Abstract
Background/Objectives: Image-derived features are routinely treated as fallible indicators of a latent attribute; whether several summaries of one segmented object behave that way is rarely tested. We asked whether a three-factor model of nuclear morphology, specified in advance, reproduces the covariance among thirty indicators. Methods: Cross-sectional secondary analysis of the Wisconsin Diagnostic Breast Cancer dataset: 569 fine-needle aspirates, ten nuclear attributes each summarised three ways, no missing cells. Multivariate normality was rejected, so the correlation matrix was fitted by maximum likelihood with a Satorra-Bentler scaled statistic and sandwich standard errors. Nine further specifications were fitted, one an unrestricted benchmark; admissibility was defined in advance. Results: The a priori model was rejected on every criterion: scaled chi-square 7288.7 on 402 degrees of freedom, p < 0.001, scaled comparative fit index 0.490, uncorrected 0.482, root mean square error of approximation 0.300 (0.297 to 0.304). Every other index failed similarly. Multitrait-multimethod alternatives fitted better, the best reaching a comparative fit index of 0.830, but it did not converge and a second implementation did not reproduce its fit statistics; every one was inadmissible, with residual variances on their bound and a radius-perimeter trait correlation of 0.995. Conclusions: Summary statistics from the same segmented objects are deterministic transforms of one another, not parallel fallible measurements, and such a measurement model has nothing left to be fallible about. Reports of latent-variable models fitted to image-derived features should give a solution's admissibility beside its fit.
Keywords
- structural equation modelling
- confirmatory factor analysis
- model fit
- improper solution
- multitrait-multimethod matrix
- nuclear morphometry
- breast neoplasms
- cross-sectional studies
Declarations
- Ethics approval
- NA
- Consent to participate
- NA
- Consent for publication
- NA
- Funding
- NA
- Competing interests
- NA
- Author contributions
- NA
- Data availability
- NA
- Registration
- NA
- Acknowledgements
- NA
Comments
Commenting requires an account and sign-in.