Use covariance-based SEM (CB-SEM) when your goal is to test or confirm an established theory and you need global model fit (for example CFI ≥ .95 and SRMR ≤ .08; Hu & Bentler, 1999); use PLS-SEM when your goal is to explain variance and predict key outcomes, when the model is complex, includes formative constructs, or the theory is less developed (Hair et al., 2019). Both are legitimate. What matters in a thesis is that the choice follows from your research objective, not from which software your department happens to license.
How they differ
| CB-SEM | PLS-SEM | |
|---|---|---|
| Goal | Confirm theory; reproduce the covariance matrix | Explain variance; predict outcomes |
| Construct logic | Common factors | Composites (weighted sums of indicators) |
| Model fit | Central: χ², CFI, TLI, RMSEA, SRMR | Limited role; SRMR sometimes reported |
| Formative constructs | Difficult to identify | Handled routinely |
| Distribution | ML estimation assumes multivariate normality (robust options exist) | No distributional assumption; bootstrapping for inference |
| Typical software | AMOS, Mplus, lavaan (R) | SmartPLS, seminr (R), ADANCO |
The arguments people get wrong
"PLS-SEM works with tiny samples." PLS-SEM converges with small samples, but convergence is not the same as adequate statistical power. Kock and Hadaya (2018) showed that the popular 10-times rule underestimates the sample needed; their inverse square root method gives 155 respondents for a minimum path of .20 at 5% significance and 80% power.
"PLS-SEM is for non-normal data." True that PLS makes no distributional assumption, but CB-SEM has robust estimators too (for example MLR in lavaan or Mplus). Non-normality alone is a weak reason to choose PLS (Hair et al., 2019).
"PLS-SEM estimates are biased." If your constructs are truly common factors, standard PLS slightly overestimates loadings and underestimates paths. Consistent PLS (PLSc) corrects this (Dijkstra & Henseler, 2015). If your constructs are composites, CB-SEM is the mismatched method.
A decision guide
- Your hypotheses come from a well-established theory and your examiner wants fit indices: CB-SEM.
- You are extending a theory, adding new constructs or paths, and want to know how much variance you explain: PLS-SEM.
- Your model has formative constructs (for example an index of firm capabilities): PLS-SEM.
- Your model has many constructs and paths and a moderate sample: PLS-SEM usually converges more easily.
- You need to compare groups with full measurement invariance testing: CB-SEM has more mature tools; PLS-SEM uses MICOM (Henseler et al., 2016).
Worked example
Two students study online shopping intention with five constructs and 260 respondents.
Student A tests the Theory of Planned Behaviour exactly as published. The research question is whether the theory holds in a new context. She uses CB-SEM, runs CFA first and reports χ²(142) = 298.4, CFI = .95, TLI = .94, RMSEA = .065 and SRMR = .049. CFI and SRMR meet Hu and Bentler's (1999) guidelines; RMSEA is slightly above their .06 but is commonly read as acceptable below .08. She then tests the structural paths.
Student B adds two new predictors, Perceived Risk and Social Media Influence, to the same model, aiming to explain more of the variance in intention. Social Media Influence is measured formatively by exposure, engagement and trust in influencers. He uses PLS-SEM, assesses the measurement model by loadings, AVE and HTMT, the formative construct by weights and VIF, and reports that R² for intention rises from .41 to .52 when the new predictors are added.
Both choices are well justified because each follows from the research objective.
How to justify your choice in the method chapter
PLS-SEM was chosen because the study aims to explain variance in the key outcome and extends an existing model with new constructs, one of which is measured formatively (Hair et al., 2019).
Run it free
Run PLS-SEM with the free PLS-SEM tool, or CFA and CB-SEM with fit indices in the CB-SEM tool (Vietnamese interface). Both are checked against R.
FAQ
Can I use both in one thesis?
Yes, some studies report CB-SEM as a robustness check. Make sure each is reported by its own standards.
Is PLS-SEM accepted by journals?
Widely, in marketing, information systems and management, provided the reporting follows current guidelines (Hair et al., 2019).
Do I need CFA before PLS-SEM?
No. PLS-SEM has its own measurement model assessment. Running a CB-SEM CFA first mixes two logics.
References
- Dijkstra, T. K., & Henseler, J. (2015). Consistent partial least squares path modeling. MIS Quarterly, 39(2), 297–316.
- Hair, J. F., Risher, J. J., Sarstedt, M., & Ringle, C. M. (2019). When to use and how to report the results of PLS-SEM. European Business Review, 31(1), 2–24. https://doi.org/10.1108/EBR-11-2018-0203
- Henseler, J., Ringle, C. M., & Sarstedt, M. (2016). Testing measurement invariance of composites using partial least squares. International Marketing Review, 33(3), 405–431. https://doi.org/10.1108/IMR-09-2014-0304
- Hu, L., & Bentler, P. M. (1999). Cutoff criteria for fit indexes in covariance structure analysis: Conventional criteria versus new alternatives. Structural Equation Modeling, 6(1), 1–55. https://doi.org/10.1080/10705519909540118
- Kock, N., & Hadaya, P. (2018). Minimum sample size estimation in PLS-SEM: The inverse square root and gamma-exponential methods. Information Systems Journal, 28(1), 227–261. https://doi.org/10.1111/isj.12131