Recent years have seen a trend in research undertaken at the U.S. Census Bureau and elsewhere to examine the multiple roles that culture plays in the research setting. Unfortunately, to date, empirical research has not fully addressed the question of whether surveys and assessment instruments are equally valid across identifiable subpopulations. For example, the possibility exists that assessment tools and studied constructs may provide differential internal validity and/or reliability across multiple populations. Measurement bias, a type of non-sampling error also labeled differential item functioning, is present when individuals equivalent on true levels of a variable, but from different groups, do not have identical probabilities of observed scores. As defined here, bias can lead to inaccurate estimates, attenuate or accentuate group differences, and affect the validity and reliability of research. When bias is present, it can be difficult, if not impossible, to interpret group differences on the construct being measured. Recent years have seen a call for model based, empirical methods to address the validity of constructs and measurement instruments across diverse populations. Latent variable models, e.g., confirmatory factor analysis (CFA), item response theory (IRT), and latent class analysis (LCA), are relatively recent entries in the research methods field and offer a powerful tool for investigating bias. The seminar presents the results of work currently undertaken at the U.S. Census Bureau utilizing latent variable models to address this issue.

Specifically, increased evidence for the internal validity of the DSM-IV alcohol abuse and dependence constructs in the general population has been demonstrated in the research literature. However, despite reported differences in the prevalence of alcohol use disorders among Hispanics, non-Hispanic African-Americans, and non-Hispanic Caucasians, less attention has been given to cross-cultural differences in the constructs across these groups. This is especially problematic when coupled with theoretical reasons to suspect that the internal validity of these constructs may differ for these groups, and that measurement bias may exist. Using data from the National Longitudinal Alcohol Epidemiological Survey (NLAES), a nationally representative survey of households, the seminar addresses these issues and presents analyses exploring violations of measurement invariance on a standardized measure of alcohol abuse and dependence across Hispanic, non-Hispanic African-Americans, and non-Hispanic Caucasian adults. The presentation describes the presence of statistically significant measurement bias and posits hypotheses for the possible sources of these differences.