NOMREG puzzle on redundancy

classic Classic list List threaded Threaded
5 messages Options
Reply | Threaded
Open this post in threaded view
|

NOMREG puzzle on redundancy

Mario Giesel-2
Dear SPSS users,

I have a question on multinomial regression via NOMREG.
My independent variables are all dummy coded (0= not mentioned / 1 = mentioned) without missing values.
Also my dependent variable is dummy coded (0/1) without missing values.
All iv's report parameter estimates for category 0 but not for category 1 saying
"This parameter is set to zero because it is redundant."

However, if I exclude the intercept with

NOMREG dv (BASE = FIRST ORDER = ASCENDING) BY @test
    /INTERCEPT = EXCLUDE
    /PRINT = SUMMARY CLASSTABLE PARAMETER MFI FIT.

one of my iv's gets a parameter calculation for category 1 as well.
I'm wondering why this single iv has not redundant information in category 1 like so:
-----------------------------------------------------------------------------------------------------------------------
Tab. 4 - Tab. 4 - Parameter Estimates                                   
Purchase Intention        B    Std. Error    Wald    df    Sig.    Exp(B)    95% Confidence Interval for Exp(B)   
                                Lower Bound    Upper Bound
Yes    [Strongly interested in champagne =,00]    -2,488    ,032    5985,243    1    ,000    ,083    ,078    ,088
      [Strongly interested in champagne =1,00]    -,602    ,044    190,042    1    ,000    ,548    ,503    ,597
a The reference category is: No.                                   
-----------------------------------------------------------------------------------------------------------------------

Does anybody have an idea how this can happen?
Comments are highly appreciated.

Thanks,
Mario
 
Mario Giesel
Munich, Germany
===================== 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
Reply | Threaded
Open this post in threaded view
|

Re: NOMREG puzzle on redundancy

Rich Ulrich

It is as if you created two variables, Male and Female, from one variable, Sex.

For Male, your values (0, 1) indicate "female, male"; for Female, (0,1) indicate "male, female".


Mathematically, Male= 1-Female;  there is no new information that you get by including

one variable after you have the other one. This is the meaning of "redundancy".


Read up on "degrees of freedom".  The number of usable, Yes/No dummy variables that you

can make from a categorical variable is (categories minus 1). 


--

Rich Ulrich


From: SPSSX(r) Discussion <[hidden email]> on behalf of Mario Giesel <[hidden email]>
Sent: Monday, June 25, 2018 11:21:12 AM
To: [hidden email]
Subject: NOMREG puzzle on redundancy
 
Dear SPSS users,

I have a question on multinomial regression via NOMREG.
My independent variables are all dummy coded (0= not mentioned / 1 = mentioned) without missing values.
Also my dependent variable is dummy coded (0/1) without missing values.
All iv's report parameter estimates for category 0 but not for category 1 saying
"This parameter is set to zero because it is redundant."

However, if I exclude the intercept with

NOMREG dv (BASE = FIRST ORDER = ASCENDING) BY @test
    /INTERCEPT = EXCLUDE
    /PRINT = SUMMARY CLASSTABLE PARAMETER MFI FIT.

one of my iv's gets a parameter calculation for category 1 as well.
I'm wondering why this single iv has not redundant information in category 1 like so:
-----------------------------------------------------------------------------------------------------------------------
Tab. 4 - Tab. 4 - Parameter Estimates                                   
Purchase Intention        B    Std. Error    Wald    df    Sig.    Exp(B)    95% Confidence Interval for Exp(B)   
                                Lower Bound    Upper Bound
Yes    [Strongly interested in champagne =,00]    -2,488    ,032    5985,243    1    ,000    ,083    ,078    ,088
      [Strongly interested in champagne =1,00]    -,602    ,044    190,042    1    ,000    ,548    ,503    ,597
a The reference category is: No.                                   
-----------------------------------------------------------------------------------------------------------------------

Does anybody have an idea how this can happen?
Comments are highly appreciated.

Thanks,
Mario
 
Mario Giesel
Munich, Germany
===================== 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
===================== 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
Reply | Threaded
Open this post in threaded view
|

Re: NOMREG puzzle on redundancy

Mario Giesel-2
To all
 
Mario Giesel
Munich, Germany


Rich Ulrich <[hidden email]> schrieb am 20:25 Montag, 25.Juni 2018:


It is as if you created two variables, Male and Female, from one variable, Sex.
For Male, your values (0, 1) indicate "female, male"; for Female, (0,1) indicate "male, female".

Mathematically, Male= 1-Female;  there is no new information that you get by including
one variable after you have the other one. This is the meaning of "redundancy".

Read up on "degrees of freedom".  The number of usable, Yes/No dummy variables that you
can make from a categorical variable is (categories minus 1). 

--
Rich Ulrich

From: SPSSX(r) Discussion <[hidden email]> on behalf of Mario Giesel <[hidden email]>
Sent: Monday, June 25, 2018 11:21:12 AM
To: [hidden email]
Subject: NOMREG puzzle on redundancy
 
Dear SPSS users,

I have a question on multinomial regression via NOMREG.
My independent variables are all dummy coded (0= not mentioned / 1 = mentioned) without missing values.
Also my dependent variable is dummy coded (0/1) without missing values.
All iv's report parameter estimates for category 0 but not for category 1 saying
"This parameter is set to zero because it is redundant."

However, if I exclude the intercept with

NOMREG dv (BASE = FIRST ORDER = ASCENDING) BY @test
    /INTERCEPT = EXCLUDE
    /PRINT = SUMMARY CLASSTABLE PARAMETER MFI FIT.

one of my iv's gets a parameter calculation for category 1 as well.
I'm wondering why this single iv has not redundant information in category 1 like so:
-----------------------------------------------------------------------------------------------------------------------
Tab. 4 - Tab. 4 - Parameter Estimates                                   
Purchase Intention        B    Std. Error    Wald    df    Sig.    Exp(B)    95% Confidence Interval for Exp(B)   
                                Lower Bound    Upper Bound
Yes    [Strongly interested in champagne =,00]    -2,488    ,032    5985,243    1    ,000    ,083    ,078    ,088
      [Strongly interested in champagne =1,00]    -,602    ,044    190,042    1    ,000    ,548    ,503    ,597
a The reference category is: No.                                   
-----------------------------------------------------------------------------------------------------------------------

Does anybody have an idea how this can happen?
Comments are highly appreciated.

Thanks,
Mario
 
Mario Giesel
Munich, Germany
===================== 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
===================== 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


===================== 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
Reply | Threaded
Open this post in threaded view
|

Re: NOMREG puzzle on redundancy

Rich Ulrich

Well, for a dichotomy, (female, male) adds nothing to (male, female). That's an

obvious meaning of "redundant"; and "adds nothing" works out in statistics, too.


The intercept exclusion? I should have mentioned that explicitly. The answer lies in the use

"Degrees of Freedom."


The residual d.f. ordinary ANOVA with two groups is N-2.  It has 1 d.f. removed for the analysis,

and another 1 d.f. removed for "the overall mean", or, in other words, "the intercept".

So, if you ask for a coefficient, there is one coefficient for the two-group difference and one

for the overall mean being different from zero.


If you do a two-group estimation with no Intercept term, then an ANOVA solution (which

minimizes residual Sums of squares) is BAD with only one term: If you have IV scores

that are far from zero, the coefficient will LARGELY reflect the overall mean ( = Intercept)

instead of the group differences... and that is a bad thing. What you get with another two

codings - which still can't be redundant - might depend on assumptions of a model, or

of the computer implementation.


The REPEATED=ID specification, on the other hand, should allow you to specify NO INTERCEPT

because that will subtract the mean for each ID separately, and conveniently.  If you are comparing

two AICs, do notice how different the d.f.'s are of the AIC.  If you are taking out the separate IDs' 

means (which I think accounts for your two plots), then the d.f.'s will be different.  In an loose or

informal way, you can "test" two AICs by regarding their difference as a chi-squared with the

d.f.  difference as the d.f.  of the chi-squared.  If your two AICs are very similar, with very-different

d.f.'s, then taking out the extra d.f.'s  (that is, the ID variance) has accounted for almost nothing.


--

Rich Ulrich


From: Mario Giesel <[hidden email]>
Sent: Monday, June 25, 2018 2:57:09 PM
To: Rich Ulrich; [hidden email]
Subject: Re: NOMREG puzzle on redundancy
 
To all
 
Mario Giesel
Munich, Germany


Rich Ulrich <[hidden email]> schrieb am 20:25 Montag, 25.Juni 2018:


It is as if you created two variables, Male and Female, from one variable, Sex.
For Male, your values (0, 1) indicate "female, male"; for Female, (0,1) indicate "male, female".

Mathematically, Male= 1-Female;  there is no new information that you get by including
one variable after you have the other one. This is the meaning of "redundancy".

Read up on "degrees of freedom".  The number of usable, Yes/No dummy variables that you
can make from a categorical variable is (categories minus 1). 

--
Rich Ulrich

From: SPSSX(r) Discussion <[hidden email]> on behalf of Mario Giesel <[hidden email]>
Sent: Monday, June 25, 2018 11:21:12 AM
To: [hidden email]
Subject: NOMREG puzzle on redundancy
 
Dear SPSS users,

I have a question on multinomial regression via NOMREG.
My independent variables are all dummy coded (0= not mentioned / 1 = mentioned) without missing values.
Also my dependent variable is dummy coded (0/1) without missing values.
All iv's report parameter estimates for category 0 but not for category 1 saying
"This parameter is set to zero because it is redundant."

However, if I exclude the intercept with

NOMREG dv (BASE = FIRST ORDER = ASCENDING) BY @test
    /INTERCEPT = EXCLUDE
    /PRINT = SUMMARY CLASSTABLE PARAMETER MFI FIT.

one of my iv's gets a parameter calculation for category 1 as well.
I'm wondering why this single iv has not redundant information in category 1 like so:
-----------------------------------------------------------------------------------------------------------------------
Tab. 4 - Tab. 4 - Parameter Estimates                                   
Purchase Intention        B    Std. Error    Wald    df    Sig.    Exp(B)    95% Confidence Interval for Exp(B)   
                                Lower Bound    Upper Bound
Yes    [Strongly interested in champagne =,00]    -2,488    ,032    5985,243    1    ,000    ,083    ,078    ,088
      [Strongly interested in champagne =1,00]    -,602    ,044    190,042    1    ,000    ,548    ,503    ,597
a The reference category is: No.                                   
-----------------------------------------------------------------------------------------------------------------------

Does anybody have an idea how this can happen?
Comments are highly appreciated.

Thanks,
Mario
 
Mario Giesel
Munich, Germany
===================== 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
===================== 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


===================== 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
Reply | Threaded
Open this post in threaded view
|

Re: NOMREG puzzle on redundancy

Mario Giesel-2
Thanks a lot for your explanation, Rich.

I now realize that the missing "redundant" mark
is not associated with a specific independent variable
but will always apply to the first iv in my set of iv's
to cope with the missing constant information.

From that point of view I can see no advantage to remove
the non significant intercept from my model as I want to
learn about the relative importance of my iv's in comparison
(btw they don't offer a REPEATED=ID specification in NOMREG).
For this I use the size of the EXP(B) statistic.

So thanks again & take care!
 
Mario Giesel
Munich, Germany


Rich Ulrich <[hidden email]> schrieb am 22:09 Montag, 25.Juni 2018:


Well, for a dichotomy, (female, male) adds nothing to (male, female). That's an
obvious meaning of "redundant"; and "adds nothing" works out in statistics, too.

The intercept exclusion? I should have mentioned that explicitly. The answer lies in the use
"Degrees of Freedom."

The residual d.f. ordinary ANOVA with two groups is N-2.  It has 1 d.f. removed for the analysis,
and another 1 d.f. removed for "the overall mean", or, in other words, "the intercept".
So, if you ask for a coefficient, there is one coefficient for the two-group difference and one
for the overall mean being different from zero.

If you do a two-group estimation with no Intercept term, then an ANOVA solution (which
minimizes residual Sums of squares) is BAD with only one term: If you have IV scores
that are far from zero, the coefficient will LARGELY reflect the overall mean ( = Intercept)
instead of the group differences... and that is a bad thing. What you get with another two
codings - which still can't be redundant - might depend on assumptions of a model, or
of the computer implementation.

The REPEATED=ID specification, on the other hand, should allow you to specify NO INTERCEPT
because that will subtract the mean for each ID separately, and conveniently.  If you are comparing
two AICs, do notice how different the d.f.'s are of the AIC.  If you are taking out the separate IDs' 
means (which I think accounts for your two plots), then the d.f.'s will be different.  In an loose or
informal way, you can "test" two AICs by regarding their difference as a chi-squared with the
d.f.  difference as the d.f.  of the chi-squared.  If your two AICs are very similar, with very-different
d.f.'s, then taking out the extra d.f.'s  (that is, the ID variance) has accounted for almost nothing.

--
Rich Ulrich

From: Mario Giesel <[hidden email]>
Sent: Monday, June 25, 2018 2:57:09 PM
To: Rich Ulrich; [hidden email]
Subject: Re: NOMREG puzzle on redundancy
 
To all
 
Mario Giesel
Munich, Germany


Rich Ulrich <[hidden email]> schrieb am 20:25 Montag, 25.Juni 2018:


It is as if you created two variables, Male and Female, from one variable, Sex.
For Male, your values (0, 1) indicate "female, male"; for Female, (0,1) indicate "male, female".

Mathematically, Male= 1-Female;  there is no new information that you get by including
one variable after you have the other one. This is the meaning of "redundancy".

Read up on "degrees of freedom".  The number of usable, Yes/No dummy variables that you
can make from a categorical variable is (categories minus 1). 

--
Rich Ulrich

From: SPSSX(r) Discussion <[hidden email]> on behalf of Mario Giesel <[hidden email]>
Sent: Monday, June 25, 2018 11:21:12 AM
To: [hidden email]
Subject: NOMREG puzzle on redundancy
 
Dear SPSS users,

I have a question on multinomial regression via NOMREG.
My independent variables are all dummy coded (0= not mentioned / 1 = mentioned) without missing values.
Also my dependent variable is dummy coded (0/1) without missing values.
All iv's report parameter estimates for category 0 but not for category 1 saying
"This parameter is set to zero because it is redundant."

However, if I exclude the intercept with

NOMREG dv (BASE = FIRST ORDER = ASCENDING) BY @test
    /INTERCEPT = EXCLUDE
    /PRINT = SUMMARY CLASSTABLE PARAMETER MFI FIT.

one of my iv's gets a parameter calculation for category 1 as well.
I'm wondering why this single iv has not redundant information in category 1 like so:
-----------------------------------------------------------------------------------------------------------------------
Tab. 4 - Tab. 4 - Parameter Estimates                                   
Purchase Intention        B    Std. Error    Wald    df    Sig.    Exp(B)    95% Confidence Interval for Exp(B)   
                                Lower Bound    Upper Bound
Yes    [Strongly interested in champagne =,00]    -2,488    ,032    5985,243    1    ,000    ,083    ,078    ,088
      [Strongly interested in champagne =1,00]    -,602    ,044    190,042    1    ,000    ,548    ,503    ,597
a The reference category is: No.                                   
-----------------------------------------------------------------------------------------------------------------------

Does anybody have an idea how this can happen?
Comments are highly appreciated.

Thanks,
Mario
 
Mario Giesel
Munich, Germany
===================== 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
===================== 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


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


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