Exploring a Person-Oriented Mediation Measure

Introduction: In psychology research, many of the statistical analyses used describe or predict relationships that exist among psychological variables.  However, these variable-oriented methods use data that has been aggregated across individuals.  Such methods often do not account for how the relationships among variables may vary for different individuals or groups. Thus it is important to use person-oriented methods that detect individual differences, along with the popular variable-oriented methods in order to create a more comprehensive understanding of the similarities and differences among human behavior.  This is especially important in intervention research, where research conclusions can affect clinical practice and government policy.  This study describes and tests the accuracy of a person-oriented measure of mediation: the percent of participants that have data consistent with mediation.  Given complete mediation where the mediated effect is positive and the treatment (X), the mediator (M), and the outcome (Y) are all binary variables, the measure is calculated by finding the proportion of people in the sample who had either had X = M = Y = 1 or X = M = Y = 0.  These configurations are consistent with a mediated effect such that X changes M and M changes Y.  Even in the presence of a significant mediated effect, it could be useful to know whether that effect was present for all, or just some individuals in the sample.  In that way, interventions may target changing the mediator more strongly for specific groups of people. 

Methods: Data were simulated from a mediation model with binary X, M, and Y.  Simulation conditions included zero, small, medium, and large path estimates, as well as sample sizes of N = 50,  N = 200, and N = 1000.  The percent consistent with mediation measure was computed for each replication and tested for significance.  Type I error rates and power are reported.

Results: Pilot data show that in general, the percent consistent with mediation measure has appropriate power and Type I error rates, except when one path is moderate or large and the other path is zero.  The measure is also demonstrated for real data samples with and without mediated effects.

Conclusions: A person-oriented measure of mediation that gives the percent of the sample that is consistent with mediation can be a useful supplement to traditional variable-oriented mediation estimates.