Proc logistics warning: Ridging has failed

[AmitKB]

Hi,
I am running proc logistics with backward selection. I am getting warning that Ridging has failed.
Can anyone explain what the warning means and some solution.

I appreciate all your help.

Thanks,

Amit

CODE:

proc logistic data=out.train1 NAMELEN=32;
model del123_target (event = '1') = &VarName
/ selection=backward fast sls=0.1;
run;

LOG:

NOTE: PROC LOGISTIC is modeling the probability that del123_target=1.
WARNING: Ridging has failed to improve the loglikelihood in Step 0. You may want to
use a different ridging technique (RIDGING= option), or switch to using
linesearch to reduce the step size (RIDGING=NONE), or specify a new set of
initial estimates (INEST= option).
WARNING: The LOGISTIC procedure continues in spite of the above warning. Results
shown are based on the last maximum likelihood iteration. Validity of the
model fit is questionable.
NOTE: Convergence criterion (GCONV=1E-8) satisfied in Step 1.
NOTE: There were 1315176 observations read from the data set OUT.TRAIN1.
NOTE: PROCEDURE LOGISTIC used (Total process time):
real time 8:19.57
cpu time 4:52.89

[jain]

Amit,

The warning in the log is just indicating that the iterative maximum likelihood optimization method for fitting the model failed to find a solution. As with any optimization program, this can happen for some models on some data sets. The message suggests some ways you can change the optimizer which *may* allow the procedure to find a solution. In fact, if the procedure does not converge, it presents the results based on the last iteration. This is not the maximum likelihood solution, so you may not even want to use the results. To fix the problem, you can try the suggested options. If they don't work, you may have to change your model... probably by reducing the number of parameters in it.

[StatDave · Accepted solution]

SELECTION=BACKWARD is usually a bad idea when there are many candidate effects in the model. The reason it is a bad idea is that the method starts with the most complex model which includes all of the candidate effects. This model is usually so complex and makes the data so sparse that the maximum likelihood solution does not exist. Better to use SELECTION=FORWARD or STEPWISE which starts with no effects (or only those forced by the INCLUDE= option) and built up a model.