beginner 8 min read
Social sciences · Topic
Reliability and Validity
probability theory · hypothesis testing · gaussian distribution
Reliability refers to the consistency of scores across time, raters, or parallel forms; validity concerns whether a test measures what it is intended to measure. Both properties are prerequisites for defensible score interpretation, and modern validity theory treats validity as a unitary concept supported by multiple sources of evidence.

Internal Consistency: Cronbach’s Alpha

Cronbach’s $\alpha$ estimates reliability from the inter-item covariance structure:

\[\alpha = \frac{k}{k-1}\left(1 - \frac{\sum_{i=1}^k \sigma^2_i}{\sigma^2_X}\right)\]

For $k$ items with equal correlations $\bar{r}$, this reduces to the Spearman-Brown form: $\alpha = k\bar{r}/[1 + (k-1)\bar{r}]$. Values $\ge 0.70$ are conventionally acceptable.

Inter-Rater Reliability

When scores depend on human judgment, rater agreement is quantified by:

Coefficient Application
Cohen’s $\kappa$ Categorical ratings, two raters
Weighted $\kappa$ Ordinal categories with partial credit
ICC(2,1) Continuous ratings, random raters, absolute agreement

The intraclass correlation (ICC) under a two-way random model is:

\[\text{ICC} = \frac{\sigma^2_P}{\sigma^2_P + \sigma^2_R + \sigma^2_{PR} + \sigma^2_e}\]

Convergent and Discriminant Validity

The Multitrait-Multimethod (MTMM) matrix tests whether a construct correlates more strongly with different methods measuring the same trait (convergent validity) than with the same method measuring different traits (discriminant validity). Campbell and Fiske’s criteria require:

  1. Validity coefficients exceed correlations in the heterotrait-monomethod triangles.
  2. Validity coefficients exceed correlations in the heterotrait-heteromethod triangles.

Formally this is now tested via a CFA comparing trait-only, method-only, and combined factor models.