Why the coefficients are 0

[bncoxuk]

Hi,

I used the PROC LOGISTIC to run a logit model. There is an independent variable called 'age' which has the levels as 1, 2, 3, 4. By default, level=4 acts as the reference group

The results showed that the parameter estimates for levels 1 and 2 are reasonable, but the estimate for level 3 is exactly 0.000 (same as level 4). I tried again by incorporating 3 into level 2, and ran the model again. But this time the estimate for level 2 also became 0.

There must be some problem with the data. Why? Please help.

[Reeza]

What does a proc freq give you with the dependent vs independend for age?

[bncoxuk]

There are two variables, age(1, 2, 3, 4) and income(1, 2, 3). Both are categorical.

if age=4 then income is always 3. That is, these two levels are perfectly correlated.

The results showed that income can be predicted nicely with estimates for level 1, 2 (level 3: 0). But for age, both levels 3 and 4 have estimates 0.000.

I don't know if this caused the problem. But age and income are not 100% correlated. They are only perfectly correlated when age=4 and income=3.

[Reeza]

If 4 is your reference level you shouldn't have an estimate for it, not sure where/why you assume it's 0.

How many observations are you working with?

Is it possible that there could be a correlation with another variable in the model?

Have you tried a different reference level?

If you grouped the 3/4 rather 2/3 what happens?

Does the log or output say anything about quasi-seperation?

There's a lot of possibilities, without access to the data/code its hard to say...

As Art suggested, post your code...

[bncoxuk]

Yeah, it is pretty hard to describle. I will try to think for a while.

Thanks.

[art297]

Post your code.  You may not have correctly submitted what you intended to submit

[bncoxuk]

There are two variables, age(1, 2, 3, 4) and income(1, 2, 3). Both are categorical.

if age=4 then income is always 3. That is, these two levels are perfectly correlated.

The results showed that income can be predicted nicely with estimates for level 1, 2 (level 3: 0). But for age, both levels 3 and 4 have estimates 0.000.

I don't know if this caused the problem. But age and income are not 100% correlated. They are only perfectly correlated when age=4 and income=3.