Using longitudinal data to estimate nonresponse bias.

In a recent survey of depressive symptoms among former prisoners of war, longitudinal data were used to estimate nonresponse bias. A predictive model was fitted to the data of current respondents and then was used to predict the scores of nonrespondents who had earlier provided similar convariate data. This analysis showed that, despite differences between respondents and nonrespondents in age, education, and severity of treatment during captivity, differences between the observed scores of respondents and the predicted scores of nonrespondents were small. For an estimation of the overall impact of nonresponse bias, revised estimates for the entire sample were calculated by combining observed data from respondents and predicted data for nonrespondents; the revised estimates differed little from those for respondents alone.