Assessment and control of nonresponse bias in a survey of medicine use by the elderly.

Health services research based on survey data is subject to potentially serious selection bias because observations are typically available only for survey respondents. This study describes a method of assessing and controlling for selection bias in the context of a survey of prescription and over-the-counter drug use by the elderly. A random sample of 6,500 Pennsylvania Medicare enrollees was sent a questionnaire regarding medicine use, insurance coverage, and health status in 1990. Applying a two-stage, limited dependent variable selection model developed by Heckman to baseline Medicare enrollment and utilization data for both respondents (70%) and nonrespondents (30%) allowed us to detect and control for negative and significant nonresponse bias in estimates of prescription drug use. Purchase of over-the-counter medication was free of such bias. The report describes how the Heckman method can be applied in other cases where health services survey samples are generated from program or organizational files that contain person-level data on all members of the sample frame.