This is another quote about the correlation issue (Page 130). The idea here is that you can randomly assign patients to group dose groups yet it is not as obvious how to randomly assign patients to SUVR amyloid reduction values as the trial is underway. In fact it is understood that various covariates will tend to influence amyloid clearance differently in the different demographics. I thought that one might be able to overcome this problem by appealing to the law of large numbers. Combining 301 and 302 did appear to average things out, though with 302 the strength of the covariates was able to overturn the strength of n. It is important to clearly understand this resolution of the correlation issue point because the lack of individual level correlation on the surface appeared to be a weakness in the aducan results.
Hmm, perhaps the MMRM should have included a term for placebos that became amyloid negative at week 78.
"Consequently, it is highly likely to achieve a balance of
prognostic factors (known and unknown) across dose groups (i.e., group-level), improving
the confidence in the underlying relationship.
On the contrary, if the patients were randomized at group-level, and the relationships
between endpoints were assessed at individual-level within a dose-level, such a balance
(in prognostic factors, known and unknown) can no longer be guaranteed (e.g., an
Reference ID: 48010710219
artificially “flat” relationship may arise because of potential imbalances in baseline
characteristics of patients). In other words, there can be multiple confounders across
different individual subjects, and when multivariate analyses are conducted, it is
extremely difficult to correctly adjust the imbalance of multiple confounders across
individual patients due to potential nonlinear relationship and complex interactions. This
challenge was well recognized in the field of pharmacometrics and FDA’s 2003 exposure response
) clearly explains these
challenges and recommends alternative methods."