How can I select variables in the data node of SAS EMiner? To define it more clearly, I would say if I have ~50 variables, how could I select / determine the strongest variables for my model?
By 'data node' I am assuming you mean the Input Data Source node. In general, you would be ill-advised to remove variables from consideration unless you knew they were not (most likely) suitable for direct use in modeling (e.g. ID information, date/timestamp information, SKU numbers, zip codes, etc...). It is also not necessary to choose variables in this node since SAS Enterprise Miner provides a wealth of methods to choose variables for your model such as the following:
* the Variable Selection node provides Regression and Tree-based methods for choosing variables
* the Tree node performs its own variable selection so it does not need prior variable selection
* the Regression node allows you to add possible terms and to perform a set of stepwise methods to perform variable selection
* the Variable Clustering node provides and alternate way of trying to remove variables which have duplicate or highly similar information
When I am running my model package, in the log I see "Pr > ChiSq" for most of the variables are 1.0000 - this explains my variables have some issue. But how to fix it. For example , I have a variable called claim_count and instead of keeping the values as continuous I grouped them in certain buckets. like 1-10,11-20.. etc
Bucketing manually without any numerical evaluation might actually do more harm than good. SAS Enterprise Miner provides a variety of bucketing algorithms which can take into account the relationship to the target variable.
* the Transform Variables node allow you to create bucket with optimal relationship to the target variable (a Tree-based method)
* Interactive Grouping allows you to create groups interactively
Please note that bucketing summarizes information and can (possibly) result in a predictor that is less capable than the original data. The buckets, however, do provide the additional capability of helping to model non-linearity which might improve how the variable information can be used. In practice, I recommend including both the original interval variable and the bucketed version prior to variable selection so that the information is used in the best possible way.
Hope this helps!
Doug
