SPSSX Discussion

Another Logistic Regression Question

Classic

List

Threaded

11 messages Options

Salbod

Another Logistic Regression Question

Dear Friends,

I’m a newbie to logistic regression analysis. I have an interpretation problem that I could use a little help with. I performed a logistic regression analysis with 4 continuous predictors. The overall chi-square was significant (p <.001) but none of the individual predicators was significance (p>.05). I noticed that there are high correlations among the predictors. Is this the problem?

Any help will be greatly appreciated.

TIA, Stephen Salbod, Pace University, NYC

SR Millis-3

Re: Another Logistic Regression Question

Collinearity is likely a problem here. Examine collinearity diagnostics: identify any condition indexes >20 and, for those, look at the variance decomposition proportions: find those that are >.50 to find the offending variables.

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: 313-993-8085
Fax: 313-966-7682

--- On Wed, 4/22/09, Salbod, Mr. Stephen <[hidden email]> wrote:

> From: Salbod, Mr. Stephen <[hidden email]>
> Subject: Another Logistic Regression Question
> To: [hidden email]
> Date: Wednesday, April 22, 2009, 7:21 PM
> Dear Friends,
> I'm a newbie to logistic regression
> analysis. I have an interpretation problem that I could use
> a little help with. I performed a logistic regression
> analysis with 4 continuous predictors. The overall
> chi-square was significant (p <.001) but none of the
> individual predicators was significance (p>.05). I
> noticed that there are high correlations among the
> predictors. Is this the problem?
>
> Any help will be greatly appreciated.
>
> TIA, Stephen Salbod, Pace University, NYC

=====================
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

SR Millis-3

Re: Another Logistic Regression Question

In reply to this post by Salbod

Although collinearity diagnostics aren't available in logistic regression in SPSS, they are for linear regression---as an option in the statistics window. Collinearity involves the independent vairables/covariates and not the response varriable/DV, so you can run the collinearity diagnostics in the linear regression module even if your response variable is binary.

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: 313-993-8085
Fax: 313-966-7682

--- On Wed, 4/22/09, DorraJ Oet <[hidden email]> wrote:

> From: DorraJ Oet <[hidden email]>
> Subject: RE: Another Logistic Regression Question
> To: [hidden email], "SPSS Syntax help" <[hidden email]>
> Date: Wednesday, April 22, 2009, 9:43 PM
> Hi,
>
>
>
> There seems no direct way in testing for collinearity in
> Logistic Regression for SPSS. Is there a way to do it?
>
>
>
> Thanks for any answer.
>
>
>
> Regards
>
> Dorraj
>
> > Date: Wed, 22 Apr 2009 16:59:36 -0700
> > From: [hidden email]
> > Subject: Re: Another Logistic Regression Question
> > To: [hidden email]
> >
> > Collinearity is likely a problem here. Examine
> collinearity diagnostics: identify any condition indexes
> >20 and, for those, look at the variance decomposition
> proportions: find those that are >.50 to find the
> offending variables.
> >
> >
> >
> > 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: 313-993-8085
> > Fax: 313-966-7682
> >
> >
> > --- On Wed, 4/22/09, Salbod, Mr. Stephen
> <[hidden email]> wrote:
> >
> > > From: Salbod, Mr. Stephen
> <[hidden email]>
> > > Subject: Another Logistic Regression Question
> > > To: [hidden email]
> > > Date: Wednesday, April 22, 2009, 7:21 PM
> > > Dear Friends,
> > > I'm a newbie to logistic regression
> > > analysis. I have an interpretation problem that I
> could use
> > > a little help with. I performed a logistic
> regression
> > > analysis with 4 continuous predictors. The
> overall
> > > chi-square was significant (p <.001) but none
> of the
> > > individual predicators was significance
> (p>.05). I
> > > noticed that there are high correlations among
> the
> > > predictors. Is this the problem?
> > >
> > > Any help will be greatly appreciated.
> > >
> > > TIA, Stephen Salbod, Pace University, NYC
> >
> > =====================
> > 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
>
> _________________________________________________________________
> More than messages–check out the rest of the Windows
> Live™.
> http://www.microsoft.com/windows/windowslive/

Sarah Carroll

ANCOVA power analysis

In reply to this post by Salbod

The design is one between-subjects factor (gender) and one within-subjects factor (time, i.e. pre vs. post), and then two continuous variables (e.g., score on a measure of self-perception and score on a subscale from a personality measure).

I have read some statistics texts and looked at the available free power analysis programs, but can't figure out how to represent the BS and WS categorical variables in combination with the continuous covariates.

I would like to try and estimate sample size based on a power of .80, an alpha of .05, and effect sizes of .10, .25, and .40.

Any suggestions on how to proceed, either with SPSS or with some other program?

Many thanks,

Sarah

Sarah Carroll, PhD

Research Director & Associate Professor, PsyD Program

JFKU Graduate School of Professional Psychology

100 Ellinwood Way, Pleasant Hill, CA 94523

ofc: 925.969.3496 email: [hidden email]

Jarrod Teo-2

Re: Another Logistic Regression Question

In reply to this post by SR Millis-3

> Date: Thu, 23 Apr 2009 04:06:54 -0700
> From: [hidden email]
> Subject: RE: Another Logistic Regression Question
> To: [hidden email]; [hidden email]
>
>
> Although collinearity diagnostics aren't available in logistic regression in SPSS, they are for linear regression---as an option in the statistics window. Collinearity involves the independent vairables/covariates and not the response variable/DV, so you can run the collinearity diagnostics in the linear regression module even if your response variable is binary.
>
> 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: 313-993-8085
> Fax: 313-966-7682
>
>
> --- On Wed, 4/22/09, DorraJ Oet <[hidden email]> wrote:
>
> > From: DorraJ Oet <[hidden email]>
> > Subject: RE: Another Logistic Regression Question
> > To: [hidden email], "SPSS Syntax help" <[hidden email]>
> > Date: Wednesday, April 22, 2009, 9:43 PM
> > Hi,
> >
> >
> >
> > There seems no direct way in testing for collinearity in
> > Logistic Regression for SPSS. Is there a way to do it?
> >
> >
> >
> > Thanks for any answer.
> >
> >
> >
> > Regards
> >
> > Dorraj
> >
> > > Date: Wed, 22 Apr 2009 16:59:36 -0700
> > > From: [hidden email]
> > > Subject: Re: Another Logistic Regression Question
> > > To: [hidden email]
> > >
> > > Collinearity is likely a problem here. Examine
> > collinearity diagnostics: identify any condition indexes
> > >20 and, for those, look at the variance decomposition
> > proportions: find those that are >.50 to find the
> > offending variables.
> > >
> > >
> > >
> > > 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: 313-993-8085
> > > Fax: 313-966-7682
> > >
> > >
> > > --- On Wed, 4/22/09, Salbod, Mr. Stephen
> > <[hidden email]> wrote:
> > >
> > > > From: Salbod, Mr. Stephen
> > <[hidden email]>
> > > > Subject: Another Logistic Regression Question
> > > > To: [hidden email]
> > > > Date: Wednesday, April 22, 2009, 7:21 PM
> > > > Dear Friends,
> > > > I'm a newbie to logistic regression
> > > > analysis. I have an interpretation problem that I
> > could use
> > > > a little help with. I performed a logistic
> > regression
> > > > analysis with 4 continuous predictors. The
> > overall
> > > > chi-square was significant (p <.001) but none
> > of the
> > > > individual predicators was significance
> > (p>.05). I
> > > > noticed that there are high correlations among
> > the
> > > > predictors. Is this the problem?
> > > >
> > > > Any help will be greatly appreciated.
> > > >
> > > > TIA, Stephen Salbod, Pace University, NYC
> > >
> > > =====================
> > > 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
> >
> > _________________________________________________________________
> > More than messages–check out the rest of the Windows
> > Live™.
> > http://www.microsoft.com/windows/windowslive/

check out the rest of the Windows Live™. More than mail–Windows Live™ goes way beyond your inbox. More than messages

Marta Garcia-Granero

Re: ANCOVA power analysis

In reply to this post by Sarah Carroll

Sarah Carroll wrote:

> The design is one between-subjects factor (gender) and one
> within-subjects factor (time, i.e. pre vs. post), and then two
> continuous variables (e.g., score on a measure of self-perception and
> score on a subscale from a personality measure).
>
>
>
> I have read some statistics texts and looked at the available free
> power analysis programs, but can't figure out how to represent the BS
> and WS categorical variables in combination with the continuous
> covariates.
>

MorePower will help you ... partially: you can set the design for the
BS&WS factors. Anyway, the covariate is going to be a problem. Perhaps
you can set the design for the factors, and then estimate the decrease
in the residual standard error (MSerror) for the different estimated
effect sizes and enter successively those estimated MSerror

>
>
> I would like to try and estimate sample size based on a power of .80,
> an alpha of .05, and effect sizes of .10, .25, and .40.
>

For which one: the BS factor, the WS factor or the covariate? Besides, I
suppose you are taking about f as effect size measure (I assume it for
the cut-off points you use, had you used partial eta-square, the
cut-offs would have been 00.01,0.06 and 0.14). Unfortunately, MorePower
uses partial eta-square.

MorePower can be downloaded from:

http://homepage.usask.ca/~jic956/work/MorePower.html

HTH,
Marta García-Granero

--
For miscellaneous SPSS related statistical stuff, visit:
http://gjyp.nl/marta/

=====================
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

Maguin, Eugene

Re: ANCOVA power analysis

In reply to this post by Sarah Carroll

Sarah,

It is possible to use the Manova procedure to produce power calculations. I
think that it could be used for the design you have in mind. There have been
discussions about this in the past couple or three years, so they are in the
archives. Look for terms like 'spss', 'power analysis' 'ANOVA' 'Manova' and
case variations of these terms. I posted some of those messages so my name
will be on a few of them. I think something about this topic is in the Data
management pdf that Ray Levesque first wrote and spss picked up. The
original source for all this was an article in a journal called Behavior
Research Methods and Instrumentation or something like that. Maybe search a
citation database on spss power analysis manova.

Here's the setup for a split plot design.
*************************************************************.
* PC1+PC2 VS COMPARE (n=64 VS n=32).
* SIGMAM(BW)~=.195 at Post for PC1 vs PC2.
*************************************************************.
matrix data variables = group rowtype_ dv1 dv2 dv3/
factor = group/format = full.
begin data
. n 96 96 96
1 mean 0.00 0.78 0.78
1 n 64 64 64
2 mean 0.00 0.00 0.00
2 n 32 32 32
. stddev 1.00 1.00 1.00
. corr 1.00 .400 .400
. corr .400 1.00 .400
. corr .400 .400 1.00
end data.

manova dv1 dv2 dv3 by group(1,2)/wsfactors dv(3)/method=unique/
error=within+residual/matrix=in(*)/power t (.05) F (.05)/
print signif(multiv averf efsize)/noprint param(estim).

I'm sorry, I don't have more time to offer better advice.

Gene Maguin

=====================
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

Hector Maletta

Re: Another Logistic Regression Question

In reply to this post by Jarrod Teo-2

Individual probabilities arising from logistic regression do not provide a prediction of individual cases. What those probabilities do is measure the expected relative frequency of an event in a group of people sharing certain characteristics.

For instance, if your predictors are sex and age group, a probability of 0.30 for males aged 30-44 means that 30 out of every 100 of those males would have the event, but there is no way of knowing WHICH individuals in the group will have the event. Individuals in the group do not have a value of 0.30: they either suffer the event (and their value is 1) or they don’t (and their value is 0). Only the group has a value of 0.30. Probabilities are relative frequencies in a group.

In a recent exchange in this forum we have discussed a similar question, in which I cited the bible of logistic regression (Hosmer and Lemeshow) to the effect that the classification table –for which you need a cutoff point—is no use to assess the goodness of fit of a model nor provides a predictive instrument.

Hector

From: SPSSX(r) Discussion [mailto:[hidden email]] On Behalf Of DorraJ Oet
Sent: 24 April 2009 03:46
To: [hidden email]
Subject: Re: Another Logistic Regression Question

Dear Millis,

Thanks for your answer.

I have another Logistic Regression.

Sometimes, during research, there will be a need to use the ROC curve to adjust the cutoff point in order to improve the event prediciton.

However, suppose if I set the cut-off point to say 0.2 for model creation on this month data, does that mean that when I want to use this model for next month's prediction, I will need to keep in mind that p<0.2 will be 0 while p>=0.2 will be 1?

Thanks again for your answer.

Cheers
Dorraj

> Date: Thu, 23 Apr 2009 04:06:54 -0700
> From: [hidden email]
> Subject: RE: Another Logistic Regression Question
> To: [hidden email]; [hidden email]
>
>
> Although collinearity diagnostics aren't available in logistic regression in SPSS, they are for linear regression---as an option in the statistics window. Collinearity involves the independent vairables/covariates and not the response variable/DV, so you can run the collinearity diagnostics in the linear regression module even if your response variable is binary.
>
> 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: 313-993-8085
> Fax: 313-966-7682
>
>
> --- On Wed, 4/22/09, DorraJ Oet <[hidden email]> wrote:
>
> > From: DorraJ Oet <[hidden email]>
> > Subject: RE: Another Logistic Regression Question
> > To: [hidden email], "SPSS Syntax help" <[hidden email]>
> > Date: Wednesday, April 22, 2009, 9:43 PM
> > Hi,
> >
> >
> >
> > There seems no direct way in testing for collinearity in
> > Logistic Regression for SPSS. Is there a way to do it?
> >
> >
> >
> > Thanks for any answer.
> >
> >
> >
> > Regards
> >
> > Dorraj
> >
> > > Date: Wed, 22 Apr 2009 16:59:36 -0700
> > > From: [hidden email]
> > > Subject: Re: Another Logistic Regression Question
> > > To: [hidden email]
> > >
> > > Collinearity is likely a problem here. Examine
> > collinearity diagnostics: identify any condition indexes
> > >20 and, for those, look at the variance decomposition
> > proportions: find those that are >.50 to find the
> > offending variables.
> > >
> > >
> > >
> > > 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: 313-993-8085
> > > Fax: 313-966-7682
> > >
> > >
> > > --- On Wed, 4/22/09, Salbod, Mr. Stephen
> > <[hidden email]> wrote:
> > >
> > > > From: Salbod, Mr. Stephen
> > <[hidden email]>
> > > > Subject: Another Logistic Regression Question
> > > > To: [hidden email]
> > > > Date: Wednesday, April 22, 2009, 7:21 PM
> > > > Dear Friends,
> > > > I'm a newbie to logistic regression
> > > > analysis. I have an interpretation problem that I
> > could use
> > > > a little help with. I performed a logistic
> > regression
> > > > analysis with 4 continuous predictors. The
> > overall
> > > > chi-square was significant (p <.001) but none
> > of the
> > > > individual predicators was significance
> > (p>.05). I
> > > > noticed that there are high correlations among
> > the
> > > > predictors. Is this the problem?
> > > >
> > > > Any help will be greatly appreciated.
> > > >
> > > > TIA, Stephen Salbod, Pace University, NYC
> > >
> > > =====================
> > > 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
> >
> > _________________________________________________________________
> > More than messages–check out the rest of the Windows
> > Live™.
> > http://www.microsoft.com/windows/windowslive/

check out the rest of the Windows Live™. More than mail–Windows Live™ goes way beyond your inbox. More than messages

Martin P. Holt-2

Re: Another Logistic Regression Question

Hector,

Thank you for your recent postings on this....well explained. One question though: where you say that a classification table is not to be used to assess goodness of fit, are you saying that the Hosmer-Lemeshow goodness of fit test should not be used ?....it's based on a classification table, isn't it ?

Thanks,

Martin

----- Original Message -----

From: [hidden email]

To: [hidden email]

Sent: Friday, April 24, 2009 3:45 PM

Subject: Re: Another Logistic Regression Question

Individual probabilities arising from logistic regression do not provide a prediction of individual cases. What those probabilities do is measure the expected relative frequency of an event in a group of people sharing certain characteristics.

For instance, if your predictors are sex and age group, a probability of 0.30 for males aged 30-44 means that 30 out of every 100 of those males would have the event, but there is no way of knowing WHICH individuals in the group will have the event. Individuals in the group do not have a value of 0.30: they either suffer the event (and their value is 1) or they dont (and their value is 0). Only the group has a value of 0.30. Probabilities are relative frequencies in a group.

In a recent exchange in this forum we have discussed a similar question, in which I cited the bible of logistic regression (Hosmer and Lemeshow) to the effect that the classification table for which you need a cutoff pointis no use to assess the goodness of fit of a model nor provides a predictive instrument.

Hector

From: SPSSX(r) Discussion [mailto:[hidden email]] On Behalf Of DorraJ Oet
Sent: 24 April 2009 03:46
To: [hidden email]
Subject: Re: Another Logistic Regression Question

Dear Millis,

Thanks for your answer.

I have another Logistic Regression.

Sometimes, during research, there will be a need to use the ROC curve to adjust the cutoff point in order to improve the event prediciton.

However, suppose if I set the cut-off point to say 0.2 for model creation on this month data, does that mean that when I want to use this model for next month's prediction, I will need to keep in mind that p<0.2 will be 0 while p>=0.2 will be 1?

Thanks again for your answer.

Cheers
Dorraj

> Date: Thu, 23 Apr 2009 04:06:54 -0700
> From: [hidden email]
> Subject: RE: Another Logistic Regression Question
> To: [hidden email]; [hidden email]
>
>
> Although collinearity diagnostics aren't available in logistic regression in SPSS, they are for linear regression---as an option in the statistics window. Collinearity involves the independent vairables/covariates and not the response variable/DV, so you can run the collinearity diagnostics in the linear regression module even if your response variable is binary.
>
> 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: 313-993-8085
> Fax: 313-966-7682
>
>
> --- On Wed, 4/22/09, DorraJ Oet <[hidden email]> wrote:
>
> > From: DorraJ Oet <[hidden email]>
> > Subject: RE: Another Logistic Regression Question
> > To: [hidden email], "SPSS Syntax help" <[hidden email]>
> > Date: Wednesday, April 22, 2009, 9:43 PM
> > Hi,
> >
> >
> >
> > There seems no direct way in testing for collinearity in
> > Logistic Regression for SPSS. Is there a way to do it?
> >
> >
> >
> > Thanks for any answer.
> >
> >
> >
> > Regards
> >
> > Dorraj
> >
> > > Date: Wed, 22 Apr 2009 16:59:36 -0700
> > > From: [hidden email]
> > > Subject: Re: Another Logistic Regression Question
> > > To: [hidden email]
> > >
> > > Collinearity is likely a problem here. Examine
> > collinearity diagnostics: identify any condition indexes
> > >20 and, for those, look at the variance decomposition
> > proportions: find those that are >.50 to find the
> > offending variables.
> > >
> > >
> > >
> > > 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: 313-993-8085
> > > Fax: 313-966-7682
> > >
> > >
> > > --- On Wed, 4/22/09, Salbod, Mr. Stephen
> > <[hidden email]> wrote:
> > >
> > > > From: Salbod, Mr. Stephen
> > <[hidden email]>
> > > > Subject: Another Logistic Regression Question
> > > > To: [hidden email]
> > > > Date: Wednesday, April 22, 2009, 7:21 PM
> > > > Dear Friends,
> > > > I'm a newbie to logistic regression
> > > > analysis. I have an interpretation problem that I
> > could use
> > > > a little help with. I performed a logistic
> > regression
> > > > analysis with 4 continuous predictors. The
> > overall
> > > > chi-square was significant (p <.001) but none
> > of the
> > > > individual predicators was significance
> > (p>.05). I
> > > > noticed that there are high correlations among
> > the
> > > > predictors. Is this the problem?
> > > >
> > > > Any help will be greatly appreciated.
> > > >
> > > > TIA, Stephen Salbod, Pace University, NYC
> > >
> > > =====================
> > > 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
> >
> > _________________________________________________________________
> > More than messagescheck out the rest of the Windows
> > Live.
> > http://www.microsoft.com/windows/windowslive/

check out the rest of the Windows Live. More than mailWindows Live goes way beyond your inbox. More than messages

Hector Maletta

Re: Another Logistic Regression Question

The Hosmer Lemeshow test is NOT based on the classification table. It is based on a 2 x g table resulting from dividing the sample into g groups of subjects with similar probabilities, and assessing the frequency of the event in the various groups. In the SPSS specification of the test, the groups are formed as deciles of the probability, i.e. g=10 groups with equal number of subjects ordered by increasing probability of the event. In their book Hosmer and Lemeshow propose also another specification (grouping the subjects by level of the probability, from 0 to 0.1, from 0.1 to 0.2 and so on). The specification used by SPSS ensures that all groups are populated and further ensures that all have approximately the same size. It would be advisable, of course, to have enough cases so that deciles have significantly more than 10 cases each: otherwise the H&L statistic would be volatile (shifting one case from one decile to the next may significantly alter the results).

The main difference, however, is not the use of a 2x2 or a 2xg table, but the way the table is used. In H&L test it is not used to check whether individual cases were well predicted, but to check whether observed relative group frequencies are similar to predicted group frequencies, not caring about individual probabilities compared to individual events.

Hector

From: SPSSX(r) Discussion [mailto:[hidden email]] On Behalf Of Martin Holt
Sent: 25 April 2009 11:28
To: [hidden email]
Subject: Re: Another Logistic Regression Question

Hector,

Thanks,

Martin

----- Original Message -----

From: [hidden email]

To: [hidden email]

Sent: Friday, April 24, 2009 3:45 PM

Subject: Re: Another Logistic Regression Question

Individual probabilities arising from logistic regression do not provide a prediction of individual cases. What those probabilities do is measure the expected relative frequency of an event in a group of people sharing certain characteristics.

For instance, if your predictors are sex and age group, a probability of 0.30 for males aged 30-44 means that 30 out of every 100 of those males would have the event, but there is no way of knowing WHICH individuals in the group will have the event. Individuals in the group do not have a value of 0.30: they either suffer the event (and their value is 1) or they don’t (and their value is 0). Only the group has a value of 0.30. Probabilities are relative frequencies in a group.

In a recent exchange in this forum we have discussed a similar question, in which I cited the bible of logistic regression (Hosmer and Lemeshow) to the effect that the classification table –for which you need a cutoff point—is no use to assess the goodness of fit of a model nor provides a predictive instrument.

Hector

From: SPSSX(r) Discussion [mailto:[hidden email]] On Behalf Of DorraJ Oet
Sent: 24 April 2009 03:46
To: [hidden email]
Subject: Re: Another Logistic Regression Question

Dear Millis,

Thanks for your answer.

I have another Logistic Regression.

Sometimes, during research, there will be a need to use the ROC curve to adjust the cutoff point in order to improve the event prediciton.

However, suppose if I set the cut-off point to say 0.2 for model creation on this month data, does that mean that when I want to use this model for next month's prediction, I will need to keep in mind that p<0.2 will be 0 while p>=0.2 will be 1?

Thanks again for your answer.

Cheers
Dorraj

> Date: Thu, 23 Apr 2009 04:06:54 -0700
> From: [hidden email]
> Subject: RE: Another Logistic Regression Question
> To: [hidden email]; [hidden email]
>
>
> Although collinearity diagnostics aren't available in logistic regression in SPSS, they are for linear regression---as an option in the statistics window. Collinearity involves the independent vairables/covariates and not the response variable/DV, so you can run the collinearity diagnostics in the linear regression module even if your response variable is binary.
>
> 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: 313-993-8085
> Fax: 313-966-7682
>
>
> --- On Wed, 4/22/09, DorraJ Oet <[hidden email]> wrote:
>
> > From: DorraJ Oet <[hidden email]>
> > Subject: RE: Another Logistic Regression Question
> > To: [hidden email], "SPSS Syntax help" <[hidden email]>
> > Date: Wednesday, April 22, 2009, 9:43 PM
> > Hi,
> >
> >
> >
> > There seems no direct way in testing for collinearity in
> > Logistic Regression for SPSS. Is there a way to do it?
> >
> >
> >
> > Thanks for any answer.
> >
> >
> >
> > Regards
> >
> > Dorraj
> >
> > > Date: Wed, 22 Apr 2009 16:59:36 -0700
> > > From: [hidden email]
> > > Subject: Re: Another Logistic Regression Question
> > > To: [hidden email]
> > >
> > > Collinearity is likely a problem here. Examine
> > collinearity diagnostics: identify any condition indexes
> > >20 and, for those, look at the variance decomposition
> > proportions: find those that are >.50 to find the
> > offending variables.
> > >
> > >
> > >
> > > 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: 313-993-8085
> > > Fax: 313-966-7682
> > >
> > >
> > > --- On Wed, 4/22/09, Salbod, Mr. Stephen
> > <[hidden email]> wrote:
> > >
> > > > From: Salbod, Mr. Stephen
> > <[hidden email]>
> > > > Subject: Another Logistic Regression Question
> > > > To: [hidden email]
> > > > Date: Wednesday, April 22, 2009, 7:21 PM
> > > > Dear Friends,
> > > > I'm a newbie to logistic regression
> > > > analysis. I have an interpretation problem that I
> > could use
> > > > a little help with. I performed a logistic
> > regression
> > > > analysis with 4 continuous predictors. The
> > overall
> > > > chi-square was significant (p <.001) but none
> > of the
> > > > individual predicators was significance
> > (p>.05). I
> > > > noticed that there are high correlations among
> > the
> > > > predictors. Is this the problem?
> > > >
> > > > Any help will be greatly appreciated.
> > > >
> > > > TIA, Stephen Salbod, Pace University, NYC
> > >
> > > =====================
> > > 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
> >
> > _________________________________________________________________
> > More than messages–check out the rest of the Windows
> > Live™.
> > http://www.microsoft.com/windows/windowslive/

check out the rest of the Windows Live™. More than mail–Windows Live™ goes way beyond your inbox. More than messages

Martin P. Holt-2

Re: Another Logistic Regression Question

Thanks, Hector. I was thinking of "Contingency" table, with "Minority Classification=No" and "Minority Classification=Yes"........hence "Classification".

BW,

Martin

----- Original Message -----

From: [hidden email]

To: [hidden email]

Cc: [hidden email]

Sent: Saturday, April 25, 2009 3:56 PM

Subject: RE: Another Logistic Regression Question

The Hosmer Lemeshow test is NOT based on the classification table. It is based on a 2 x g table resulting from dividing the sample into g groups of subjects with similar probabilities, and assessing the frequency of the event in the various groups. In the SPSS specification of the test, the groups are formed as deciles of the probability, i.e. g=10 groups with equal number of subjects ordered by increasing probability of the event. In their book Hosmer and Lemeshow propose also another specification (grouping the subjects by level of the probability, from 0 to 0.1, from 0.1 to 0.2 and so on). The specification used by SPSS ensures that all groups are populated and further ensures that all have approximately the same size. It would be advisable, of course, to have enough cases so that deciles have significantly more than 10 cases each: otherwise the H&L statistic would be volatile (shifting one case from one decile to the next may significantly alter the results).

The main difference, however, is not the use of a 2x2 or a 2xg table, but the way the table is used. In H&L test it is not used to check whether individual cases were well predicted, but to check whether observed relative group frequencies are similar to predicted group frequencies, not caring about individual probabilities compared to individual events.

Hector

From: SPSSX(r) Discussion [mailto:[hidden email]] On Behalf Of Martin Holt
Sent: 25 April 2009 11:28
To: [hidden email]
Subject: Re: Another Logistic Regression Question

Hector,

Thank you for your recent postings on this....well explained. One question though: where you say that a classification table is not to be used to assess goodness of fit, are you saying that the Hosmer-Lemeshow goodness of fit test should not be used ?....it's based on a classification table, isn't it ?

Thanks,

Martin

----- Original Message -----

From: [hidden email]

To: [hidden email]

Sent: Friday, April 24, 2009 3:45 PM

Subject: Re: Another Logistic Regression Question

Individual probabilities arising from logistic regression do not provide a prediction of individual cases. What those probabilities do is measure the expected relative frequency of an event in a group of people sharing certain characteristics.

For instance, if your predictors are sex and age group, a probability of 0.30 for males aged 30-44 means that 30 out of every 100 of those males would have the event, but there is no way of knowing WHICH individuals in the group will have the event. Individuals in the group do not have a value of 0.30: they either suffer the event (and their value is 1) or they dont (and their value is 0). Only the group has a value of 0.30. Probabilities are relative frequencies in a group.

In a recent exchange in this forum we have discussed a similar question, in which I cited the bible of logistic regression (Hosmer and Lemeshow) to the effect that the classification table for which you need a cutoff pointis no use to assess the goodness of fit of a model nor provides a predictive instrument.

Hector

From: SPSSX(r) Discussion [mailto:[hidden email]] On Behalf Of DorraJ Oet
Sent: 24 April 2009 03:46
To: [hidden email]
Subject: Re: Another Logistic Regression Question

Dear Millis,

Thanks for your answer.

I have another Logistic Regression.

Sometimes, during research, there will be a need to use the ROC curve to adjust the cutoff point in order to improve the event prediciton.

However, suppose if I set the cut-off point to say 0.2 for model creation on this month data, does that mean that when I want to use this model for next month's prediction, I will need to keep in mind that p<0.2 will be 0 while p>=0.2 will be 1?

Thanks again for your answer.

Cheers
Dorraj

> Date: Thu, 23 Apr 2009 04:06:54 -0700
> From: [hidden email]
> Subject: RE: Another Logistic Regression Question
> To: [hidden email]; [hidden email]
>
>
> Although collinearity diagnostics aren't available in logistic regression in SPSS, they are for linear regression---as an option in the statistics window. Collinearity involves the independent vairables/covariates and not the response variable/DV, so you can run the collinearity diagnostics in the linear regression module even if your response variable is binary.
>
> 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: 313-993-8085
> Fax: 313-966-7682
>
>
> --- On Wed, 4/22/09, DorraJ Oet <[hidden email]> wrote:
>
> > From: DorraJ Oet <[hidden email]>
> > Subject: RE: Another Logistic Regression Question
> > To: [hidden email], "SPSS Syntax help" <[hidden email]>
> > Date: Wednesday, April 22, 2009, 9:43 PM
> > Hi,
> >
> >
> >
> > There seems no direct way in testing for collinearity in
> > Logistic Regression for SPSS. Is there a way to do it?
> >
> >
> >
> > Thanks for any answer.
> >
> >
> >
> > Regards
> >
> > Dorraj
> >
> > > Date: Wed, 22 Apr 2009 16:59:36 -0700
> > > From: [hidden email]
> > > Subject: Re: Another Logistic Regression Question
> > > To: [hidden email]
> > >
> > > Collinearity is likely a problem here. Examine
> > collinearity diagnostics: identify any condition indexes
> > >20 and, for those, look at the variance decomposition
> > proportions: find those that are >.50 to find the
> > offending variables.
> > >
> > >
> > >
> > > 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: 313-993-8085
> > > Fax: 313-966-7682
> > >
> > >
> > > --- On Wed, 4/22/09, Salbod, Mr. Stephen
> > <[hidden email]> wrote:
> > >
> > > > From: Salbod, Mr. Stephen
> > <[hidden email]>
> > > > Subject: Another Logistic Regression Question
> > > > To: [hidden email]
> > > > Date: Wednesday, April 22, 2009, 7:21 PM
> > > > Dear Friends,
> > > > I'm a newbie to logistic regression
> > > > analysis. I have an interpretation problem that I
> > could use
> > > > a little help with. I performed a logistic
> > regression
> > > > analysis with 4 continuous predictors. The
> > overall
> > > > chi-square was significant (p <.001) but none
> > of the
> > > > individual predicators was significance
> > (p>.05). I
> > > > noticed that there are high correlations among
> > the
> > > > predictors. Is this the problem?
> > > >
> > > > Any help will be greatly appreciated.
> > > >
> > > > TIA, Stephen Salbod, Pace University, NYC
> > >
> > > =====================
> > > 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
> >
> > _________________________________________________________________
> > More than messagescheck out the rest of the Windows
> > Live.
> > http://www.microsoft.com/windows/windowslive/

check out the rest of the Windows Live. More than mailWindows Live goes way beyond your inbox. More than messages