Estimation Methods and SEM Fit Indices

1,644 words · 7 min read · 4 references cited

Structural equation modeling (SEM) reports its goodness of fit through a small set of indices that have, by convention, hardened into thresholds. The Hu and Bentler (1999) cutoffs — CFI ≥ 0.95, TLI ≥ 0.95, RMSEA ≤ 0.06, SRMR ≤ 0.08 — are now the de facto standard against which SEM models are judged in psychology and education journals. The implicit assumption is that these cutoffs apply uniformly: a CFI of 0.94 is borderline regardless of how the model was estimated, and a model that fails the threshold under one estimator should also fail it under another. Shi and Maydeu-Olivares (2020), in Educational and Psychological Measurement, demonstrate that this assumption is wrong, with consequences for how SEM results should be interpreted and reported.

CFI (≥ better)≥ 0.95TLI (≥ better)≥ 0.95RMSEA (≤ better)≤ 0.06SRMR (≤ better)≤ 0.0800.20.40.60.8Cutoff value
Figure 1. The Hu and Bentler (1999) fit-index cutoffs that have hardened into de facto SEM standards.

Why the estimator matters at all

SEM fits a parameterized covariance matrix to an observed data covariance matrix by minimizing some discrepancy function between the two. The standard estimators differ in what discrepancy they minimize. Maximum likelihood (ML) minimizes a function derived from the multivariate normal log-likelihood, which is appropriate when data are continuous and approximately normally distributed. Unweighted least squares (ULS) minimizes the sum of squared differences between observed and model-implied covariances, treating every covariance equally. Diagonally weighted least squares (DWLS) — implemented as WLSMV in Mplus and as the default robust estimator for categorical SEM in lavaan — weights each covariance by its asymptotic variance, which is appropriate for ordered-categorical or non-normal data.

The fit indices are computed from the same fitted model, but the values they return depend on the discrepancy function the estimator minimized. RMSEA is built around the chi-square statistic associated with the fitted model; CFI compares this chi-square to a baseline-model chi-square; SRMR is computed directly from the standardized residual covariances. When the estimator changes, the chi-square changes, the residual covariances may change, and the indices follow.

This is not a subtle effect. Beauducel and Herzberg (2006), looking at ordered-categorical SEM, showed that ML estimation applied to ordinal data produces systematically inflated chi-square values relative to the WLSMV alternative, leading to RMSEA values that look like poor fit when the underlying model is correctly specified. Their finding was a major step in establishing WLSMV as the recommended estimator for categorical SEM, but it also flagged the more general issue that fit-index values are not estimator-invariant.

What Shi and Maydeu-Olivares (2020) found

Shi and Maydeu-Olivares ran simulation studies that crossed estimator (ML, ULS, DWLS) with several types of model misspecification (incorrect dimensionality, omitted cross-loadings, ignored residual correlations) at various sample sizes. They tracked how RMSEA, CFI, and SRMR responded to each misspecification under each estimator.

Estimator (ML / ULS /DWLS)Model chi-squareStd. residualcovariancesRMSEA & CFI(sensitive)SRMR (robust)
Figure 2. Shi and Maydeu-Olivares (2020) found RMSEA and CFI shift with the estimator via the chi-square, while SRMR bypasses it and stays comparatively stable.

The principal result: RMSEA and CFI behaved differently across estimators. The same true model and the same true misspecification produced different RMSEA and CFI values depending on whether ML, ULS, or DWLS was the estimator. The difference was large enough to change the verdict under the conventional Hu-Bentler thresholds: a model that passed CFI ≥ 0.95 under one estimator could fail under another, with the underlying data and the underlying misspecification unchanged. The implication is that “the model fits” or “the model does not fit” depends on the estimator, not only on the model and the data.

SRMR, by contrast, was substantially less sensitive to estimator choice. Because SRMR is computed from the standardized residual covariances directly, without going through the chi-square machinery that the other indices depend on, it is more nearly estimator-invariant. Shi and Maydeu-Olivares argue that SRMR therefore deserves more interpretive weight than the conventions allow, and that RMSEA and CFI should be interpreted with the estimator explicitly in mind.

The misspecification analysis sharpened the picture further. RMSEA and CFI were sensitive to misspecification type as well as estimator: omitted cross-loadings produced different fit-index responses than ignored residual correlations, and the patterns of response interacted with the estimator. The cleanest practical reading is that the Hu-Bentler thresholds are derived under specific conditions (ML estimation, continuous indicators, particular misspecification patterns) and do not transfer reliably to other estimator-data-misspecification combinations.

What this means for SEM practice

The actionable recommendations distill to:

  • Report the estimator explicitly. “We fit the model in lavaan with ML estimation” is the minimum disclosure; “We used WLSMV with robust standard errors and a robust scaled chi-square” is more useful. The reader needs to know which estimator was used to interpret the fit indices.
  • Treat the Hu-Bentler thresholds as estimator-specific guidance, not universal cutoffs. The original Hu and Bentler (1999) simulation used ML estimation on continuous data with specific misspecification patterns. Using their thresholds for DWLS or WLSMV on categorical data extrapolates beyond the original derivation.
  • Weight SRMR more heavily. Shi and Maydeu-Olivares (2020) and several preceding studies have shown SRMR to be more robust to estimator choice than RMSEA or CFI. SRMR-based assessment is more comparable across studies that used different estimators.
  • For ordered-categorical data, prefer WLSMV (DWLS) over ML. Beauducel and Herzberg (2006) established that ML on categorical data inflates the chi-square; the inflation distorts every chi-square-derived fit index. WLSMV is the standard alternative.
  • Run a sensitivity analysis when feasible. Refitting the model under an alternative estimator and reporting how the fit indices change provides direct evidence about the estimator-dependence of the conclusion. If the conclusion is robust across estimators, the model fit is more credible.

The deeper lesson is that fit-index reporting is one of the points where automated SEM software produces a single answer per fit and the user is encouraged to read it as definitive. The fit-index value is conditional on the estimator, the data structure, and the misspecification pattern; treating it as an unconditional verdict on model adequacy ignores all three conditioning structures.

Where this connects to broader latent-variable methodology

The estimator-fit-index interaction is one of several places where SEM practice has to be more careful than the standard reporting templates assume. The choice of fit-index cutoffs (Hu and Bentler conventional vs alternative thresholds), the treatment of factor-retention decisions in EFA that feed into the SEM specification, the handling of rotation indeterminacy in multidimensional models, and the treatment of item-distribution constraints in reliability all interact with the estimator choice. Each of these is documented in the methodological literature; each is ignored or mishandled in routine empirical practice.

The unifying recommendation is sensitivity analysis. SEM results that survive variation across reasonable methodological choices are more credible than results that exist only under one specific configuration of estimator, prior, fit-index threshold, and rotation criterion. The cost of the sensitivity analysis is mostly computational; the cost of failing to do it is publishing results that may not replicate when another team chooses a different default.

For high-stakes applications — clinical assessment validation, educational measurement, policy-relevant social science — the sensitivity standard should be the operating norm rather than a methodological extra. The Shi-Maydeu findings make a specific case for sensitivity to the estimator; they fit into a broader pattern of methodological decisions whose effects are non-trivial and whose proper handling is to disclose, vary, and report.

Frequently Asked Questions

Why is RMSEA sensitive to the estimator?

RMSEA is derived from the chi-square statistic of the fitted model. Different estimators (ML, ULS, DWLS) minimize different discrepancy functions and produce different chi-square values for the same model and data. The chi-square dependence carries through to RMSEA. CFI is similarly affected because it normalizes the model chi-square against a baseline-model chi-square.

Why is SRMR more robust?

SRMR is computed directly from the standardized residual covariances — the difference between observed and model-implied covariances after fitting. It does not go through the chi-square machinery and therefore does not inherit chi-square’s estimator dependence. It still depends on the estimator (different estimators produce different fitted parameters), but the dependence is much weaker.

Should I always use WLSMV for categorical data?

For ordinal data with five or fewer response categories, WLSMV (DWLS in lavaan, WLSMV in Mplus) is the methodological default. It treats the underlying continuous trait correctly and avoids the chi-square inflation that ML produces on categorical data. For continuous data with approximately normal distributions, ML is appropriate and slightly more efficient.

Are the Hu-Bentler cutoffs still useful?

As guidance for ML estimation on continuous data with the misspecification patterns Hu and Bentler studied, yes. As universal thresholds applicable across estimators and data types, no. They are conventionally cited as universal but the original derivation does not support that usage. Treat them as defaults that may not apply to the specific configuration.

How should I report SEM fit?

Report the estimator explicitly, all major fit indices (chi-square, RMSEA with confidence interval, CFI, TLI, SRMR), the sample size, and any robustness corrections (robust standard errors, scaled chi-square). When the estimator is non-default for the data type, justify the choice. When fit indices disagree across estimators in a sensitivity analysis, report the disagreement rather than picking the most favorable answer.

References

  • Beauducel, A., & Herzberg, P. Y. (2006). On the performance of maximum likelihood versus means and variance adjusted weighted least squares estimation in CFA. Structural Equation Modeling, 13(2), 186–203. https://doi.org/10.1207/s15328007sem1302_2
  • Browne, M. W., & Cudeck, R. (1992). Alternative ways of assessing model fit. Sociological Methods & Research, 21(2), 230–258. https://doi.org/10.1177/0049124192021002005
  • Hu, L., & 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. https://doi.org/10.1080/10705519909540118
  • Shi, D., & Maydeu-Olivares, A. (2020). The effect of estimation methods on SEM fit indices. Educational and Psychological Measurement, 80(3), 421–445. https://doi.org/10.1177/0013164419885164

Related questions

What is item response theory?

Every time you take a standardized test — an IQ assessment, a college entrance exam, a professional certification — the questions have been calibrated using sophisticated statistical models that most test-takers never learn about. Item Response Theory (IRT) is the mathematical framework behind virtually all modern psychological and educational testing, and understanding its basics illuminates why tests work the way they do. Read more →

What is attenuation-corrected reliability?

Most psychometrics textbooks teach the classical "correction for attenuation" — Spearman's century-old technique for estimating what the correlation between two psychological constructs would be if the tests measuring them were perfectly reliable. The technique is simple: divide the observed correlation by the square root of the product of the two reliabilities. The technique is also limited: it adjusts the relationship between two scales, but assumes the reliability values plugged into the denominator are themselves accurate. A 2022 paper by Jari Metsämuuronen in Applied Psychological Measurement argues that this assumption is broken in practice. Reliability estimates produced by Cronbach's alpha and similar formulas are themselves attenuated by the same mechanical errors that attenuate correlations — and in some datasets, alpha may be deflated by 0.40–0.60 units of reliability. Metsämuuronen's contribution is a class of deflation-corrected reliability estimators that apply the classical attenuation logic inside the reliability formula rather than only to correlations between scales. Read more →

How are JCCES General Knowledge items structured?

The General Knowledge (GK) subtest of the Jouve Cerebrals Crystallized Educational Scale (JCCES) measures factual breadth — the accumulated stock of information about the world that crystallized intelligence theory treats as a core component of acquired cognitive ability. A subtest of this kind has to satisfy two structural requirements: items should span a meaningful range of difficulty, and they should order along a single underlying continuum of factual breadth rather than tapping multiple unrelated dimensions. This study examined the item structure of the JCCES GK subtest using multidimensional scaling (MDS) on response data from 588 respondents, and recovered the empirical signature of a well-ordered unidimensional construct: a horseshoe-shaped scaling pattern that is the canonical evidence of a Guttman simplex — items ordered cleanly along a single difficulty continuum. Read more →

How well does coefficient alpha perform with non-normal data?

Cronbach's coefficient alpha is the most-reported reliability statistic in psychology and educational measurement. It is also one of the most-misunderstood. The classical formula assumes that test items measure a single construct with equal factor loadings (tau-equivalence), uncorrelated errors, and continuously distributed scores. Real psychological measurement rarely meets all three assumptions: most scales use Likert responses (discrete), have items with unequal contributions to the construct (congeneric), and produce score distributions that depart from normality. The natural question is how badly alpha breaks under these violations and which alternatives perform better. A 2023 simulation study by Xiao and Hau in Educational and Psychological Measurement provides a systematic answer, with implications for the routine reliability reporting that fills psychometric methods sections. Read more →

Why does the choice of estimator matter in SEM?

SEM fits a parameterized covariance matrix to an observed data covariance matrix by minimizing some discrepancy function between the two. The standard estimators differ in what discrepancy they minimize. Maximum likelihood (ML) minimizes a function derived from the multivariate normal log-likelihood, which is appropriate when data are continuous and approximately normally distributed. Unweighted least squares (ULS) minimizes the sum of squared differences between observed and model-implied covariances, treating every covariance equally. Diagonally weighted least squares (DWLS) — implemented as WLSMV in Mplus and as the default robust estimator for categorical SEM in lavaan — weights each covariance by its asymptotic variance, which is appropriate for ordered-categorical or non-normal data.

What did Shi and Maydeu-Olivares (2020) find?

Shi and Maydeu-Olivares ran simulation studies that crossed estimator (ML, ULS, DWLS) with several types of model misspecification (incorrect dimensionality, omitted cross-loadings, ignored residual correlations) at various sample sizes. They tracked how RMSEA, CFI, and SRMR responded to each misspecification under each estimator.