A Method for Nonresponse Correction in the Analysis of Mail Survey Data

Even when researchers using mail surveys go to great lengths to maximize response rates, there will always be some portion of the intended sample that does not respond, either at all, or sufficiently -completely to allow inclusion in the estimating sample. This paper examines nonresponse and its apparent consequences for a survey of water-based recreational participation conducted in the Northwest US. We describe how zip codes can be used in combination with special software to determine distances from each address in our intended sample of 7034 households to each of the recreational sites featured in our survey. Zip codes also allow us to merge our intended sample with 1990 Census data aggregated to the zip code level. We demonstrate how to model the survey response/nonresponse decision explicitly, and show that statistical corrections for nonresponse can have a potentially important effect on the apparent inferences from our models. We strongly advocate, based on these.results, that any researcher using a mail survey can and should explore analogous response/nonresponse models and corrections before making any claims as to the robustness of empirical results to non-random sample selection. * This data set employed in this paper was collected as part of a study managed by Hagler Bailly Consulting, Inc., under contract with the U.S. Army Corps of Engineers (contract # TCN 93357).