Moldova - Value Chain Training

The study population includes farm operators in approximately 88 communities--48 treatment communities, 32 control communities, and 8 A-list communities (high priority sites that were purposively selected to receive training). To be included in the study, farmers must have cultivated targeted crops (which, for each community, were identified in advance by the implementer). Across these 88 communities, about 2100 farmers were interviewed in the 2012-2013 FOS.
1. Sample frame
For the sample frame, the survey contractor developed a list of all farm operators cultivating crops in targeted value chains in the 80 study communities (treatment and control) and 8 A-list communities (high priority sites that were purposively selected to receive training). This list included information about farm size and which of the targeted crops the farm operator cultivated. In three communities, the survey contractor did not identify any farmers cultivating targeted crops, so the final sample frame included 77 study communities and 8 A-list communities. Information on total farm size was used to draw separate samples for farms of different sizes.

2. Drawing the sample
For small farms (less than 10 hectares), we drew a random sample of farm operators in targeted value chains in each community. To determine the number of farmers to select in each community and to select farmers, we implemented the following steps:

·We allocated the total small-farm sample across communities in proportion to their size (the number of small-farm operators in targeted value chains). For example, if one community had twice as many treatment small-farm operators as another, we allocated twice as many small-farm operators to that community. To ensure that very small communities were adequately represented and that very large communities do not drive the impact estimates, no community's sample could be below a minimum of 20 or above a maximum of 150 small farmers. Allocating the sample in this way ensured that the sample was balanced across communities but still close to self-weighting.

·We drew the sample in each community using implicit stratification by value chain. We used implicit stratification by value chain (sorting farmers in each community by value chain and selecting the sample so that it was evenly spread across this ordered list) to ensure that the randomly-selected sample provided proportional representation of the different value chains in each community.

For medium (between 10 and 100 hectares) and large (100 hectares or larger) farms, we determined that there were relatively few farms in the value chain training sample frame (174 medium farms and 77 large farms). We therefore attempted to interview all operators of these farms so that we would have precise estimates for these groups.

3. Use of replacements
In some cases, the survey contractor was unable to conduct an interview with a selected farm operator. This occurred for various reasons, such as refusal to participate or ineligibility for the survey (if it was determined that the operator did not cultivate the targeted value chains). To account for this, we developed a list of replacement farmers in each community at the same time that we selected our initial sample. Because all medium and large farmers were selected for the sample, the replacement list included only small farmers. These procedures were designed to help ensure that we reached our target sample sizes for the analysis while maintaining the representativeness of the sample to the extent possible and keeping the replacement procedure reasonably straightforward.


The analysis sample does not include all respondents to the survey. The analysis sample excludes farmers from one stratum that had five treatment communities and three control communities. This stratum was excluded because it contained virtually no control farmers. As a result, the analysis sample includes 902 farmers in 41 treatment communities, 563 farmers in 28 control communities, and 200 farmers in 8 A-list communities.


The overall response rate was 83 percent in treatment and control communities.


Our sampling strategy attempted to create a survey sample that was as close to self-weighting as possible. However, we still need to apply weights to ensure that our analysis sample is representative of farm operators in the targeted value chains in the treatment and control communities. We constructed weights to account for:

·Differences in sampling probabilities across farmers. We drew the sample of eligible small farmers using implicit stratification in each community. The sampling probability for small farmers in a given community was therefore determined by the fraction of small farmers sampled in that community. Because the community allocations were roughly proportional to the number of eligible farmers in each community (except for small deviations due to the minima and maxima we imposed), this sampling probability was similar for most small farmers. Nevertheless, we need to adjust for the small deviations in this probability. We surveyed all medium and large farmers; therefore, their sampling probability was one. The inverse of the sampling probability was used to obtain a farm-level sampling weight for each farmer.

·Possible differential nonresponse across different types of farmers. To adjust for possible systematic nonresponse among certain types of farmers, we computed response rates within cells that we defined by random assignment stratum, treatment status, and farm size (small, medium, or large). We used the inverse of the response rate to obtain a nonresponse weight for all farmers in a given cell.

We then multiplied these weights to yield preliminary farm-level weights. In addition, to ensure that treatment status was not correlated with random assignment stratum, we reweighted the control farms in each stratum so that their (weighted) sum was equal to the (weighted) sum of treatment observations in that stratum. Finally, we normalized these adjusted weights so that their sum was equal to the number of observations for each farm size group (small, medium, and large).

Baseline differences between the treatment and control groups were estimated in a regression framework. This regression model enabled us to account for the features of the evaluation design, specifically the stratified random assignment. In addition, because the unit of random assignment is the community, to obtain the correct standard error for the baseline differences we had to account for the fact that outcomes in the same communities are likely correlated. The regression model enabled us to account for this using the “cluster” correction in Stata, with the community as the level of clustering.