📋 What This Is About
A measurement model describes how observable items relate to a latent variable (the construct):
every item i loads on the factor with λi and has an
error variance θi. Example: factor analysis just estimated loadings for the
8-item cognitive test (V1–V4 verbal, S1–S4 spatial) — which model (parallel / tau-equivalent / congeneric)
underlies it, and which reliability coefficient is therefore correct? Exactly which assumptions you make about λ
and θ decides that. This is the missing bridge between factor analysis and reliability (α vs. ω).
① Model Class & Reliability
Gap ωtotal − α
—
underestimation by α
—
② Measurement Model — Path Diagram
③ Implied Covariance Matrix
Concepts
The Model Hierarchy
Parallel ⊂ tau-equivalent ⊂ congeneric —
each model relaxes one assumption: parallel requires equal λ and equal θ, tau-equivalent only equal λ,
congeneric leaves both free. The more general, the more realistic — but the more important the correct coefficient.
Why α Needs Tau-Equivalence
The reason: α is based on the unweighted sum score — every item counts
equally, with no individual weighting. That's optimal exactly when all items measure the construct equally well,
i.e. have equal loadings (tau-equivalence). If loadings are unequal, α still treats strong and
weak items the same and wastes information → it underestimates reliability.
ω weights each item by its loading and therefore stays correct even for congeneric data.
The Matrix Reveals the Model
Parallel: all diagonals equal, all covariances equal.
Tau: covariances (λiλj) equal, variances unequal.
Congeneric: everything unequal. The model is directly readable from the implied covariance matrix.
Bridge: EFA → Reliability
Factor analysis estimates the loadings from real data; here you see
which model underlies them and which coefficient applies. This then feeds into the α-vs-ω tool.
→ Factor Analysis (EFA)
→ Factor Analysis (EFA)
Must λ and θ Be Exactly Equal?
In reality, never — the models are idealizations, real values always vary a
bit (the tool uses a small tolerance; minimal slider movements therefore don't immediately flip the class).
What matters is which equality counts for what:
• Loadings λ decide α vs. ω — only with (approximately) equal λ is α exact, otherwise it underestimates.
• Error variances θ only separate parallel from tau-equivalent. Since tau-equivalence alone already suffices for α = ω, θ-equality is irrelevant for the reliability number — it only matters when you need true interchangeability.
In practice: test equality assumptions via CFA (χ² difference test, CFI/RMSEA) — or just report ω. When in doubt, ω.
• Loadings λ decide α vs. ω — only with (approximately) equal λ is α exact, otherwise it underestimates.
• Error variances θ only separate parallel from tau-equivalent. Since tau-equivalence alone already suffices for α = ω, θ-equality is irrelevant for the reliability number — it only matters when you need true interchangeability.
In practice: test equality assumptions via CFA (χ² difference test, CFI/RMSEA) — or just report ω. When in doubt, ω.
Why Even Bother with Parallel?
For the reliability number, parallel adds nothing over tau-equivalent — there,
α = ω already holds exactly. The payoff of parallel is interchangeability: equal loadings and
equal error variances make all items (and entire test halves or forms) true equivalent replicates.
Only then can you build parallel test forms (Form A/B), split the test into equivalent
halves in any way (split-half exact regardless of how it's divided), and swap or drop items
without changing properties — every item measures with identical precision (same SEM). The price: parallel is the
strictest and least realistic assumption.