EFA Before CB-SEM: Analyze Items Together, by Block, or Proceed Directly to CFA?
Independent, mediating and dependent roles belong to the structural model. EFA examines item covariance, so joint or separate analyses should follow measurement questions and the confirmation design.
Distinguish new, adapted and previously validated scales; choose joint or justified blockwise EFA, address method effects, split EFA/CFA samples, transition to CFA, and transparently report any same-sample analysis.
EFA: decisions need several layers of evidence
- Data
Check items and the correlation matrix
- Factor retention
Parallel analysis, scree and theory
- Rotation
Choose for the factor relationships
- Interpret loadings
Review cross-loadings and item meaning
Conceptual diagram. Do not delete items solely to achieve a loading cutoff or better index.
Source: Lê Minh · Conceptual illustration, not software output.1. Ask the right question before analysis
In CB-SEM, independent, mediating and dependent describe latent-variable roles in the structural model. EFA uses item correlations or covariances and does not know which variables will predict others. No statistical rule therefore requires splitting items according to their construct’s position on the diagram.
Ask which items should be examined together to detect structure, cross-loadings and conceptual boundaries. EFA explores; CFA evaluates a prespecified measurement model (Fabrigar et al., 1999; Brown, 2015).
- Specify exploration or confirmation as the objective.
- Classify each scale as new, adapted or previously validated.
- Identify shared wording, sources and response formats.
- Identify constructs sharing a content domain.
- Plan independent CFA data if EFA changes the model.
2. Three situations require different strategies
Scale development accumulates evidence; alpha, EFA and CFA are not a mandatory sequence for every study. DeVellis and Thorpe (2022) emphasize theory, item development and independent validation throughout the process.
| Situation | Main strategy | Avoid |
|---|---|---|
| New scale | EFA followed by CFA on independent data | Finalizing the model from one sample |
| Translated or substantially revised scale | Cognitive interviews, pilot and EFA/CFA according to the changes | Assuming the original structure remains intact |
| Scale validated in the same context | May proceed directly to CFA | Treating EFA as mandatory |
| Several instruments from different domains | Justified blockwise analysis may be appropriate | Splitting by causal role |
3. When is joint EFA useful?
Joint EFA helps when constructs share an instrument or domain, or respondents may not distinguish them. Including items in one matrix reveals cross-loadings, wording factors and unexpected construct content.
If service quality, satisfaction and trust are assessed together, joint EFA may show satisfaction items loading on quality. Separate analyses conceal that evidence. Worthington and Whittaker (2006) treat factor analysis as part of scale development and refinement, linking item decisions with content.
It does not establish discriminant validity by CFA standards; it identifies items that fail to cluster as expected when presented together.
4. When can blockwise analysis be justified?
Separate blocks may suit clearly different domains, administration stages, response formats or independent measurement questions. A clinical symptom scale and a knowledge test, for example, may not need one EFA.
State the rationale before inspecting loadings. Do not split blocks to avoid cross-loadings. If constructs will enter a joint CFA measurement model, assess their boundaries afterward.
| Reason | Acceptable? | Additional evidence |
|---|---|---|
| Different domains and response tasks | Yes | Joint CFA or subsequent relationship assessment |
| Different instruments, administered separately | Conditionally | Describe each block’s source and objective |
| Insufficient sample for all items | Not a default remedy | Redesign sample size |
| Removing cross-loadings from the output | No | Analyze jointly and address content |
5. Wording and method can create artificial factors
Reversed wording, shared stems, item position or rating sources can create clusters unrelated to substantive constructs. DeVellis and Thorpe (2022) warn that similar wording may produce an apparent factor. Check wording and method before naming a factor.
Do not automatically delete every reversed item. Check reverse coding, cognitive interviews and competing models. In CFA, method factors or correlated uniquenesses require an explanation rather than modification-index-driven fit repair.
6. Why can same-sample EFA and CFA be overoptimistic?
EFA uses data to choose factors, items and loadings. CFA on that sample evaluates a model already optimized to its particular noise. Its fit is not independent confirmation.
Anderson and Gerbing (1988) separate measurement and structural assessment in a two-step approach. When EFA changes measurement, confirmation needs new data or a subset not used for selection. Otherwise, describe findings as internal or preliminary.
Renaming EFA as CFA does not create independent data. Decisions based on the exploratory sample need testing outside it.
7. Designing an EFA/CFA sample split
Random splitting is defensible only if both parts remain adequate and important groups are preserved. In stratified samples, split within strata to retain rare groups. Prespecify the seed and splitting rule.
If N cannot support two stable halves, it is often better to use the full sample for one main objective and collect validation data later. Finch et al. (2016) emphasize that one study is not final evidence of construct validity.
| Design | Advantages | Limitations |
|---|---|---|
| Random EFA/CFA split | Fast, internally independent data | Reduced N in each part |
| New CFA sample | Stronger replication evidence | Time and cost |
| Group-based cross-validation | Assesses generalization | Requires sufficient N per group |
| Same unsplit sample | Uses full N | Only preliminary claims are justified |
8. From an EFA solution to a CFA model
Fix the EFA solution
Preserve items, factor membership and all deletion decisions.
Specify CFA
Assign items to intended constructs; do not reproduce incidental cross-loadings without justification.
Choose the estimator
Select ML/MLR or WLSMV according to data type and distribution.
Assess fit
Inspect overall fit, standardized residuals, loadings and errors.
Compare competing models
Consider one-factor, correlated-factor or higher-order models when theory requires them.
Evaluate scores
Assess reliability and validity evidence appropriate to the structure.
Proceed to the structural model
Test paths only after defending measurement.
Preserve versions
Justify all CFA changes and distinguish confirmatory from post hoc decisions.
9. Hypothetical example: three constructs in one questionnaire
These numbers are illustrative. A study has 24 items measuring supervision quality, research confidence and continuation intention. The 520 participants are split by faculty into 260 for EFA and 260 for CFA.
Joint EFA suggests three factors. Two confidence items cross-load with quality because both concern supervisor support. One distinct item is retained for rewriting and one duplicate removed. CFA on the other half compares correlated three-factor and one-factor models. The former fits better; structural paths are estimated only after measurement stabilizes.
| Finding | Decision | Further evidence needed |
|---|---|---|
| Two cross-loadings share support wording | Revise wording and remove the duplicate | CFA on the other sample half |
| Three moderately correlated factors | Retain the oblique solution | Correlated-factor CFA |
| Poor one-factor fit | Do not combine into a total score | Report competing models |
10. When can CB-SEM proceed without EFA?
Proceed directly to CFA when measurement is clearly specified and content, language, response format and target population have not changed substantially. Justify this decision rather than treating CFA as inherently superior.
If CFA fits poorly, do not return to EFA and then CFA on the same data as though independent. Consider ESEM, competing models or explicitly exploratory analysis with new-data confirmation. Brown (2015) treats CFA as hypothesis modeling, not unrestricted searching.
11. Seven common mistakes
| Mistake | Problem | Correction |
|---|---|---|
| Splitting by independent/dependent role | Causal role is not measurement structure | Split according to measurement questions |
| EFA for every validated scale | Unnecessary procedure | State the objective and use CFA where appropriate |
| Splitting to avoid cross-loadings | Conceals distinction problems | Analyze shared-domain constructs jointly |
| Same-sample EFA and CFA | Overoptimistic fit | Use independent data or limit claims |
| Relying only on KMO | Does not assess structure | Use PA, rotation, pattern and content |
| Repeated MI-driven CFA changes | Capitalization on chance | Theoretically justified competing models |
| Estimating structural paths too early | Paths depend on weak measurement | Fix measurement first |
12. APA 7 reporting template and checklist
Reporting example in English
Items were analyzed [jointly/in blocks] because [measurement rationale], not their independent or dependent roles. The [N] sample was [randomly/stratifiably/not] split into EFA n = [ ] and CFA n = [ ]. EFA used [correlations, extraction, rotation, retention method]. The [number]-factor solution led to [item decisions]. CFA on [independent data/the same sample with limitations] tested [model] using [estimator] against [competing model]. Structural analysis followed only after adequate measurement evidence.
- Distinguish measurement from structural roles.
- State prior scale status.
- Justify joint or blockwise analysis.
- Report splitting rules and seed where applicable.
- Report correlations, extraction and rotation.
- Describe factor-count selection.
- Record removed, revised and retained items.
- Assess method effects.
- State the CFA estimator.
- Compare competing models where appropriate.
- Do not call same-sample analysis independent confirmation.
- Fix measurement before structural analysis.
- State limitations if validation data are unavailable.
Conclusion
Do not split EFA by independent or dependent labels. Examine together the items whose boundaries need investigation, splitting only for genuinely different instruments, domains or response tasks.
EFA generates measurement hypotheses; CFA tests them. The process gains strength from independent data and decision records, not merely running every named technique.
References
- Anderson, J. C., & Gerbing, D. W. (1988). Structural equation modeling in practice: A review and recommended two-step approach. Psychological Bulletin, 103(3), 411-423.https://doi.org/10.1037/0033-2909.103.3.411
- Brown, T. A. (2015). Confirmatory factor analysis for applied research (2nd ed.). Guilford Press.https://www.guilford.com/books/Confirmatory-Factor-Analysis-for-Applied-Research/Timothy-Brown/9781462515363
- DeVellis, R. F., & Thorpe, C. T. (2022). Scale development: Theory and applications (5th ed.). SAGE.https://us.sagepub.com/en-us/nam/scale-development/book269114
- Fabrigar, L. R., Wegener, D. T., MacCallum, R. C., & Strahan, E. J. (1999). Evaluating the use of exploratory factor analysis in psychological research. Psychological Methods, 4(3), 272-299.https://doi.org/10.1037/1082-989X.4.3.272
- Finch, W. H., Immekus, J. C., & French, B. F. (2016). Applied psychometrics using SPSS and AMOS. Information Age Publishing.https://www.infoagepub.com/products/Applied-Psychometrics-using-SPSS-and-AMOS
- Worthington, R. L., & Whittaker, T. A. (2006). Scale development research: A content analysis and recommendations for best practices. The Counseling Psychologist, 34(6), 806-838.https://doi.org/10.1177/0011000006288127
Methodological note: If EFA changes factors or items, CFA needs independent data to serve as confirmation. Without a validation sample, moderate the strength of the claims.