Bioequivalence and Bioavailability Forum

Hi Balaji,

❝ However could you please elaborate on the below . .

❝

❝ ❝ In the worst case the study is done and the optimizer fails to converge (independent from the software)

❝

❝ - What do you meanly optimiser fails to converge ?

As I wrote above:

❝ ❝ ❝ ❝ […] the partial replicate is a lousy design (since T is not repeated and the FDA’s model is over-specified).

My wording was not precise. The partial replicate design is not lousy per se – only model recommended by the FDA’s to evaluate if. The model attempts to estimate the intra-subject variability of T which is not uniquely contained in the data. Do you remember our conversation about intra-subject variability in a parallel design? Similar story.

❝ When does this happen ?

If one is not allowed to scale (s_wR <0.0294) and use the FDA’s mixed-effects model for ABE with the FA0(2) specification of the covariance structure according to the guidance. Such a situation is not uncommon: RSABE possible for C_max but not for AUC.

1. If one is lucky the software only throws a warning:
   Convergence criteria met but final hessian is not positive definite. (SAS)
   Newton's algorithm converged with modified Hessian. Output is suspect.
Model may be over-specified. A simpler model could be tried. (Phoenix/WinNonlin)
   However, the estimated s²_wT is nonsense. This is just annoying (the agency is not interested in the variability of the test product and one doesn’t have to report it). The PE and its 90% CI is still correct.
   
2. In certain cases the optimizers fails to converge. See John’s data set of AUCinf. In Phoenix/WinNonlin the same warning as above but in SAS:
   WARNING: Did not converge.
WARNING: Output 'Estimates' was not created.
   This is terrible because one gets no result for ABE at all. Study done, statistician blamed for it…

It is impossible to know beforehand whether one will end up with #1 or #2; search the forum for details. Ways out:

• State in the protocol that you will deviate from the guidance and use FA0(1) instead.
  
• Stick to the code given in the guidance but use one of the fully replicated designs (RTRT|TRTR, RTTR|TRRT, RTR|TRT, or RTT|TRR) instead. IMHO, much better.
  Furthermore, 4-period replicate designs require less treatments than 3-period replicates (see above). Unless one runs into problems with the total blood volume I don’t see any reason for performing 3-period replicate studies.

Note that these problems are apparent only with the FDA’s model for ABE. The fixed-effects model preferred by the EMA does not have any issues.

I asked numerous times here and in Certara’s Extranet why one wants to use a partial replicate design – and never received an answer.
May I ask: Why do you want to use it?