Assessing whether items are biased against minority groups is very important for the validity of measurement scales in cancer control and in health sciences in general. Although we have scales to measure beliefs (based on the Health Belief Model) that are predictive of cancer screening behavior, we do not know the extent of cultural bias. Thus, we are missing a critical component in the assessment of psychometric validity. The statistical techniques for detecting item bias are called procedures for detecting differential item functioning (DIF). Under certain conditions, two of the most commonly-used DIF procedures, logistic regression and Mantel-Haenszel, can inaccurately display inflated Type I error. The primary purpose of the present methodological proposal is to improve power and Type I error of both the logistic regression and Mantel-Haenszel procedures by using alternative matching scores other than traditional summed-item score. Using existing data from two large NIH-funded randomized intervention trials to increase mammography screening (ideal because of equal percentages of Caucasians and African-Americans), we will address these aims: (1) Using computer simulations based on the data structure for generalizing results to breast cancer screening belief scales, we will evaluate the extent to which innovative modifications of logistic regression and Mantel-Haenszel DIF procedures control Type I error inflation and improve power, and (2) Using the best modified logistic regression and Mantel-Haenszel DIF procedures (derived from Aim 1), assess items for DIF in the two existing data sets based upon ethnicity, age, education and income in six scales commonly used to predict cancer screening (barriers, benefits, susceptibility, fatalism, self-efficacy, and fear). Improved accuracy of DIF procedures will increase the validity of cancer control scales, providing a firm measurement foundation for advancing the science of cancer control and prevention. Additionally, the findings will have broad applicability to detecting item bias in other scales.
