Response surface Profiler Graph

[#1]

I've estimated a response surface with one DV and six IVs. A three-dimensional plot of the DV against two IVs is produced via red triangle > Factor Profiling > Surface Profiling. I would like to produce instead a graph of the DV against one IV, holding all others constant--i.e., I want a scatterplot of DV vs one IV, holding others constant, not a three-d plot of the DV and two IVs. Any ideas?

[#2]

Isn't what your after effectively the regular prediction profiler? Each plot visualizes the relationship between that factor and the response while holding all other factors constant.  Is it that you also need to plot design points on there too that match the same fixed settings for all your other factors?

[#3]

Thanks, Cameron. I need a publication quality graph, not a diagnostic tool.

For each of the IVs, I want to plot the scatterplot of DV vs IV, holding constant the other IVs at the values of the other IVs associated with the maximum value of the response surface...

[#4]

You can make a scatterplot for one of your IVs. Then add each of the other IVs to a local data filter. That will allow you to get every combination possible. You could then even specify a column switcher for your X-axis with all of the IVs. This is a very interactive approach, but will allow you to quickly specify whatever combination you want.

[#5]

Ok, I'm still a bit confused at the scatterplot part though.  Do you have actual data at your optimized set point, much less points covering the range of values for each factor being plotted while all other factors are held constant at their optimal set point?  Unless you have a full factorial and an optimal set point at one of your design points, I'm not sure what points you would plot.

If you want to just plot the prediction formula conditional on all but one factor at a time being held constant, I guess you could either do custom graphs that you script up from scratch, or edit the prediction formulas to hard code in the optimal set points for all factors but 1 and plot that function (again with the Profiler, but with just 1 plot at a time).

I've published using JMP graphics before, and I think you could polish up the Profiler to look publication-worthy.  Stretch it to a good size, edit the appearance of the axes, labels, titles, etc to your liking and I think it looks good.

[#6 · Accepted solution]

You've identified the problem I'm facing, Cameron. This goes back to my earlier post regarding retrieving both the validation R2 and training R2 from the boosted tree platform (thanks so much for your help!) Now that I have both R2s, I want to show how R2s changes with the learning rate, all else constant. The graph that you provided in my other post gets me started, but it shows how R2s change with learning rate for differing values of the tuning parameters (e.g., row sampling rate, col sampling rate, etc.) as stipulated by the design table.

So this is my latest thought. Fit Model > Validation R2 as the Y variable > the tuning parameters as the X variables > created second-order product terms of all tuning parameters using the Response Surface Macro > Save Prediction Formula. I could then change the prediction formula so that each of the other tuning parameters (except the learning rate) are set to their values from the maximum R2 from the design results.

Any reactions for me?

[#7]

Yep, that should work!