Webinar Day 2: SmartPLS3 for Data Analysis - Basic and Advance Analysis (See Description)

SmartPLS3 is presented as a practical workflow for analyzing survey data with structural equation modeling—starting from how to enter and code questionnaire responses, then moving through measurement-model quality checks (reliability and validity), and finally testing structural relationships with bootstrapping. The core message is that credible SmartPLS results depend less on clicking “run” and more on systematically validating constructs (outer model) before interpreting paths (inner model), including careful handling of discriminant validity and higher-order constructs.

The session begins with the basics of survey data preparation: after distributing a questionnaire to hundreds of people and receiving completed responses, the data must be entered into analysis software. Excel and SPSS are offered as common options for data entry and anomaly checking, but SmartPLS3 is positioned as the main tool for the modeling stage. Variables are defined as columns, with each questionnaire item mapped to an indicator (e.g., ethical responsibilities coded as ER1–ER8). A practical recommendation is to label rows with questionnaire numbers so any data-entry problems can be traced back to the original respondent’s choices.

SmartPLS3 is then framed as software for structural equation modeling (SEM), combining ideas from factor analysis and regression. The instructor contrasts covariance-based SEM (e.g., Amos) with PLS-SEM, noting that PLS-SEM is often used when normality assumptions are not reliable and that it can produce stronger loadings in practice. SEM’s logic is split into two parts: the measurement model (outer model) evaluates how well indicators represent latent constructs, while the structural model (inner model) tests hypothesized relationships among latent constructs.

For the measurement model, the workflow emphasizes construct reliability and validity. Reliability is assessed using Cronbach’s alpha and composite reliability, with a commonly used threshold of 0.70. Factor loadings are treated as a key diagnostic: low-loading indicators may be removed, but not blindly—content validity must be protected, and the session warns against “deleting sprees” that chase statistical thresholds at the expense of construct meaning. Convergent validity is checked via Average Variance Extracted (AVE), with AVE expected to exceed 0.50. Discriminant validity is evaluated using multiple criteria: the Fornell–Larcker approach (square root of AVE should exceed inter-construct correlations), cross-loadings (items should load highest on their own construct), and HTMT (heterotrait–monotrait ratio), with conservative guidance such as HTMT < 0.85 (or ≤ 0.90).

After the outer model passes quality checks, the structural model is tested using bootstrapping. Path coefficients, t-statistics, and p-values determine whether relationships are significant, and R² indicates how much variance in dependent constructs is explained. Mediation is handled through bootstrapped indirect effects, then classified as no mediation, partial mediation, or complete mediation depending on whether direct and indirect effects remain significant. Moderation is tested by adding an interaction term (e.g., role ambiguity moderating the CC → organizational performance link) and interpreting the sign of the moderation coefficient, supported by slope analysis.

Finally, the session addresses higher-order constructs and the reflective vs formative distinction. It explains why higher-order constructs often require special treatment (e.g., two-stage approaches) and why discriminant validity statistics may not apply directly to the higher-order layer. For reflective–reflective higher-order models, the session uses a two-stage approach by exporting latent variable scores and re-importing them as indicators. For reflective–formative cases, it highlights problems like near-zero explained variance under standard modeling and recommends disjoint two-stage approaches, then validates higher-order constructs using collinearity (VIF) and significance of outer weights, while again cautioning that indicator deletion can harm content validity.