#SmartPLS4 Webinar Day 4: Complex Modelling Example using SmartPLS

Structural equation modeling in SmartPLS starts with disciplined data screening and then moves through a two-stage quality check: measurement model assessment (reliability and validity) before any hypothesis testing in the structural model. The workflow begins by cleaning the dataset—checking minimum/maximum values, handling missing data, identifying outliers, and screening for respondent misconduct using standard-deviation logic. Once the data are fit, the measurement model is evaluated using factor loadings, reliability (Cronbach’s alpha and composite reliability), and construct validity. Convergent validity is assessed via Average Variance Extracted (AVE), with AVE needing to exceed 0.50. Discriminant validity is then tested to ensure constructs are distinct rather than overlapping, using HTMT (with conservative guidance often <0.85) and the Fornell–Larcker criterion (square root of AVE should exceed inter-construct correlations). Cross-loadings can also be used as a practical diagnostic when discriminant validity looks weak.

After the measurement model is sound, SmartPLS can handle more complex research designs—higher-order constructs, mediators, moderators, and comparisons across groups. For multi-group analysis (MGA), the key idea is to test whether path relationships differ across groups such as male vs. female. The process runs bootstrapping separately for each group and then compares path coefficients. In the example, collaborative culture shows a negative effect on role ambiguity in both groups, but the effect is insignificant for females (driven by higher standard error and smaller group representation) and stronger for males. MGA also checks whether mediation differs across groups; here, collaborative culture does not mediate through the proposed pathway in either group. A second MGA approach—bootstrap MGA—tests whether differences in path coefficients are statistically significant; most differences are not significant except for one relationship involving how IM influences CC.

The webinar then scales up to a complex model combining reflective and formative constructs, higher-order constructs, mediators, and a higher-order moderator. The model is built by first placing lower-order constructs (the subdimensions) on the canvas, then linking them to their higher-order constructs. For reflective–reflective higher-order constructs (e.g., internal service quality), validation follows the same logic as ordinary measurement models: outer loadings, reliability, AVE, and discriminant validity. For reflective–formative higher-order constructs (e.g., internal marketing and role stress), the process adds collinearity checks (VIF) and bootstrapping to validate formative weights. Even when a formative weight’s p-value is not significant, the guidance is not to automatically delete indicators; instead, confirm that outer loadings are significant and that the construct meets the broader measurement-quality thresholds.

Finally, the structural model is bootstrapped (e.g., 10,000 resamples) to test direct effects, mediation, and moderation. Direct paths are interpreted using p-values and effect sizes (beta coefficients). Mediation is classified as partial when both direct and indirect effects are significant, and as full when the direct effect becomes insignificant while the indirect effect remains significant. Moderation is handled via interaction effects; if moderation is insignificant, slope analysis is unnecessary. The workflow can also incorporate control variables (age, gender, job rank), with the recommendation to avoid dummy variables for binary or ordinal variables and to compare results with and without controls to confirm whether the model changes meaningfully. The overall takeaway is that complex SmartPLS modeling is less about shortcuts and more about a repeatable quality-control pipeline: clean data, validate measurement models at every construct level, then interpret structural hypotheses with bootstrapped inference.