The impact of the VCC platform will be evaluated using a cluster RCT spanning 92 cooperatives in Nepal. Treatment was assigned at the cooperative level through a stratified randomization design as described in detail below. Treatment effects will be estimated at both the household and cooperative levels. The cooperatives in our study cover a wide geographic area. Based on conversations with HPIN, we understand cooperatives largely sell to local markets and buyers, such that broader macroeconomic price impacts (extending beyond the geographic borders within which the cooperative operates) are not anticipated. Cooperatives are geographically based, and members of a cooperative are only able to sell through their cooperative. Because cooperative membership comes through SHG membership, and SHG membership is extremely local (tole, or neighborhood), nonmembers would only be able to join their local cooperative and not a cooperative in a different treatment group.
To form strata prior to randomization, we first used the baseline data from both the household and cooperative leader survey to identify what we expect to be strong predictors of our outcomes of interest, including region (Terai and Mid-Hills), household goat revenue over the past 12 months, number of cooperative members, and cooperative revenue. Household goat revenue was top coded, replacing obvious outliers with reported number of goats sold multiplied by the median revenue per goat sold. The strata were then created using the following steps:
1. Dropping a single cooperative that was used in a pilot of the VCC app, leaving 107 cooperatives in total.
2. Sorting cooperatives within each region by average household-level goat revenue.
3. Splitting each region at its 33rd and 66th percentiles of average household goat revenue, yielding six bins.
4. Sorting cooperatives in each of the six bins by number of cooperative members.
5. Splitting each bin at the median number of cooperative members, yielding a total of 12 bins.
6. Sorting each of the 12 bins by cooperative revenue.
7. Splitting each bin at the median of cooperative-level revenue (from any source), yielding a total of 24 bins ranging in size from four to six cooperatives.
Within each cooperative, we generated a uniform random variable and assigned the two (in the case of strata with four or five members) or three (in the case of strata with six members) cooperatives with the greatest value of the random variable to treatment. For strata with five members, the odd-numbered cooperative was assigned to treatment with 50% probability using a uniform random variable.
Two complications arose around the time of treatment assignment. First, HPIN informed the research team that cooperatives in a single district could not be part of the randomization because the cooperatives work together to market goats. At that point, the randomization was complete and cooperatives had already been contacted to coordinate VCC app training. Therefore it would have been problematic to perform the randomization again. Instead, we verified that balance for our chosen indicators was maintained after dropping the cooperatives from the aforementioned district.2 The final treatment assignment includes 92 cooperatives in total, with 45 assigned to treatment and 47 assigned to control.
Second, a coding error occurred during the randomization that resulted in 39 of 92 cooperatives being assigned to strata using the wrong value of total cooperative revenue. Fortunately, the damage caused by the error was minimal. Total cooperative revenue was used at the lowest level of stratification, and no household-level variables were affected. Total cooperative revenue is almost identical on average in treatment and control cooperatives.