significant effect to the magnitude of b.

Randomization using coin flip by the research team

district

Yes

38 districts (groups of villages)

384 small-scale tilapia pond farmers

192 farmers per treatment arm

We used clustersampsi in Stata to compute for the minimum number of clusters needed and the number of sample per cluster. The mean of the outcome indicator (productivity (kg/m2)) is 1.14, and the standard deviation is 0.52. Using complete randomization, minimum sample size per arm would have been 67 with 80% power and 5% level of significance. Given the randomization done at cluster level, we have to compute for the intraclass correlation (ICC), which is 0.10. Given that the districts have different number of farmers (therefore different cluster size), we have to further adjust the sample size to detect the minimum effect we are interested in. The cluster sizes range from 3-37, and the coefficient of variation (of cluster sizes) is 0.82. The design effect should be inflated by 2.57, accounting for the ICC and varying cluster sizes. The study is limited by the number of districts at least 3 small-scale active tilapia pond farmers. On the 4 major-producing regions (Brong Ahafo, Ashanti, Eastern and Volta), there are only 38 districts with at least 3 small-scale active tilapia pond farmers. The total number of farmers in these districts is 384. For 1 treatment (aquaculture training), we have 19 clusters and 192 farmers available for the study. With this design effect inflator, the sample size of 192 farmers in 19 clusters in each treatment arm can detect a minimum of 18% increase in the productivity (kg per m2). Local partners predict and aims an increase in 20%, so we have enough power to detect the predicted outcome increase. Last, given that we are doing a pair matching method before randomization, we can further deflate the design effect to capture this increase in efficiency and power.