Modeling NML Using the Area Frame Survey

This report outlines the results of various experiments to model the probability of an area-frame farm not being on the census mailing list (NML) using covariates such as total sales, type of farm, acreage, various operator characteristics (gender, Hispanic status, race, and whether the primaryl occupation of the principal operator is farming), number (if any) of equine on the farm, and, optionally, Area-Frame-Survey stratum. Three sets of experiments were conducted. The first used California data only. The second used data from three states Illinois, Indiana, and Iowa, while the third used data from the entire 48 contiguous states (there is no Area Frame Survey in either Alaska or Hawaii). The statistical methodology employed was logistic regression with a modification of stepwise regression for variable selection. Standard errors were estimated using designbased linearization methods. Conclusions are drawn about the nature of 2002 NML farms. Some are as expected (farms with small total sales are more likely to be NML). Others are surprising (holding all other factors constant, point farms are not sgnificantly more likely to be NML than other farms with less than $2,500 in annual sales). A number of methodological problems needing further research are suggested as a result of this study. ∗The empirical investigations described in this report where conducted by Professor Chang while he was an ASA research fellow at NASS from January to July of 2004. Although much was accomplished, there are limitations in the analysis due to the compact time frame.