Using Respondent Prediction Models to Improve Efficiency of Incentive Allocation

In an effort to increase response rates, the National Agricultural Statistics Service (NASS) began experimenting with monetary incentives in 2004. Follow-up assessments of the monetary incentive in 2005 demonstrated that ATM cash cards are beneficial in increasing Agricultural Resource Management Survey Phase III (ARMS III) response rates and decreasing survey costs; however, it is unknown which sampled units would have responded without the incentive. This paper discusses the use of data mining to identify likely ARMS III respondents. A series of models were built using 2002 Census of Agriculture data to predict several years of ARMS III sample respondents before (2003-2004) incentives were introduced. These models were applied to the years after incentives were introduced (2005-2007) to confirm that they continued to identify likely respondents. The respondent prediction models discussed in this report enable NASS to flag persons likely to respond given no incentive. Providing incentives to these respondents requires substantial costs, but likely does not increase overall response rates. In addition, if providing them incentives does increase response rates, it may increase them in such a way that NASS estimates are further biased if only more of the same type of operations opt to respond