Measurement invariance tests whether measurement models function identically across groups. Levels include configural (same structure), metric (equal loadings), scalar (equal intercepts), and strict (equal residuals). SEM procedures test increasingly restrictive models; partial invariance (some parameters equal, some free) often best represents reality. Without invariance, group comparisons are problematic.
Your work on differential item functioning (DIF) gave you a tool for asking, at the item level, whether a specific test question performs differently across groups after controlling for the underlying trait. Measurement invariance extends this logic to the level of the entire measurement model: does the construct you're measuring have the same meaning, captured through the same measurement structure, in both groups? If it doesn't, comparing group means on your scale is like comparing distances measured with slightly different rulers — the numbers don't mean what you think they mean.
The levels of invariance form a hierarchy of increasingly restrictive constraints, and each level is easiest to understand in terms of what a factor model actually does. In a CFA (confirmatory factor analysis) model, each observed item score is related to the latent factor through two parameters: a factor loading (the slope — how much item scores change per unit increase in the latent trait) and an intercept (the item's baseline value when the latent factor is at zero). Configural invariance requires only that the same general factor structure — which items load on which factors — holds in both groups. This is the minimum: both groups are measuring *something analogous*. Metric invariance adds the requirement that factor loadings are equal across groups, meaning items respond to the factor with the same sensitivity in both groups. The yardstick has the same unit size. Scalar invariance further requires equal intercepts: not only is the unit size the same, but the zero point is the same. This level is required before you can meaningfully compare latent mean differences between groups. Strict invariance adds equal residual variances — seldom required and seldom achieved.
In practice, partial invariance — where some loadings or intercepts are constrained equal and others are freed — is common and often defensible. If three of four intercepts are invariant, you can still compare latent means if you anchor the comparison on the invariant items and acknowledge that the non-invariant item may be functioning differently (perhaps reflecting a genuine cultural difference in how a concept is interpreted, not just measurement artifact). The key is to test rather than assume, and to report what you find honestly.
The testing procedure involves fitting a sequence of nested CFA models with progressively tighter constraints and comparing fit at each step. Start with the configural model (most free), then add metric constraints, then scalar constraints. At each step, compare fit using chi-square difference tests or fit index changes (ΔCFI ≥ .010, ΔRMSEA ≥ .015 signal meaningful misfit from the added constraints). When a constraint fails, examine modification indices to identify which specific loadings or intercepts are non-invariant. This gives an empirically grounded answer to a question that was previously left to assumption.
The stakes in applied research are high. A researcher comparing depression scores between cultures without testing measurement invariance may report a mean difference that is a measurement artifact rather than a true difference in depression. Conversely, establishing scalar invariance before reporting cross-group comparisons provides strong evidence that the comparison is fair and interpretable. Measurement invariance is therefore not a technical footnote — it is the empirical precondition for the most common use case in applied psychology: asking whether two groups differ on a construct of interest.