Miraculouser oracle [BE/BA News]
❝ ...
❝ Partial Tests of Model Effects
❝ Hypothesis Numer_DF Denom_DF F_stat P_value
❝ ----------------------------------------------------------------------
❝ int 1 17 1088.59 0
❝ Stage 1 17 17.3504 0.000648307
❝ Sequence 1 17 0.369037 0.551559
❝ Sequence*Stage 1 17 2.22777 0.153875
❝ Sequence*Stage*Subject 16 17 24.9524 1.04115E-08
❝ Stage*Period 2 17 0.967076 0.400153
❝ Treatment 1 17 0.0454252 0.833759
Seems to me like all the F-tests are done with MSE as denominator.
My beasty SAS gives almost the same:
Source DF Type III SS Mean Square F Value Pr > F
treatment 1 0.00208481 0.00208481 0.05 0.8338
period(stage) 2 0.08876896 0.04438448 0.97 0.4002
sequence 1 0.01436992 0.01436992 0.31 0.5831
stage 1 0.79630537 0.79630537 17.35 0.0006
sequence*stage 1 0.10224489 0.10224489 2.23 0.1539
subjec(sequen*stage) 16 18.32323177 1.14520199 24.95 <.0001
The question is: Are these tests appropriate?
Like the test for the sequence effect in the classical 2x2x2 crossover I would suppose to test stage and sequence as between-subject effects and also the new introduced interaction sequence*stage with subject(stage*sequence) as denominator. But I'm not quite sure.
The as bogus suspected
RANDOM
statement in Proc GLM
may assist is in getting the 'right' F-tests. But as always when the Almighty miraculous oracle prophesies difficulties are just around the corner (at least for statistical sermons).Formulating
RANDOM stage sequence subject(sequence*stage) / test;
gives me the anticipated denominator
Source DF Type III SS Mean Square F Value Pr > F
treatment 1 0.002085 0.002085 0.05 0.8338
period(stage) 2 0.088769 0.044384 0.97 0.4002
subjec(sequen*stage) 16 18.323232 1.145202 24.95 <.0001
Error: MS(Error) 17 0.780224 0.045896
Source DF Type III SS Mean Square F Value Pr > F
sequence 1 0.014370 0.014370 0.01 0.9122
* stage 1 0.796305 0.796305 0.70 0.4166
sequence*stage 1 0.102245 0.102245 0.09 0.7689
Error 16 18.323232 1.145202
Error: MS(subjec(sequen*stage))
* This test assumes one or more other fixed effects are zero.
But! The F-test for stage is only valid if sequence*stage is assumed zero.
Formulating
RANDOM stage sequence sequence*stage subject(sequence*stage) / test;
gives:
Source DF Type III SS Mean Square F Value Pr > F
treatment 1 0.002085 0.002085 0.05 0.8338
period(stage) 2 0.088769 0.044384 0.97 0.4002
subjec(sequen*stage) 16 18.323232 1.145202 24.95 <.0001
Error: MS(Error) 17 0.780224 0.045896
Source DF Type III SS Mean Square F Value Pr > F
sequence 1 0.014370 0.014370 0.16 0.7743
Error 0.7141 0.062612 0.087676
Error: 1.014*MS(sequence*stage) - 0.014*MS(subjec(sequen*stage)) - 22E-16*MS(Error)
Source DF Type III SS Mean Square F Value Pr > F
stage 1 0.796305 0.796305 7.79 0.2190
Error 1 0.102245 0.102245
Error: MS(sequence*stage)
Source DF Type III SS Mean Square F Value Pr > F
sequence*stage 1 0.102245 0.102245 0.09 0.7689
Error 16 18.323232 1.145202
Error: MS(subjec(sequen*stage))
No warning, but ... Rather wild denominators in the F-tests for sequence and stage!
Especially unique for sequence with non-interger df's

Any opinion out there?
BTW: The whole mess is "für die Katz" (strictly for the birds). The confidence intervals are not affected at all.
with sequence*stage 101.454363 (88.445194 ... 116.377016)
without 101.454363 (88.445194 ... 116.377016)
Regards,
Detlew
Complete thread:
- EMA: Q&A update (Two-Stage designs) Helmut 2013-03-08 14:47
- EMA: Q&A update (Two-Stage designs) ElMaestro 2013-03-08 15:12
- EMA: Q&A update (Two-Stage designs) d_labes 2013-03-08 15:40
- Almighty miraculous oracle Helmut 2013-03-08 16:43
- Miraculouser oracled_labes 2013-03-12 12:12
- Almighty miraculous oracle Helmut 2013-03-08 16:43