Weird behaviour of gamlj 2.5.3 GLM

[mcfanda@gmail.com]

Hi,
thanks, but maybe I'm not following your reasoning well. In general, in the general linear model, SS_model is not necessarily equal to SS_effects. It is not a bug.
If you have a balanced design for Genre and Volture, try removing the covariate Nbr and you'll see that SS_model=SS_effects. If Genre and Volture are not balanced or a covariate is in the model, SS_model is not necessarily equal to SS_effects.

[mcfanda@gmail.com]

Here is an example from Cohen and Cohen exercise example. Two continuous variables and their interaction (raw variables).

spss.png (28.8 KiB) Viewed 10122 times

here the same in gamlj

gam.png (24.23 KiB) Viewed 10122 times

You can see the results are identical (granted rounding).

You can also see the SS_model=5887, whereas SS_effects=2161+387+1136=3684
You can also see that SS_total=SS_model+SS_res=5887+22675=28561.

[mcfanda@gmail.com]

To finish up, here is an example with balanced data. In the first analysis, we just have two balanced factors, so SS_model=SS_effects

gam2.png (23.52 KiB) Viewed 10122 times

Now, adding a covariate change the sums and we get SS_model != SS_effects

gam3.png (23.16 KiB) Viewed 10122 times

[mcfanda@gmail.com]

The same is SPSS

spss2.png (22.27 KiB) Viewed 10122 times

spss3.png (22.77 KiB) Viewed 10122 times

[Arnaud Mortier]

My goodness, this is a part of the theory that I'd always completely overlooked. What a shame. Thank you so much for your help!

[Arnaud Mortier]

Hi,
So there is definitely some part of the theory that I overlooked, but for the record I found a computer with the older 2.6.1 still in place, so I can provide a proof that at some point 2.6.1 used to show SS_model = SS_effects.
Thanks again!
Arnaud

Current version and older version:

current.png (71.22 KiB) Viewed 10091 times

older.png (65.89 KiB) Viewed 10091 times

[mcfanda@gmail.com]

No problem, the important thing is that it is fine now. Thanks for the feedback