Integrating the results of a nonresponse follow-up survey into the survey from which its items were selected

Abstract

A nonresponse follow-up (NRFU) survey was conducted for the National Pilot of the Residential Energy Consumption Survey (RECS), an address-based sample survey of potential primary residences in the US enumerated by web and mail. Virtually all unit (i.e., whole-record) nonrespondents to the National Pilot were sent a short mail questionnaire containing 18 key items from the full survey. Here, we first compare two ways of adjusting variables collected on the NRFU for unit nonresponse. In one, only the weights for respondents to the full National Pilot survey were adjusted to compensate for nonresponse using a calibration weighting procedure that assumes response to be a logistic function of variables known for the entire sample (the NRFU sample was ignored). In the other, only the NRFU-survey respondents’ weights were adjusted for nonresponse using an analogous calibration weighting scheme, while weights for the respondents to the full survey were not adjusted. The resulting two national estimates for many of the NRFU variables were then compared. When the two were significantly different, the latter estimate was treated as unbiased and added as a calibration variable when adjusting (a second time) for unit nonresponse to the full sample. When they were not significantly different, both were deemed unbiased, and the mean of the two added as a calibration variable when readjusting for nonresponse to the full sample. The theory behind this practice and its repercussions are discussed.