My Data Collection is Over! How do I start the Data Analysis using #SmartPLS4?

SmartPLS 4 is presented as a practical workflow for moving from raw survey data to a fully tested structural equation model—covering measurement quality (loadings, reliability, validity), higher-order constructs, mediation, and moderation—using a single end-to-end process.

The session starts with model setup and data preparation. After collecting data, the first step is coding indicators consistently: each construct uses initials plus item numbers (e.g., Vision items V1–V5), with demographic variables like age, gender, and employment rank included alongside the questionnaire items. SmartPLS 4 accepts data in CSV/Excel or SPSS formats, and the workflow emphasizes data screening: checking minimum/maximum values and indicator statistics to catch discrepancies before analysis. Once imported, the project is organized around a workspace folder, and a new project is created with a PLS-SEM model type.

The core modeling work begins with the measurement model, which is treated as two layers: lower-order constructs first, then higher-order constructs. In the example, Internal Marketing and Internal Service Quality are higher-order constructs. Internal Marketing is modeled as reflective-formative at the higher level: Vision, Development, and Rewards are reflective subdimensions that combine to form the higher-order construct. Internal Service Quality is modeled as reflective-reflective: its subdimensions are reflective and interchangeable in the sense that removing one subdimension would not invalidate the higher-order construct.

On the SmartPLS canvas, constructs and indicators are dragged in, then connected to reflect the hypothesized relationships. The measurement model is run using the PLS-SEM algorithm, with reporting centered on outer loadings (indicator quality), construct reliability (Cronbach’s alpha and composite reliability, targeting >0.70), and construct validity. Convergent validity is checked via AVE (Average Variance Extracted), with a threshold of >0.50. Discriminant validity is assessed using HTMT and Fornell–Larcker logic: HTMT values should be below 0.90, and the square root of AVE for each construct should exceed its correlations with other constructs.

Higher-order constructs require extra validation steps. For reflective-reflective higher-order constructs, the same outer loading/reliability/validity checks apply. For reflective-formative higher-order constructs, the workflow shifts: collinearity is evaluated with VIF (acceptable when <5), then bootstrapping is used to test outer weights and outer loadings for the formative indicators (with significance and magnitude thresholds such as >0.50 for loadings). Latent variable scores are generated and re-imported so the higher-order constructs can be validated at the correct level.

After measurement quality is established, the structural model is assessed through bootstrapping. Direct effects are evaluated via path coefficients and one-tailed significance (p < 0.05), with attention to which relationships are significant (green) versus not (red). Mediation is handled by inspecting indirect effects; in the example, perceived organizational support is not a mediator while other paths through mediators are significant. Moderation is implemented by creating interaction effects between the predictor and moderators, then re-running bootstrapping to test whether interaction terms significantly affect the dependent variable. The session also demonstrates how to plot moderation using a Johnson-Neyman-style slope approach via the “stats tool package” by James G. Kline.

Finally, the session ties results to reporting practice: measurement model tables (factor loadings, reliability, convergent and discriminant validity including HTMT and Fornell–Larcker), structural model metrics (VIF, R², Q² via PLS predict), and a results section organized as direct effects, mediation, then moderation. The takeaway is a repeatable checklist for producing publishable SmartPLS 4 outputs, including the special handling required for higher-order constructs and interaction effects.