I am helping a student use a traditional two-way ANOVA with JMP 7. We are using 'fit model' and two categorical variables and their interaction. I am using the "Analysis of variance" output for the model SS, the error SS and the total SS. I am using the "effect tests" to get SS and F and p for each variable. (Is this correct so far?).
The problem is that the SS for the parameters is not adding up to the "model" SS. It is pretty close, but not exact. Is a small discrepancy to be expected? Do people typically report SS for parameters these days (I usually use general linear models than traditional ANOVA's so feel out of touch on this one).
Thanks in advance for any assistance!
putting it in simple terms, the different sum of squares you mention are not of the same "Type". for a detailed explenation look up Type 1 sum of squares and type 3 sum of squares.
sticking to a simple explanation, the SS in the effect test table represents the "clean contribution" of that effect to the model. in order to evalue it from the model SS you would need to run the mode with and without the specific effect. in the two different models estimated you can compare the SS of the "full" model (with the effect) and of the "base" model (without the effect) and notice that the difference in SS (of the whole model) is exactly equivalent to the SS of the effect.
hope this gives you a first clue.
Your answer is helpful and I did find the Type 1 vs Type 3 (and a few other types) and reference to 'partial SS' (gosh these words ring dim memories of stats class a few decades ago). Perhaps it is most appropriate to just give the F-score for the full model and the F and p-values for each parameter. It seems the SS won't really give useful information beyond that.
since the effect SS is the "net" contribution to the model i use it to rank the effects. i.e. the effect with the highest SS is the most predictive within the context of the specific model in hand.
a little tip: right click on any output table to sort it by column. in this case sorting by SS will result in the effects ranked by net contribution to explanation of the variance in Y.