Heywood Cases in CFA and SEM
Causes and Countermeasures
by Arndt Regorz, MSc.
November 22, 2026
Heywood cases are improper solutions in a CFA or SEM model estimation. Parameters violate logical or mathematical constraints.
Typical examples are negative variances (variances are squared measures and can never be negative) or communalities greater than one.
This blog post will address possible causes for Heywood cases and countermeasures you can use if you have a Heywood case as a result of your model estimation. It is primarily based on the recent journal article by Cheng et al. (2026) that focussed on Heywood cases in CFA.
Reasons for Heywood cases in CFA and SEM
There are several different reasons why you could get a Heywood case as a result of your CFA or SEM model.
Heywood cases can be the result of sampling variability. Especially in small samples this is possible if the true value lies at the boundary of the admissible value space.
Another possible cause is model misspecification. There are several ways to misspecify a model leading to Heywood cases:
- Omitting relevant cross-loadings
- Fixing parameters to zero that are different from zero
- Equality constraints that are not true in the population
- Indicators to the wrong factors
- Too many factors (overextraction)
Model underidentification can also lead to Heywood cases. That can be theoretical underidentification or empirical underidentification.
Finally, poor data quality can result in Heywood cases:
- Extreme multicollinearity
- Low variability
- Severe measurement error
Countermeasures against Heywood cases
One key countermeasure is to prevent model misspecification. Questions you should ask yourself: Are there relevant cross-loadings missing from the model? Do you need additional error covariances? Are there restrictive assumptions (e.g. equality constraints) you could relax?
Another area to focus on is data preprocessing: Dealing with missing data, deleting extreme outliers (after inspection), modeling possible nonlinearity.
In the case of a negative variance estimate it could also help to fix the variance to zero. Here it is important not simply to report a zero instead of the estimated negative variance, but instead to rerun the model with the problematic variance fixed to zero, because fixing the variance will lead to other changes in the model estimate as well.
You could also use other estimation methods that are less susceptible to Heywood cases: Regularized common factor analysis (RCFA), unweighted least squares (ULS), and Bayesian methods.
Finally, there are some newer methods to deal with Heywood cases: Exponential transformation on unique variances, logistic transformation on factor loadings, and maximum a posteriori (MAP) estimation for factor loadings. You can find more information about these approaches in the article by Cheng et al. (2026).
Reference
Cheng, Y., Lyu, W., & Astivia, O. L. O. (2026). Possible solutions to Heywood cases in confirmatory factor analysis. Multivariate Behavioral Research. Advance online publication. https://doi.org/10.1080/00273171.2026.2741681
Citation
Regorz, A. (2026, November 22). Heywood cases in CFA and SEM: Causes and countermeasures. Regorz Statistik. https://www.regorz-statistik.de/blog/Heywood_cases.html