The Measurement of the Mediator and Its Influence on Statistical Mediation Conclusions

• Published in: Psychological methods 2021-02, Vol.26 (1), p.1-17
• Main Authors: Gonzalez, Oscar, MacKinnon, David P.
• Format: Article
• Language: English
• Subjects: Auditory Stimulation; Computer Simulation; Data Interpretation, Statistical; Factor Analysis, Statistical; Human; Humans; Mediation; Models, Statistical; Noise Effects; Psychology - methods; Psychometrics; Psychometrics - methods; Reproducibility of Results; Statistical Measurement; Statistical Validity; Test Interpretation; Test Reliability; Test Validity
• ISSN: 1082-989X; 1939-1463
• DOI: 10.1037/met0000263

In psychology, the causal process between 2 variables can be studied with statistical mediation analysis. To make a causal interpretation about the relation between variables, researchers who use the statistical mediation model make many assumptions about the variables in the model, among which are measurement assumptions about the mediator. For example, researchers often assume that the measure of the mediator yields scores that are reliable and that have a valid interpretation. In this article, we address how several measurement challenges affect the conclusions of statistical mediation analysis, and how researchers can use different psychometric models to study theoretically different causal processes. We use simulated data sets to illustrate how 10 well-fitting and theoretically sound statistical mediation models could significantly detect the indirect effect or miss it entirely depending on how the mediator is represented in the model. In the example, power to detect the indirect effect varied by the amount of true mediator variance that the psychometric model of the mediator was able to isolate. Different strategies to incorporate psychometric methods into mediation research are discussed and future directions are considered. Translational Abstract Statistical mediation analysis is a procedure commonly used to explain the relation between 2 variables or to study how an intervention led to a change in an outcome. Researchers use the statistical mediation model to uncover the intermediate mechanisms (known as mediators) that transmit the effect from an independent variable to an outcome. To obtain meaningful estimates, researchers have to make several assumptions about the variables in the mediation model. One of these assumptions is that the measure of the mediator yields scores that are reliable and that can be ascribed a valid interpretation. However, in social science research, many of the measures used to assess mediating constructs are prone to measurement error or misrepresentation of the dimensions that underlie the measure. In this article, we use simulated data sets to illustrate how the reliability and the representation of the mediator affect the probability of finding evidence supporting statistical mediation. Also, we provide recommendations and discuss several stages where researchers can incorporate measurement theory when they are testing for mediation.