topic Best practices for modeling on an ordinal response target in SAS Data Science

Best practices for modeling on an ordinal response target

MZM — Mon, 27 Mar 2017 12:43:48 GMT

Hi Modellers,

I am trying to build a model on a dataset with 2 million observations and an ordinal response variable with 53 levels with a normal distribution ranging from -26 to +26.

Both SAS/Base 9.4 and Enterprise Miner 13.2 are available for use.

I am looking for any suggestions on modeling techniques that could be used especially in EM.

Was also wondering if there's a way to use cumulative logit link function in EM.

Does that make sense to consider the response variable interval and then to use GLM?

Thanks,

Re: Best practices for modeling on an ordinal response target

PaigeMiller — Thu, 23 Mar 2017 15:38:31 GMT

Unless your 53 levels are highly non-linear, I would treat them as a continuous variable and perform partial least squares regression modelling (which in my opinion is probably the best way to model 900 independent variables) in PROC PLS. I do not know if this is available in Enterprise Miner as I don't use it. I would not use PROC GLM with 900 independent variables.

Re: Best practices for modeling on an ordinal response target

WendyCzika — Thu, 23 Mar 2017 15:36:43 GMT

When you use the Regression node in Enterprise Miner with an ordinal target, it does use the cumulative logit link function.