1. When a test measures correlated latent traits, no matter the items are designed in simple or mixed structure, adopting the multidimensional model is efficient in estimating person proficiency, not item parameters.
2. In table 8, I think the ARMSE of estimated IRF derived from the SST cannot be compared to that derived from the MST. As defined in Equation 14, in the SST approach the discrepancy between true and estimated IRF denotes as a square measure, whereas in the MST approach it is actually a volume.
3. The Example 1 section on page 394 seems an impossible case. No one unless God is capable to design a set of items measuring a composite ability mixed of multiple latent traits with fix proportion because all slope parameters should be estimated. It could only take place in simulation.