Logistic regression question

[Brian J. Hall]

Dear List,
I've run across a problem while running a hierarchical logistic regression.
This is the output I get when I add a continuous variable to the model. The previous several model steps are fine (a combination of categorical and continuous variables), so it appears that there is something amiss with this one variable.
B            SE            Wald   DF  Sig     Exp(b)                   Lower  upper
22.016    3762.157    .000    1    .995    3643607134.112    .000    .

I have tried the following:
1. ran correlations to see if this variable was multicolinear with other variables: the largest correlation was .40
2. I dichotomized the variable, and this did not help. See below
3. I noted the range of the variable is 0 - 27. For one level of the DV the range is 0 - 20; for the other, the range is 10 - 26.
4. trichomomizing the variable does not help either, considering that one level of the DV represents only those in the extreme high values on the continuous variable.

Any assistance regarding what to do with this variable would be welcomed!

Thanks in advance,
--
Brian
_____________________________________________________________
Brian J. Hall, M. A.
Research Study Coordinator, Rush Medical College
Department of Behavioral Sciences
Rush University Medical Center
1653 W. Congress Parkway, 310 Rawson
Chicago, IL 60612

[SR Millis-3]

I suspect that there is a high degree of collinearity between the continuous variable that you added and at least one of more of the other variables/covariates.

You'll need to run collinearity diagnostics but you have to do in with the multiple regression module---it's not option in logistic regression in SPSS.

Scott Millis
~~~~~~~~~~~
"Kunst ist schön, macht aber viel Arbeit."

Scott R Millis, PhD, ABPP (CN,CL,RP), CStat, CSci
Professor & Director of Research
Dept of Physical Medicine & Rehabilitation
Dept of Emergency Medicine
Wayne State University School of Medicine
261 Mack Blvd
Detroit, MI 48201
Email:  [hidden email]
Tel: [phone]
Fax: [phone]


--- On Thu, 10/8/09, Brian J. Hall <[hidden email]> wrote:

> From: Brian J. Hall <[hidden email]>
> Subject: Logistic regression question
> To: [hidden email]
> Date: Thursday, October 8, 2009, 6:09 PM
> Dear List,
> I've run across a problem while running a hierarchical
> logistic regression.
> This is the output I get when I add a continuous variable
> to the model. The previous several model steps are fine (a
> combination of categorical and continuous variables), so it
> appears that there is something amiss with this one
> variable.
>
> B            SE            Wald
> DF  Sig     Exp(b)
> Lower  upper
> 22.016    3762.157
> .000    1    .995    3643607134.112
> .000    .
>
> I have tried the following:
> 1. ran correlations to see if this variable was
> multicolinear with other variables: the largest correlation
> was .40
>
> 2. I dichotomized the variable, and this did not help. See
> below
> 3. I noted the range of the variable is 0 - 27. For one
> level of the DV the range is 0 - 20; for the other, the
> range is 10 - 26.
> 4. trichomomizing the variable does not help either,
> considering that one level of the DV represents only those
> in the extreme high values on the continuous variable.
>
>
> Any assistance regarding what to do with this variable
> would be welcomed!
>
> Thanks in advance,
> --
> Brian
> _____________________________________________________________
> Brian J. Hall, M. A.
> Research Study Coordinator, Rush Medical College
>
> Department of Behavioral Sciences
> Rush University Medical Center
> 1653 W. Congress Parkway, 310 Rawson
> Chicago, IL 60612
>
>
>

=====================
To manage your subscription to SPSSX-L, send a message to
[hidden email] (not to SPSSX-L), with no body text except the
command. To leave the list, send the command
SIGNOFF SPSSX-L
For a list of commands to manage subscriptions, send the command
INFO REFCARD

[Anthony Babinec]

Other possibilities are that you have a 0 cell somewhere or you have

separation. Regarding the 0 cell, it could be something simple to detect,

or it could be that for some combination of the other variables and the covariate

in question the cases fall in only one category of the target variable.

Regarding separation, it could be that some combination of the variables

in the model including the covariate in question completely separates the outcome

groups.

 

See Hosmer and Lemeshow, Applied Logistic Regression, 2^nd edition, section 4.5

on numerical problems.

 

Tony Babinec

[hidden email]

 

"Be the change you want to see in the world." Gandhi

[Steve Simon, P.Mean Consulting]

Brian J. Hall writes:

> I've run across a problem while running a hierarchical logistic
> regression. This is the output I get when I add a continuous variable
> to the model. The previous several model steps are fine (a
> combination of categorical and continuous variables), so it appears
> that there is something amiss with this one variable
> B SE Wald DF Sig Exp(b) Lower upper
 > 22.016 3762.157 .000 1 .995 3643607134.112 .000 .

If I'm reading this right, your confidence interval goes from zero to
missing. It's kind of like the Buzz Lightyear phrase "To infinity and
beyond."

In addition to the comments already made, I would suggest that you look
at rescaling your continuous variable. Sometimes revising your
independent variable to be in units of kilograms instead of grams can
make a big difference in your model. If you think about a simple example
relating birth weight in grams to mortality during the first 30 days of
life, does it really make sense to estimate an odds ratio that
represents the decrease in risk of mortality associated with a 1 gram
change in weight?

If your odds ratio is way too small or way too large to make sense, it
may be worthwhile to change the scale of your independent variable.

Now scaling would not explain why you also have a problem when you
dichotomize your independent variable. The other comment on collinearity
may apply here. If there is a problem with collinearity, try a simple
model with only the new independent variable and add back in, one at a
time, each of the other independent variables. The spot at which your
model misbehaves may give a hint as to the source of the collinearity.
--
Steve Simon, Standard Disclaimer
Sign up for The Monthly Mean at www.pmean.com/news

=====================
To manage your subscription to SPSSX-L, send a message to
[hidden email] (not to SPSSX-L), with no body text except the
command. To leave the list, send the command
SIGNOFF SPSSX-L
For a list of commands to manage subscriptions, send the command
INFO REFCARD