AMOS vs SmartPLS for SEM

What is the difference between CB-SEM and PLS-SEM?

Both estimate a measurement model (how items relate to constructs) and a structural model (how constructs relate to each other). They do it in different ways, for different purposes.

Covariance-based SEM treats each construct as a common factor: something that causes its items to vary together, with the rest being error. It estimates the whole model at once and asks whether the covariances it implies match the covariances in your data. That is why it gives you global fit statistics. AMOS is the best-known program for it in Indian universities, largely because it comes with IBM SPSS licences. Others include the free R package lavaan, and the SEM modules in JASP and jamovi, which use lavaan underneath.

PLS-SEM builds each construct as a weighted composite of its items, then estimates the paths between composites using a series of regressions. It aims to maximise the variance explained in the outcome constructs. SmartPLS is the most widely used program; alternatives include the R packages seminr and cSEM, ADANCO and WarpPLS.

The software and the method are not quite the same thing. SmartPLS 4 can now also run CB-SEM, and R can do both. So “we used SmartPLS” does not tell an examiner which method you used. When you write the thesis, name the method first and the software second.

Which should you choose for your thesis?

Start from the research aim, then look at the model.

• Your aim is to test an established theory and you want to know whether the whole model fits the data: CB-SEM (AMOS).
• Your aim is to explain or predict a key outcome, or you are extending a theory with new constructs: PLS-SEM (SmartPLS).
• Your model has formative constructs, such as an index built from distinct components: PLS-SEM is simpler.
• Your model has a feedback loop (X affects Y and Y affects X): CB-SEM.
• Your model is very large, with many constructs and paths: PLS-SEM usually estimates more easily.
• Your field’s top journals mostly use one method for this kind of model: that is a legitimate factor, as long as it is not your only reason.

If none of these points clearly applies, and your model is a standard reflective model tested with a few hundred survey responses, both methods are defensible and usually give similar conclusions. In that case, the tradition of your field and your supervisor’s expertise are reasonable tie-breakers. What matters is that you can explain the choice.

A note on hypotheses

A thesis that frames everything as confirmatory hypothesis testing of an established theory, and then chooses PLS-SEM “for prediction”, invites an obvious question at the viva. Make the aim you state and the method you choose agree.

Which justifications do examiners accept?

A large share of PLS-SEM papers justify the method by small sample size and non-normal data alone. Those two reasons have been criticised strongly. Rönkkö and Evermann (2013) argued that PLS offers no real advantage with small samples, and Hair and colleagues, who are among PLS-SEM’s main advocates, have themselves said that these should not be the main reasons for choosing it.

Weak on their own

• “The sample is small.” PLS will run with a small sample, but that does not make the estimates reliable. You still need enough power for the effects you expect. Our sample size guide covers the inverse square root method used for PLS.
• “The data is not normal.” Likert data is rarely normal. But CB-SEM has well-established remedies (bootstrapping in AMOS, the MLR estimator with corrected standard errors in lavaan), so non-normality alone does not force PLS.
• “It gave better results.” This is not a methodological reason, so leave it out of both the decision and the write-up.

Stronger

• The study aims to predict or explain variance in a key outcome, for example intention to adopt a technology.
• The model includes formative constructs.
• The model is complex, with many constructs, mediators and moderators.
• The study extends an existing theory with new constructs, so it is partly exploratory.
• For CB-SEM: the aim is to confirm a well-established theory, and overall model fit is needed to judge it.

Cite the method literature in your justification. Hair, Risher, Sarstedt and Ringle (2019) set out when PLS-SEM is appropriate and what to report; for CB-SEM, Kline’s textbook Principles and Practice of Structural Equation Modeling is a standard reference.

What do you report for each method?

Reviewers and examiners expect a standard set of results for each method. Leaving out an expected item is one of the most common reasons SEM chapters come back with corrections.

CB-SEM in AMOS

• Measurement model (CFA): standardised loadings, composite reliability, AVE, discriminant validity, and fit indices (chi-square with degrees of freedom, CFI, TLI, RMSEA with its confidence interval, SRMR).
• Any modifications, such as correlated error terms, with a theoretical reason for each.
• Structural model: fit indices again, standardised path coefficients, significance, R² for each outcome.
• Indirect effects tested by bootstrapping, with confidence intervals.
• How missing data and non-normality were handled.

PLS-SEM in SmartPLS

• Measurement model: loadings, Cronbach’s alpha, composite reliability (rho_A and rho_C), AVE, and HTMT for discriminant validity. For formative constructs: weights, their significance, and collinearity (VIF).
• Structural model: collinearity (VIF), path coefficients with bootstrapped confidence intervals, R², f², and predictive power using PLSpredict (Shmueli and colleagues, 2019).
• Bootstrapping settings: the number of subsamples (5,000 or more is commonly used and reported) and the confidence interval method.
• SRMR if you report it, with a note that its use as a fit measure in PLS is debated.

In both cases, put the measurement model results before the structural model. If the measurement fails, the path results mean little.

Practical points: licences, learning and common errors

Licences

AMOS is licensed through IBM, and many Indian universities include it in their SPSS site licence; ask your computer centre or department before buying. SmartPLS sells its own licences, with different terms for students and institutions and a trial period. Check the current terms on each vendor’s website. If cost is a problem, R (lavaan for CB-SEM, seminr for PLS-SEM) and JASP are free, and increasingly accepted in journals.

Use a properly licensed copy of either program. Unlicensed copies breach the licence terms, and journals increasingly ask you to state the software and version you used.

Learning curve

Both programs let you draw the model, which makes them easy to start with and easy to misuse. The drawing is the least of the work. Budget time to learn what the output means, not just where to click.

Errors we see often

• Reporting Fornell–Larcker only, without HTMT.
• In AMOS, adding correlated errors between items of different constructs until fit becomes acceptable.
• In SmartPLS, deleting items one by one until every loading exceeds 0.70, leaving two-item constructs.
• Testing mediation with the Baron and Kenny steps instead of bootstrapped indirect effects (see mediation and moderation analysis).
• Calling a cross-sectional PLS model “causal” in the discussion.
• Pasting whole software output pages into the thesis instead of formatted tables.

How do you write the justification in the methodology chapter?

One or two paragraphs are enough. State the method, the reasons tied to your aim and model, and the software with its version. A workable pattern:

The study used partial least squares structural equation modelling (PLS-SEM), because its primary aim was to explain and predict variance in adoption intention, and because the model included two formative constructs (Hair et al., 2019). The analysis was carried out in SmartPLS [version]. Significance was assessed by bootstrapping with 10,000 subsamples.

or

Covariance-based SEM was used because the study tested an established theory, and assessing how well the theorised model fitted the data was central to the research questions. The analysis was carried out in IBM SPSS AMOS [version] using maximum likelihood estimation, with bootstrapped standard errors because several indicators departed from normality.

Replace the bracketed parts, and make sure the reason you give matches what your introduction says the study is for.