Logistic Propensity Models to Adjust for Nonresponse in Physician Surveys

1. Introduction Logistic propensity models for nonresponse adjustments have been used for various studies (Little 1986). The logistic models used to compute the scores reflect the propensity to respond based on attributes of both respondents and nonrespondents. Propensity scores can be used to compute explicit adjustments factors or to form weighting cells. Recent work has shown the benefits of using a limited number of weighting classes (Eltinge et al. 1997). In the current paper, we look at the advantages of using a different number of weighting cells or the propensity scores to adjust for nonresponse in a Physician Survey. The Community Tracking Study (CTS), which is funded by the Robert Wood Johnson Foundation, is designed to provide a sound information base for decision making by health leaders. It does so by collecting information on the United States health system, and how it is evolving, as well the effects of those changes on people. Begun in 1996, the CTS, is a longitudinal project that relies on periodic site visits and surveys of households, physicians, and employers. This survey consists of two samples, a site sample and a supplemental sample. The site sample is a national survey of 60 locations in the United States: 48 large Metropolitan Statistical Areas (MSAs), 3 small MSAs, and 9 non-MSAs. The supplemental sample includes all 48 contiguous states stratified in 10 different regions, as described in Potter et al. (2000). In the Physician Survey, we had three different subgroups of physicians for the site sample, and for the supplemental sample based on their Round Two interview status: (1) Round Two interviews (reinterviews); physicians who completed the Round Two interview, (2) Round Two noninterviews (noninterviews); physicians who were selected for the Round Two sample but who did not complete the interview for reasons such as ineligible, refusals or not located, and (3) new sample (new); physicians in the Round Three sampling frame who were not selected for the Round Two sample. For the physician survey of the Round Three, we used weighted logistic models to adjust for