Armenia - Water to Market Farmer Training

After interviewers completed each questionnaire, the interviewers reviewed the questionnaire entries and submitted them to the field coordinator for cross-editing. During data entry in SPSS, mistakes were corrected using visual and program control. In the analysis phase, subsequent edits were made to logically impute data where appropriate.

The data was entered in SPSS format by 4 specialists. Each set of responses for a questionnaire was entered by 2 specialists independently to cross-check skips and prevent mechanical mistakes. The first thousand and final five thousand entries were reviewed by Mathematica and MCA-Armenia, who compared the data entries to the hardcopy questionnaires and provided feedback on the data entry process. These data were transmitted to Mathematica for analysis.

After receiving the data, Mathematica merged the FPS3 and FPS1 data. While analyzing the data, Mathematica identified several inaccurate records of farming households. These farmers were identified systematically based on their reported amounts harvested and sold at baseline versus follow-up. First, Mathematica identified specific crop harvests and amounts sold where the farmer's report changed by over 200 tons from baseline to follow-up. This identified fourteen farmers with harvests and sale amounts for barley, grape, peach, sweet cherry, potato, red beet, haricot, and gramma. None of the 14 identified harvests and sale amounts were accompanied by large changes in crop land area or revenues. Mathematica concluded that these results were likely to be outliers and replaced the outlying number based on the information about land and crop revenues. For many of these 14 harvests, this consisted of treating a reported amount sold as the revenues for that crop. This is plausibly a data recording error in that the FPS3 records crop revenues next to crop harvest amounts. Seven additional records were similarly recoded because they implied implausible prices per unit sold.

A second approach was used to address outliers for which there was insufficient evidence to conclusively determine if the reported value was accurate. The approach was to systematically censor outcome and baseline measures of income, production, cultivated land area at the 98th percentile for each measure, or the 2nd-highest value for that measure if the 98th percentile was less than or equal to zero. This process also helps de-identify any individuals with especially large amounts of income, production, or land.

The censored variables were used to constructed nonresponse weights to adjust for differences in observed characteristics between households who did and did not respond to the FPS3. Nonresponse weights were calculated using the procedure described in Appendix A of the Water-to-Market Evaluation report. The code to construct these weights are located in the Stata program "1_armenia_construct.do". These materials are provided as external resources.