Bioequivalence and Bioavailability Forum

Dear Raveendranath,

❝ I tried with the following procedure.

❝

❝ proc glm data=one outstat=tanova;


❝ class trt per seq sub;


❝ [...]


❝ run;

❝

❝ But I didn't get Intra CV for test & Reference separately.

Correct. There is no way to specify different residuals for groups within Proc GLM. This is because all estimates and tests rely not on such different variabilities but only on the pooled error MSE.

You have two options:

1. Use the sequence group TT and RR to calculate differences T-T or R-R within each subject of the corresponding sequence group, respectively. The variance of these differences have the expectation 2*sigma²_{within T} for the sequence group TT and 2*sigma²_{within R} for the sequence group RR. So you can estimate the intra-subject variance by simply calculating the sample variance of the appropriate differences. This method is sometimes called "Method of moments". To get rid of the factor 2 you can start the calculations by
   (T1-T2)/sqrt(2) and (R1-R2)/sqrt(2).
   
2. Use Proc MIXED instead of Proc GLM. The code in Appendix E of the FDA guidance "Statistical approaches ..." works also for the Balaam design. But the standard estimation method for Proc MIXED is REML (resticted maximum likelihood est.) and is somewhat different to the least square estimation within Proc GLM. For balanced designs and no missing values the differences are only marginal.
   Since the FDA model is in most cases somewhat over-specified you can consider the simpler model of the between subject variances with compound symmetry.
   In this model the subject by treatment interaction is considered as zero.

BTW: My recommended code for Proc GLM is found here.
BTW2: What was the reason that you performed your study according to the Balaam design? Just to cite from the FDA guidance:
" ... For the majority of drug products, two-period replicated crossover designs such as the Balaam design (which uses the sequences TR, RT, TT, and RR) should be avoided ...".