Linear Constraint Breaking when Model Doesn't Include Parameter

Question

Currently I have 10 variables in my system and 2 responses. I also have a linear constraint placed that uses all 10 variables.

The first response is a function of all 10 variables (everything is significant) but the 2nd response is just a function of 5/10 of the variables. When I use the prediction profiler for the 2nd response I get the error

I assume this is because I’m not using all 5 variables in the model for the 2nd response so it can’t load the constraint. Is there anyway to keep a constraint on 10 variables (because it’s needed for 1 response) but still have it work when 5/10 of the variables are used.

cc: @ChristianStopp

ChristianStopp · Accepted Answer

@EndogenousCame1

One way I can think that may work would be to

* save the prediction formulas via the red triangle option
* modify the formula for the 2nd response to include each of the removed terms but with a 0 for the weighting (i.e. parameter estimate), such as “ + 0 * :Factor6”. 
* then use those formulas in the Profiler found under Graphs

I tested this out on a very simple example (3 factors, 1 response, main effects only model, and a constraint including all 3 factors), getting the same error message you saw when I reduced a term from  my model.  When I moved to the stand alone Profiler after modifying the formula as described above, the profiler worked as I had expected so hopefully it still retains functionality in a more complex set up like yours.  If your model included more complex terms (interactions, quadratics), I expect adding the zero-weighted terms to replace those removed in model reduction would be needed too.

I've attached the simple example with 3 add'l scripts to reflect the full model w/ constraints shows the profiler working, another with the reduced model and broken profiler, and the standalone working profiler using the modified prediction formula including the zero weighting. hope that helps!

Victor_G · Accepted Answer

Hi @EndogenousCame1,

I might have found a simple solution to help by mimicking your situation. I have the same problem when I try to use the Prediction Profiler and only 7 out of the 10 variables are in the Prediction formula/model :

Here is the trick: You can save the prediction formula of your response with 5 variables, and then modify manually this prediction formula to add other removed variables and give them a coefficient of 0 like this :

Running the Prediction Profiler on this modified prediction formula should remove the warning message and lets you use it normally :

Attached is the sample dataset used to mimick your situation with the two prediction formula saved (original and modified).

Hope this answer will help you and solve your problem,