CBSEM using #SmartPLS4 | 14 | Mediation Analysis using CBSEM Model in SmartPLS4

Mediation analysis hinges on separating an effect into two parts: a direct path from an independent variable to a dependent variable, and an indirect path that runs through a mediator. In practice, the indirect effect is computed as the product of two links—X→M (often labeled A) and M→Y (labeled B)—so the indirect effect equals A×B. The total effect is then the sum of the direct effect (the X→Y relationship when the mediator is included, labeled C′) and the indirect effect (A×B). This decomposition matters because it clarifies *how* influence travels, not just whether it exists.

The transcript lays out mediation types based on which components are statistically significant. Partial mediation occurs when both the indirect effect (A×B) and the direct effect (C′) are significant, meaning some influence flows through the mediator while some still bypasses it. Full mediation is the opposite pattern: the indirect effect is significant, but the direct effect becomes insignificant once the mediator is in the model—indicating the mediator carries the effect. It also distinguishes complimentary mediation (direct and indirect effects move in the same direction, such as both positive) from competitive mediation (direct and indirect effects differ in sign, such as a negative direct effect paired with a positive indirect effect).

For testing mediation, the discussion contrasts the classic Baron and Kenny (1986) four-step approach with later refinements. Baron and Kenny required significance checks on the total effect (X→Y), then on X→M, then on both X→M and M→Y, and finally on the direct and indirect paths together. That framework has been criticized because it relied on unstandardized coefficients and significance testing (including Sobel tests) that can miss mediation when suppressor effects exist or when only the product A×B matters. The transcript highlights a key shift: significance of A or B individually is no longer treated as a requirement, since the indirect effect is the product A×B.

The recommended method uses bootstrapping to test the indirect effect’s confidence interval. Bootstrapping repeatedly resamples the dataset (with replacement) to generate thousands of pseudo-samples (commonly 5,000–10,000; here, 1,000 is used) and then checks whether the indirect effect’s confidence interval excludes zero. This approach directly targets the statistical uncertainty around A×B.

A SmartPLS4 walkthrough then demonstrates the workflow in a single structural model. After ensuring the measurement model is assessed first, the model fit is checked, and bootstrapping is run (CBM bootstrapping with one-tailed testing set because relationships are hypothesized as positive). The results are interpreted through standardized path coefficients and p-values. The indirect effect of organizational commitment (OC) on organizational performance (OP) through collaborative culture (CC) is reported as significant (p < 0.05), with a T statistic consistent with the one-tailed setup. To determine mediation type, the direct effect OC→OP in the presence of CC is also examined: because it remains significant, the transcript concludes partial mediation. It further reports the specific indirect effect magnitude (standardized indirect effect around 0.189) and uses bias-corrected confidence intervals to confirm the indirect effect does not cross zero, supporting the mediation hypothesis. Finally, it notes how to structure reporting in structural equation modeling: measurement model reliability/validity first, then direct effects, then mediation (and moderation if present).