CB-SEM or PLS-SEM: choosing before you click
A surprising number of SEM problems start with the wrong family of method. Covariance-based SEM (AMOS, lavaan, Mplus) tests whether your whole model reproduces the observed covariance matrix, so global fit indices matter and reflective, well-established scales are the norm. Partial least squares SEM (SmartPLS, seminr) maximises explained variance in the endogenous constructs and is often chosen for prediction-oriented or exploratory work, formative constructs or complex models.
| CB-SEM | PLS-SEM | |
|---|---|---|
| Main goal | Confirm a theory | Explain and predict |
| Typical software | AMOS, lavaan, Mplus | SmartPLS, seminr |
| Fit assessment | CFI, TLI, RMSEA, SRMR, chi-square | R², Q², PLSpredict; SRMR is optional |
| Formative constructs | Awkward | Handled directly |
Neither is a fix for small samples or weak measures. We help you pick the approach that suits your research question and justify it in your methodology chapter.
The two-step process we follow
Step one: the measurement model
Before any hypothesis is tested, we check that each construct is measured well. In AMOS that means a confirmatory factor analysis with standardised loadings, composite reliability and average variance extracted (AVE). In SmartPLS it means indicator loadings, Cronbach's alpha, rho_A and composite reliability, and AVE. Discriminant validity is assessed with the HTMT ratio, with the Fornell-Larcker criterion reported alongside if your field still expects it.
Step two: the structural model
Only once measurement holds do we estimate paths. We check collinearity (inner VIF), bootstrap the indirect effects for mediation rather than relying on the Sobel test, and set up moderation through interaction terms or multi-group analysis. Effect sizes (f²) and predictive relevance are reported where they apply.
Common problems we help resolve
- Poor fit in AMOS. We look for the cause, such as a cross-loading item or an omitted path, rather than adding error covariances simply because modification indices suggest them.
- Heywood cases such as negative error variances, usually linked to small samples or constructs with only two indicators.
- Discriminant validity failures where two constructs are effectively measuring the same thing.
- Common method bias concerns, addressed with a full collinearity test, a marker variable or a latent method factor, depending on design.
- Non-normal or ordinal data, where we switch to robust estimators or bootstrapping instead of ignoring the issue.
If your questionnaire is still being built, getting the scales right at that stage through our survey and questionnaire design support prevents many of these problems.
What to send us
Share your conceptual model or hypotheses, the questionnaire with item wording, your data file (SPSS, Excel or CSV), and any output you have already produced. Mention which software your university licenses and what your supervisor expects to see. A focused review of an existing model often takes a few days; a full measurement and structural analysis typically takes one to two weeks. You can start through a free quote request.
It also helps to tell us how the data were collected. Whether responses came from a single survey at one time point, whether any items were reverse-worded, and how missing values were handled all affect estimator choice and how much weight a common method bias test can carry. Small details like these often decide whether a model converges cleanly at the first attempt.
Analysis you can stand behind
SEM output is dense, and examiners know it. They often ask why an item was dropped, why HTMT was used, or how a mediation effect was tested. We document each decision and its justification so that you, as the researcher, can answer those questions confidently. Your model and conclusions remain your own work, and our role follows the principles in our academic integrity policy.
What you get
- A model specification diagram that matches your hypotheses
- Measurement model results: loadings, reliability, convergent and discriminant validity
- Structural model results with path coefficients, bootstrapped confidence intervals and R²
- Mediation and moderation tests set up and interpreted correctly
- APA-style tables and a results write-up template for you to adapt in your own words
- Saved model files (.amw, SmartPLS project or lavaan script) so every result can be rerun
What we don't do
- Adjusting the model purely to chase better fit statistics without theoretical justification
- Writing your discussion chapter; we explain the findings so you can interpret them yourself
- Guaranteeing that a hypothesis will be supported
How it works
Describe your project
Level, deadline and what you need help with.
Agree a written scope
What's included, the price and the timeline.
Work in stages
Check-ins so you can steer the work.
Review and ask
Explanations and follow-up questions included.