Comparing coefficients for nested logistic regressions

dnh

New Member
#1
Hi,

maybe one out there can help. I am interested in comparing the coefficients of two binary logistic models. In one nested in the other one, only one IV is the difference between the two models.

I have the theoretically based argument that some coefficients become smaller after including this one extra variable.

In Stata, I computed the two models and then did a suest (seemingly unrelated estimation) with clustered standard errors (it's a stacked dataset) and then compared the coefficients using the test command. It works all fine and the results are neat - but is it the correct approach?

Thanks in advance for any help!

Cheers,
dnh
 

Masteras

TS Contributor
#2
if you have two binary logistic regression models why didn't you just did logistic regresion? I did not understand the reason fot the SURE but anyway. The deviance of the one model minus the deviance of the other model follow a chi-square with the some df. the df is the df of the one deviance - eh df of the other deviance.
 

dnh

New Member
#3
Hi,

thanks for your reply - but that would be an omnibus test of the entire model? That's not what I need - I need a comparison of each single IV coefficient.

If that is feasible at all ... ?

Cheers,
dnh
 

dnh

New Member
#5
I wouldn't be surprised if I don't get it at once! ;-)

How would I then proceed if I find that an IV has coefficient X1 in Model 1 and coefficient X2 in Model 2? How would I compare these two?

Cheers,
dnh
 

Masteras

TS Contributor
#6
ok, x1 is good in a model and X2 is a good in a model. then in a third model x1 and x2 are both good. then it is over. i do not think you have explained me what is it that you want. meybe you made a mistake somewhere and you ask the wrong questions.
 

dnh

New Member
#7
I see, I am unclear.

Say, I have Model 1 with the IVs X1...X10. Say, most of them are significant, as they should. Then, I compute Model 2 with the same IVs but one extra, X1....X11. I then argue that some of the coefficients for the variables in both models, i.e. X1 ... X10, are smaller in size in Model 2 than in Model 1 (because X11 takes some of their effect in Model 2), but not necessarily insignificant (which of course would be neater).

Does that make sense? Thanks for your patience!

Cheers,
DNH
 

Masteras

TS Contributor
#8
yes, now it does. the size does not matter, the significance matters. if you want them to becomparable however, compare the standardized coefficients.
You want to see the significance of X11? or th significance of some of the X1.. X10 in the new model?
 

dnh

New Member
#9
This is what I did now: I computed y-standardized coefficients using Long and Feese's spost in Stata. I understand from Mood 2010 that this is the way to compare standardized coefficients across models. But how can I then make a significance test of the change in the y-standardized coefficients across models?