Skip to content

Measurement & Index Lab

An index is reliable when its items vary together. Cronbach's α compares the sum of the item variances with the variance of the total: the more the items covary, the larger the total is relative to its parts. Drop a weak item or reverse a negatively worded one and watch α move.

α = 0.690

Variance picture: Σ item variances vs variance of the total

Σ s²ᵢ7.83s²_total18.412 Σ cov = 10.58 · α = 0.690
Variance picture: Σ item variances vs variance of the total: Σ s²ᵢ = 7.83, s²_total = 18.41, α = 0.690.
Item statistics
itemmeanSDr_it (corrected)α if deleted
T13.661.160.5000.622
T23.451.110.5360.611
T3 (R)3.471.110.4820.629
T43.421.160.6040.584
T5 (R)3.601.090.4950.625
T63.151.21-0.0030.782

Formula

α = k/(k−1) · (1 − Σ s²ᵢ / s²_total)s²_total = Σ s²ᵢ + 2 Σ_{i<j} cov(i, j)

Forgetting to reverse a negatively worded item makes it covary negatively with the rest: the total's variance shrinks and α can fall below zero. That is a coding error, not low reliability.

Method: Course data sets: ds.index-120 (module 02 responses) and ds.efficacy-120 (one-factor model, seed 120). α, corrected item-total r, α-if-deleted and standardised α as SPSS Reliability Analysis. Attenuation: seeded normal error on every item; the criterion is Y = 0.6·z(clean index) + 0.8·noise, so its own reliability is 1.

Validation: α, α-if-deleted, corrected item-total r and standardised α match base R (covariance form) on four keyed item matrices to 1e-12, and the module 02 implementation agrees (tests/unit/measurement-lab.test.ts); a 20 000-case simulation confirms r_obs = r_true·√ρ_xx within 0.01. · numerically validated