PLS-SEM has become the default in management, information systems and marketing research, and a great deal of it is reported incompletely. The most common reason is that people jump to the path coefficients before establishing that the constructs are measured properly at all.
PLS-SEM or CB-SEM?
The choice is not about sample size alone, despite how often it is described that way. It is about what you are trying to do.
| Consideration | PLS-SEM (SmartPLS) | CB-SEM (AMOS, Mplus, lavaan) |
|---|---|---|
| Primary goal | Prediction and explaining variance | Theory testing and model fit |
| Sample size | Works with smaller samples | Generally needs larger samples |
| Formative constructs | Handled naturally | Difficult to specify |
| Model complexity | Handles many constructs and indicators | Complexity strains estimation |
| Distribution | No normality assumption | Assumes multivariate normality (ML) |
| Global fit test | Limited (SRMR, NFI) | Full battery (χ², CFI, TLI, RMSEA) |
Justify the choice in your methodology chapter with reference to your research aim. "We used PLS-SEM because our sample was small" is the weakest available justification.
Stage one: the measurement model
Nothing in the structural model means anything until the measurement model holds. For reflective constructs, four things are assessed.
- Indicator reliability — outer loadings above 0.708. Items between 0.40 and 0.708 are removed only if doing so raises AVE or composite reliability above the threshold.
- Internal consistency — composite reliability between 0.70 and 0.95. Above 0.95 suggests redundant items rather than excellent measurement.
- Convergent validity — AVE of 0.50 or higher, meaning the construct explains at least half the variance in its indicators.
- Discriminant validity — HTMT below 0.85 (or 0.90 for conceptually similar constructs). HTMT is now expected; Fornell-Larcker alone is no longer sufficient for most reviewers.
Formative constructs are assessed differently
If your indicators cause the construct rather than reflect it, the reflective criteria do not apply and using them is a conceptual error.
- Check convergent validity by redundancy analysis against a global single-item measure.
- Check collinearity between indicators — VIF below 3 is the usual guidance.
- Assess the significance and relevance of each outer weight; a non-significant weight with a high loading may still be retained on conceptual grounds.
Stage two: the structural model
- Check collinearity between predictor constructs (inner VIF below 3).
- Run bootstrapping — 5,000 subsamples is the current standard — for path significance.
- Report R² for each endogenous construct, with the usual rough guidance of 0.25 weak, 0.50 moderate, 0.75 substantial, interpreted against your own field.
- Report f² effect sizes: 0.02 small, 0.15 medium, 0.35 large.
- Report Q² via PLSpredict for predictive relevance — a value above zero indicates the model predicts out of sample.
- Report SRMR as an approximate fit measure; below 0.08 is the conventional threshold.
What reviewers most often send back
- HTMT not reported, only Fornell-Larcker.
- Path coefficients reported without f² effect sizes.
- PLS-SEM chosen with no justification beyond sample size.
- Items dropped to improve fit without saying which or why.
- Bootstrap subsamples not stated, or left at the software default of 500.
- Formative constructs assessed with reflective criteria.
Questions this raises
The old "ten times the largest number of arrows" rule is widely criticised. Run an inverse square root power calculation or a Monte Carlo power analysis based on your model and expected effect sizes, and report it.
PLS-SEM does not produce the covariance-based fit indices. SRMR and NFI are available as approximations. Reporting CFI or RMSEA from a PLS model indicates a misunderstanding of the method.
The two constructs are not empirically distinct as measured. Examine the item wording for overlap, consider whether they are genuinely separate concepts, and either revise the instrument or merge them with a theoretical justification.
Still stuck after reading this? That is usually the point at which it is worth asking someone. Describe your project or ask on WhatsApp.
