Bioequivalence and Bioavailability Forum

Dear sample size geeks, nerds, and simul-ants (weirdos for short),

for each combination of design, power, GMR, and CV: Red circles = sample sizes of the Two Lászlós (if >201 none). Green bars = sampleN.scABEL(..., nsims = 1e5) or sampleN.RSABE(..., nsims = 1e5); always balanced sequences. Blue circles = twenty runs simulated with 10,000 sim’s each (random seeds). Starting from the respective result, I decreased the sample size whilst maintaining power. This might give unbalanced sample sizes (even sample sizes for the partial replicate and odd ones for the full replicate).

ABEL (EMA)

• 3-period 3-sequence partial replicate [TRR|RTR|RTT]
  • Power 80%
    
    
    
  • Power 90%
    
  
  
• 4-period 2-sequence full replicate [TRTR|RTRT]
  • Power 80%
    
    
    
  • Power 90%
    

Now a comparison of the reported sample sizes (y-axis) vs. PowerTOST’ (x-axis). Blue from the tables, and red up-rounded to the next complete sequence. Do the linear fits closer to the unity line? Sometimes yes, sometimes not.

• 3-period 3-sequence partial replicate [TRR|RTR|RTT]
  • Power 80%
    
    
    
  • Power 90%
    
  
  
• 4-period 2-sequence full replicate [TRTR|RTRT]
  • Power 80%
    
    
    
  • Power 90%
    

RSABE (FDA)

• 3-period 3-sequence partial replicate [TRR|RTR|RTT]
  • Power 80%
    
    
    
  • Power 90%
    
  
  
• 4-period 2-sequence full replicate [TRTR|RTRT]
  • Power 80%
    
    
    
  • Power 90%
    

Comparison of the reported sample sizes (y-axis) vs. PowerTOST’ (x-axis). Blue from the tables, and red up-rounded to the next complete sequence. Do the linear fits get closer to the unity line? Generally yes.

• 3-period 3-sequence partial replicate [TRR|RTR|RTT]
  • Power 80%
    
    
    
  • Power 90%
    
  
  
• 4-period 2-sequence full replicate [TRTR|RTRT]
  • Power 80%
    
    
    
  • Power 90%
    

The upper part of the RSABE sample sizes looks disturbing first. Explanation: Since the sample sizes are generally smaller than for ABEL, the many imbalanced sequences have a larger impact. However, rounding up to the next complete sequence (red circles) in the lower part gives a better match (closer to the unity line).

Maybe you can make sumfink out of it. I still hold that 10,000 simulations are too few.