Structural Equation Modelling (SEM) is one of the most popular techniques in management, marketing, education and behavioural research. Most scholars quickly meet two options: PLS-SEM (usually run in SmartPLS) and CB-SEM (usually run in AMOS, LISREL or lavaan in R). Choosing the wrong one — or failing to justify your choice — is a common reason for examiner and reviewer comments.
The core difference
CB-SEM (covariance-based) tests how well your theoretical model reproduces the observed covariance matrix. It is a confirmatory, theory-testing approach and reports overall model-fit indices such as CFI, TLI and RMSEA.
PLS-SEM (partial least squares) maximises the explained variance (R²) of the dependent constructs. It is prediction-oriented and well suited to complex models and theory development.
When CB-SEM (AMOS) is the better choice
- Your goal is to confirm or reject an established theory
- All constructs are reflectively measured
- Your sample is reasonably large and data are close to multivariate normal
- Your discipline or target journal expects global model-fit statistics
When PLS-SEM (SmartPLS) is the better choice
- Your goal is prediction or extending an existing theory
- The model is complex, with many constructs, paths, mediators or moderators
- You have formative constructs, or a mix of formative and reflective
- Your sample is smaller or data are non-normal
- You plan to use latent variable scores in follow-up analyses
Side-by-side comparison
| Aspect | CB-SEM (AMOS) | PLS-SEM (SmartPLS) |
|---|---|---|
| Main goal | Theory confirmation | Prediction & theory development |
| Measurement models | Mainly reflective | Reflective and formative |
| Sample size | Larger samples preferred | Works with smaller samples |
| Distribution | Assumes multivariate normality (ML) | Non-parametric, bootstrapping |
| Key outputs | CFI, TLI, RMSEA, SRMR | R², f², Q², HTMT, PLSpredict |
What to report in your thesis
Whichever you choose, examiners expect a clear justification linked to your research objectives, plus the standard two-step reporting:
- Measurement model — indicator loadings, internal consistency (Cronbach's alpha, composite reliability), convergent validity (AVE) and discriminant validity (HTMT or Fornell-Larcker).
- Structural model — path coefficients with significance (bootstrapped in PLS), R², effect sizes and, where relevant, mediation or moderation results.
Common mistakes to avoid
- Choosing PLS-SEM only because "the sample is small" without any other justification
- Reporting model-fit indices from PLS as if they were CB-SEM fit statistics
- Skipping common method bias checks for single-source survey data
- Ignoring HTMT when discriminant validity is in question
The bottom line
Neither method is "better" in general. Start from your research question: confirming a well-established theory points to CB-SEM; predicting outcomes or exploring a complex model points to PLS-SEM. Justify the choice in your methodology chapter and report the measurement and structural models in full.
