I have a question regarding bioequivalence assessment for semi-replicate 3 sequence design (ABB, BAB, BBA). Bioequivalence is (for EU) normally assessed with ANOVA factors period, sequence, treatment and subject(sequence) as fixed factors.
If I understand correctly, if subject(sequence) is a fixed factor, only subjects with data for both treatments (A and B ) are included in the analysis?
If this is true, why do PEs with their 90%CI differ in the following case?
If a subject with sequence BAB misses his/her 2. period and returnes to the 3. period, he/she has data just for treatment B (without data for treatment A - BXB). So if I do ANOVA with all four factors as fixed (including “subject(sequence)” as fixed) I get different PEs and their 90% depending on whether this subject with missing period 2 is included in the dataset or not. Why is this so? Shouldnt the results (PE with 90% CI) be the same whether or not this subject is included in the dataset or not because if we have factor “subject(sequence)” as fixed, this should exclude subjects that dont have data for both treatments from analysis?
LinMix is not pure linear model engine, it uses mixed modelling under the hood even if there are no random effects specified.
Citing Helmut:
PHX/WNL in its default setting (incomplete data included) will recover information of the between-subject variance (REML estimate of a mixed-effects model). In SAS-speak that’s Proc MIXED instead of Proc GLM. In other words, the variance will be differently split into their between-, within-, and residual components. Therefore, the CI (and CVintra) will be affected as well
Interestingly, I also get two different 90% CI with SAS PROC GLM for this case (whether or not subject with missing period 2 in sequence BAB is included in analysis or not). Is this also expected? :huh:
Sorry for delay but I’ve had some time off due to a family bereavement.
so the solution in this case is to manually exclude the whole subject’s data before the analysis?
I wouldn’t jump to exclude data before considering how you would explain to a reviewer/regulator, it depends on your statement in Protocol/SAP. At least it should be clearly and unambiguously noted in PK Report (population for BE analysis), so everyone can reproduce it.
Interestingly, I also get two different 90% CI with SAS PROC GLM for this case (whether or not subject with missing period 2 in sequence BAB is included in analysis or not). Is this also expected?
I think for the current situation SAS can get some contrasts for period effect of current subject and since period is an intraindividual effect, it makes the difference.
These are the sort things I would defer to statistical experts but generally I think it better to keep all data;
Hi! I’m also interested in bioequivalence assessment for semi-replicate 3 sequence design. I understand that if subject(sequence) is a fixed factor, only subjects with data for both treatments (A and B are included in the analysis. Could you help me, how does this affect the power of the analysis?