Variance of a mean of correlated measures. Consider the mean of four measures (j=1,2,3,4) within group i, xi1, xi2, xi3, xi4, mutually correlated ICC=r because they share a common environment, and the variance of that mean. mean xi. = (xi1 + xi2 + xi3 + xi4)/4 var(xi.) = var([xi1 + xi2 + xi3 + xi4]/4) = (1/16)* var(xi1 + xi2 + xi3 + xi4) = (1/16)*{var(xi1)+var(xi2)+var(xi3)+var(xi4) +cov(xi1,xi2) + cov(xi1,xi3) cov(xi2,xi1) + cov(xi2,xi3) cov(xi3,xi1) +cov(xi3,xi2) cov(xi4,xi1) +cov(xi4,xi2) + cov(xi4,xi3) and assuming homoscedasticity of variances and covariances = (1/16)*{4xvar(X) + (42-4)xcov(Xij,Xik) } = var(X)/4 * { 1 + (4-1) xcov(Xij,Xik)/var(X) } = var(X)/4 * { 1 + (4-1) x ICC } or more generally, for m members in the group, var(xi.) = var(X)/4 * { 1 + (m-1) x ICC} + cov(xi1,xi4) + cov(xi2,xi4) + cov(xi3,xi4) } Common model (compound symmetry, exchangeability) xij = ui + eij xik = ui + eik where ui~N(0,τ2) and eij, eik ~ indepN(0,σ2res) so var(X)= (σ2res + τ2), cov(xij,xik) = τ2, and ICC = τ2/(σ2res + τ2), DEFF = 1+(m-1)ICC for Design Effect From the analysis of variance perspective, the natural measure of homogeneity is the variance component ratio, VCR = τ2/σ2res. Some algebra shows var(xi.) = σ2res/4 * { 1 + m x VCR) RVIF = 1 + m VCR for Residual Variance Inflation Factor. In the analysis of the school MATH data the MIXED model analysis of 311 students in 20 schools (mean number students=15.6) gave estimates of the components of variance school(cond) 262.67 residual 2150.33 with the ration (VCR) = 0.1222 and so Total variance 2413.00 ICC = 0.1222/(1+0.1222) = 0.1089 DEFF = 1 + 14.6*0.1089 = 2.590 inflating total variance 2413 so variance of the mean is Var(Xavg) =2413*2.59/311 = 20.0950 RVIF= 1+15.6*0.1222 = 2.906 inflating residual variance 2150.33 Var(Xavg) = 2150.33*2.906/311 = 20.0950