Detecting measurement confounding in epidemiological research: construct validity in scaling risk behaviours: based on a population sample in Minnesota, USA.

Abstract

STUDY OBJECTIVE: The purpose of this study was to construct and validate a behavioral risk index scale to determine if the scaled variable could be used to study possible latent dimensions of risk behaviour. DESIGN: Data from the Minnesota Behavioral Risk Factor Surveillance System (BRFSS) were used to construct and validate a behavioral risk index using item response theory methods. PARTICIPANTS: A representative sample of 3303 Minnesota adults 18 years of age and older. RESULTS: Massive evidence was found against the construct validity of the index scale as a measure of risk behaviour. Seven commonly studied risk behaviours could not be scaled into a valid construct of health behaviour for either men or women. Tests of scalability, homogeneity, and item independence were rejected. In addition, item bias was found for all of the items in relation to important exogenous variables, especially age and education. CONCLUSIONS: The risk behaviours do not represent sufficiently similar types of phenomena to form an additive scale of health related risk taking. Not only do the practices fall on different, undefined dimensions of behavior, subgroup differences in risk taking would be hidden by data reductions summing the behavioral practices into additive scales. The findings indicate that the behaviors have quite distinct meanings that should be studied separately.