Bioequivalence and Bioavailability Forum

Hi BEQool

❝ I have a question regarding df in partial replicate 2x3x3 design. 44 subjects were included in the PK/STAT analysis - 43 of them had data for all periods while 1 subject had data only for 2 periods (the missing period was with reference treatment B so the subject had data for both treatment A and treatment B). Both in SAS and Phoenix WNL the df of the design (Error) is 84. Why is this so? According to the formula 2*n-3 with 44 subjects included in the analysis, the df should be 2*44-3= 85. Or is this discrepancy (84 vs 85 df) a conseqeunce of one missing period for one subject (although the subject has data for both A and B and is included in the analysis)?

The residual df's are really just the number of observations minus the number of constants (=model effects) you try to estimate via your stats model.
In your example, you lost one observation but you are actually estimating the same number of constants or model effects. Therefore there will be one less in the residual DF than in the ideal case, cf 2n-3 or something along such lines.
To say it differently: Your model matrix has the same number of columns, but one less row.