topic Re: Question about input vars in a proc logistic in Statistical Procedures

Question about input vars in a proc logistic

juanvg1972 — Fri, 13 Apr 2018 09:11:56 GMT

I have some question about input vars in a proc logistics procedures:

1) I have a var that have int values from 10 to 30 (10,11,12...30). I don't if use it as 'quantitative' or 'classification'. I use 'quantitaive' for vars like sales (from 0 to 1000 and with decimals). Is there any criteria??

2) I have a qualitatitve var that have values 'A', 'D', 'T'. I can use as input var as 'classification'. But I have noticed that the target value changes a lot with the changes of values like this: A ( var target ascend), D (var target descend), T (var targer strongly ascend), I am thinking in use this mapping D ->1, A -> 2, T -> 3 and then I have a linear correlation between the target var and the input var. Can I use then this transformed var as a 'quantitaive' var. Does it make sense?

3) last question: can I get the cumulative lift chart or similar as output in proc logistics. I can see the 'ROC curve' but not lift chart

Any advice will be greatly appreciated

Re: Question about input vars in a proc logistic

PaigeMiller — Fri, 13 Apr 2018 10:50:28 GMT

1) If many consecutive integers as the levels, treat it as continuous

2) You need to include an interaction between the classification variable and other continuous variables

3) I don't know

Re: Question about input vars in a proc logistic

Ksharp — Fri, 13 Apr 2018 13:08:22 GMT

1) Both could be accepted.

2)No Need to transform.

3)SAS/EM 's Linear Regression Node can get lift chart.

Re: Question about input vars in a proc logistic

StatDave — Fri, 13 Apr 2018 20:23:13 GMT

1. A numeric variable with many distinct levels is usually treated as continuous, not as a CLASS variable.

2. It's not clear what you are saying. Is the variable with values A,D,T your target (response) variable that you are modeling? If so, then you need to fit a multinomial logistic model. If you can consider the levels as ordered (such as descend, ascend, strongly ascend), then you just need to make sure that they appear in that order (or the reverse if you want to model the probability of ascending rather than descending) in the Response Profile table. Otherwise, the results are meaningless.

3. These can be created using this macro.