Statistics Glossary → Confirmatory Factor Analysis (CFA)
Confirmatory Factor Analysis (CFA)analysis
An analysis that tests whether the items of a scale load on latent factors specified in advance from theory as expected, i.e. whether a proposed measurement model fits the data.
Concept
Concepts such as intelligence, anxiety or job satisfaction cannot be measured directly; they are latent variables captured through several observed items. Exploratory factor analysis discovers how items group on their own, whereas confirmatory factor analysis (CFA) tests a structure the researcher has built beforehand: which item measures which factor is stated in advance, and the question is how well this model reproduces the correlations in the data. This makes CFA the core evidence of construct validity in scale development and adaptation.
The model writes each item as the effect of a factor plus an item-specific error: x = λ·F + e, where the loading λ shows how well the item represents the factor. From this structure a model-implied covariance matrix is derived and compared with the observed one. Fit is judged with the chi-square test and indices such as CFI, TLI, RMSEA and SRMR; common guidelines treat CFI and TLI above .90-.95, RMSEA below .06-.08 and SRMR below .08 as good fit.
With large samples the chi-square test reacts even to small misfits, so it is read alongside the indices rather than used alone. Sufficiently high standardised loadings, composite reliability (CR) and average variance extracted (AVE) complete the picture of measurement quality. Changes made to improve fit (residual covariances, cross-loadings) should rest on theory and be confirmed in a new sample; otherwise the model is merely tailored to the data at hand.
When to use it
- When the factor structure of a scale is known from theory or earlier studies and you want to test whether it holds in a new sample (scale adaptation, construct validity).
- If there is no prior expectation about the structure, start with exploratory factor analysis; rather than exploring and confirming on the same data, use two separate samples.
- To test directional relations between latent factors (one construct predicting another) move on to structural equation modelling; whether the scale works the same way in different groups is tested with measurement invariance analysis.
An everyday example
A school designs a nine-question survey on pupils' school life: three questions are written to measure belonging, three teacher support and three peer relations. CFA tests whether the answers of 300 pupils really reflect this three-dimensional structure. If, say, the question 'I feel lonely at this school' turns out to be tied more to peer relations than to belonging, model fit suffers, and the question is either reworded or its move to another dimension is debated on theoretical grounds.
Parameters in MerQur
Number of factors (1-6)
Concept: The number of latent factors is a theoretical claim about how many distinct concepts the scale measures. In CFA this number and the items belonging to each factor are specified in advance, not discovered from the data; a factor usually needs at least three items to be identified soundly.
Everyday example: For the school survey three factors are defined, belonging, teacher support and peer relations, each with its own three questions. Changing the number of factors until the results look good turns a test into an exploration.
What it does: Sets the number of latent factors in the measurement model; a separate variable selector opens for each factor (at least 3 indicators per factor). The covariances between factors are free.
Effect on the result: The number of parameters, degrees of freedom and fit. The number of factors should come from theory or a prior exploratory analysis (Brown, 2015).
On the validation data: 1 factor (SAQ q01-q08, n = 2571) chi-square(20) = 554.19, CFI = 0.871. 3 factors (HS) chi-square(24) = 85.306. 4 factors (HS 24 tests) chi-square = 316.79; 6 factors 478.93.
Structural paths (SEM)
Concept: In classical CFA the relations between factors are undirected correlations. Defining structural paths assumes that one factor predicts another (a regression), and the analysis becomes a simple structural equation model.
Everyday example: Instead of saying only that belonging, teacher support and peer relations are correlated, you can make a directional claim: 'teacher support and peer relations predict belonging'.
What it does: Defines directed relationships between latent factors ('F3 ~ F1 + F2'); if empty, the covariances between factors are free (classical CFA).
Effect on the result: Regression coefficients are reported instead of covariances; with a saturated structural part the chi-square does not change.
On the validation data: HS 'F3 ~ F1 + F2' chi-square = 85.306 (the same as the CFA), F3 ~ F2 = 0.0531 (p = 0.319). PoliticalDemocracy chi-square(35) = 38.125, F2 ~ F1 = 1.4824.
Estimator (ML, MLM, MLR)
Concept: The estimator determines the principle by which model parameters are computed and how their uncertainty is measured. Maximum likelihood (ML) assumes the data are multivariate normal; its robust variants give the same estimates but correct the standard errors and chi-square for non-normality.
Everyday example: If the survey questions are answered on a 1-5 scale and most pupils pick 4 or 5, the answers pile up at one end. The loadings stay the same, but a robust estimator stops the results from looking more precise than they are.
What it does: ML is normal-theory maximum likelihood; MLM gives the same estimates with the Satorra-Bentler scaled chi-square and robust SE (complete data); MLR Huber-White sandwich SE and the Yuan-Bentler scaled chi-square. Under non-normality MLM/MLR, with missing data MLR (+FIML) is used.
Effect on the result: Does not change the estimates; changes the standard errors, z, p, confidence intervals and the scaled fit indices.
On the validation data: HS: ML chi-square = 85.306; MLM scaled 80.872 (scale 1.0548), x2 loading SE 0.1033 instead of 0.0997; MLR scaled 87.132 (scale 0.9790), x2 SE 0.1321.
Missing data (Listwise deletion, FIML)
Concept: Listwise deletion removes anyone with a missing item from the analysis entirely. Full information maximum likelihood (FIML) uses whichever items each person answered; when missingness can be explained by the observed information (MAR), it is less biased and more efficient.
Everyday example: If 40 of 300 pupils skipped a single question, listwise deletion throws away their other eight answers too. FIML keeps the eight questions those pupils did answer.
What it does: Listwise deletion drops the row with missing values; FIML lets each case contribute to the likelihood with its observed part (MAR assumption). MLM and FIML cannot be used together.
Effect on the result: n, the estimates, SE and chi-square; under FIML the mean structure is also estimated (AIC/BIC change).
On the validation data: bfi 5 factors: listwise n = 2436, chi-square(265) = 4165.47; FIML n = 2800, chi-square = 4674.26. In the 'missing' example data the FIML factor correlation is 0.5014. With complete data FIML gives SE from the observed information (HS x9 SE 0.1951).
RMSEA denominator (N, N - 1)
Concept: RMSEA measures how badly the model approximately misfits in the population, per degree of freedom. Using N or N - 1 for the sample size in its denominator is merely a convention that differs between programs; with a large sample it changes the result negligibly.
Everyday example: With 300 pupils, the difference between N and N - 1 moves RMSEA only in the third decimal. The option is useful when results have to match another program exactly.
What it does: Chooses whether D in RMSEA = sqrt(max(chi-square - df, 0) / (df x D)) is N or N - 1; the default is N.
Effect on the result: Only the point RMSEA (the confidence interval and PCLOSE always use N).
On the validation data: HS: 0.0921 (N), 0.0923 (N - 1).
Confidence Level (%90, %95, %99, custom)
Concept: Sets how much assurance the confidence intervals for factor loadings and structural paths carry. A higher level gives a wider interval.
Everyday example: If the belonging factor's loading on a question is 0.72, you can say with 95% confidence that the true loading lies between, say, 0.64 and 0.80; asking for 99% widens the range.
What it does: Sets the level of the Wald confidence intervals of loadings and structural paths (estimate +- z x SE). The RMSEA confidence interval is always 90%.
Effect on the result: Only the interval width.
On the validation data: HS x2 loading 95% [0.358, 0.749]; intervals are also given at the 90%, 99% and 85.5% levels.
Measurement invariance (group)
Concept: Measurement invariance tests whether a scale measures the same concept in the same way across groups. The steps become progressively stricter: the same factor structure (configural), equal loadings (metric), equal intercepts (scalar) and equal residual variances (strict). Meaningful comparison of group means requires at least scalar invariance.
Everyday example: Before comparing girls' and boys' sense of belonging, you must be sure that the question 'I feel at home at school' means the same thing to both groups. Without invariance, a difference in means may reflect the question being read differently rather than a real difference in belonging.
What it does: Fits the configural, metric (equal loadings), scalar (+ equal intercepts) and strict (+ equal residual variances) models in turn for the selected group; gives the chi-square difference, CFI and RMSEA differences and score tests (Chen, 2007: CFI difference <= -0.01).
Effect on the result: Does not change the main CFA result; adds an invariance table.
On the validation data: HS, school group: configural chi-square(48) = 115.851, metric (54) = 124.044 (difference 8.192, p = 0.224), scalar (60) = 164.103 (difference 40.059, CFI difference -0.038), strict (69) = 181.511. Under MLR the difference is Satorra-Bentler (2001) scaled (metric 6.567).
→ Measurement invariance (configural, metric, scalar, strict)
Partial invariance (free parameters)
Concept: Partial invariance means freeing the parameters of the few items that behave differently across groups, when full invariance fails, while keeping the rest equal. The group comparison can then still be made without relying on the problematic items.
Everyday example: If one peer-relations question works differently for girls and boys, only that question's intercept is freed, and the comparison continues with the other eight questions held equal across groups. The choice of which question to free should also be defensible on theoretical grounds.
What it does: Specifies the parameters exempt from equality constraints ('F1 =~ x2; x7 ~ 1'). Should be used for a few items indicated by score tests, with a theoretical rationale (Byrne et al., 1989).
Effect on the result: In the metric, scalar and strict steps the listed parameters stay free across groups.
On the validation data: With 'F1 =~ x2; x7 ~ 1' scalar chi-square(58) = 148.587 (full scalar 164.103), difference(5) = 25.231.
→ Measurement invariance (configural, metric, scalar, strict)
Additional model syntax
Concept: Lets you add relations beyond the core structure predicted by theory: correlated errors between two items (a residual covariance) or an item also loading on a second factor (a cross-loading). Every added parameter improves fit somewhat, so additions need a justification.
Everyday example: If two questions are worded almost identically ('My teacher helps me' and 'My teacher supports me'), their errors are expected to be related, which can be modelled with a residual covariance. Adding relations at random just to raise fit makes the result untrustworthy.
What it does: Adds a residual covariance ('x7 ~~ x8') or a cross-loading ('F1 =~ x9'); should be used only with a theoretical rationale (a change based on modification indices requires cross-validation).
Effect on the result: The number of parameters and fit.
On the validation data: HS: 'F1 =~ x9' chi-square(23) = 52.382 (loading 0.4371); 'x7 ~~ x8' chi-square(23) = 53.272, covariance 0.3527 (SE 0.0673). These two candidates are the largest modification indices of the default model (36.41 and 34.15).
Additional output: Factor variance = 1 identification (std.lv)
Concept: Latent factors have no unit of their own; to identify the model, either the loading of each factor's first item or the factor's variance is fixed to 1. The second route gives standard errors for all loadings and shows the relations between factors directly as correlations; model fit is the same either way.
Everyday example: If belonging's 'unit' is tied to the first question, that question's loading always shows as 1 and cannot be tested. Fixing the factor variance to 1 shows all nine questions' loadings with their standard errors.
What it does: Fixes the factor variances to 1 and gives all loadings (including the first item) and the factor correlations with SE.
Effect on the result: Fit does not change; only the scale. It is not computed when there is a structural path, and the output contains a warning line saying so.
On the validation data: HS x1 loading 0.8996 (SE 0.0808), F1-F2 correlation 0.4585 (SE 0.0638). SAQ std.lv loadings 0.485 ... 0.424.
Additional output: Multivariate normality (Mardia)
Concept: Mardia's tests check whether the items are jointly multivariate normal, in terms of skewness and kurtosis. If normality fails, ML standard errors and chi-square become unreliable and a robust estimator is preferred.
Everyday example: If most pupils answered 'strongly agree' to every question, the answers are not normally distributed. When Mardia's test shows this, switching to a robust estimator is the right decision.
What it does: Tests the normality assumption of ML with the multivariate skewness (chi-square) and kurtosis (z) tests; if significant, MLM/MLR is recommended (Mardia, 1970).
Effect on the result: Additional table only.
On the validation data: HS: b1p = 6.807, skewness chi-square(165) = 341.48 (p < .001), b2p = 102.904, kurtosis z = 2.407 (p = 0.016).
Additional output: Residual correlation matrix
Concept: Residual correlations are the differences between the observed correlation of two items and the correlation the model predicts. Large residuals show which item pairs the model fails to explain, i.e. where the misfit comes from.
Everyday example: If the question 'I feel lonely at this school' correlates with the peer-relations questions much more strongly than the model predicts, the gap stands out in the residual matrix and signals that the question's factor should be reconsidered.
What it does: Gives the difference between the observed and model-implied correlations; lists those with |r| > 0.10 (Kline, 2016).
Effect on the result: Additional table only.
On the validation data: HS: |r| > 0.10 in 5 pairs, the largest x2-x7 = -0.189.
Example on real data
Data: Holzinger and Swineford 1939, lavaan package
Variables: 3 factors: F1 visual (x1-x3), F2 textual (x4-x6), F3 speed (x7-x9)
- Question: Do the nine mental ability tests (x1-x9) fit the prespecified three-latent-factor structure (visual, textual, speed)?
- Model fit: χ²(24, N = 301) = 85.31, p < .001; the model does not fit the data exactly (the chi-square is sensitive in large samples).
- Fit indices: CFI = .931 (acceptable), TLI = .896 and RMSEA = .092, 90% CI [.071, .114] (poor), SRMR = .065 (good); the overall assessment is a weak fit.
- Loadings: All free loadings are significant (p < .001); the standardized loadings range from .42 (x2) to .86. The weakest indicator is x2 (λ = .42).
- Reliability and validity: F2 (textual) is strong (CR = .89, AVE = .72); F1 (CR = .63, AVE = .37) and F3 (CR = .69, AVE = .43) are below the criteria (CR >= .70, AVE >= .50). The correlations between factors range from .28 to .47.
- Additional finding: The largest modification indices are F1 =~ x9 (MI = 36.41) and x7 ~~ x8 (MI = 34.15); these parameters should be freed only with a theoretical rationale.
- APA: χ²(24, N = 301) = 85.31, p < .001, CFI = .931, TLI = .896, RMSEA = .092, 90% CI [.071, .114], SRMR = .065.
Related concepts
Measurement model and structural modelMeasurement invariance (configural, metric, scalar, strict)Robust ML estimators (MLM, MLR; Satorra-Bentler)Model fit indices (χ², CFI, TLI, RMSEA, SRMR)Full information maximum likelihood (FIML)Factor (component) loadingsModel identification (std.lv, unit loading)Mardia's test of multivariate normality
The numbers in the examples are MerQur's output on independent, publicly available validation data; see the MerQur Technical Note for details.