CB-SEM vs PLS-SEM: How to Choose the Right Structural Equation Modeling Approach

Structural equation modeling (SEM) combines two things: a measurement model, which links observed indicators such as questionnaire items to the constructs they represent, and a structural model, which specifies the hypothesized relationships between those constructs. Both CB-SEM and PLS-SEM estimate these two parts, which is why they are so often confused. The difference lies in how the constructs are represented and what the estimation procedure tries to achieve.

Covariance-based SEM (CB-SEM) treats each construct as a common factor: an unobserved variable that causes its indicators, with each indicator also carrying its own measurement error. Estimation, usually by maximum likelihood, searches for parameter values that make the covariance matrix implied by your model as close as possible to the covariance matrix observed in your data. The model is therefore a claim about the whole covariance structure, and you can test how well that claim holds.

Partial least squares SEM (PLS-SEM) represents each construct as a composite: a weighted combination of its indicators. The algorithm alternates between estimating those weights and running a series of ordinary least squares regressions, with the aim of maximizing the variance explained in the endogenous (dependent) constructs. It is often described as a prediction-oriented, variance-based approach. Because it does not fit the entire covariance matrix in one step, it does not produce a single global test of model fit in the way CB-SEM does.

Neither approach is universally better. They were designed for different purposes, and a large methodological literature, including a long and sometimes heated debate between advocates and critics of PLS, has clarified when each is appropriate. The practical question for you is which assumptions match your theory and which evidence your examiners, reviewers or target journal will expect.

The table below summarizes the differences that matter most when you are planning a study or justifying a choice in a methodology chapter. Treat it as a map, not a set of rules: several rows describe tendencies that depend on model complexity, data quality and the specific estimator used.

Many students choose PLS-SEM because a colleague used SmartPLS or because a sample seemed too small for AMOS. That is a weak justification, and examiners who know the field will notice. A stronger choice starts from what your study is trying to demonstrate.

If your goal is to test an established theory, where the model is fully specified in advance and you want to know whether the data are consistent with it, CB-SEM is the natural choice. The global fit test is exactly the evidence that a theory-testing study needs: it can tell you that the model as a whole is misspecified, even if individual paths are significant.

If your goal is to explain as much variance as possible in an outcome such as intention, satisfaction or performance, or to predict outcomes for new cases, PLS-SEM fits the aim well. It is also frequently used when researchers extend an existing model with new constructs whose relationships are less settled, or when the model includes formatively measured constructs such as an index built from distinct components.

The nature of your constructs matters too. If you think of a construct as a latent trait that causes its indicators, such as anxiety measured by several interchangeable items, a common factor model is conceptually appropriate. If the construct is defined by its components, such as a socioeconomic status index built from income, education and occupation, a composite is closer to the idea. Recent guidance encourages researchers to state explicitly which measurement philosophy they adopt and to choose the estimator that matches it; in PLS software this may mean using consistent PLS for reflective common-factor constructs.

• Theory confirmation with reflective constructs and an adequate sample: CB-SEM is usually the better fit.
• Prediction or explanation of key outcomes, especially with formative constructs: PLS-SEM is a defensible choice.
• Mixed measurement (reflective and formative) with a prediction focus: PLS-SEM, reporting PLSc results for reflective parts where appropriate.
• Very small samples: neither method rescues an underpowered study; consider simplifying the model or collecting more data.

The most persistent myth about PLS-SEM is that it works with any sample size. The algorithm will usually converge with small samples, but convergence is not the same as trustworthy estimates. Small samples produce wide confidence intervals, low statistical power to detect modest effects, and results that may not replicate. The old ten-times rule, which suggested ten cases per arrow pointing at the most-targeted construct, has been widely criticized as too lenient.

Better practice for PLS-SEM is to base the minimum sample on the smallest path coefficient you expect to detect and the desired power, using methods such as the inverse square root approach or a formal power analysis. Current PLS software includes sample size and power tools that support this kind of planning.

For CB-SEM there is no single number either. Required sample size grows with the number of parameters, the strength of factor loadings, the amount of missing data and the degree of non-normality. Simulation-based power analysis, for example with Monte Carlo facilities in Mplus or simulation packages in R, gives a model-specific answer. Whatever method you use, report how you arrived at your target sample, not only the final number.

A CB-SEM analysis normally proceeds in two steps. First, a confirmatory factor analysis (CFA) tests the measurement model on its own. Second, the structural paths are added and the full model is estimated. Separating the steps helps you see whether any misfit comes from the measurement of constructs or from the proposed relationships between them.

Global fit is judged with the chi-square test and approximate fit indices. The chi-square test is sensitive to sample size, so with large samples it often rejects models that fit reasonably well; for that reason researchers also report the comparative fit index (CFI), Tucker-Lewis index (TLI), root mean square error of approximation (RMSEA) with its confidence interval, and the standardized root mean square residual (SRMR). Commonly cited cutoffs, such as CFI near 0.95 or above and RMSEA near 0.06 or below, come from simulation studies under specific conditions and should be treated as guidelines rather than pass or fail thresholds.

At the measurement level you report standardized loadings, composite reliability (or McDonald's omega), average variance extracted, and discriminant validity, increasingly assessed with the heterotrait-monotrait ratio (HTMT) rather than only the Fornell-Larcker criterion. If fit is poor, modification indices can suggest changes, but every modification should be theoretically justified and reported; free respecification driven only by indices capitalizes on chance.

• Report the estimator (for example robust ML or WLSMV for ordinal items) and how missing data were handled.
• Report both unstandardized and standardized path estimates where your field expects them.
• Test indirect effects with bootstrapped confidence intervals rather than the older causal-steps approach.
• If you compare groups, establish measurement invariance before comparing paths or means.

Because PLS-SEM does not rely on a global fit test, evaluation follows a structured sequence of criteria. For reflectively measured constructs you examine indicator loadings (values of about 0.708 or higher are often recommended, because the construct then explains at least half of the indicator's variance), internal consistency reliability (Cronbach's alpha, composite reliability and rho_A), convergent validity through average variance extracted (0.50 or higher), and discriminant validity using HTMT.

Formatively measured constructs are assessed differently: you check convergent validity through redundancy analysis where possible, collinearity among indicators using variance inflation factors, and the significance and relevance of indicator weights. Deleting a formative indicator changes the meaning of the construct, so decisions about weak indicators need theoretical reasoning, not only statistics.

The structural model is then evaluated through collinearity checks, the size and bootstrapped significance of path coefficients, the coefficient of determination (R-squared) for endogenous constructs, effect sizes (f-squared), and predictive performance. Out-of-sample prediction with PLSpredict, and model comparison tests such as the cross-validated predictive ability test (CVPAT), have become standard parts of a thorough report.

SmartPLS and other packages do report SRMR and exact-fit statistics for PLS models. Their interpretation for composite models is still debated, so most guidance recommends reporting them with caution rather than treating them as the main evidence for the model.

Whichever method you choose, the same few problems account for most critical feedback. Reviewing this list before you finalize your analysis plan can save a round of revisions.

• Justifying PLS-SEM only by non-normal data or small sample size, without linking the choice to the research aim or construct type.
• Using CB-SEM fit cutoffs as rigid pass or fail rules, or reporting only the indices that look good.
• Treating a formative construct as reflective (or the reverse), which distorts reliability and validity checks.
• Deleting many indicators to improve fit or AVE without reporting what was removed and why.
• Ignoring common method variance when all constructs come from one self-report survey at one time point.
• Claiming causal effects from cross-sectional survey data without acknowledging the design limitation.
• Testing mediation with the causal-steps method instead of bootstrapped indirect effects.
• Running a multigroup comparison without first checking measurement invariance (MICOM in PLS, invariance testing in CB-SEM).

Some researchers run both CB-SEM and PLS-SEM and report them together, presenting the second as a robustness check. This can be useful when results agree, because it shows that conclusions do not depend on the estimation philosophy. It can also cause confusion if you do not explain why estimates differ. Path coefficients from standard PLS tend to be somewhat different from CB-SEM estimates when constructs are truly common factors, because composites include measurement error; consistent PLS reduces that difference.

If you do report both, choose one as the primary analysis based on your aim and present the other briefly, with a clear statement of what it adds. Avoid choosing whichever method gives more significant results, and never run both and report only the more favorable one: that is a form of analytic flexibility that undermines the credibility of the findings.

A good justification paragraph does four things. It states the research aim (confirmatory, explanatory or predictive). It describes how constructs are conceptualized (common factors, composites or a mix). It explains how sample size was planned and whether the data meet the estimator's requirements. It names the software and version, the estimator, and the key settings such as the number of bootstrap subsamples.

In the results section, follow the evaluation sequence of your method: measurement model first, then structural model, then any mediation, moderation or group comparisons. Present tables that a reader can check, for example loadings, reliability and validity statistics in one table, HTMT or correlations in another, and path estimates with confidence intervals in a third. If your model includes indirect or conditional effects, dedicated mediation and moderation analysis techniques will shape how those effects are tested and reported.

Finally, discuss limitations honestly. Cross-sectional data, single-source measurement, and model modifications made after seeing the data all affect how strongly you can state your conclusions. Stating them clearly shows methodological maturity rather than weakness.

Use the questions below as a final check before you commit. If most of your answers point in one direction, that method is probably the right primary analysis for your study.