Reliability: Cronbach's α vs. McDonald's ω

Dr. R. Düsing · Osnabrück University
📋 What This Is About
Cronbach's α is the most-reported reliability coefficient — and the most frequently misused. It requires tau-equivalence (equal loadings) and unidimensionality. If these are violated, α can underestimate reliability (unequal loadings) or feign unidimensionality (multidimensionality) — exactly as with the 8-item cognitive test (V1–V4 verbal, S1–S4 spatial): two factors, but a single α. McDonald's ω is the more robust alternative — in two variants: ωtotal and ωhierarchical. This tool shows when each number misleads. Prerequisite: the measurement models (parallel/tau/congeneric).
← CTT Foundations: reliability as a proportion of variance
①The Three Coefficients
Cronbach's α
—
requires tau-equiv. & 1 dim.
ωtotal
—
total shared variance
ωhierarchical
—
general factor only
—
②Variance Decomposition of the Sum Score
What Does the Total Score's Variance Consist of? — and Where Do α, ωt, ωh Sit
③Split-Half — α ≈ Mean of All Test Halvings
Distribution of All Possible Split-Half Reliabilities (Spearman-Brown Corrected)
Concepts
When α Underestimates
With unequal loadings (congeneric), α — as an unweighted sum score — treats all items the same and wastes information: α < ωtotal. α is then a lower bound on the true reliability. Solution: report ωtotal.
When α Deceives
With multidimensionality (strong group factors), α and ωtotal also count the group variance as "reliable." A high α then suggests a homogeneity that doesn't exist. ωhierarchical shows how much really goes to the single general factor — often alarmingly little.
α = Mean of the Split-Halves
Split-half divides the test into two halves and correlates them (Spearman-Brown corrected). The result depends on how you split. Cronbach's α is exactly the mean over all possible halvings — panel ③ makes this visible.
KR-20 & Practice
For dichotomous (0/1) items, KR-20 is nothing other than Cronbach's α. The methodological literature (Revelle, McDonald, Flora) recommends reporting ω by default and, when unidimensionality is in doubt, additionally ωh.
The Measurement Chain
This tool stands at the end of the chain: CTT Foundations (what is reliability) → Factor Analysis (estimate loadings from data) → Measurement Models (which assumption: parallel/tau/congeneric) → here (which coefficient is correct). The loadings ω computes with come directly from a factor model.
→ CTT Foundations → Factor Analysis → Measurement Models
Reliability α vs. ω — Background
The Model

The items measure a general factor (the actual construct) with loading λg,i. Optionally there are two group factors (items 1…k/2 and the rest) with loading λs — standing in for multidimensionality / content clusters / correlated errors. Items are standardized (Var = 1), so θi = 1 − λg,i² − λs².

The Three Coefficients
Var(S) = variance of the sum score = ΣΣ Cov(X_i,X_j) ω_total = [ (Σλg)² + Σ_groups(Σλs)² ] / Var(S) ω_hier = (Σλg)² / Var(S) α = k/(k−1) · ( 1 − Σ Var(X_i) / Var(S) )

ωtotal = the proportion of variance attributable to all common factors (general + group) = total reliability. ωhierarchical = the proportion attributable to the general factor alone — the measure of how much a single sum score can carry. α only uses item variances and covariances, without knowing the factor structure.

Three Scenarios

Tau-equivalent (equal λg, λs = 0): α = ωt = ωh. Here α is exactly correct.
Congeneric (unequal λg, λs = 0): α < ωt = ωh. α underestimates — the more unequal the loadings, the larger the gap.
Multidimensional (λs > 0): ωh << ωt. α and ωt stay high, but only a small part is attributable to the general factor — a high α feigns unidimensionality.

Why α Does Not Mean Unidimensionality

A common misconception: "high α = homogeneous, unidimensional scale." Wrong — α already rises with the number of items and with any positive correlations, regardless of where they come from. Several correlated subdimensions drive α upward even though the scale is not unidimensional. Only ωh (or an explicit factor analysis) reveals this.

Split-Half (Panel ③)

Splitting the test into two halves and correcting the half-correlation via Spearman-Brown (rsh = 2r/(1+r)) gives a split-half measure — but its value depends on the split chosen. Cronbach's α is the average over all possible halvings. The histogram shows this distribution; α (orange) sits at its center, ωt (green) usually above it.

Recommendation

Report ωtotal instead of α by default (or in addition). If multidimensionality is suspected, report ωhierarchical and check the factor structure. α isn't "forbidden," but under tau-equivalence it's a special case of ω and otherwise misleading.

Literature

Cronbach, L. J. (1951). Coefficient alpha and the internal structure of tests. Psychometrika, 16, 297–334.
McDonald, R. P. (1999). Test Theory: A Unified Treatment. Erlbaum.
Revelle, W. & Zinbarg, R. E. (2009). Coefficients alpha, beta, omega, and the glb. Psychometrika, 74, 145–154.
Flora, D. B. (2020). Your coefficient alpha is probably wrong. AMPPS, 3(4), 484–501.
Moosbrugger, H. & Kelava, A. (Eds.) (2020). Testtheorie und Fragebogenkonstruktion (3rd ed.). Springer.