Logistic Propensity Models to Adjust for Nonresponse in MD Surveys

• Activity Number: 10
• Type: Topic Contributed
• Date/Time: Sunday, August 11, 2002 : 2:00 PM to 3:50 PM
• Sponsor: Section on Survey Research Methods*
• Author(s): Nuria Diaz-Tena*+ and Stephen Williams and Frank Potter and Michael Sinclair
• Affiliation(s): Mathematica Policy Research, Inc. and Mathematica Policy Research, Inc. and Mathematica Policy Research, Inc. and Mathematica Policy Research, Inc.
• Keywords: nonresponse ; weighting ; propensity modeling ; Community Tracking Study ; physician surveys

Abstract

Unit nonresponse adjustments in sample surveys generally rely on the inverse of the propensity to respond either in defined classes (the weighting class method) or at the unit level (logistic propensity scores). The physician surveys in the Community Tracking Study (CTS) are national samples of more than 12,000 physicians, and we used weighted logistic models to predict the probability of locating a physician (propensity score for location) and other weighted logistic models to predict the probability that a located physician responds (propensity score for response). The models used data from the sampling frames (AMA and AOA files) and, for physicians selected from the prior round, past survey status or responses. In both round two (1998-1999) and round three (2000-2001) CTS physician surveys, the adjustment factor to the sampling weights was the inverse of the propensity score (that is the inverse of the predicted probability to locate a physician and then the inverse of the predicted probability of a located physician to respond). In this presentation, we will describe the models used for round thee and compare these to the models developed for the second round of the CTS survey.