Tests of Mediating Effects

Mediator variables are common in disciplines such as psychology, sociology, management, education, political science, and public administration. A mediator variable (referred to hereafter as a mediator) transmits the effects of an independent variable to a dependent variable. This is illustrated in Figure 1a, which shows a causal chain involving an independent variable (Z1), a mediator (Z2), and a dependent variable (Z3). In instances of actual (as opposed to assumed) mediation, two types of effects are possible. We illustrate these with reference to Figures 1a and 1b. In complete mediation, there is only an indirect effect of the independent variable on the dependent variable (see Figure 1a). However, in partial mediation, an independent variable has both a direct effect on the dependent variable and an indirect effect (see Figure 1b).

There are many examples of mediating effects. For instance, behavioral intentions are assumed to mediate the relation between attitudes and behavior, and stress is hypothesized to mediate the relation between stressors and strain.

Research Design Issues and Tests of Mediation

Tests of mediation may be based on data from studies using three major types of designs: randomized experiments, quasi-experiments, and nonexperiments. Inferences about mediation that stem from such tests rest on a relatively firm foundation when they are based on data from randomized experiments. For example, in testing for the mediating effect of Z2 on the relation between Z1 and Z3, one randomized experiment can be used to show that Z1 causes Z2, and another can be conducted to demonstrate that Z2 causes Z3.

Unfortunately, when tests of mediation are based on data from quasi-experimental studies, inferences about mediation rest on a much weaker foundation. Moreover, when such tests rely on data from non-experimental research, inferences about mediating effects are almost never justified. One of the major [Page 996]reasons for this is that when tests of mediation are based on data from nonexperimental research, all that a researcher can confidently conclude is that an observed pattern of relations (e.g., covariances, correlation coefficients) among the assumed independent, mediator, and dependent variables is consistent with a hypothesized model (e.g., that shown in Figure 1b). However, the same pattern of relations may also be consistent with a number of alternative models, including those shown in Figures 1c to 1e. Thus, the results of nonexperimental research almost always provide highly equivocal evidence about mediation. Moreover, as noted below, the validity of inferences about mediation is not at all improved by the use of various statistical procedures that purport to test causal models. One important reason for this is that in quasi-experimental or nonexperimental research, mediation models assume a specific set of causal connections between (among) variables (e.g., the model shown in Figure 1b), and statistical procedures are incapable of providing credible evidence of the correctness of the assumed pattern of relations.

Figure 1 Some Possible Causal Relations Among Three Variables

Source: Stone-Romero & Rosopa (2004).