nav[aria-label="Primary Navigation"] { padding: 0; & ul { list-style: none; width: 100%; display: flex; flex-direction: row; justify-content: start; align-items: start; gap: 30px; padding: 0; & li { margin: 0; } & ul li { list-style: none; } } }

Creating two factor interaction plot (without the full matrix profile)

CanonicalHazard

How do I have JMP give a clearer, bigger, simpler two factor interaction plot for just two factors that I hand select for the model that has been fit with Fit Model?

JMP gives a profiler, like so, but it is a giant matrix and frankly not particularly readable if one wants to focus well on one particular two factor interaction.

I see no way to tell it I don't want to see the whole matrix and to select the factors as desired.

Creating in graph builder using the raw data gives a different result and I'm not clear why (I'm an engineer not a statistician), but I suspect it has something to do with the fact the model in the interaction profile in the Fit Model area doesn't include all the possible model terms as many were insignificant. Here is the AB interaction. It is bigger and clearer, but according to the Fit Model profiler, it is wrong, very wrong.

Accepted answers

CanonicalHazard

Hi @Victor_G ,

Appear to have the solution, though it is rather involved. From JMP Technical Support:

Since you have all continuous variables that you wish to exclude, I think the following approach should work.

If you do the following I think you will have the same Interaction Plot in Graph Builder that you see in your Least Squares report...

-- Run your model as you normally would and turn on the Interaction Plots.
-- Click the hotspot in the top left and choose Factor Profiling > Profiler to turn on the Prediction Profiler (if it is checked it is already on).
-- Now go down to the Profiler and click directly in the plot for A while holding down the Alt key (to specify Factor Settings).
-- Check the Lock Factor Setting box and click OK.
--Continue to do the same to Lock Factor Settings for every factor except the two you are interested in (D and I).
-- Now click the hotspot associated with the Prediction Profiler and select Output Grid Table.
-- Open Graph Builder and build the interaction plot from this new table (i.e., Response as Y, D as X, I as Overlay).
-- This should be the same plot you see in the Least Squares report, except that the axes may be different and there are traces for each of the curves. You can copy and paste the axes from the original Interaction Plot to Graph Builder if you like:
-- Right click on the x-axis of the original Interaction Plot and choose Edit > Copy Axis Settings.
-- Right-click on the x-axis of your plot in Graph Builder and choose Edit > Paste Axis Settings.
-- Do the same for y-axis.

An example below of response Y3 for the two factor D x I interaction -- I had transformed Y3 with a log (time to failure):

Using local data filter:

The message suggests something else may happen when the model contains non-continuous predictors. I'll deal with that when it happens.

All comments

Victor_G

Hi @CanonicalHazard,

It should be possible to do it with Graph Builder, provided that you saved and used the Prediction formula of your response from your model.

As an example, I used the dataset "2x3x4 Factorial" available in the JMP sample index.

When building the model, it seems that two interactions effects are significant :

The interactions can be seen in the "Interaction Profiles" panel :

I save the Prediction formula for the response Y, and use it in the Graph Builder, use X1 as X and X2 as "overlay" role variable in the Graph Builder and I am able to find the same interaction plot than seen before :

I can also do it with X2 as X and X1 as overlay variable, even if X2 is a categorical factor, as this factor has a "Value Order" property, so JMP knows in which order to display the levels. I end up with this graph, very similar to what was shown previously in "Interaction Profiles" panel :

And here with X3 as X variable (and removing points from Graph Builder to make the visualization more simple) :

Of course you can customize the graph, removing the points or keeping them, changing the colors, visualisation styles, ... Take also care that the Y range of the graph is the same than the predicted Y values range, in order to have the same slope and behaviour for your interaction (and not have misleading visualization).

I hope this answer will help you,

CanonicalHazard

Thanks. I see that works for a number of data sets. For some unknown reason it does not work with mine.

Victor_G

Hi @CanonicalHazard,

Would it be possible to provide the dataset in an anonimized way ?
It would be much easier to investigate this topic with the same data to see what is going on.

CanonicalHazard

Hi @Victor_G ,

Attached is an experiment, including because it has the additional element of a 3rd level for each factor. Check out the CG interaction in the "Fit Least Squares" interaction profiler. How to pop that out separately without the entire matrix of all the factors is my goal. There must be a command somewhere to do this. Saving the Y predicted formula then using graph builder does not recreate what is shown in the profiler.

Profiler:

Graph builder of Y predicted formula:

Interaction Plot Example.jmp

hogi

Please note that the curves depicted via Prediction Profiler depend on the current setting in the profiler.

E.g. for "2x3x4 Factorial", with 0, y, C, the curves look similar to the curves in
https://community.jmp.com/t5/Discussions/Creating-two-factor-interaction-plot-without-the-full-matrix/m-p/665355/highlight/true#M85368 :

here are some other interaction profiles for the setting: 1, x, D

Jmp uses the full prediction model to depict the interactions.
Variables which are not specified by the position in the matrix (group by) or on the x axis are taken from the settings of the prediction profiler.

On the other hand, when you generate a Interaction Plot in Minitab, it takes the raw values.

Variables which are not specified by the position in the matrix (group by) or on the x axis are removed by taking the mean over all existing values.

So, if you select some "mean" values in the prediction profiler, the Interaction Profiles in Jmp might look similar to the Interaction Profiles in Minitab.

hogi

The "Minitab version" can be created with the below code.
Instead of using the raw values, one could also use the Predicted values (-> column switcher).

To get the functionality of the Jmp Interaction Profiles, one would need a functionality to specify fixed values (from prediction profiler) for the third (or in general: additional) dimension(s).

dt = Open( "$SAMPLE_DATA/Design Experiment/2x3x4 Factorial.jmp" );

	myFit= dt <<
	Fit Model(
		Y( :Y ),
		Effects( :X1, :X2, :X3, :X1 * :X2, :X1 * :X3, :X2 * :X3 ),
		Personality( "Standard Least Squares" ),
		Emphasis( "Effect Screening" ),
		Run(
			Profiler(
				1,
				Confidence Intervals( 1 ),
				Desirability Functions( 1 ),
				Term Value(
					X1( 0, Lock( 0 ), Show( 1 ) ),
					X2( "x", Lock( 0 ), Show( 1 ) ),
					X3( "A", Lock( 0 ), Show( 1 ) )
				)
			),
			:Y 		)
	);
myFit  << Predicted Values;


dtLabels= New Table( "labels",
	Add Rows( 3 ),
	New Column( "variable", Character,Nominal,
		Set Values( {"X1", "X2", "X3"} )
	)
);

myDT = dt << Join(
	With( dtLabels ),
	Cartesian Join,
);

myDT << New Column( "Legend",
	Character, 
	Nominal,
	set each value( Char(As Column( :variable[] )[]) )
);

myDT << Graph Builder(
	Show Control Panel( 0 ),
	Variables(
		X( :X1 ),
		X( :X2 ),
		X( :X3 ),
		Y( :Y ),
		Group Y( :variable ),
		Overlay( :Legend )
	),
	Elements(
		Position( 1, 1 ),
		Points( X, Y, Legend( 1 ) ),
		Smoother( X, Y, Legend( 6 ) )
	),
	Elements(
		Position( 2, 1 ),
		Points( X, Y, Legend( 2 ) ),
		Smoother( X, Y, Legend( 3 ) )
	),
	Elements(
		Position( 3, 1 ),
		Points( X, Y, Legend( 4 ) ),
		Smoother( X, Y, Legend( 5 ) )
	),
	Column Switcher( :Y, {:Y, :Predicted Y}, Title( "choose column" ) )
);