Concept map / Measurement invariance
Share
Measurement invariance
Change one item's loading and intercept in a second group and see when a gap in observed scores reflects a real difference in the trait.
1Group B scores 0.30 higher: the trait or one item?
Two groups take the same four-item scale, and group B’s average total is 0.30 points higher. Group B may have more of the trait, or one item may work differently in group B. The totals alone cannot say which.
Measurement invariance means the link between the trait and the item scores is the same in every group. It is checked in steps with the factor model: configural (the same items load on the same factor), metric (equal loadings) and scalar (equal loadings and intercepts). Observed means can be compared as trait means only under scalar invariance.
Items 1, 3 and 4 work the same way in both groups. Change item 2’s loading λ and intercept ν in group B, and move group B’s latent mean κ.
- ConfiguralSame items load on the same factor in both groups
- MetricEqual loadings
- ScalarEqual loadings and intercepts
Item 2 has loading 0.70 in group A and 0.70 in group B, and intercept 2.80 and 2.80. It reaches scalar invariance. The two groups have the same latent mean, and group B’s expected total score is the same as group A’s. Read as if every item were invariant, the totals suggest a trait gap of 0.00, which matches the real one.
More controls
The other loadings are 0.80, 0.90 and 0.60, and both latent variances are 1.
2A weaker loading in group B: does the item mean drop?
Both groups have latent mean 0 and variance 1. In group B, item 2’s loading drops from 0.70 to 0.40, and its intercept stays at 2.80. How does group B’s mean on item 2 compare with group A’s?
3An item mean is the intercept plus loading times κ
In a one-factor model, item i in group g has expected score νig + λigη at trait level η. Averaging over the group replaces η with the group’s latent mean κg. Point at a symbol to find it in the chart.
For item 2: group A’s mean is 2.80 + 0.70 × 0 = 2.80, and group B’s is 2.80 + 0.70 × 0.00 = 2.80. The other three items have the same parameters in both groups, so each of their means moves by its loading times 0.00.
So a total-score gap mixes the latent mean difference, weighted by the loadings, with any intercept differences. Here it is 0.00 from item 2’s intercept plus 3.00 × 0.00 from the latent means, 0.00 in all. Only with equal intercepts does the gap reflect the latent means alone.
4Check yourself
The tasks use the three sliders in section 1.
- Give item 2 metric but not scalar invariance: the same loading in both groups and a different intercept.Now: Loadings 0.70 and 0.70; intercepts 2.80 and 2.80
- Keep the latent means equal and still produce a total-score gap of at least 0.30 points in either direction.Now: Latent mean gap 0.00, total-score gap 0.00
- Hide a real difference: move group B’s latent mean at least 0.50 away from group A’s, yet make the two groups’ means on item 2 equal.Now: Latent mean gap 0.00, item 2 means 2.80 and 2.80
5Configural, metric and scalar fits in lavaan
The code simulates 500 people per group from the page’s model, with item 2’s intercept raised by 0.30 in group B and equal latent means, the state the second task asks for. It fits configural, metric and scalar models with lavaan::cfa, using group = and group.equal, and compares them with likelihood ratio tests. With this seed the total-score gap is 0.32. The metric model fits as well as the configural one (chi-square difference 3.91 on 3 degrees of freedom, p = 0.27), and the scalar model fits worse (27.67 on 3, p < 0.001). Freeing item 2’s intercept with group.partial gives partial scalar invariance and a group B latent mean of 0.02, against 0.05 when every intercept is forced equal.
Configural, metric and scalar models with lavaanShow the code
library(lavaan)
# Four items, one factor. Item 2's intercept is 0.30 higher in
# group B; both groups have latent mean 0 and variance 1.
set.seed(1)
lam <- c(0.8, 0.7, 0.9, 0.6)
nu_a <- c(3.0, 2.8, 3.2, 2.9)
sim <- function(n, nu) {
eta <- rnorm(n)
sapply(1:4, function(i)
nu[i] + lam[i] * eta + sqrt(1 - lam[i]^2) * rnorm(n))
}
d <- data.frame(rbind(sim(500, nu_a),
sim(500, nu_a + c(0, 0.3, 0, 0))))
names(d) <- paste0("y", 1:4)
d$group <- rep(c("A", "B"), each = 500)
# Observed total-score gap, B minus A
round(diff(tapply(rowSums(d[, 1:4]), d$group, mean)), 2)
m <- "f =~ y1 + y2 + y3 + y4"
eq <- c("loadings", "intercepts")
configural <- cfa(m, d, group = "group")
metric <- cfa(m, d, group = "group", group.equal = eq[1])
scalar <- cfa(m, d, group = "group", group.equal = eq)
lavTestLRT(configural, metric, scalar)
# Group B latent mean: full scalar vs item 2 intercept freed
partial <- cfa(m, d, group = "group", group.equal = eq,
group.partial = "y2~1")
round(c(scalar = coef(scalar)[["f~1.g2"]],
partial = coef(partial)[["f~1.g2"]]), 2)Uses lavaan. Install once with install.packages(c("lavaan")).
What invariance protects
- A gap in observed scores is a gap in the trait only when the items relate to the trait the same way in both groups.
- The steps add constraints one at a time: the same pattern of loadings, then equal loadings, then equal intercepts. Each step is tested against the one before.
- A loading difference changes variances and covariances, and it changes item means only away from the point where the two lines cross. An intercept difference shifts the mean directly.
- An intercept shift on a single item can make a total-score gap between groups with equal trait means, or hide a real difference. Freeing that one intercept while keeping the rest equal, called partial invariance, can still give comparable latent means.
- In item response theory the same question is asked item by item as differential item functioning, and invariance is part of the validity argument for comparing groups.
Recommended reading
- Confirmatory Factor Analysis for Applied ResearchTimothy A. Brown2nd ed., the chapter on measurement invariance and latent mean differences