Image

PLS-SEM Compared with CB-SEM (and More)

PLS-SEM and CB-SEM are complementary rather than competing methods for estimating structural equation models. When the same technology acceptance model (TAM) is estimated on the same dataset with nine different approaches, maximum likelihood CB-SEM, consistent PLS-SEM (PLSc-SEM), and PLS-SEM produce closely aligned results — while alternative CB-SEM estimators such as GLS, ULS, and ADF diverge substantially. The choice of estimation method therefore matters more than the simple "PLS-SEM vs. CB-SEM" framing suggests. Instead of committing to a single estimation tradition, researchers should consider a multimethod SEM approach that uses factor-based and composite-based estimators side by side to establish the robustness of their findings (Hair et al., 2026; Sarstedt et al., 2024; Sharma et al., 2024).
We deliberately do not use the term "PLS-SEM vs. CB-SEM". Both methods are complementary rather than competitive. Even though this issue is well-known (see, for example, Jöreskog and Wold, 1982), many researchers still focus on comparing the differences of model estimations when using covariance-based structural equation modeling (CB-SEM) and partial least squares structural equation modeling (PLS-SEM). Instead of distinguishing between common factor models and composite models (Henseler et al., 2014), these discussions focus on PLS-SEM's capabilities to mimic CB-SEM. But PLS-SEM in its original form (Wold, 1982; Lohmöller, 1989) has not been created to mimic CB-SEM! PLS-SEM researchers should follow Rigdon's (2012) call and begin emancipating the method from its CB-SEM sibling (also see Sarstedt et al., 2014; Sarstedt et al., 2016; Rigdon, 2014; Rigdon et al., 2017).
For guidance on how to select the method that fits your study, see our decision guide on choosing between PLS-SEM and CB-SEM. The most important reason to select CB-SEM or PLS-SEM is the research goal (structure or prediction): "The primary purpose of the ML approach is to study the structure of the observables .... The primary purpose of the PLS approach is to predict the indicators by means of the components expansion (1)." (Jöreskog and Wold, 1982; p. 266). In line with this notion, Hair et al. (2011; p. 144) recommend:
  • If the goal is predicting key target constructs or identifying key 'driver' constructs, select PLS-SEM.
  • If the goal is theory testing, theory confirmation, or comparison of alternative theories, select CB-SEM.
  • If the research is exploratory or an extension of an existing structural theory, select PLS-SEM.
However, Bentler and Huang (2014), Dijkstra (2014), and Dijkstra and Henseler (2015) introduced methods that provide consistent PLS-SEM estimations. These consistent PLS-SEM (PLSc-SEM) estimations of common factor models have been designed to mimic CB-SEM (Sarstedt et al., 2016). Thereby, researchers can also use PLS-SEM to study structure (Rigdon et al., 2017). As a result, we see two developments of PLS-SEM: one direction uses PLS-SEM for prediction-oriented studies, and another direction uses PLS-SEM (via PLSc-SEM) to mimic CB-SEM for studies that focus on analyzing and testing the model structure.

The TAM Example: One Model, One Dataset, Nine Estimations

In face of these developments, we use an application of the well-known technology acceptance model (TAM; Davis, 1989). The model estimation uses a dataset with 1,190 responses and the SmartPLS (Ringle et al., 2024) and AMOS (Arbuckle, 2006) software (alternative structural equation modeling software solutions are, for example, EQS, LISREL, and MPLUS). You can download the TAM example, which can be imported as a project into the SmartPLS software, from the resources of this webpage.
Similar to the study presented by Sarstedt et al. (2024), we used CB-SEM and PLS-SEM to estimate the model. However, we also used alternative CB-SEM estimators and added the results of the sum score regression and equal weighting.

Results at a Glance

The following table summarizes the standardized path coefficients of the TAM structural model across all nine estimations (values as shown in the model figures below; EOU = perceived ease of use, USEF = perceived usefulness, ATT = attitude toward using, BI = behavioral intention to use, Use = actual system use):
PathML (AMOS)ML (SmartPLS)PLSc-SEMPLSSum scoresEqual weightsGLS (AMOS)ULS (AMOS)ADF (AMOS)
EOU → USEF0.510.5090.5000.4450.4430.4440.580.750.88
USEF → BI0.470.4660.4590.4010.3810.3900.560.871.00
USEF → ATT0.260.2640.2760.2670.2680.2680.200.401.14
EOU → ATT0.290.2850.2720.2530.2510.2500.330.15-0.60
ATT → BI0.160.1600.1770.1730.1690.1700.19-0.10-0.29
ATT → Use0.290.2870.3170.2710.2770.2760.350.350.33
BI → Use0.190.1910.1600.1340.1100.1240.080.220.21
Key findings:
  • The results of ML-CBSEM in AMOS and SmartPLS are identical.
  • PLSc-SEM and PLS-SEM come closest to the ML-CBSEM results — and, to some extent, although by a slightly larger margin, so do sum score regression and PLS-SEM using equal weights.
  • In contrast, it is surprising to note the large differences in results when using alternative CB-SEM estimation methods (e.g., GLS, ULS, and ADF). The ADF estimation even reverses the sign of two structural paths.
Therefore, the one-sided view pitting CB-SEM against PLS-SEM seems to be misleading. Rather, both model estimation methods should be considered as complementary approaches to SEM (Jöreskog & Wold, 1982). Differences within the CB-SEM family of estimators can be far larger than the differences between ML-CBSEM, PLSc-SEM, and PLS-SEM.

Why Results Vary: Three Sources of Uncertainty

Sarstedt et al. (2024) asked leading SEM experts to estimate the same prespecified model on the same dataset using the estimators they developed or mastered (CB-SEM, PLS, PLSc-SEM, GSCA, and GSCAM). Even under these tightly controlled conditions, the outcomes varied considerably — in effect sizes, in significance levels, and in the analytical workflows the experts followed. The study identifies three major sources of results variability:
  1. Methodological uncertainty arises from the choice of the SEM method itself. Factor-based and composite-based estimators embed different assumptions about the nature of constructs and different optimization routines, so a preference for one method necessarily comes with assumptions about unknown entities in the model.
  2. Model estimation uncertainty arises from the analytical decisions researchers make along the way — whether to modify the model, which evaluation criteria to emphasize (model fit vs. predictive power), algorithm settings, bootstrapping options, and so on.
  3. Interpretational uncertainty arises from how researchers read the estimates — for example, treating a path as "present" or "absent" based on a significance threshold, even when point estimates across methods barely differ.
The practical conclusion is not to search for the one "correct" estimator, but to make analytical decisions transparent, acknowledge the uncertainty of results, run robustness checks across methods, and document alternative analytical workflows (Sarstedt et al., 2024). This is precisely what a multimethod SEM approach delivers.

The Multimethod SEM Framework

Building on this evidence, Hair et al. (2026) formalize a multimethod SEM framework that applies factor-based (CB-SEM) and composite-based (PLS-SEM) estimators to the same structural model. The framework rests on two ideas.
First, most theories in business and social science research aim to both explain and predict (Gregor, 2006; Sharma et al., 2024). Explanation provides theoretical relevance and is assessed with in-sample criteria (path significance, effect sizes, model fit); prediction provides practical relevance and is assessed with out-of-sample criteria such as PLSpredict and the cross-validated predictive ability test (CVPAT). Because a model that fits well does not necessarily predict well (and vice versa), both assessments are needed. Sharma et al. (2024) show that combining in-sample explanatory and out-of-sample predictive assessment also enhances the replicability of research findings.
Second, CB-SEM and PLS-SEM rest on different conceptual and statistical assumptions (common factor vs. composite model). A structural path that remains stable across both estimation paradigms is unlikely to be an artifact of one method's assumptions — convergence reinforces confidence, while divergence signals that a finding is estimator-sensitive and requires closer scrutiny.

The Multimethod Workflow

Adapted from Hair et al. (2026), the workflow proceeds as follows:
  1. Specify the model based on theory-based deduction.
  2. Check the data (distributional properties, missing values, extreme observations, Heywood cases) as relevant to each SEM type.
  3. Establish measurement model quality using the established assessment guidelines.
  4. Validate the overall model: assess model fit with CB-SEM and explained variance (R²) with PLS-SEM; assess overall predictive validity with out-of-sample tests such as CVPAT.
  5. Assess the structural paths in-sample with CB-SEM (significance and size of the path coefficients).
  6. Assess the structural paths in-sample with PLS-SEM.
  7. Compare the in-sample results across both estimators and interpret convergence and divergence.
  8. Test each focal path out-of-sample: use CVPAT to compare the hypothesized model with a nested model that excludes the path, isolating its predictive contribution.
  9. Integrate explanation and prediction and report estimator convergence or divergence transparently.

Interpreting the Results: Theoretical and Practical Relevance

At the path level, the joint evidence from in-sample (explanatory) and out-of-sample (predictive) testing leads to four cases (adapted from Sharma et al., 2024, and Hair et al., 2026):
Out-of-sample: not significantOut-of-sample: significant
In-sample: not significantThe path is neither theoretically nor practically relevant. Consider removing it or adjusting the theory.The path is practically relevant but lacks explanatory support — a signal of hidden mechanisms and an opportunity for theory development.
In-sample: significantThe path is theoretically relevant but adds no predictive value — a possible sign of overfitting to sample-specific noise.The path is theoretically and practically relevant — the strongest form of confirmatory evidence.
In addition, the in-sample comparison across estimators matters: when both CB-SEM and PLS-SEM find a path significant and the predictive test confirms its relevance, the evidence is strongest. When the estimators disagree, the in-sample support is estimator-contingent and should be flagged as provisional; researchers can then triangulate with methods that bridge both paradigms, such as consistent PLS-SEM (PLSc-SEM) or integrated GSCA, and consider replication or model respecification (Hair et al., 2026).

Multimethod SEM in SmartPLS

SmartPLS supports the complete multimethod workflow in one software environment: PLS-SEM and PLSc-SEM for composite-based estimation, CB-SEM and CFA for factor-based estimation, and PLSpredict and CVPAT for out-of-sample predictive assessment. The TAM example on this page can serve as a template: researchers can download the project and reproduce the multimethod comparison themselves.

Alternative Model Estimations

CB-SEM maximum likelihood (ML) results using AMOS (standardized coefficients)

TAM estimated with CB-SEM maximum likelihood (ML) in AMOS: standardized path coefficients

CB-SEM maximum likelihood (ML) results using SmartPLS (standardized coefficients)

TAM estimated with CB-SEM maximum likelihood (ML) in SmartPLS: identical results to AMOS ML

PLS-SEM mimicking CB-SEM results via consistent PLS-SEM (PLSc-SEM) using SmartPLS

TAM estimated with consistent PLS-SEM (PLSc-SEM) in SmartPLS: results close to ML-CBSEM

PLS-SEM results using SmartPLS

TAM estimated with PLS-SEM in SmartPLS: standardized path coefficients and R² values

Sum score regression results using SmartPLS

TAM estimated with sum score regression in SmartPLS

Equal weights PLS-SEM results using SmartPLS

TAM estimated with equal indicator weights in SmartPLS

CB-SEM generalized least squares (GLS) results using AMOS (standardized coefficients)

TAM estimated with CB-SEM generalized least squares (GLS) in AMOS: results diverge from ML

CB-SEM unweighted least squares (ULS) results using AMOS (standardized coefficients)

TAM estimated with CB-SEM unweighted least squares (ULS) in AMOS: results diverge strongly from ML

CB-SEM asymptotically distribution-free (ADF) results using AMOS (standardized coefficients)

TAM estimated with CB-SEM asymptotically distribution-free (ADF) estimation in AMOS: strongly diverging results with sign reversals

Frequently Asked Questions

Do PLS-SEM and CB-SEM produce different results?

When CB-SEM uses maximum likelihood (ML) estimation, PLS-SEM and especially consistent PLS-SEM (PLSc-SEM) produce very similar structural model results, as the TAM example on this page shows. Larger differences occur within the CB-SEM family itself: alternative estimators such as GLS, ULS, and ADF can diverge substantially from ML — including sign reversals of structural paths.

What is a multimethod SEM approach?

A multimethod SEM approach estimates the same structural model with both factor-based (CB-SEM) and composite-based (PLS-SEM) estimators and compares the results at the path level, complemented by out-of-sample predictive tests such as CVPAT (Hair et al., 2026). Paths that are significant under both estimators and predictively relevant provide the strongest confirmatory evidence; diverging results flag estimator-sensitive findings that require closer scrutiny.

Why do results vary even when experts analyze the same model and data?

Because uncertainty enters through three channels: the choice of the SEM method (methodological uncertainty), the analytical decisions made during estimation and evaluation (model estimation uncertainty), and the interpretation of estimates and thresholds (interpretational uncertainty) (Sarstedt et al., 2024). Multimethod estimation and transparent reporting of the analytical workflow make this variability visible and manageable.

Why should I assess prediction in addition to model fit?

Because in-sample fit does not guarantee out-of-sample predictive power — the best-fitting model is not necessarily the best-predicting model, and vice versa. Combining in-sample explanatory assessment with out-of-sample predictive assessment (e.g., PLSpredict, CVPAT) confirms both the theoretical and the practical relevance of a model and enhances the replicability of the findings (Sharma et al., 2024).

What is consistent PLS-SEM (PLSc-SEM)?

Consistent PLS-SEM (PLSc-SEM; Dijkstra & Henseler, 2015) corrects the PLS-SEM estimates of reflectively measured constructs for attenuation so that they mimic common factor model results. Researchers can thereby use PLS-SEM to study model structure, similar to CB-SEM.

Why do GLS, ULS, and ADF results differ so much from ML?

These CB-SEM estimators optimize different discrepancy functions and rely on different assumptions (e.g., ADF requires very large samples to work well). With the same model and the same data, they can produce substantially different parameter estimates — in this TAM example, ADF even reverses the sign of two paths. This underlines that "CB-SEM" is not a single, uniform set of results.

Which method should I use for my study?

The choice depends on the research objective: prediction and explanation favor PLS-SEM, theory testing and global model fit favor CB-SEM, and a multimethod strategy combines both. See our guide on choosing between PLS-SEM and CB-SEM for practical decision rules and justification examples.

Can I reproduce this comparison myself?

Yes. Download the TAM example project and import it into SmartPLS. You can then estimate the model with PLS-SEM, PLSc-SEM, CB-SEM (ML), sum scores, and equal weights in one environment.

References

SmartPLS GmbH
Rheinpromenade 2
D-40789 Monheim am Rhein
Germany

+49 02173 2643460

Crowdin | Agile localization for tech companies

Company

About usUsage TermsPrivacyImprint

Learn

CoursesGetting StartedFAQSubmit a ticket

Software & Licenses

PricingPurchaseDownloads30 Days Free TrialSubmit Withdrawal