Figure 3:

Validity $(\mathit{\text{va}}{r}_m\left(\widehat{\theta }\right)/\mathit{\text{va}}{r}_s\left(\widehat{\theta }\right))$ of Root 1 for both cases, for three designs and a variety of true ACC’s. The three classic designs (M=2, M=3, and M=6 sequences) each have K=20 observations per cluster per time period. Throughout, ${\sigma }_t^2=1$. Each ACC is achieved in three different ways by adjusting the balance of ${\tau }_t^2$ and ${\gamma }_t^2$ or ${\eta }_t^2/2$. If ${\gamma }_t^2$ or ${\eta }_t^2/2={\tau }_t^2$, 50% of the average between-cluster variance (the numerator of the ACC) comes from random intercepts. Similarly, if ${\gamma }_t^2$ or ${\eta }_t^2/2={\tau }_t^2/10$, then 90% of the variance comes from random intercepts and if ${\gamma }_t^2$ or ${\eta }_t^2/2={\tau }_t^2*10$, then 10% of the variance comes from random intercepts.