ADAPTIVEREG contour plots

[cl5]

Hi SAS

Any help would be appreciated. I’m running SAS ondemand for academics using PROC ADAPTIVEREG. When I run model 1:

proc adaptivereg data=temp11 plots=fit;

model numtoday = date factor1;

run;

I get results 1 as attached. When I run model 2 with an additional factor being considered:

proc adaptivereg data=temp11 plots=fit;

model numtoday = date factor1 compound;

run;

I get results 2 as attached. The contour plot in results 2 looks completely different from that in results 1 and seem interpretable, even though the final models have the same factors retained with some difference in basis functions. 

I am interested that the "compound" factor gets dropped when running the model 2 statements but the strange plot makes me question the results. Can anyone shed light on why there's a discrepancy in the plots and what that means?

[Rick_SAS]

In addition to the fit plot, you should look at the diagnostic plots and the plots of the functional components. Because ADAPTIVEREG fits splines, it is hard to interpret by looking at tables.

I have two suggestions:

1. Use a different nonparametric procedure (such as PROC GAMPL) to see if the overall features in the predicted value are reproducible. You can look at the article "Nonparametric regression for binary response data in SAS,"  to see an example when the response variable is binary, but the same process will apply to models that have a normal response.

2. Your two models are most different in regions for which there is no data. You shouldn't worry about what the model does outside of the data region because that is extrapolation, which is generally not valid for this kind of regression. See the article "Truncate response surfaces" in which I discuss models that are nonsensical away from the data.