topic Re: Logistic regression result with a Point Estimate = "."? in SAS Data Science

Logistic regression result with a Point Estimate = "."?

okla — Mon, 09 Oct 2017 05:32:17 GMT

Hi All,

I was building a logistic regression, and part of the Odds Ratio result is shown below..

Odds Ratio Estimates

Point

Effect Estimate

var_Finance_Insurance 0 vs NoCC 1.729

var_Finance_Insurance 1 vs NoCC 2.334

var_Gender F vs M 1.128

var_Health_GeneralHealth 0 vs NoCC 0.586

var_Health_GeneralHealth 1 vs NoCC .

var_Leisure_BookWorm 0 vs NoCC 1.304

var_Leisure_BookWorm 1 vs NoCC .

I am just wondering for var_Health_GeneralHealth (1 vs NoCC) and var_Leisure_Bookworm (1 vs NoCC), why does it have a Odd ratio of "."?

Is it because it has a same % as var_Finance_insurance for the NoCC? Below are my distribution for var_Health_General, var_Leisure_Bookworm and var_Finance_insurance.

var_Leisure_BookWorm	Count(*)	Sum(TARGET)	%
1	4163	1068	0.25654576
0	25451	6239	0.245137716
NoCC	16142	2993	0.185416925


var_Finance_Insurance	Count(*)	Sum(TARGET)	%
1	9910	2872	0.289808274
0	19704	4435	0.225081202
NoCC	16142	2993	0.185416925

var_Health_GeneralHealth	Count(*)	Sum(TARGET)	%
1	4792	1502	0.313439065
0	24822	5805	0.23386512
NoCC	16142	2993	0.185416925

Re: Logistic regression result with a Point Estimate = "."?

WendyCzika — Mon, 09 Oct 2017 16:00:26 GMT

If you are using GLM effect coding of your class inputs, then you will always have 1 level that has a missing estimate since the estimates are calculated in reference to one of the levels (the reference level) and the parameterization is singular. If you are using the Regression node in SAS Enterprise Miner, you can change the Input Coding property to Deviation to get estimates for all levels compared to the average effect of all levels.

Re: Logistic regression result with a Point Estimate = "."?

okla — Tue, 10 Oct 2017 01:13:17 GMT

Hi,

The result i shown was actually based on "Deviation" setting. I tried GLM as well, same thing happen. I am using EM12.1

Re: Logistic regression result with a Point Estimate = "."?

WendyCzika — Tue, 10 Oct 2017 14:49:28 GMT

Ahhh, ok. Then as you first said, I would think there is another input or coded effect perfectly correlated with the coded effect for the levels of those inputs.