Rotating correlated factors should not change how much of each item the fitted model explains. get_communalities() treats those factors as independent, which can report impossible negative uniqueness.
Example
With factor_analyzer 0.5.1, a three-factor minres/promax fit to the bundled 2,571-row, 23-item questionnaire reports q08 communality 1.103082 and uniqueness −0.103082. The model-implied values from diag(Λ Φ Λᵀ) are 0.590760 and 0.409240. q06 is also negative under the released calculation. The largest communality error across items is 0.512322.
Cause and checks
The code sums squared pattern loadings, valid for orthogonal factors only. With an oblique structure matrix, the row sum of loadings * structure_ supplies the factor correlations. Corrected promax, oblimin and quartimin outputs match the unrotated model within 7.8e-16 for minres and ML extraction; varimax is an unchanged negative control. The corrected questionnaire values match the R reference fixture within 2.2e-6. The full 102-test suite passes with a two-line correction and regression.
Pinned source: de933d26808039d965c117874c67260ec5149690. Reproducer, input, results, patch and tests: https://github.com/dnncha/cheerfulduck/blob/main/projects/research/content/findings-20260908.json (factor-analyzer-oblique-communalities); evidence archive SHA-256 663e8f6ccbec7fa35aa3027ad9b94017285259fa9a3f5dd2e550eb7442805b15.
This demonstrates changed diagnostics on a real questionnaire, not a changed fitted score, paper or scientific conclusion. No matching report was found in bounded issue/PR searches on 27 September 2026; #92 and PR #51 cover different questions. Prepared with AI assistance; not independently reviewed.
Rotating correlated factors should not change how much of each item the fitted model explains.
get_communalities()treats those factors as independent, which can report impossible negative uniqueness.Example
With factor_analyzer 0.5.1, a three-factor minres/promax fit to the bundled 2,571-row, 23-item questionnaire reports q08 communality 1.103082 and uniqueness −0.103082. The model-implied values from
diag(Λ Φ Λᵀ)are 0.590760 and 0.409240. q06 is also negative under the released calculation. The largest communality error across items is 0.512322.Cause and checks
The code sums squared pattern loadings, valid for orthogonal factors only. With an oblique structure matrix, the row sum of
loadings * structure_supplies the factor correlations. Corrected promax, oblimin and quartimin outputs match the unrotated model within 7.8e-16 for minres and ML extraction; varimax is an unchanged negative control. The corrected questionnaire values match the R reference fixture within 2.2e-6. The full 102-test suite passes with a two-line correction and regression.Pinned source:
de933d26808039d965c117874c67260ec5149690. Reproducer, input, results, patch and tests: https://github.com/dnncha/cheerfulduck/blob/main/projects/research/content/findings-20260908.json (factor-analyzer-oblique-communalities); evidence archive SHA-256663e8f6ccbec7fa35aa3027ad9b94017285259fa9a3f5dd2e550eb7442805b15.This demonstrates changed diagnostics on a real questionnaire, not a changed fitted score, paper or scientific conclusion. No matching report was found in bounded issue/PR searches on 27 September 2026; #92 and PR #51 cover different questions. Prepared with AI assistance; not independently reviewed.