A "devil" hidden in the details: The effects of measurement error in regression analysis

Regression analysis procedures are frequently used in social work research. One of the assumptions on which regression analysis is based is the assump tion that all independent variables have been measured without error. This assumption is probably ignored by most researchers. However, the consequenc es of violations of this assumption are profound. Measurement error in a si ngle variable in a regression model can bias all parameters estimates in la rgely unpredictable ways. This means that researchers ignore this assumptio n at the peril of the validity of their results. In this article we discuss the consequences of measurement error in regression analysis. We then illu strate these consequences through several data analyses in which we first a ssume that there is no measurement error, and then use methods in which mea surement error is explicitly included. These illustrations suggest that man y of the results reported by social work researchers may be artifacts of me asurement error. We conclude with a number of recommendations that may be u sed by social work researchers to decrease the likelihood that measurement error will lead them to erroneous conclusions.