I'm using JMP to verify my own calculations, but have some results I'm unable to find out why they are different in JMP.
I've made a full quadratic model for a function with five variables, but the parameter estimates for the main (linear) effects (and intercept) are much higher in JMP than the ones I calculate myself. For the quadratic and interaction terms I get the same values.
If use the estimates from JMP, the model will not give the right values. When using my own estimates it will.
Can anybody explain or guide me to a reference, that explains the reason to this difference. I see a logic in the t ratio (which I'm interested in calculation to evaluate the variables) being larger for the main effects since these also have an influence on the quadratic and interaction terms, but haven't been able to find out how this influences the t ratio calculation from the references I've used.
Thanks in advance
Solved! Go to Solution.
I would guess it has to do with the way JMP constructs the contrasts between categories for ordinal and nominal factors or with polynomial centering which is the default in the fit model platform.
one option is to compare predicted values of alternative calculations to see whether they are different. in JMP it is very easy to use the prediction profiler for extracting a few predicted values. if predicte values are the same across calculations then it is just an issue of contrasts structure.
Thanks for your answer. I'm very new to statistics and all the different terms, but if I understand you right, you refer to discrete variables. I only have continuous variables in the model.
I've just found some information about the generel regression significance test/extra sum of squares method. Is this the right way to find the "real" significance of the variables. The books I use, doesn't mention anything about interactions, but I've found some information online, that it's not possible to use the "normal" t-test when there are interactions. This is only valid when the interaction terms are zero.
Thanks again. That gives me the right estimates.
I still don't understand how (and precisely why) the calculations are done. Do you know of any references that explains how and why?
Thanks. I found that in the meantime, but that's only the why (which I think I understand).
Then there's the how, since I do my calculations manually. Haven't been able to find anything by googling.
When you have fit your model in JMP, try saving the prediction formula. Now look at that formula from your data table. You will see how JMP has done the coding/centering. If you want to "uncode" the model, click the red triangle in the formula editor and choose simplify.