The Concept, Theory, and Process of Mediation in Structural Equation Modelling (SEM)

Mediation in structural equation modeling (especially PLS path models) hinges on one idea: an exogenous construct can affect an endogenous construct partly or wholly through an intervening mediator construct. When the mediator sits between X and Y, the model produces three related effects—direct, indirect, and total—that determine not just whether mediation exists, but what kind it is. The direct effect reflects the X→Y path while accounting for the mediator (often labeled as the “c complement” path). The indirect effect captures the X→M→Y chain, calculated as the product of the X→M coefficient (a) and the M→Y coefficient (b). The total effect combines both: direct plus indirect.

Different mediation patterns follow from which components are statistically meaningful. Partial mediation occurs when the indirect effect is significant and the direct effect remains significant too—so some influence runs through the mediator and some bypasses it. Full mediation appears when the indirect effect is significant but the direct effect is insignificant, meaning the X→Y relationship is effectively transmitted only through M. The sign of effects further splits mediation into complimentary versus competitive. Complimentary mediation happens when the direct and indirect effects move in the same direction (for example, both positive). Competitive mediation arises when they oppose each other—such as a negative direct effect paired with a positive indirect effect.

Testing mediation has evolved beyond the classic Baron and Kenny (1986) framework. Baron and Kenny proposed four steps: first show X significantly affects Y (the c path), then show X affects the mediator M (the a path), then show both a and b are significant when M predicts Y, and finally test the direct and indirect relationships together to determine whether mediation is present. Over time, several assumptions in that approach were criticized. The requirement that the c path be significant was relaxed because suppressor effects can mask a total effect while a meaningful indirect effect still exists. Likewise, the need for both a and b to be individually significant was challenged because the indirect effect is fundamentally the product a×b; an indirect effect can be significant even if one component is not.

The modern, more accepted approach uses bootstrapping to test the indirect effect’s significance. Bootstrapping repeatedly resamples the data (often thousands of times, such as 5,000) with replacement to build an empirical sampling distribution for the indirect effect. Significance is then assessed via confidence intervals: if the indirect effect’s interval excludes zero, mediation is supported. This method avoids reliance on flawed significance tests tied to the Sobel test.

Before running mediation tests, the mediation model must meet the same quality checks as any PLS SEM model. Measurement models—especially reflective mediator constructs—need adequate reliability and validity, because poor reliability can shrink estimated indirect effects. Structural checks also matter: collinearity must be controlled, since high collinearity can bias path coefficients, flip signs, and create misleading conclusions about mediation type. Discriminant validity problems can also distort indirect effects.

Once measurement and structural criteria are satisfied, the mediation decision follows a significance logic. If the paths X→M (P1) and M→Y (P2) are significant, then the direct X→Y path (P3) determines the mediation type: significant P3 implies partial mediation (with complimentary or competitive classification based on the signs of direct and indirect effects), while non-significant P3 implies full mediation. If P1 and P2 are not significant, mediation is not supported. If P3 is not significant and P1/P2 are significant, the model indicates no direct effect—only mediated influence. If P3 is significant without P1/P2, the result points to a direct-only relationship with no mediation.