Ridging has failed to improve likelihood function: Proc Logistics

[vishal_prof_gmail_com]

Hi ,

I am running Proc Logoctics procedure.I got the message "Ridgin failed to improve the likelihood function". I changed the option ridging = absolute and this also gave the same message.

With Ridging = none  , the no. of iteration increased , but still the model ended with same warnning . This time the message was to use halfmax=option and increase no. of steps oe specify a new set of initial parameter estimates using INEST option. What does this mean?.

I have 26 predictor variables in the model , is this because of large no. of variables. If so then how do we reduce the variables.

Vishal

[JacobSimonsen]

I had same experience in PROC PHREG. Same message: "ridging failed to improve likelihood".
When I wrote "ridging=none" the model converged.

[pearsoninst]

As per as my thought are concerned it is due to a bad model .You data might have to much of outliers ,super multicollinearity,data not selected correctly ,too many variables and so on ....There is no way machine can fix the issue.
Reduce the variable , do some one to one regression analysis ,check VIF etc . there is no clear cut answer for this as of now at least from my side

[Kazzie]

As already pointed out, this is probably a data issue. 

I encountered the same warning and for me the problem was so-called perfect separation (a lot can be found by simply googling this term). Loosely explained, this means that for a certain categorical variable, one or more categories / groups only have data for either your success or non-success group. 

For example: 

Variable color that can take on Blue, red, yellow. 

Y = 1 (success): Blue 100, red 50, yellow 50

Y = 0 (fail / non-success): Blue = 50, red = 100, yellow = 0 

For the Yellow group, there are only successes. 

HTH

[StatDave]

It is not really possible to know exactly what caused this, but you can try various combinations of the TECH= and RIDGING= options. You could also try using the FIRTH option which maximizes a penalized likelihood. Since you have a fair number of predictors, you might be flirting with separation issues, so the the FIRTH option might help. Also you could consider using a model selection method (but don't use BACKWARD). The more modern Lasso method is available in PROC HPGENSELECT.