Confirmatory Factor Analysis

Confirmatory factor analysis (CFA) is a statistical method used to evaluate whether observed variables represent a theoretically specified structure of underlying latent factors. Unlike exploratory factor analysis (EFA), where the factor structure is investigated with relatively fewer prior restrictions, CFA begins with an explicit measurement model derived from theory, previous research or earlier exploratory analysis.

For example, previous research may propose that customer experience consists of three distinct but related constructs: service quality, perceived value and customer trust. Each construct is measured using several questionnaire items. CFA allows the researcher to specify which items are expected to represent each construct and then evaluate how adequately this proposed measurement structure corresponds to the observed data.

The purpose of CFA is not simply to obtain acceptable CFI, TLI, RMSEA or SRMR values. Model evaluation requires consideration of global fit, parameter estimates, residuals, theoretical coherence and the plausibility of any model modifications. Recent methodological work particularly cautions against treating universal fit-index cutoffs as mechanical pass/fail rules because their behaviour varies with characteristics of the model and data. (PubMed Central (PMC))

On this page:

• Confirmatory factor analysis explained simply
• What CFA is
• CFA vs EFA
• CFA and measurement models
• Latent factors and observed indicators
• Model specification
• Model identification
• Estimation
• Factor loadings
• Evaluating model fit
• Chi-square
• CFI and TLI
• RMSEA and SRMR
• Why good model fit does not prove a model is correct
• Residuals and local model fit
• Modification indices
• Model respecification and overfitting
• CFA, reliability and construct validity
• CFA and structural equation modelling
• Dudovskiy CFA Model Evaluation Framework
• Application example
• Advantages and limitations
• Common mistakes
• CFA in business research
• CFA in the age of AI
• When to use CFA
• Dissertation example
• Exam tip

Methodological question	CFA implication
Is the proposed factor structure specified beforehand?	CFA may be appropriate
Which indicators represent each construct?	Specify these relationships in the measurement model
Can the model parameters be uniquely estimated?	Establish model identification
Does the model reproduce the observed covariance structure adequately?	Examine multiple sources of model-fit evidence
Are indicators meaningfully related to their specified factors?	Examine factor loadings
Where might the model be misspecified?	Examine residuals and diagnostics
Should a modification index automatically change the model?	No; modifications require substantive justification
Do acceptable fit indices prove the model is correct?	No
Is CFA the same as SEM?	No; CFA is part of the broader SEM family

Confirmatory Factor Analysis Explained Simply

Suppose a hotel group uses a questionnaire containing twelve statements. Based on established theory, four items are intended to measure Service Quality, four measure Perceived Value, and four measure Customer Trust.

The researcher is not asking the data to discover how many factors exist from scratch. A three-factor structure has already been proposed:

Service Quality
→ Items 1–4

Perceived Value
→ Items 5–8

Customer Trust
→ Items 9–12

CFA asks whether the observed relationships among the twelve questionnaire items are reasonably consistent with this theoretically specified measurement model.

If the model is supported, the evidence strengthens the argument that the proposed measurement structure provides a plausible representation of the data. If it fits poorly, the researcher needs to investigate why. Perhaps an item does not represent its intended factor well, two constructs are less distinct than expected, an omitted relationship matters, or the theoretical model itself requires reconsideration.

The key point is that CFA starts with a model and evaluates it against the data.

Dudovskiy Research Assistant

Not sure if confirmatory factor analysis is the correct methodology for your dissertation?

Enter your topic to receive a free, tailored methodology preview.

Dissertation topic

Generate Free Methodology Preview → Free · No account required

Learn more about Dudovskiy Research Assistant

What Is Confirmatory Factor Analysis?

Confirmatory factor analysis is a latent-variable modelling technique in which researchers specify hypothesized relationships between observed indicators and underlying factors and evaluate that measurement model using empirical data.

The factors represent constructs that cannot be observed directly. Customer trust, organizational commitment and perceived usefulness, for example, are theoretical concepts rather than directly measurable quantities. Researchers instead observe responses to questionnaire items intended to represent those constructs.

CFA formalizes this relationship. The researcher specifies which observed variables are indicators of which latent factors, whether factors may correlate and other theoretically relevant relationships. The model then implies a particular covariance structure among the observed variables, which can be compared with the covariance structure found in the data.

CFA is therefore fundamentally model-based and theory-driven. Brown’s widely used treatment of CFA emphasizes both its conceptual and applied role in latent-variable modelling rather than reducing it to formulas or software procedures. (Guilford Press)

Confirmatory Factor Analysis vs Exploratory Factor Analysis

CFA and EFA address related but different methodological questions.

Exploratory Factor Analysis (EFA)	Confirmatory Factor Analysis (CFA)
Explores an uncertain latent structure	Evaluates a specified latent structure
Factor structure is not completely specified beforehand	Measurement model is specified beforehand
Helps identify possible factors	Tests a theoretically proposed factor configuration
Often used in scale development	Often used in scale validation
Requires decisions about extraction and rotation	Requires decisions about model specification, identification, estimation and evaluation
Generates a plausible factor structure	Evaluates a hypothesized factor structure

Suppose a researcher develops 20 items to investigate consumers’ perceptions of autonomous retail technology but does not know their underlying dimensional structure. EFA may be appropriate for exploring that structure.

If EFA and theory subsequently suggest four factors—technological trust, privacy concern, perceived convenience and perceived risk—the researcher could specify this four-factor measurement model and evaluate it with CFA, ideally using independent data.

The distinction is therefore:

EFA → What factor structure might explain these relationships?

CFA → How adequately does this specified factor structure represent these relationships?

This does not mean CFA eliminates researcher judgment. It relocates much of that judgment to model specification, evaluation and theoretically defensible interpretation.

CFA and the Measurement Model

A measurement model describes how observed indicators relate to latent constructs.

Consider a study examining employee attitudes toward hybrid working. The theoretical model contains three constructs:

Work Flexibility
→ Flexibility Item 1
→ Flexibility Item 2
→ Flexibility Item 3

Digital Collaboration
→ Collaboration Item 1
→ Collaboration Item 2
→ Collaboration Item 3

Organizational Support
→ Support Item 1
→ Support Item 2
→ Support Item 3

CFA allows this proposed structure to be represented statistically. The model specifies that the three flexibility items indicate Work Flexibility, the collaboration items indicate Digital Collaboration, and the support items indicate Organizational Support.

The latent factors may also be permitted to correlate when theory suggests relationships among the constructs.

The objective is not merely to draw three circles connected to nine rectangles. The measurement model represents substantive claims about what is being measured and how observed responses relate to theoretical constructs.

Latent Factors and Observed Indicators

An observed indicator is a directly measured variable, such as a participant’s response to a questionnaire statement.

A latent factor is an unobserved construct inferred through its relationships with those indicators.

Suppose the latent factor is Brand Trust. Indicators might include responses to statements such as:

• “I consider this brand dependable.”
• “This brand keeps its promises.”
• “I trust information provided by this brand.”
• “I feel confident purchasing from this brand.”

CFA does not directly observe Brand Trust. It evaluates a model in which the covariance among these observed responses is partly represented through their relationship with the latent construct.

This distinction matters because CFA is not simply a sophisticated method for grouping questionnaire items. It evaluates a theoretical measurement relationship between indicators and constructs.

Model Specification

Before CFA is estimated, the researcher specifies the proposed measurement model.

Specification includes decisions about which indicators load on which factors, which loadings or other parameters are constrained, whether latent factors can correlate, and whether any additional relationships are theoretically justified.

These decisions should normally follow from substantive theory, established measurement instruments, previous research or clearly documented exploratory work. If relationships are added simply after inspecting the current dataset, the analysis begins to move away from strict confirmation and toward exploration.

Suppose previous research proposes that four items measure Customer Trust and another four measure Customer Satisfaction. The researcher should specify that structure because there is a theoretical reason for doing so—not because trying several alternative configurations eventually produced attractive fit statistics.

CFA therefore starts before the software runs. The quality of the analysis depends partly on the quality of the model being proposed.

Model Identification

A CFA model must be identified before its parameters can be uniquely estimated.

Identification concerns whether the information available from the observed data is sufficient to estimate the unknown parameters in the specified model. A model that is not identified cannot produce unique parameter estimates regardless of how theoretically appealing it appears.

Researchers commonly establish a scale for each latent variable by fixing one factor loading or fixing the factor variance, while the broader identification of the model depends on its complete structure.

Identification can appear highly technical, but its methodological meaning is straightforward:

The model must contain enough information and appropriate constraints for the proposed parameters to be estimable.

Statistical software may flag obvious identification failures, but researchers should understand why their model is identified rather than treating the absence of an error message as methodological justification.

Estimation in CFA

After specification and identification, the model parameters are estimated from the observed data.

Maximum likelihood (ML) and variants of maximum likelihood are widely used, but the appropriate estimator depends on characteristics of the indicators and data. Questionnaire responses based on ordered categories, substantial non-normality, missing data and other features can influence the choice of estimator.

This matters because the estimator affects standard errors, test statistics and model-fit calculations. A method appropriate for approximately continuous, normally distributed indicators cannot automatically be assumed appropriate for every Likert-scale dataset.

The methodology chapter should therefore identify the estimator used and, where relevant, explain why it was suitable for the measurement level and distributional characteristics of the data.

Factor Loadings in CFA

Factor loadings represent the relationships between observed indicators and their specified latent factors.

Suppose an item intended to measure Customer Trust has a standardized loading of 0.82 on the Customer Trust factor. This indicates a strong relationship between the indicator and the latent construct within the specified model.

Another item might load only 0.32. That weaker relationship deserves investigation.

Researchers often use conventional loading thresholds to guide interpretation, but no numerical threshold should replace substantive reasoning. The importance of a loading depends on the measurement context, sample, model and theoretical role of the indicator.

A weak item should not automatically be deleted merely because removing it improves the model. Researchers should consider whether the item is poorly worded, whether it captures an important but underrepresented part of the construct, whether the model is misspecified, and whether deletion would narrow the conceptual meaning of the scale.

Evaluating CFA Model Fit

Model fit concerns how adequately the covariance structure implied by the specified CFA model corresponds to the covariance structure observed in the data.

Researchers commonly examine several types of evidence rather than relying on one statistic. Frequently reported measures include:

• chi-square test of model fit;
• Comparative Fit Index (CFI);
• Tucker-Lewis Index (TLI);
• Root Mean Square Error of Approximation (RMSEA); and
• Standardized Root Mean Square Residual (SRMR).

These measures evaluate model fit from different perspectives. RMSEA and SRMR, for example, quantify forms of discrepancy or approximation error, while CFI and TLI evaluate the proposed model relative to a baseline model. (PubMed Central (PMC))

Using multiple indices does not mean collecting enough statistics until one supports the preferred conclusion. The aim is to develop a coherent assessment of where and how the model fits or fails to fit.

Chi-Square Test

The chi-square test evaluates the hypothesis of exact fit between the model-implied and population covariance structures.

A statistically significant chi-square indicates evidence against exact model fit. In applied research, however, exact fit is often an extremely demanding standard, and the chi-square statistic is sensitive to factors including sample size and the magnitude of model discrepancies.

Consequently, researchers commonly supplement the chi-square test with approximate fit indices rather than treating statistical significance alone as the final verdict.

This does not mean the chi-square result should simply be ignored whenever it is inconvenient. Instead, it should be interpreted as one part of a broader model-evaluation process.

CFI and TLI

The Comparative Fit Index (CFI) and Tucker-Lewis Index (TLI) assess the specified model relative to a baseline model, typically one representing very limited relationships among the observed variables.

Higher values generally indicate better relative fit.

Researchers frequently encounter guidelines such as 0.90 or 0.95 when interpreting these indices. Such values can provide orientation, but they should not be treated as universal laws. The performance of fit indices depends on factors including factor loadings, sample size, model complexity, number of indicators and estimation conditions. (PubMed Central (PMC))

The methodological question should therefore not be reduced to:

“Is CFI above 0.95?”

A stronger question is:

“Taken together with the model, data, other fit evidence and parameter estimates, what does the CFI tell us about the adequacy of this measurement model?”

RMSEA and SRMR

The Root Mean Square Error of Approximation (RMSEA) evaluates approximate model fit while accounting for model complexity through degrees of freedom. Lower values generally indicate less approximation error.

The Standardized Root Mean Square Residual (SRMR) summarizes standardized differences between observed and model-implied relationships. Lower values generally indicate smaller residual discrepancies.

As with CFI and TLI, familiar thresholds such as RMSEA below 0.06 or 0.08 and SRMR below 0.08 are often encountered in applied research. These values should be understood as guidelines developed under particular methodological conditions rather than universal boundaries separating valid from invalid models.

A major contemporary review of CFA practice emphasizes that fixed cutoff rules can behave differently according to data and model characteristics and warns against overgeneralizing thresholds derived from relatively narrow simulation conditions. (PubMed Central (PMC))

Why Good Model Fit Does Not Prove a Model Is Correct

This is one of the most important principles in CFA.

Suppose a three-factor model produces apparently excellent CFI, TLI, RMSEA and SRMR values. It may be tempting to conclude:

“The CFA proves that the three-factor model is correct.”

That conclusion goes beyond the evidence.

Fit indices evaluate aspects of correspondence between the specified model and observed data. They do not establish that the model is the only possible explanation, that every theoretical assumption is correct, or that the latent constructs possess unquestionable validity.

Different models can sometimes provide plausible representations of the same covariance structure. A model may also show attractive global fit while containing problematic local relationships or theoretically questionable parameters. Research on CFA evaluation explicitly distinguishes quantifying model misfit from demonstrating that a proposed model is the true population model. (PubMed Central (PMC))

The defensible conclusion is therefore closer to:

The evidence indicates that the specified measurement model provides an adequate or plausible representation of the observed relationships under the conditions examined.

That is a stronger methodological statement precisely because it does not claim more than CFA can establish.

Residuals and Local Model Fit

Global fit indices summarize model performance, but they can conceal specific areas of misspecification.

Researchers can therefore examine residuals, which represent discrepancies between observed relationships and those implied by the model. Large residuals can indicate that the specified measurement model fails to reproduce particular relationships adequately.

Suppose global fit appears acceptable, but two questionnaire items have a much stronger empirical relationship than the model predicts. This may suggest overlapping wording, omitted relationships, method effects or a conceptual feature not represented by the specified model.

Recent methodological research highlights this distinction between global fit and local problems: models can display good overall fit while individual parameter patterns remain difficult to interpret. (PubMed Central (PMC))

Global model fit should therefore not prevent researchers from asking where the model fits poorly.

Modification Indices

A modification index (MI) estimates how much model fit could improve if a currently constrained parameter were freely estimated.

For example, a modification index might suggest allowing an indicator to load on another factor or permitting residuals of two indicators to correlate.

Modification indices can be diagnostically useful. They can help identify relationships that the specified model may have omitted. However, they do not explain why the relationship exists or whether adding it makes theoretical sense. Methodological guidance therefore recommends interpreting them together with substantive reasoning rather than automatically following the largest value. (PubMed Central (PMC))

A large modification index should trigger the question:

“Why might this relationship exist?”

not:

“What should I add to make my CFI higher?”

Model Respecification and Overfitting

A poorly fitting CFA model sometimes needs reconsideration. An indicator may have been assigned to the wrong construct, two items may share wording that creates additional covariance, or the theoretical measurement structure may simply not characterize the sample adequately.

Model respecification can therefore be legitimate when supported by substantive reasoning.

The danger arises when researchers repeatedly inspect modification indices, free parameters and delete items until the model achieves desired fit statistics. Because each modification responds to patterns in the same sample, the final model can become increasingly tailored to sample-specific noise.

Research on CFA explicitly warns that post-hoc specification searches shift analysis toward exploration and create a risk of overfitting. Modified models should ideally be evaluated using independent data. (PubMed Central (PMC))

This creates an important methodological boundary:

Theory-guided diagnosis and respecification ≠ unrestricted fit optimization.

If extensive data-driven modifications become necessary, researchers should acknowledge that the analysis has become more exploratory rather than continuing to describe it as purely confirmatory.

CFA, Reliability and Construct Validity

CFA, reliability and validity address related but distinct measurement questions.

Reliability coefficients such as Cronbach’s alpha and McDonald’s omega concern consistency or reliability of scale scores under particular assumptions. CFA evaluates a specified measurement structure.

A model with good CFA fit does not automatically demonstrate that scores are sufficiently reliable. Likewise, a scale with high reliability does not establish that the intended factor structure is correct.

CFA can contribute important evidence relevant to construct validity, including evidence about factor structure, relationships among constructs and patterns of indicator loadings. But construct validity is broader than model fit. It accumulates from theoretical reasoning and multiple forms of empirical evidence rather than being proven by one CFA.

The measurement argument is therefore better represented as:

Construct definition → Measurement model → Factor structure → Reliability evidence → Validity evidence

rather than:

Good CFA fit = valid scale.

Convergent and Discriminant Validity in CFA

CFA is frequently used as part of evaluating whether indicators behave consistently with theoretical expectations about related and distinct constructs.

Convergent validity concerns whether indicators intended to represent the same construct provide evidence of convergence. Factor loadings and related measurement evidence can contribute to this assessment.

Discriminant validity concerns whether theoretically distinct constructs are empirically distinguishable. Extremely high correlations between latent factors, for example, may raise questions about whether two proposed constructs are sufficiently distinct.

Neither should be reduced to a single automatic statistic. The broader question is whether the pattern of evidence supports the proposed interpretation of the constructs.

CFA therefore provides an important framework for investigating validity evidence, but the researcher remains responsible for connecting statistical findings to the theoretical meaning of the measures.

CFA and Structural Equation Modelling

CFA belongs to the broader family of structural equation modelling (SEM) techniques.

CFA primarily concerns the measurement model: how observed indicators relate to latent constructs and how those constructs relate to one another within the measurement structure.

SEM can extend this framework by specifying structural relationships among latent variables. For example, after establishing a defensible measurement model for Service Quality, Customer Satisfaction and Customer Loyalty, the researcher might investigate whether:

Service Quality → Customer Satisfaction → Customer Loyalty

CFA therefore frequently precedes analysis of a structural model.

The distinction can be summarized as:

CFA → Are the latent constructs being represented adequately by their indicators?

Structural model → How are those constructs hypothesized to relate to one another?

This makes CFA an important methodological bridge between measurement development and full structural equation modelling.

Dudovskiy CFA Model Evaluation Framework

The Dudovskiy CFA Model Evaluation Framework synthesizes established CFA principles into a practical sequence for evaluating a theoretically specified measurement model. It does not propose a new statistical form of CFA. Its purpose is to prevent model evaluation from becoming a mechanical exercise in achieving preferred fit-index thresholds.

Theoretical Construct → Measurement Model → Model Identification → Estimation → Global Model Fit → Parameter & Residual Diagnostics → Theory-Guided Respecification → Defensible Measurement Model

The central principle is:

A CFA model should be judged through converging theoretical and statistical evidence—not by whether one fit index crosses a preferred cutoff.

1. Theoretical Construct

Begin with the construct rather than the statistical model. Researchers should define what each latent variable represents and explain why the selected indicators are expected to measure it.

Without this theoretical foundation, even an excellent-fitting model can be difficult to interpret meaningfully.

2. Measurement Model

Translate the theoretical argument into a specified measurement structure. Identify which indicators represent each factor and which relationships among factors are theoretically permitted.

This specification should precede examination of model results if the analysis is genuinely confirmatory.

3. Model Identification

Establish that the specified parameters can be uniquely estimated. Appropriate scaling and sufficient information are required before meaningful parameter estimation can occur.

Identification is therefore a prerequisite rather than another fit statistic.

4. Estimation

Select an estimation approach appropriate to the characteristics of the indicators and data. The estimator should correspond to measurement level, distributional properties and other relevant conditions.

Researchers should know what estimator produced their CFA results and why it was appropriate.

5. Global Model Fit

Evaluate multiple forms of global fit evidence, potentially including chi-square, CFI, TLI, RMSEA and SRMR.

These measures should be interpreted collectively and contextually rather than converted into a simplistic pass/fail checklist. Contemporary evidence shows that fit-index performance depends on features of the measurement model and data. (PubMed Central (PMC))

6. Parameter and Residual Diagnostics

Examine factor loadings, factor correlations, residuals and other local evidence.

This stage asks whether apparently acceptable overall fit conceals problematic indicators, unexpected relationships or local areas of misspecification.

7. Theory-Guided Respecification

If problems emerge, determine whether any modification has a substantive theoretical justification. Modification indices can help diagnose possible misspecification, but they should not dictate model changes.

Extensive data-driven modification changes the character of the analysis and increases the danger of overfitting.

8. Defensible Measurement Model

The final model should combine adequate statistical evidence with theoretical coherence and transparent reporting.

The researcher should be able to explain not only whether the model fits, but why the model was specified in that form, what evidence supports it, what problems were investigated and what modifications were made.

The framework therefore changes the objective from:

“How can I make my CFA model fit?”

to:

“What combination of theoretical and statistical evidence makes this measurement model defensible?”

Application of Confirmatory Factor Analysis: an Example

Consider a researcher investigating customer trust in digital financial services. Previous literature proposes that trust consists of three related dimensions: institutional trust, technological trust and transaction security. An established questionnaire provides four indicators for each dimension.

Unlike a researcher conducting EFA, this researcher already has a theoretically specified three-factor measurement model. CFA is therefore used to evaluate whether the proposed structure provides an adequate representation of responses collected from users of digital financial services.

The researcher first specifies the twelve indicators according to their theoretically intended factors and permits the three latent factors to correlate. After establishing model identification and selecting an estimator appropriate to the data, the researcher evaluates global fit using several indices rather than one preferred statistic.

The initial model shows reasonable overall fit but reveals a large residual involving two similarly worded technological-trust items. A modification index also suggests correlated residuals between them. Rather than immediately freeing the covariance, the researcher examines the wording and recognizes that both items explicitly refer to system failure.

There is therefore a substantive explanation for their additional shared variance. If the residual covariance is added, the modification is transparently reported as theoretically justified rather than presented as part of the original model.

The revised model demonstrates adequate global fit, interpretable loadings and plausible factor correlations. The researcher nevertheless describes it as a supported measurement model rather than claiming that CFA has proven the three-factor conceptualization of trust to be uniquely correct.

Advantages and Limitations of Confirmatory Factor Analysis

CFA allows researchers to translate theoretical measurement propositions into explicit statistical models. This is particularly valuable when established theory predicts that particular indicators should represent particular latent constructs. Instead of merely observing correlations among questionnaire items, researchers can evaluate a coherent model connecting observed measurements to theoretical concepts.

The method also provides detailed diagnostic information. Researchers can examine factor loadings, relationships among latent variables, residuals and multiple dimensions of global model fit. This makes CFA useful not only for evaluating whether a measurement model performs adequately but also for understanding where it may fail.

Its confirmatory appearance can nevertheless create excessive confidence. A model with attractive fit indices is not necessarily the true measurement model, and fixed cutoff values do not operate identically across all models and datasets. Recent methodological work shows that characteristics such as sample size, factor loadings, number of indicators, missing data and model complexity can affect fit indices and their ability to detect misspecification. (PubMed Central (PMC))

CFA is also vulnerable to post-hoc optimization. Researchers can often improve fit by correlating residuals, freeing cross-loadings or deleting troublesome indicators. Each change may appear statistically beneficial while gradually moving the analysis away from the theory it was intended to test. The resulting model may fit one sample exceptionally well yet replicate poorly elsewhere.

The strength of CFA therefore comes from the combination of explicit theory and empirical evaluation. Removing either component weakens the methodological argument.

Common Mistakes When Using Confirmatory Factor Analysis

Treating fit-index thresholds as universal acceptance criteria is one of the most consequential errors. Statements such as “CFI exceeded 0.95, therefore the model was validated” compress a complex model-evaluation problem into a single number. Fit indices are useful evidence, but their behaviour depends on characteristics of the model and data. (PubMed Central (PMC))

A related problem is ignoring local evidence once global fit appears acceptable. Strong overall CFI or RMSEA values can coexist with weak factor loadings, problematic residuals or theoretically implausible parameter estimates. A model should therefore be examined both globally and locally.

Modification indices can become particularly seductive. Following the largest modification index, rerunning the model and repeating the process will almost inevitably improve fit, but it can transform confirmatory analysis into sample-specific model searching. Methodological guidance warns that such changes should be theoretically justified and, where possible, tested with independent data. (PubMed Central (PMC))

Deleting indicators merely because they reduce fit is similarly problematic. An item may represent an essential part of a construct’s conceptual domain. Improving statistical fit at the cost of construct coverage can produce a cleaner but substantively poorer measurement model.

Finally, CFA should not be described as proving validity. It can provide important evidence concerning measurement structure and construct relationships, but validity requires a broader theoretical and empirical argument.

Confirmatory Factor Analysis in Business Research

CFA is particularly valuable in business research because many important variables are latent constructs measured through multi-item scales. Brand trust, customer satisfaction, perceived value, organizational commitment, entrepreneurial orientation and technology acceptance cannot usually be observed directly.

Suppose a researcher studies the relationship between perceived service quality, customer satisfaction and loyalty intentions. Before testing structural relationships among these constructs, the researcher needs confidence that the questionnaire indicators provide a defensible representation of the proposed latent variables.

CFA can evaluate that measurement model before the researcher interprets relationships among the constructs themselves.

This is especially important when established scales are applied in new industries, populations or cultural settings. Previous validation provides valuable evidence, but it does not guarantee that a measurement structure will behave identically in every new context.

CFA therefore helps business researchers connect abstract theoretical constructs with the empirical indicators used to represent them.

Confirmatory Factor Analysis in the Age of AI and Digital Research

Generative AI substantially reduces the technical barrier to CFA. Researchers can ask AI systems to produce R, Python or other statistical code, explain fit indices, interpret factor loadings and diagnose software errors. This can make sophisticated latent-variable analysis accessible to students who previously found the software intimidating.

The danger is that CFA output is unusually easy to make appear authoritative. An AI system can confidently state that a model has “excellent fit” because several conventional thresholds are satisfied without examining whether those thresholds are appropriate for the particular model or whether local evidence reveals substantial problems. Contemporary CFA research provides good reason to avoid this mechanical interpretation of fixed cutoffs. (PubMed Central (PMC))

AI can also accelerate modification-index searching. A researcher can provide model output and ask an AI system how to improve fit, receiving suggestions to correlate residuals, remove indicators or free parameters. Technically, those suggestions may improve the numerical results. Methodologically, repeated data-driven modifications can turn CFA into an exploratory optimization exercise and increase overfitting risk.

The most valuable role for AI is therefore not to make the model “pass.” It is to help the researcher interrogate the model: explain why a particular residual might be large, compare alternative interpretations, check syntax, identify assumptions and challenge weak theoretical justifications.

CFA remains confirmatory only to the extent that theory constrains what the researcher is willing to do after seeing the data.

When to Use Confirmatory Factor Analysis

CFA may be appropriate when:

• theory or previous research specifies a hypothesized factor structure;
• an established measurement scale needs to be evaluated in a new sample or context;
• an exploratory factor structure is ready for independent confirmatory evaluation;
• researchers need to test whether particular indicators represent specified latent constructs;
• competing theoretically plausible measurement models need to be evaluated;
• measurement quality needs to be established before testing structural relationships among latent variables;
• researchers need evidence concerning relationships among constructs as part of a broader validity argument;
• the sample and data are adequate for estimation of the proposed latent-variable model.

CFA is less appropriate when the researcher genuinely does not know the underlying factor structure and needs the data to reveal plausible dimensions. EFA is generally more consistent with that objective.

Dissertation Example

A dissertation titled “The Impact of Sustainable Brand Trust on Customer Loyalty in the Fashion Industry” measures three latent constructs: sustainable brand credibility, sustainable brand trust and customer loyalty. Each construct is operationalized through multiple questionnaire items adapted from previously validated scales.

Because the measurement structure is theoretically specified before analysis, the methodology chapter justifies confirmatory factor analysis rather than EFA as the primary method for evaluating the measurement model. The researcher specifies which indicators represent each construct, explains the estimator used and evaluates global model fit using chi-square together with CFI, TLI, RMSEA and SRMR rather than relying on a single statistic.

Factor loadings, latent-factor correlations and residual evidence are subsequently examined to assess the model locally. One modification index suggests correlating the residuals of two similarly worded trust items. The researcher does not automatically make the change; the shared wording and theoretical rationale are considered before any respecification, and the decision is reported transparently.

The final measurement model shows adequate overall fit and theoretically interpretable parameters. The dissertation describes this as evidence supporting the proposed measurement structure rather than claiming that CFA proves construct validity. Reliability evidence is considered separately, and only after the measurement model has been evaluated does the researcher proceed to examine the hypothesized structural relationships among sustainable brand credibility, trust and loyalty.

Exam Tip

If asked to distinguish CFA from EFA, do not say merely that “EFA explores and CFA confirms.” That is a useful starting point but an incomplete methodological explanation.

A stronger answer states that CFA begins with a theoretically specified measurement model in which relationships between observed indicators and latent factors are defined in advance. The researcher then evaluates how adequately that model represents the observed covariance structure.

Remember the logic:

Theory → Specify → Identify → Estimate → Evaluate → Diagnose → Respecify only when justified → Defend

Also remember that CFA is not a contest to achieve CFI above 0.95 or RMSEA below a particular number. Fixed cutoff values can be useful guidelines, but methodological research shows why they should not be treated as universal pass/fail criteria. (PubMed Central (PMC))

The principle to remember is:

Good fit supports the plausibility of a specified model; it does not prove that the model is uniquely correct.

Build a methodology you can explain and defend

Confirmatory factor analysis requires more than reporting a set of fit indices. Dudovskiy Research Assistant can help you determine whether CFA fits your dissertation, structure your measurement-model evaluation and explain the methodological decisions you need to justify.

References

Brown, T.A. (2015). Confirmatory Factor Analysis for Applied Research. 2nd ed. New York: Guilford Press. (Guilford Press)

Goretzko, D., Siemund, K. and Sterner, P. (2024). Evaluating Model Fit of Measurement Models in Confirmatory Factor Analysis. Educational and Psychological Measurement, 84(1), 123–144. (PubMed Central (PMC))

Hu, L. and Bentler, P.M. (1999). Cutoff criteria for fit indexes in covariance structure analysis: Conventional criteria versus new alternatives. Structural Equation Modeling, 6(1), 1–55.

MacCallum, R.C., Roznowski, M. and Necowitz, L.B. (1992). Model modifications in covariance structure analysis: The problem of capitalization on chance. Psychological Bulletin, 111(3), 490–504.

Brown’s text remains a major applied reference on CFA and explicitly covers related developments including measurement invariance, bifactor analysis and exploratory structural equation modelling. (Guilford Press)