Measurement Error Obfuscates Scientific Knowledge: Path to Cumulative Knowledge Requires Corrections for Unreliability and Psychometric Meta-Analyses
Industrial and Organizational Psychology (7,4), 2014, pp. 507-518.
12 Pages Posted: 17 Jan 2015
Date Written: December 1, 2014
All measurements must contend with unreliability. No measure is free of measurement error. More attention must be paid to measurement error in all psychological research. The problem of reliability is more severe when rating scales are involved. Many of the constructs in industrial-organizational (I-O) psychology and organizational behavior/human resource management research are assessed using ratings. Most notably the organizationally central construct of job performance is often assessed using ratings (Austin & Villanova, 1992; Borman & Brush, 1993; Campbell, Gasser, & Oswald, 1996; Viswesvaran, Ones, & Schmidt, 1996; Viswesvaran, Schmidt, & Ones, 2005). The reliability of its assessment is a critical issue with consequences for (a) validation and (b) decision making. For over a century now, it has been known that measurement error obfuscates relationships among variables that scientists assess. Again for over a century, it has been known that statistical corrections for unreliability can help reveal the true magnitudes of relationships being examined. However, until mid-1970s, corrections for attenuation were hampered by the fact that the effect of sampling error is magnified in corrected correlations (Schmidt & Hunter, 1977). Only with the advent of psychometric meta-analysis was, it possible to fully reap the benefits of corrections for attenuation because the problem of sampling error was diminished by averaging across many samples and thereby increasing sample sizes. Since the advent of psychometric meta-analysis 38 years ago, scientific knowledge in the field of I-O psychology has greatly increased. Hundreds of meta-analyses have established basic scientific principles and tested theories.
Against this backdrop, LeBreton, Scherer, and James (2014) have written a focal article that distrusts corrections for unreliability in psychometric meta-analyses. They question the appropriateness of using interrater reliabilities of job performance ratings for corrections for attenuation in validity generalization studies. Because of length limitations on comments in this journal, we will address only major errors, not all errors in LeBreton et al.'s strident article. The focal article is unfortunately more emotional than rational in tone and conceptually and statistically confused. In our comment, we address only the two latter problems.
We have organized our comment in five major sections: (a) purpose of validation and logic of correction for attenuation, (b) reliability of overall job performance ratings, (c) validity estimation versus administrative decision use of criteria, (d) accurate validity estimates for predictors used in employee selection, and (e) correct modeling of job performance determinants.
Suggested Citation: Suggested Citation