Common Questions/Misconceptions About Structural Equation Modelling Answered.

TL;DR

Include independent variables, mediators, moderators, dependent variables, and relevant constructs together in a full SEM model rather than reporting separate measurement models for each construct.

Briefing Cornell Notes

Briefing

Structural equation modelling (SEM) work should keep models intact rather than stripping out variables just because they’re not “independent.” When using covariance-based SEM (e.g., AMOS) or variance-based SEM (e.g., Smart PLS), the recommended practice is to include all relevant constructs—independent variables, mediators, moderators, dependent variables, and control variables—inside a complete measurement model and then estimate the full model. Running constructs separately can help with initial checks of reliability and validity, but reporting separate measurement models for each construct doesn’t serve the goal of SEM, which is to estimate a coherent system of relationships and assess how constructs relate to one another.

Mediation and moderation should also be handled within SEM rather than split into simpler, non-SEM workflows. If mediation is removed and the analysis is reduced to only independent variables (IV) and dependent variables (DV), or if moderation is removed and the model is broken into separate pieces (e.g., IV–DV and mediator–DV models), then SEM tools become unnecessary. The core strength of SEM is its ability to estimate complex models in one framework, so mediators and moderators should be included and assessed together to produce clearer, more interpretable results.

A key distinction is made between placing moderators in the measurement model versus testing moderating effects. Moderators should be treated as latent variables within the measurement model so their reliability and validity are evaluated like any other construct. However, the moderating effect itself—how a moderator changes the strength or direction of a relationship—belongs to the structural part of the model, not the measurement model. In other words, the measurement model checks whether the moderator is measured well; the structural model tests whether it actually moderates an endogenous relationship.

The transcript also addresses a common reporting question: whether to assess and report the effect of a moderator on an endogenous variable. The guidance is conditional. If the moderator is a proper antecedent and aligns with the study objectives—such as examining how role ambiguity affects the relationship between collaborative culture and organizational performance—then the moderator’s impact on the dependent outcome should be assessed and reported.

Finally, the discussion pushes back against the habit of deleting “insignificant” paths. Insignificant results don’t automatically mean the study is wrong, especially in complex models where some relationships may genuinely show no effect in a given context. Instead of removing these paths, researchers should report them and justify them in the discussion. One example is shared from prior work where a negative effect was found despite expectations of a positive relationship; rather than dropping the result, the researchers investigated and explained why the observed pattern occurred. The overall message: keep SEM models comprehensive, assess measurement quality for moderators, test moderating effects in the structural model, and report both significant and non-significant findings with context-based interpretation.

Cornell Notes

SEM results are most credible when the analysis keeps the model complex and complete. For covariance-based SEM (e.g., AMOS) and variance-based SEM (e.g., Smart PLS), constructs should be included together in a full measurement model and then estimated as an integrated system, rather than reporting separate models for each construct. Mediation and moderation should be tested within SEM, not by stripping them out and using simpler IV–DV or mediator–DV comparisons. Moderators belong in the measurement model to evaluate reliability and validity, while moderating effects belong in the structural model. Insignificant paths should be reported and interpreted, not removed, because complex models often yield non-significant relationships that still need context-based explanation.

Should researchers put every variable (independent, mediator, moderator, dependent, controls) into a single SEM run, or analyze pieces separately?

The guidance is to include all relevant constructs in the SEM model. Separate construct-by-construct measurement checks can be useful initially, but reporting separate measurement models for each construct doesn’t match SEM’s purpose. For covariance-based SEM (e.g., AMOS), start by checking constructs individually if needed, then build a complete measurement model where all constructs are linked. For Smart PLS, the same principle applies: include mediators, moderators, and other variables in the measurement/structural assessment. Controls can be handled via two models—one without controls and one with controls—so comparisons reflect the role of controls.

Is it appropriate to perform mediation analysis separately from the rest of the SEM workflow?

No. Mediation should be assessed within SEM when SEM is being used for its strength: handling complex models. Removing mediation and then only testing IV–DV relationships (or splitting the model into separate parts) undermines the reason to use SEM tools in the first place. The recommendation is to keep mediators in the model and estimate the full system.

Where do moderators belong: the measurement model or the structural model?

Moderators should be included in the measurement model so their reliability and validity can be assessed like any other latent variable. But the moderating effect—the way the moderator changes an IV-to-DV relationship—should not be treated as a measurement-quality criterion. That moderating effect belongs in the structural model, because it tests interrelationships between constructs.

If a study includes a moderator, should the moderator’s effect on the endogenous variable always be reported?

Only when it fits the study’s objectives and the moderator is a proper antecedent for the relationship being tested. For example, if collaborative culture predicts organizational performance and role ambiguity is hypothesized to change that relationship, then the moderator’s impact on the dependent outcome should be assessed and reported.

What should researchers do with insignificant SEM results—remove them or report them?

Report them. Insignificant paths don’t mean the study is automatically wrong, particularly in complex models where some relationships may show no effect in a specific context. The key is to justify the outcome in the discussion—either explaining why an effect might be absent or providing context for unexpected directions. An example described involves finding a negative effect when a positive one was expected; the researchers investigated and explained the reasons rather than deleting the result.

Review Questions

When would it make sense to run two SEM models (with and without control variables), and what should be compared between them?
How do reliability/validity checks for a moderator differ from testing a moderating effect in SEM?
Why is deleting insignificant paths often discouraged in complex SEM models, and what should replace that deletion in the write-up?

Key Points

1
Include independent variables, mediators, moderators, dependent variables, and relevant constructs together in a full SEM model rather than reporting separate measurement models for each construct.
2
Use two-model comparisons only when control variables are the focus—one model without controls and one with controls—then report results based on the comparison.
3
Test mediation within SEM rather than stripping mediation out and relying on simpler IV–DV or mediator–DV-only models.
4
Place moderators in the measurement model to evaluate reliability and validity, but test moderating effects in the structural model.
5
Report the effect of a moderator on the endogenous variable when it aligns with study objectives and the hypothesized relationship.
6
Do not remove insignificant paths; interpret and justify them in the discussion using context-specific explanations.

Highlights

SEM’s strength is estimating complex systems; removing mediation or moderation and splitting the model into simpler parts defeats that advantage.

Moderators should be measured and validated like other latent constructs, but the moderating effect itself belongs in the structural relationships.

Insignificant results aren’t proof of failure—complex models naturally yield some non-significant paths that still require explanation.

Unexpected directions (like a negative effect when a positive one was expected) should be investigated and reported rather than omitted.

Topics

SEM Model Specification
Mediation in SEM
Moderation vs Moderating Effect
Measurement Model Validity
Handling Insignificant Results

Mentioned

AMOS
Smart PLS
SEM
IV
DV
SCM
SPSS