Lower level mediation in multilevel models.

Abstract

Multilevel models are increasingly used to estimate models for hierarchical and repeated measures data. The authors discuss a model in which there is mediation at the lower level and the mediational links vary randomly across upper level units. One repeated measures example is a case in which a person's daily stressors affect his or her coping efforts, which affect his or her mood, and both links vary randomly across persons. Where there is mediation at the lower level and the mediational links vary randomly across upper level units, the formulas for the indirect effect and its standard error must be modified to include the covariance between the random effects. Because no standard method can estimate such a model, the authors developed an ad hoc method that is illustrated with real and simulated data. Limitations of this method and characteristics of an ideal method are discussed.

Clarifying conditions and decision points for mediational type inferences in Organizational Behavior

“…Mediators (M) in this context could be relatively stable level-2 work attitudes (e.g., organizational commitment) assessed at a single point in time, or temporally changing level-1 variables (e.g., moods) assessed using time sampling techniques. Plenty of good work is currently being advanced along these lines (e.g., Judd et al, 2001;Kenny, Korchmaros, & Bolger, 2003). In summary, multi-level designs expand the scope of mediational inferences to incorporate relationships that reside within levels of analysis, traverse levels of analysis, and unfold over time.…”

Pathways to change in one-session exposure with and without cognitive intervention: An exploratory study in spider phobia

“…Hofmann et al, 2007). Time served as the Level 1 unit, which was nested within participants (Level 2) (e.g., Kenny, Korchmaros, & Bolger, 2003). We performed separate analyses in which FSQ…”

Rumination, emotion, and forgiveness: Three longitudinal studies.

“…The traditional Sobel test for mediation ignores the so-called ab covariance (Kenny, Korchmaros, & Bolger, 2003), which is covariation between the random effects for the association of the predictor with the mediator (the a path) and the random effects for the unique association of the mediator with the outcome while controlling for the direct effect of the predictor (the b path). Failure to consider this ab covariance when testing mediational hypotheses in multilevel models can lead to underestimates or overestimates of mediated effects.…”

“…The Sobel tests, whose t values appear in Table 3, also support the conclusion that anger (but not fear) toward the transgressor mediated the associations of rumination with avoidance and revenge. Kenny et al (2003) wrote that incorporating the ab covariance for multilevel mediational models is only necessary when the random effects variance for the a path (i.e., the coefficient representing the regression of the putative mediator on the putative x variable) and the random effects variance for the b path (i.e., the coefficient representing the association of the outcome variable y with the putative mediator variable when then putative x variable is simultaneously controlled) are significantly different from zero. These conditions were not fulfilled in our data-that is, for all mediational models, the a path, the b path, or both lacked significant random effects variance-so Kenny et al's calculations were not necessary to accurately estimate mediation.…”