Factor-Structure or Latent-Model Check¶
A model — instantiates Construct–Proxy–Signal Validity Alignment
Fits a latent-variable model to item responses to test whether their internal structure matches the construct's theorized dimensions — internal-structure evidence.
If a set of items really measures the construct you say it does, the pattern of how they move together should match the construct's theorized shape. Factor-Structure or Latent-Model Check tests that. It fits a latent-variable model — factor analysis, a confirmatory factor model, or an item-response model — to the responses and asks whether the items cohere into the hypothesized number of dimensions, loading where theory predicts and staying separate where it predicts distinctness. Its defining contribution is internal-structure evidence: it works entirely within one method, from the item covariances alone, and it is the mechanism that catches a scale silently splitting into two constructs, or two supposedly distinct subscales collapsing into one. Unlike a multi-method matrix, it does not need a second instrument — it interrogates the shape hidden inside a single set of responses.
Example¶
A high-school "science self-efficacy" scale has twelve items and is treated as a single number. A confirmatory factor model asks whether those twelve items really form one self-efficacy factor. The fit says otherwise: the items split cleanly into "confidence doing hands-on inquiry" and "confidence passing science tests," and several items load nearly as strongly on a general self-confidence factor that has nothing specifically to do with science. The internal structure does not match the one-dimensional construct the scale claimed — reporting a single self-efficacy score would blend two distinct things and contaminate both with generic confidence. The model has surfaced a dimensionality and discriminant problem no total score would ever reveal.
How it works¶
- Specify the theorized structure. State how many dimensions the construct has and which items should load on which — a confirmatory test rather than a fishing expedition.
- Fit the latent model and read the loadings. Items should load strongly on their intended factor (structural convergence) and weakly elsewhere; cross-loadings and unexpected factors are misfit.
- Check discriminant separation and dimensionality. Factors that are supposed to be distinct must not correlate near-unity; misfit means the internal structure contradicts the construct theory.
Tuning parameters¶
- Exploratory vs. confirmatory — letting the structure emerge versus testing a pre-stated one; confirmatory is a real test, exploratory a hypothesis generator.
- Number of factors — how many dimensions the model allows; over-factoring fragments the construct, under-factoring hides real multidimensionality.
- Fit-index thresholds — how good the fit must be to accept the structure; looser thresholds pass weaker evidence.
- Item-retention rule — how aggressively poorly-loading items are dropped; aggressive trimming improves fit but risks overfitting the sample.
When it helps, and when it misleads¶
Its strength is a rigorous, quantitative test of dimensionality and internal discriminant structure from data you already have — the clearest way to catch a construct that has quietly fractured or fused.
Its central trap is that good fit is not right meaning[1]: a model can fit beautifully and still be a well-measured wrong construct, because internal coherence says the items belong together, not what they belong to. Fit indices also invite chasing — trimming items until the numbers look acceptable, overfitting to one sample. And a structure confirmed in one group need not hold in another. The discipline is to cross-validate on fresh data, treat fit as one strand of evidence rather than the verdict, and never let a clean structure substitute for evidence about what the factor means.
How it implements the components¶
validity_evidence_matrix— it produces the internal-structure cell of the evidence matrix: the dimensionality-and-loadings strand of the overall case.convergent_discriminant_check— item loadings on their intended factor are a within-method convergent test, and factor separation a discriminant one.
It does not assemble those strands into the overall argument (that is the Construct Validity Argument), and it tests structure within a single group only — checking that the structure holds equivalently across subgroups is the Measurement Invariance Audit.
Related¶
- Instantiates: Construct–Proxy–Signal Validity Alignment — it supplies internal-structure evidence, one of the archetype's core validity strands.
- Sibling mechanisms: Multi-Trait Multi-Method Matrix · Measurement Invariance Audit · Construct Validity Argument · Known-Groups or Contrast-Case Test · Content-Domain Review Panel · Construct-to-Proxy Traceability Table · Cognitive Interview or Response-Process Probe · Proxy Drift and Goodhart Audit · Validity Limitation Memo
Editorial Notes¶
Form Classification¶
Form family: Analysis, Modeling & Optimization
Rationale: Factor-Structure or Latent-Model Check operates as a computation, comparison, model, or analytic representation used to infer, estimate, or choose because it fits a latent-variable model to item responses to test whether their internal structure matches the construct's theorized dimensions — internal-structure evidence.
Independent corroboration: The frozen evidence defines Factor-Structure or Latent-Model Check as 'Fits a latent-variable model to item responses to test whether their internal structure matches the construct's theorized dimensions — internal-structure evidence', so its operative form is Analysis, Modeling & Optimization.
Review outcome: Independent reviewer agreement; high confidence.
Origin Attribution¶
Primary origin: Statistics & Experimental Design
Origin pattern: Single lineage
Present-day reach: Specialized
Rationale: Factor analysis and latent-variable model checks are canonical statistical and psychometric methods for internal-structure validity.
Related originating lineages:
- Psychology — Psychometrics materially originated their use for theorized psychological constructs and item responses.
Review outcome: Independent reviewer agreement; high confidence.
Notes¶
Its natural partner and contrast is the Multi-Trait Multi-Method Matrix: the matrix tests validity across methods, this model tests structure within one. A construct that passes internal-structure checks can still be dominated by method variance the single-method model cannot see — so the two are complementary, not redundant.
References¶
[1] Cronbach, L. J., & Meehl, P. E. "Construct Validity in Psychological Tests". Psychological Bulletin 52(4), 281–302 (1955). Explains that internal consistency alone does not establish what construct a test measures. registry ↩