Adam C. Carle, MA, PhD, Statistical Research Division, U.S. Census Bureau, Center for Survey Methods Research, Washington, DC 20233-9100, 301-763-1863, adam.c.carle@census.gov

The possibility exists that assessment tools are biased and differentially valid and/or reliable across multiple populations. Measurement bias, a form of non-sampling error, occurs when individuals identical on a construct being measured can be expected to have different observed scores as a function of group membership. Measurement invariance holds when individuals equivalent on the construct, but from different populations, have the same probability of achieving a given observed score on the instrument. When measurement bias is present, it is 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 measurement instruments across diverse populations. Latent variable models, such as confirmatory factor analysis (CFA) and others, are relatively recent entries in the research methods field and are a powerful tool for investigating bias. Unfortunately, epidemiological research rarely considers bias, and latent variable models are virtually unutilized. With training in these models lacking, researchers find it difficult to understand studies employing them and issues of measurement bias are often ignored. Using data from the National Longitudinal Alcohol Epidemiological Survey (NLAES), a nationally representative household survey of 42,692 adults, the current paper develops the CFA model to assess measurement bias in the context of epidemiological research. The general mathematical CFA model is developed and a method for exploring measurement bias is described. Finally, results are used to demonstrate the importance of establishing measurement invariance prior to making epidemiological estimates.

Learning Objectives: