ERROR: Model hierarchy is violated with INCLUDE=1 in PROC LOGISTIC

[Haris]

Can't find anything in SAS Documentation about this error.  What does it even mean?

Here's the PROC LOGISTIC that gives the error:

proc logistic data=AnalyzeData desc;

  class Covar11 Covar12;

  model Outcome= Covar1*Covar2 ... Covar40

  / Selection=stepwise Include=1 SLStay=.2 SLEntry=.2

           LackFit NoDummyPrint RSquare;

run;

Note that the first argument is an interaction term.  Removing the interaction from the covariate list or switching the selection method to BACKWARD eliminates the problem.

[stat_sas]

Hi,

The INCLUDE option directs PROC LOGISTIC to keep the first variable listed in the MODEL Statement in every model it creates during execution of this task. Covar1 appears as a covariate with Covar2. Try putting like Covar1 Covar1*Covar2....may solve the problem.

[Haris]

I want to force the interaction, not Covar1.  BTW, replacing the string 'Covar1*Covar2' with 'Covar1|Covar2' also eliminates the error.  Even though the INCLUDE=1 statements specifies only 1 variable, all three components of the Covar1|Covar2 get forced into the model.  This happens to be what I want in this particular case, but what if I only wanted to force the multiplicative term? 

I still don't understand what Hierarchy is violated.  Do the simple effects of the interaction term have to be in the model for it to run?  I can see how that makes sense statistically but not from the strict programming point of view.

[SteveDenham]

In fact, the simple effects of the interaction term MUST be in the model, using the algorithms that SAS implements for model selection. See the HIERARCHY option for details, where HIERARCHY=SINGLE is given as the default. HIERARCHY=NONE ought to set things so that it would include the interaction but not the simple effects.

Steve Denham

[Haris]

Thanks!  It would be helpful if SAS included the reference to the HIERARCHY option in the error message.

[SteveDenham]

Well, if they did that consistently across procedures, I guess half of the traffic in this forum would disappear :smileymischief:

Steve Denham

[Haris]

You're right, Steve.  What would we do then?  Actual work!?!?!:smileyshocked:

[SteveDenham]

This is what I do while PROC GLIMMIX is running

And running

And running

Steve Denham

[Haris]

I recommend that you switch to PROC NLMIXED.  You'll have a lot more time for this

[SteveDenham]

I have tried.  I usually spend a lot more time writing code with NLMIXED, so I can't just set something up and let it cook for a couple of hours.

I was afraid you were going to point me at MCMC... (or even scarier point MCMC at me.

Oh yeah, one other thing.  Stopping Stepwise by Flom and Cassell.

Steve Denham