Detecting item bias in latent construct between group comparisons: An illustrative example using multi-sample covariance structural equations modeling

Analyzing, responding to, and managing group differences in consumer behavior are critical for the successful formulation and execution of strategy. By understanding such distinctions, managers can design and deliver differentiated product offerings and promotions tailored to the needs of specific demographic, socioeconomic, and psychographic segments in the marketplace. From a methodology standpoint, comparing group differences involves two broad approaches. First, qualitative techniques such as interview protocols and ethnographies can generate unique insights about both within and between group phenomena. Second, managers and marketers typically use statistical tools such as t-tests and multivariate analysis of variance (MANOVA) techniques to obtain quantitative estimates of group differences. While MANOVA techniques are useful, they can also distort true group differences especially when latent constructs (e.g., satisfaction) are involved. In particular, MANOVA models for latent constructs will yield optimal results only if measures possess identical psychometric properties (e.g., item to construct relationships) across groups. However, the notion of psychometric equivalence is rarely tested in practice. Hence, biased items and measurement artifacts can confound true between group differences in MANOVA models. This paper discusses how item bias can be explicitly accounted for and controlled while making between group comparisons using MANOVA. Specifically, it describes how the multi-sample covariance structural equations modeling (SEM) method can provide researchers with a better basis to assess and address item bias. We illustrate the multi-sample SEM method by analyzing empirical data on perceptual latent constructs (performance ambiguity and input uncertainty) collected from automotive repair managers. The results show that by using the SEM and MANOVA techniques in tandem, researchers can control the deleterious effects of item bias and obtain robust and meaningful estimates of between group differences. The implications of our study for future research are also discussed.