ACADEMIC WRITING GUIDE

How to Conduct and Report Confirmatory Factor Analysis in a Dissertation (With Examples)

By Acadelyra Editorial TeamPublished

Confirmatory factor analysis (CFA) helps you evaluate whether a measurement structure specified before analysis is consistent with the relationships among your observed indicators. In a dissertation, that means explaining why particular questions represent particular constructs, estimating the model appropriately, and showing what the results support and leave uncertain.

A convincing CFA section does more than display a diagram and announce that the fit indices are acceptable. It connects theory, item content, data quality, estimation, model diagnostics and interpretation. A model can fit reasonably while some indicators remain weak or important validity questions remain unanswered.

This guide explains how to conduct and report confirmatory factor analysis in a dissertation, using one coherent academic-support example. The worked example uses a deliberately constructed synthetic covariance matrix and an illustrative sample size. It contains no observations or findings from real participants. Its numerical calculations are reproducible; its reporting paragraphs are teaching models that must be adapted to analyses you actually perform.

When should you use CFA rather than EFA?

Use CFA when you have a defensible measurement hypothesis before examining the confirmatory results. That hypothesis might come from an established instrument, a theoretical account or a separate exploratory study. Specify the number of factors, the indicators assigned to each factor and the relationships you allow between factors and residuals.

Exploratory factor analysis (EFA) addresses a different question: what common-factor structure might describe this item set? Our guide to conducting and reporting EFA in a dissertation explains factor retention, extraction and rotation. CFA starts with an explicit structure to evaluate; rotation and parallel analysis are not its central decision steps.

Choosing between exploratory and confirmatory factor analysis
DecisionEFACFA
Starting questionWhat common dimensions might underlie these items?Does this specified measurement model adequately represent these items?
Item assignmentsInvestigated through the loading patternSpecified as free or constrained parameters before estimation
Main decisionsExtraction, retention, rotation and interpretationSpecification, identification, estimator and evaluation of restrictions
Evidence producedA candidate structure needing further evaluationEvidence about a particular model in a particular sample
Main reporting riskCalling an exploratory pattern definitive validationCalling a model confirmed after repeatedly changing it to fit the same data

Running EFA and then CFA on the same observations does not provide independent confirmation. If you can collect another suitable sample, use it to evaluate the proposed structure. A planned split sample can also help, provided each portion is adequate for its analysis. Where neither is feasible, disclose the reuse of data and describe the analysis as provisional model development.

An instrument validated elsewhere still needs scrutiny in your setting. A new language, population or administration mode can alter how items function. Conversely, CFA is not automatically necessary for every dissertation: the analysis must answer a measurement question relevant to your design.

Understand the measurement model before opening software

A latent factor is an unobserved construct represented statistically through observed indicators. In a reflective CFA, the factor accounts for covariance among its indicators. The indicators also contain residual variation, which can include item-specific influences and measurement error.

For a continuous indicator, the basic relationship is:

indicator = intercept + loading × factor + residual

The loading describes how the indicator relates to the factor. The residual variance describes variation left unexplained by the specified factor structure. A correlation between factors describes their association; it does not establish that one causes the other.

Do not assume every set of questionnaire questions is reflective. If indicators jointly define an index, such as distinct components of material disadvantage, treating them as interchangeable effects of one common factor may misrepresent the construct. Establish the measurement rationale before choosing CFA.

Specify the academic-support example

Suppose theory distinguishes Support access from Help-seeking comfort. The proposed model has two correlated factors with three indicators each. The descriptions below give the intended content; they are not a validated questionnaire ready for field use. For the numerical demonstration, the indicators are treated as continuous scores.

The prespecified two-factor teaching model
IndicatorIntended contentAssigned factor
A1Knowing where academic help is availableSupport access
A2Identifying the appropriate support serviceSupport access
A3Finding learning resources when neededSupport access
C1Feeling comfortable requesting helpHelp-seeking comfort
C2Discussing academic difficulties openlyHelp-seeking comfort
C3Approaching a lecturer about a problemHelp-seeking comfort

Each indicator loads on its assigned factor. Cross-loadings are fixed to zero, residual covariances are fixed to zero, and the factor correlation is freely estimated. These restrictions are substantive assumptions. For example, fixing residual covariances to zero assumes that the model leaves no additional association between indicators after accounting for their factors.

Write down those choices, plausible alternatives and the rules for handling problematic indicators in your dissertation data analysis plan. Recording them before inspecting fit makes later departures easier to explain honestly.

Check assumptions and data quality

Begin with the actual analytic sample and item-level data. Confirm coding, reverse scoring, response categories, eligibility rules and missing-value codes. Inspect distributions, unusual response patterns, influential observations, very low item variability and near-duplicate indicators. An apparently poor loading can reflect a coding error rather than a weak construct.

Conventional single-level CFA assumes independent observations. Students nested within classes, repeated responses from the same person or a complex survey design may require an estimator, standard-error correction or multilevel model that reflects that dependence. A robust estimator for nonnormality alone does not automatically resolve clustering.

For continuous indicators, consider whether linear factor relationships and the chosen distributional assumptions are reasonable. Normal-theory maximum likelihood relies on assumptions about the joint distribution for its usual tests and standard errors. For ordinal indicators, investigate category frequencies and sparse response combinations; the underlying-response model has its own assumptions.

Check whether the covariance or correlation matrix is suitable for estimation. A non-positive-definite matrix, redundant variables or a nearly singular information matrix requires investigation. KMO and Bartlett's test, often discussed for EFA, do not establish that a particular CFA model is identified or correct.

Missing-data treatment belongs in the methods section. Explain the pattern, plausible mechanism and procedure used, rather than allowing software defaults to decide silently. Our guide to handling missing survey data explains the assumptions behind common approaches. Mean substitution can distort covariances and should not be used as an automatic preparation step.

Justify sample size for this model

There is no universal minimum that makes every CFA trustworthy. Necessary sample size depends on the number of factors and indicators, loading strength, estimator, missingness, category distributions and the precision needed for the intended interpretation.

Muthén and Muthén's Monte Carlo demonstration explains why simple participant-to-item rules can mislead. Prefer a model-specific simulation or another defensible planning approach where feasible, examining convergence, parameter recovery and uncertainty as well as statistical power. Our dissertation power-analysis guide introduces planning principles; its two-group formulas do not calculate a CFA sample requirement.

The illustrative N = 300 used below makes the calculations concrete. It is not a recommendation for your model and is not evidence that the teaching questionnaire would work in a real sample.

Establish model identification and factor scaling

Identification concerns whether the model's free parameters can be uniquely recovered under its restrictions. A software run that terminates is not, by itself, sufficient evidence of identification or a trustworthy solution.

Because a factor has no directly observed unit, its scale must be fixed. Common approaches fix one marker loading or fix the factor variance. The example fixes each factor variance to one while estimating all assigned loadings. This sets the latent scale; it does not force every observed indicator to have unit variance in a real dataset.

For a covariance-only continuous model with six indicators, the distinct observed variances and covariances total 6 × 7 / 2 = 21. The example estimates 6 loadings + 6 residual variances + 1 factor correlation = 13 free parameters. Its model degrees of freedom are 21 − 13 = 8. Factor variances are fixed, and all unassigned loadings and residual covariances remain fixed at zero.

Positive degrees of freedom are a necessary counting check for this overidentified model, not a general proof of identification. The arrangement of indicators, restrictions and parameter values also matters. Three informative indicators per factor are a useful starting design, but fewer can sometimes work with justified constraints; simply counting indicators cannot settle every case. UCLA's CFA seminar explains scaling and parameter counting.

Adding a mean structure or modelling ordinal thresholds changes the bookkeeping. A just-identified model cannot be meaningfully assessed through global fit because it has no degrees of freedom to test its covariance restrictions. Do not interpret perfect-looking fit from such a model as strong measurement evidence.

Select an estimator that matches the indicators

Choose the estimator from the measurement scales, distributional properties, sampling design and missing-data plan. State the exact estimator and software version in a completed dissertation.

Estimator choices and reporting implications
SituationPossible approachWhat to explain
Approximately continuous indicators under suitable distributional assumptionsMaximum likelihood, MLLikelihood convention, assumptions and missing-data procedure
Continuous indicators with problematic nonnormalityRobust maximum likelihood, such as MLRRobust standard errors and the specific scaled or robust test and indices reported
Binary or ordered categorical indicatorsA categorical estimator such as WLSMVWhich variables are ordered, threshold modelling and treatment of sparse categories
Incomplete continuous observationsML with an appropriate full-information missing-data methodThe missingness assumptions and implementation used
Clustered or repeated observationsA method accommodating dependenceCluster definition, model level and standard-error approach

The lavaan estimator documentation distinguishes ML variants and their handling of incomplete continuous data. Its categorical-data documentation explains that declaring ordered outcomes selects WLSMV, using diagonal weights for estimation and fuller information for robust inference. It also states that full-information ML is not currently supported for those ordered outcomes.

A numbered Likert response is still ordinal. Treating it as approximately continuous needs a reasoned justification; choosing WLSMV also requires adequate category information and an appropriate underlying-response model. Neither approach guarantees reliable estimates from unsuitable data.

Do not combine an ordinary chi-square statistic with unrelated robust fit indices and present them as one undifferentiated output. Report the actual variants used. For nested comparisons involving robust ML or WLSMV, use the software's appropriate corrected comparison procedure rather than subtracting corrected chi-square values by hand.

Evaluate the solution before interpreting fit indices

Read warnings, convergence information, standard errors, residual variances and latent correlations first. A negative residual variance, a latent correlation outside its allowable range or unstable estimates can signal an improper solution. Standardized loadings above one also require investigation. A reassuring CFI cannot repair those problems.

Distinguish an unstandardized loading, expressed in the model's chosen units, from a fully standardized loading. In this simple model, squaring a fully standardized loading gives the proportion of that indicator's model-implied variance explained by its assigned factor. This shortcut needs reconsideration when indicators have cross-loadings or the model is otherwise more complex.

Avoid deleting every item below a favourite threshold. Evaluate the item's conceptual role, uncertainty, wording, distribution and implications for content coverage. A statistically significant loading can still be too weak for the intended use; a relatively modest loading may reflect an important part of a broad construct.

Verified standardized loadings in the synthetic example
IndicatorStandardized loadingExplained varianceResidual variance
A10.800.640.36
A20.700.490.51
A30.600.360.64
C10.800.640.36
C20.700.490.51
C30.600.360.64

For A1, 0.80² = 0.64, so its assigned factor explains 64% of its model-implied variance; 1 − 0.64 = 0.36 remains residual variance. For A3, the explained proportion is 0.60² = 0.36, or 36%. A3 therefore deserves closer attention than A1, but these numbers alone do not justify removing it.

The factor correlation is 0.50. This suggests related dimensions under the model, rather than interchangeable labels or a demonstrated causal relationship. A real report should provide the loadings and factor correlations with appropriate standard errors or confidence intervals, not only the rounded point estimates shown for teaching.

Interpret CFI, TLI, RMSEA and SRMR together

Global fit assesses how well the model represents the observed covariance structure under its assumptions. Local diagnostics show where discrepancies remain. Both are needed, alongside substantive interpretation.

What commonly reported CFA fit statistics tell you
StatisticMain informationPractical caution
Model chi-square, df and pTests exact fit under the estimation assumptionsSensitive to sample size and departures from assumptions; nonsignificance is not proof
CFICompares discrepancy with an independence baseline modelIts meaning depends partly on that baseline and the estimator
TLICompares model and baseline discrepancy relative to degrees of freedomCan behave unusually in small or difficult models; do not interpret it alone
RMSEA with its confidence intervalSummarises approximate misfit relative to model degrees of freedomCan be imprecise or problematic with few degrees of freedom
SRMRSummarises standardized residual discrepanciesAn average can conceal a large local residual

Commonly cited reference values include CFI and TLI near or above .95, RMSEA near or below .06, and SRMR near or below .08. These are contextual guides, not universal pass/fail standards. Goretzko, Siemund and Sterner examine the limitations of fixed cutoffs and the need for fuller reporting and model-sensitive evaluation.

Predefine an evaluation strategy suitable for your design and read the pattern of evidence. Do not switch standards after seeing results or ignore an unfavourable statistic because another one is favourable. Inspect residuals and theoretically plausible alternatives where needed.

Worked interpretation of the synthetic model

The same covariance example gives the following ordinary ML results. It uses the Wishart likelihood convention, with N − 1 as the likelihood multiplier. The displayed RMSEA and its noncentral-chi-square interval also explicitly use an N − 1 denominator. Software conventions can differ slightly, so identify the convention and retain the values produced by your chosen software in your dissertation.

Verified global-fit results for the teaching example
QuantityValueInterpretation
Illustrative sample size300An input to this synthetic demonstration
Model chi-square12.000Interpreted with the model df and assumptions
Model df8The covariance restrictions provide testable degrees of freedom
Model p-value0.151Exact fit is not rejected; it is not established as true
Independence-baseline chi-square467.720Reference discrepancy for CFI and TLI
Independence-baseline df15All distinct off-diagonal covariances are restricted
CFI0.991Favourable relative fit against this baseline
TLI0.983Favourable relative fit with a df adjustment
RMSEA0.041A small point estimate under the stated convention
RMSEA 90% confidence interval0.000 to 0.085Appreciable uncertainty remains about approximate misfit
SRMR0.027A small average standardized residual discrepancy

Using unrounded inputs, CFI is calculated as 1 − (model chi-square − model df) / (baseline chi-square − baseline df), giving 0.991. TLI uses the corresponding chi-square-to-df comparisons, giving 0.983. The stated RMSEA calculation is sqrt((12 − 8) / (8 × 299)), giving 0.041. These simplified expressions apply to this ordinary ML example; robust variants require their own corrections.

The overall pattern is encouraging, but the RMSEA interval extends to 0.085. Report that uncertainty rather than describing the point estimate as conclusive evidence of excellent fit. A p-value of 0.151 also does not mean there is a particular probability that the model is correct.

SRMR here is the root mean square of the lower-triangle standardized differences between the observed and fitted covariance matrices, including the diagonal. The average is small, but a real analysis must also inspect individual residuals to identify concentrated misfit.

Evaluate reliability and validity separately

Reliability concerns the consistency or precision of a specified score under a measurement framework. Validity concerns the evidence supporting interpretations and uses of scores. Good CFA fit does not establish either automatically.

For a simple continuous congeneric subscale with unit factor variance, uncorrelated residuals and the stated indicator scaling, a model-based reliability coefficient for the sum can be calculated as:

CR = (sum of loadings)² / ((sum of loadings)² + sum of residual variances)

For each example subscale, the loading sum is 0.80 + 0.70 + 0.60 = 2.10 and the residual sum is 0.36 + 0.51 + 0.64 = 1.51. Thus CR = 4.41 / (4.41 + 1.51) = 0.745, approximately. The semTools composite-reliability documentation explains why the exact coefficient, score definition, scaling and model assumptions matter, especially for ordinal outcomes or correlated errors.

Cronbach's alpha addresses related questions under different assumptions; it does not demonstrate unidimensionality. Our Cronbach's alpha dissertation guide explains why a high coefficient cannot substitute for a defensible measurement model. Avoid reporting one reliability estimate for all indicators when your scoring plan treats them as distinct subscales.

Convergent and discriminant evidence

Average variance extracted (AVE) summarises how much indicator variance is accounted for by the factor under the specified calculation. For these standardized indicators, AVE = (0.64 + 0.49 + 0.36) / 3 = 0.497, approximately. The semTools AVE documentation describes the model-based formula and its restrictions.

This is close to, but below, the frequently used .50 reference value. Rounding it upwards and declaring convergent validity established would conceal the actual result. The subscale's reliability estimate can be reasonably encouraging while the average explained indicator variance remains borderline. Consider loading uncertainty, content coverage and the intended use before deciding whether revision is warranted.

For discriminant evidence, consider whether the factors are substantively and empirically distinguishable. Here their squared correlation is 0.50² = 0.25, below each factor's AVE. That comparison is one limited check; it does not settle discriminant validity. Examine the correlation's uncertainty, competing explanations and other suitable diagnostics. Content evidence and theoretically expected relationships with external variables also remain relevant.

Use modification indices to investigate, not to chase fit

A modification index estimates the potential improvement in model chi-square from freeing a constrained parameter, under local approximations. The expected parameter change helps describe the proposed parameter's size. Neither statistic supplies a substantive reason to change the model. The lavaan modification-index guide distinguishes the index from expected parameter changes.

Suppose diagnostics suggest a residual association between A1 and A2. Shared wording might plausibly explain it. Before freeing that covariance, review the wording, residuals, theory and consequences for the meaning of Support access. Alternatively, an omitted cross-loading might signal that the simple structure is too restrictive.

Do not add every suggested path, remove difficult items until a threshold is met, or claim that a revised model was specified in advance. Retain the original model results, document each change and explain whether it was planned or suggested by the observed data. Changes made on the confirmatory sample require new evaluation, ideally using independent data.

No modification is added to the worked model. The purpose of the example is to interpret its existing evidence, including the borderline AVE, rather than engineer a uniformly flattering set of results.

Reproduce the worked covariance example

The following optional R example supplies the synthetic covariance matrix directly. It represents continuous indicators, complete covariance information and an illustrative N = 300. It cannot diagnose real response distributions, missingness or recruitment quality. The matrix is rounded for display; small differences beyond the reported precision can occur.

library(lavaan)

S <- matrix(c(
  1.000000000, .571377653, .479757040, .296707429, .231500025, .252111456,
  .571377653, 1.000000000, .393424111, .255700796, .280456060, .212954084,
  .479757040, .393424111, 1.000000000, .333353506, .243935399, .183513870,
  .296707429, .255700796, .333353506, 1.000000000, .561514955, .477412876,
  .231500025, .280456060, .243935399, .561514955, 1.000000000, .420597672,
  .252111456, .212954084, .183513870, .477412876, .420597672, 1.000000000
), nrow = 6, byrow = TRUE)
dimnames(S) <- list(c("A1", "A2", "A3", "C1", "C2", "C3"),
                    c("A1", "A2", "A3", "C1", "C2", "C3"))

model <- '
  access  =~ A1 + A2 + A3
  comfort =~ C1 + C2 + C3
  access ~~ comfort
'
fit <- cfa(model, sample.cov = S, sample.nobs = 300,
           std.lv = TRUE, estimator = "ML", likelihood = "wishart")
summary(fit, fit.measures = TRUE, standardized = TRUE)
standardizedSolution(fit)

# Match the article's explicit N-1 RMSEA convention.
chisq <- unname(fitMeasures(fit, "chisq"))
df <- unname(fitMeasures(fit, "df"))
rmsea_article <- sqrt(max(chisq - df, 0) / (df * 299))
nc_bound <- function(target) {
  if (pchisq(chisq, df) < target) return(0)
  uniroot(function(ncp) pchisq(chisq, df, ncp = ncp) - target,
          c(0, 200))$root
}
rmsea_ci_article <- sqrt(sapply(c(.95, .05), nc_bound) / (df * 299))
rmsea_article
rmsea_ci_article
packageVersion("lavaan")

The official lavaan CFA example explains its model syntax and default constraints. With actual ordinal responses, use the appropriate categorical specification and estimator instead of applying this continuous covariance example unchanged. Keep your syntax and complete output so readers can distinguish the model specified from the interpretation written afterward.

Dissertation-ready methodology and results paragraphs

These paragraphs demonstrate structure and restraint. They describe the synthetic teaching analysis, not research you may claim to have conducted. In a real dissertation, add the actual recruitment and analytic sample, screening results, missing-data procedure, planning justification, software version and uncertainty estimates.

Example methodology paragraph

“A confirmatory measurement model distinguished Support access from Help-seeking comfort, with each factor represented by its three prespecified indicators. Cross-loadings and residual covariances were fixed to zero, the factor correlation was freely estimated, and factor variances were fixed to one for scaling. For this teaching demonstration, a synthetic continuous-indicator covariance matrix was analysed using normal-theory maximum likelihood with the Wishart likelihood convention and an illustrative sample size of 300. Evaluation considered identification, admissibility, standardized loadings, the exact-fit test, CFI, TLI, RMSEA with its interval, and standardized residual discrepancies. The reported RMSEA calculations used an N − 1 denominator. Reliability and average variance extracted were evaluated separately from global fit.”

Example results paragraph

“The synthetic two-factor model had 8 degrees of freedom and an admissible solution. Overall fit was encouraging: chi-square(8) = 12.000, p = 0.151; CFI = 0.991; TLI = 0.983; RMSEA = 0.041, 90% CI [0.000, 0.085]; and SRMR = 0.027. Standardized loadings ranged from 0.60 to 0.80, and the factor correlation was 0.50. Model-based reliability was approximately 0.745 for each subscale, while AVE was approximately 0.497. These results illustrate how favourable global fit can coexist with borderline average explained indicator variance. No post hoc modifications were added. The synthetic demonstration does not establish validity in a real population or support comparisons between demographic groups.”

Our guide to reporting statistical results in a dissertation explains how to connect tables, estimates and interpretation. Copy the organisation of these examples while replacing every claim with evidence from your own analysis.

CFA reporting inventory for a completed dissertation
Reporting elementIncludeAvoid
Measurement rationaleInstrument provenance, constructs, indicators and specified restrictionsDescribing a diagram without explaining its meaning
Sample and preparationAnalytic N, design, distributions, coding and missing-data procedureUnexplained exclusions or default treatment
Identification and estimationScaling, free parameters, df, estimator and software versionNaming only the software
Global and local evidenceTest statistic, df, p, index variants, RMSEA interval and residual diagnosticsDeclaring success from one index
Parameter resultsLoadings with uncertainty, factor correlations, residual variances and any residual covariancesReporting only significant or favourable estimates
Reliability and validityNamed coefficients, score definitions, assumptions and wider evidenceTreating reliability or fit as complete validation
Revisions and limitsOriginal and revised models, rationale, exploratory decisions and replication needsPresenting data-driven revisions as prespecified confirmation

Limit what your CFA conclusions claim

A well-fitting CFA does not prove that the model is uniquely correct. Other structures can explain similar covariance patterns. Cross-sectional factor correlations do not establish causal relationships, and one sample cannot show that measurement behaves identically everywhere.

If your dissertation compares groups or time points, evaluate measurement invariance at a level appropriate to the planned comparison. For continuous indicators, configural, loading and intercept restrictions address different questions; ordinal models also involve thresholds and require an appropriate invariance strategy. The lavaan multiple-groups tutorial explains why latent-mean comparisons require more evidence than separately fitting the same diagram in each group.

If model fit is poor, report it and investigate the restrictions, item content, data quality and sampling limitations. A transparent account of a weak measurement model is more informative than an unexplained sequence of deletions. Subsequent regressions or group comparisons cannot make uncertainty about the underlying measurement disappear.

Frequently asked questions

Do I have to conduct EFA before CFA?

No. A sufficiently specified model from theory or prior instrument research can justify CFA directly. EFA is useful when the structure needs exploration. When both are used, distinguish model development from independent evaluation and explain whether observations were reused.

How many participants do I need for CFA?

There is no universal number. Justify the sample using the proposed model, expected loading strength, item types, estimator, missingness and precision goals. A model-specific simulation is more informative than treating the example's illustrative sample as a requirement.

Is CFI above .95 enough to accept the model?

No. Examine the solution's admissibility, other fit statistics, residuals, parameter uncertainty and substantive rationale. Reference values are not universal rules, and a favourable CFI does not establish reliability or validity.

Can I delete an item with a loading below .70?

There is no automatic deletion rule. Review its content, uncertainty, response distribution and contribution to the construct. Report any removal, the rationale and its effect on the measurement model and scoring plan.

Can I report CFA when RMSEA and CFI disagree?

Yes, provided you report the disagreement accurately and investigate it. Model size, degrees of freedom, baseline fit, estimator and sample characteristics can affect indices differently. Do not suppress the statistic that makes your preferred conclusion less comfortable.

Does acceptable CFA fit establish construct validity?

It supplies evidence about the specified internal structure in the analysed sample. Content, response processes, reliability, external relationships, consequences of use and replication may still need evaluation. State exactly which interpretation your evidence supports.

What should I do if the model does not converge?

Investigate coding, sparse categories, redundant items, scaling, identification, starting values and model misspecification. Inspect warnings before interpreting estimates. Increasing the iteration limit alone does not demonstrate that the underlying problem has been resolved.

Turn the analysis into a clear dissertation account

Explain why the proposed structure matters, show how the estimator matches the data, report both global and local evidence, and distinguish planned decisions from revisions. Preserve the complete output and retain limitations even when the headline fit indices look favourable.

If you would value help interpreting methodological feedback or improving the clarity of your own CFA methods, tables and results narrative, explore Acadelyra's academic support services. Support should help you understand and communicate your research while meeting your institution's authorship and assistance requirements.

Sources and further reading