CB-SEM and PLS-SEM: choose from measurement theory to research objectives
Choosing SEM from sample size or software branding can produce a model that runs while answering the wrong question. First identify the target quantity, how the construct is measured, and the evidence needed to assess the result.
CB-SEM and PLS-SEM differ in construct representation and estimation. Small samples, nonnormal data or many paths do not select the method automatically; consider evaluation, prediction and the data-generating design together.
Choose a SEM approach from purpose and measurement
- Purpose
Testing, explanation or prediction?
- Concept
Common factor or composite?
- Data
Variable types, distributions and design
- Estimation
Choose a defensible method
Conceptual diagram. Do not select CB-SEM or PLS-SEM solely from available software.
Source: Lê Minh · Conceptual illustration, not software output.1. Start with the question and the construct
A study of trust might represent trust as a latent characteristic responsible for observed responses. A study of digital readiness might define readiness through a combination of infrastructure, skills and procedures. These need not imply the same measurement model. Explain the construct before choosing an algorithm (Kline, 2023) (Hair et al., 2022).
State the primary objective in one sentence: estimate relations among factors while modeling measurement error, evaluate a covariance structure, or predict a target outside the estimation sample. Multiple objectives are possible, but explain which objective governs the design. An arrow diagram cannot supply that argument.
Do not equate AMOS with the entire CB-SEM family or SmartPLS exclusively with PLS-SEM. Current SmartPLS guidance also covers covariance-based estimation. Software names cannot establish what the construct represents (SmartPLS, n.d.).
2. Factors, composites and arithmetic means
Common-factor models distinguish a latent variable from indicator-specific variation. A composite combines indicators through weights to form a representative quantity. The composite-based estimation perspective does not mean that all measurement specifications in PLS must be formative (Hair et al., 2022).
Common factor: xⱼ = λⱼη + εⱼ. Illustrative composite: c = Σⱼwⱼxⱼ.
| Quantity | Construction | Required justification |
|---|---|---|
| Common factor | A latent variable accounts for indicator covariance | Factor structure, measurement errors and identification |
| Composite | A weighted combination of indicators | Content coverage and the meaning of weights |
| Mean score | Equal weighting under a declared scoring rule | Direction, missing-item rule and defensible aggregation |
| Factor score | An estimated score from a fitted factor model | Scoring method and uncertainty of individual representation |
In a hypothetical two-item example, scores of 4 and 2 have a mean of 3. A raw weighted sum using weights 0.8 and 0.2 is 3.6. This demonstrates that weighting changes the quantity. It is not fitted PLS output: actual algorithms also specify scaling and estimation. Nor is 3.6 a factor score without a fitted factor model.
3. Compare models and evidence at the right level
| Aspect | Factor-based CB-SEM | Typical PLS-SEM |
|---|---|---|
| Measurement | Loadings, unique variances, structure and validity | Assessment appropriate to reflective or formative specifications |
| Structure | Parameters, uncertainty and alternative specifications | Paths, uncertainty, explanatory results and diagnostics |
| Fit | Model tests, fit indices and residuals where defined | Diagnostics suited to the estimator rather than copied CB-SEM rules |
| Prediction | Can be evaluated out of sample when relevant | Requires out-of-sample evaluation; in-sample R² is insufficient |
| Dependence | Address clustering, repeated measures or multilevel design | Still requires appropriate inference; individual resampling does not fix clustering |
These are categories of evidence, not a compulsory universal index list. PLS applications can be confirmatory, and CB-SEM applications can address prediction. The distinction is not an exclusive divide between exploration and confirmation (Rigdon et al., 2017).
4. Two hypothetical selection exercises
Scenario A concerns managerial support, trust and intention to stay. The constructs are theorized as reflective factors, and the researcher wants to evaluate measurement structure and structural relations. CFA/CB-SEM with an estimator suited to the indicators is a candidate. Merely having 180 or 300 observations does not establish its suitability.
Scenario B defines digital readiness through infrastructure, skills and procedures and aims to predict operational performance. PLS-SEM may be appropriate if the composite interpretation and relations are justified and a prediction design is specified. Obtaining a smaller p value is not a methodological rationale.
In both scenarios check support for the variable types, missing-data strategy and dependent observations. Different measurement specifications can target different quantities; estimates from two approaches are not automatically estimates of the same parameter (Kline, 2023) (Rigdon et al., 2017).
5. Avoid convenient selection shortcuts
Switching to PLS does not create information in a small sample. Plan precision and power for the important paths and check solution behavior. Nonnormality alone does not exclude CB-SEM because alternative estimators or adjustments may be available for the model and software (Kline, 2023).
A categorical predictor, a binary outcome and an ordinal indicator require different decisions. An estimator cannot repair incorrectly coded data or an unsuitable measure. Consider complexity, missingness and group dependence during design rather than only after an error message.
PLSc applies a correction for appropriate reflective common-factor specifications. It does not turn every composite into a factor or automatically resolve endogeneity and causal identification (Hair et al., 2022).
6. A reproducible selection workflow
- Define each construct and justify a factor or composite representation.
- Identify predictors, outcomes, indicators and grouping variables with their scales.
- Select estimation, missing-data handling and inference before examining p values.
- Plan sample size, precision and solution diagnostics.
- Evaluate measurement before interpreting structural paths.
- Specify an evaluation sample and benchmark when prediction is an objective.
- Archive software versions, settings, syntax or model files and complete output.
When two methods form a robustness exercise, document differences in measurement, scaling, the analyzed sample, constraints and effect definitions. Explain these differences before attributing coefficient differences to an algorithm.
7. Report the rationale and its limits
Reporting template to adapt
The study selected [method] because constructs were represented by [measurement model] and the primary objective was [objective]. Data features [details] were addressed using [estimator/rule]. Evaluation included [measurement evidence], [structural evidence] and [prediction assessment where relevant]. Interpretation is limited by [design, sample and assumptions].
Claims that a method is better require a criterion and a model class. Stable coefficients and acceptable fit in a cross-sectional study do not establish causal direction or rule out omitted variables.
Continue: categorical variables in SEM. Prediction with PLSpredict and CVPAT.
Conclusion
A defensible choice connects concepts, data, estimation and evaluation. Once these are clear, software implements a decision that can be checked rather than deciding the research question.
References
- Hair, J. F., Jr., Hult, G. T. M., Ringle, C. M., & Sarstedt, M. (2022). A primer on partial least squares structural equation modeling (PLS-SEM) (3rd ed.). SAGE https://us2.sagepub.com/en-us/nam/a-primer-on-partial-least-squares-structural-equation-modeling-pls-sem/book270548
- Kline, R. B. (2023). Principles and practice of structural equation modeling (5th ed.). The Guilford Press https://www.guilford.com/books/Principles-and-Practice-of-Structural-Equation-Modeling/Rex-Kline/9781462551910
- Rigdon, E. E., Sarstedt, M., & Ringle, C. M. (2017). On comparing results from CB-SEM and PLS-SEM: Five perspectives and five recommendations. Marketing ZFP, 39(3), 4–16. https://doi.org/10.15358/0344-1369-2017-3-4
- SmartPLS (n.d.). PLS-SEM and CB-SEM: When to use which SEM method. https://smartpls.com/documentation/choosing-pls-sem/choosing-between-pls-sem-and-cb-sem/