Using Cognitive Test Scores in Social Science Research

Date of Original Version




Abstract or Table of Contents

A standard problem in social science attempts to better understand the large wage disparities between black and white workers in U. S. labor markets. Social scientists have conducted hundreds of studies of observed racial wage gaps, seeking to understand the extent to which they are driven by differences in human capital or disparate treatment by employers. In order to get an unbiased estimate of such effects, it is necessary to include in the regression equations measures of human capital. While years of schooling has traditionally been used as a measure of human capital, social scientists are increasingly turning to cognitive test scores, as a more direct measure. Most social science research that uses cognitive test scores as an independent variable models the test score as fixed and without error. However, since test scores have measurement error, modeling the test score in this way can produce biased results which can result in incorrect policy conclusions. Current methods for modeling the test score with error are limited to single point in time analysis with a fixed cognitive assessment administered to all subjects, and situations in which the measurement error is homogeneous across all subjects. In response to these drawbacks, a new model called the Mixed Effects Structural Equations (MESE) model is developed. The MESE model is demonstrated using data from the National Adult Literacy Survey by analyzing black-white wage gaps in married men, single men, and single women. Three important findings are of note. First, much of the black-white wage gap can be attributed to a black-white disparity in skills suggesting that more attention ought to be focused on the development of skills. Second, comparisons of the the MESE model to a model with no measurement error demonstrate the importance of modeling the measurement error. Third, comparisons of the MESE model to a model using current methodology suggest the MESE model may solve some of the drawbacks noted in the other current methods.