Evaluating assumptions of weighting class methods for partial response using a selection model

In survey sampling, information about the prevalence of a health outcome Y for a defined target population is frequently obtained using a two‐stage data collection process. In the first stage, households that have members of the target population are identified and socio‐demographic data that are believed to be associated with Y are collected. At the end of the first stage of data collection, permission is requested to contact the member's health providers so that accurate information about Y can be obtained. When permission is obtained, a second phase of data collection is conducted in which those health providers are contacted and Y is obtained. A ‘complete response’ results when data are obtained from both the first and the second phases of the survey. A ‘partial response’ results when data are collected from the first phase, but Y is not obtained in the second phase. To adjust for selection bias in estimating the prevalence of Y caused by partial responders' missing Y values, potential differences between complete and partial responders are typically taken into account by using weighting class methods. These methods assume that missing Y values are missing at random (MAR). This paper describes statistical tests for evaluating whether missing data are missing completely at random or MAR. Copyright © 2008 John Wiley & Sons, Ltd.