SPSSX Discussion

Ordinal Regression in SPSS

Classic

List

Threaded

6 messages Options

Keval Khichadia

Ordinal Regression in SPSS

Hi,
I am trying to do an ordinal regression on the results of a Student Satisfaction Survey (Noel Levitz). Maybe someone working in Institutional Research would be familiar with this?
I have as my dependent variable (SAT100- Rate your overall satisfaction with your experience here thus far). Responses are on a scale (1-Not satisfied at all to 7 - Very Satisfied).
I want to include as predictor variables: Gender, Age, Ethnicity, ClassLevel, Enrollment (Day, Evening,Weekend), CourseLoad (FullTime, Part-Time), Educational Goal (Bachelors Degree, Masters Degree, etc.), Employment(Full-Time, Part-Time,etc), Residence (Residence Hall, Fraternity, Off-Campus,etc.), Geographic (In-State, Out-of-State, International), Disability Level, Choice (1st choice, 2nd choice, 3rd choice).
I used the follwing procedure in SPSS:
PLUM
SAT100 BY GENDER AGE ETHNIC CURENR CURLOAD CLASSLEV CURRGPA EDUGOAL
EMPLOY CURRRES RESCLASS DISAB CHOICE
/CRITERIA = CIN(95) DELTA(0) LCONVERGE(0) MXITER(100) MXSTEP(5) PCONVERGE
(1.0E-6) SINGULAR(1.0E-8)
/LINK = LOGIT
/PRINT = FIT PARAMETER SUMMARY .
There are 1746 (85.4%) cells (i.e., dependent variable levels by combinations of predictor variable values) with zero frequencies.Unexpected singularities in the Fisher Information matrix are encountered. There may be a quasi-complete separation in the data. Some parameter estimates will tend to infinity.
The PLUM procedure continues despite the above warning(s). Subsequent results shown are based on the last iteration. Validity of the model fit is uncertain..
I have not had too much practice with ordinal regression so I am not sure how I can resolve this problem so I can interpret the results correctly. Also, I am not sure if I am using the independent variables in the best way (i.e, some independent varibales (Ethnicity) have 7 different groups (Caucasian, Latino, Asian-Amnerican, etc.). Is it better to create a binary variable here or can I still interepret results with a large number of levels?
Thanks,
Keval

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

David Hitchin

Re: Ordinal Regression in SPSS

Quoting Keval Khichadia <[hidden email]>:

> Hi,I am trying to do an ordinal regression on the results of a
> StudentSatisfaction Survey, etc

Your problem is not something specific to ordinal regression, but is
often found in ordinary regression or analysis of variance when (1) you
haven't observed all combinations of the data (empty cells) and (2) your
data is unbalanced (different numbers of cases in different cells).

Consider a very simple case. You have a small firm with male managers
whose average salary is 40,000 and female operatives whose average
salary is 15,000. One explanation is that managers re paid more 25,000
more than operatives (regardless of gender) and another is that men are
paid 25,000 more than women (regardless of status), but there are many
intermediate explanations, e.g. that men are paid 5,000 more than women,
and managers are paid 20,000 more than operatives. You can't distinguish
between these from the data, and the only way that you could find out
the real explanation is if you also have some male operatives and some
female managers.

One possibility is that you can just add the gender and status effects
to explain the averages (a "main effects model", but another is that
when a man is promoted to manager he gets a larger rise than a woman who
is promoted (which would be, in technical terms, an "interaction").

In a sense you have been lucky, because SPSS has told you that you have
a problem. Quite often, given similar data, you may get what looks like
a simple solution, but in fact this is a rather arbitrary example from a
vast family of different solutions, each of which fits the model equally
well in a mathematical sense, but so you have a vast range of possible
interpretations.

You might make some progress by drastically simplifying your equation.
You CAN'T get satisfactory results from your data because there ISN'T
enough of it.

However, to use regression or analysis of variance methods for
unbalanced and / or incomplete data requires a lot of statistical
expertise, and even with that, I'm sorry to say, many of the most
interesting and important research questions cannot be answered by such
methods. Just to make this absolutely clear, it is NOT that the
statistical methods are inadequate, but that the data that you have just
doesn't contain sufficient information.

David Hitchin

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

Christian Deindl

Re: Ordinal Regression in SPSS

In reply to this post by Keval Khichadia

I would suggest two practical approaches:

1) try if you can fit a linear regression model
one find it quite often in satisfaction research (with a 11 point scale)

2) but more important: build your model im small steps an enter one
independent variable after the other
then you might be able to find your problem

am I understanding your model correctly that "ethnic" is a categorical
variable?
Then you have to build dummy-variables

christian

Keval Khichadia schrieb:

> Hi,
> I am trying to do an ordinal regression on the results of a Student Satisfaction Survey (Noel Levitz). Maybe someone working in Institutional Research would be familiar with this?
> I have as my dependent variable (SAT100- Rate your overall satisfaction with your experience here thus far). Responses are on a scale (1-Not satisfied at all to 7 - Very Satisfied).
> I want to include as predictor variables: Gender, Age, Ethnicity, ClassLevel, Enrollment (Day, Evening,Weekend), CourseLoad (FullTime, Part-Time), Educational Goal (Bachelors Degree, Masters Degree, etc.), Employment(Full-Time, Part-Time,etc), Residence (Residence Hall, Fraternity, Off-Campus,etc.), Geographic (In-State, Out-of-State, International), Disability Level, Choice (1st choice, 2nd choice, 3rd choice).
> I used the follwing procedure in SPSS:
> PLUM
> SAT100 BY GENDER AGE ETHNIC CURENR CURLOAD CLASSLEV CURRGPA EDUGOAL
> EMPLOY CURRRES RESCLASS DISAB CHOICE
> /CRITERIA = CIN(95) DELTA(0) LCONVERGE(0) MXITER(100) MXSTEP(5) PCONVERGE
> (1.0E-6) SINGULAR(1.0E-8)
> /LINK = LOGIT
> /PRINT = FIT PARAMETER SUMMARY .
> There are 1746 (85.4%) cells (i.e., dependent variable levels by combinations of predictor variable values) with zero frequencies.Unexpected singularities in the Fisher Information matrix are encountered. There may be a quasi-complete separation in the data. Some parameter estimates will tend to infinity.
> The PLUM procedure continues despite the above warning(s). Subsequent results shown are based on the last iteration. Validity of the model fit is uncertain..
> I have not had too much practice with ordinal regression so I am not sure how I can resolve this problem so I can interpret the results correctly. Also, I am not sure if I am using the independent variables in the best way (i.e, some independent varibales (Ethnicity) have 7 different groups (Caucasian, Latino, Asian-Amnerican, etc.). Is it better to create a binary variable here or can I still interepret results with a large number of levels?
> Thanks,
> Keval
>
>
>
>
> ===================
> 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
>

Keval Khichadia

Re: Ordinal Regression in SPSS

In reply to this post by Keval Khichadia

Hi,

Thank you for the responses. I have decided to use linear regression for the student satisfaction survey.

I have as my dependent variable (SAT- So far, how has your college experience met your expectations).

I coded most of the predictor variables as dummy (0,1) and for the categorical variables with many levels such as age and gpa I created AGE1-AGE5, GPA1-GPA5 and coded them (1,0) depending if they fell into that group.

I ran the stepwise regression below, and choice2, gpa4 were identified as significant. REGRESSION
/MISSING LISTWISE
/STATISTICS COEFF OUTS R ANOVA
/CRITERIA=PIN(.05) POUT(.10)
/NOORIGIN
/DEPENDENT SAT99
/METHOD=STEPWISE Gender2 Ethnic2 CurEnr2 CurLoad2 ClassLev2 employ2
currres2 resclass2 disab2 choice2 age1 age2 age3 age4 age5 GPA1 GPA2 GPA3
GPA4 GPA5 GPA6
/PARTIALPLOT ALL
/SCATTERPLOT=(*ZRESID ,*ZPRED )
/RESIDUALS HIST(ZRESID) NORM(ZRESID)
/SAVE RESID .
examine

variables=res_1
/plot boxplot stemleaf histogram npplot.
The stepwise procdure lists 2 models in the output. Are the residuals and output charts based on the last model (the model that includes all the prdictor variables chosen in stepwise procedure)?
I am getting normality tests significant(Kolmogrov-Smirnov, Shapiro-Wilk) but from the stem and leaf plot and the noraml q-q plot of unstandardized residuals the data appear to be normal. Can I still trust the results?
----- Original Message ----
From: Christian Deindl <[hidden email]>
To: [hidden email]
Sent: Monday, June 2, 2008 2:52:36 AM
Subject: Re: Ordinal Regression in SPSS

I would suggest two practical approaches:

1) try if you can fit a linear regression model
one find it quite often in satisfaction research (with a 11 point scale)

2) but more important: build your model im small steps an enter one
independent variable after the other
then you might be able to find your problem

am I understanding your model correctly that "ethnic" is a categorical
variable?
Then you have to build dummy-variables

christian

Keval Khichadia schrieb:

Matthew Reeder

Re: Ordinal Regression in SPSS

Hi Keval,

You'd mentioned applying stepwise as your selection method. Without knowing much about your particular situation, use of stepwise methods is often advised against on several grounds, both statistical (e.g., capitlization on sampling error) and theoretical (aside from your initial choice of variables, it's entirely data-driven). I'm guessing someone else will chime in with additional information or a counter, but you may be better off considering another method (e.g., Enter) if you're going to stick with linear regression.

If you're curious to learn more, one pretty common reference is shown below. There was also a recent thread in the listserv (though focusing more specifically on the issue of multicollinearity; see 'collinearity and stepwise regression' a few weeks back). I'm sure there are other threads, as well, if you search.

Thompson, B. (1995). Stepwise regression and stepwise discriminant analysis need not apply here: A guidelines editorial. Educational and Psychological Measurement, 55, 525-534.

Best of luck,
Matt

Keval Khichadia <[hidden email]> wrote:
Hi,

Thank you for the responses. I have decided to use linear regression for the student satisfaction survey.

I have as my dependent variable (SAT- So far, how has your college experience met your expectations).

I coded most of the predictor variables as dummy (0,1) and for the categorical variables with many levels such as age and gpa I created AGE1-AGE5, GPA1-GPA5 and coded them (1,0) depending if they fell into that group.

I ran the stepwise regression below, and choice2, gpa4 were identified as significant. REGRESSION
/MISSING LISTWISE
/STATISTICS COEFF OUTS R ANOVA
/CRITERIA=PIN(.05) POUT(.10)
/NOORIGIN
/DEPENDENT SAT99
/METHOD=STEPWISE Gender2 Ethnic2 CurEnr2 CurLoad2 ClassLev2 employ2
currres2 resclass2 disab2 choice2 age1 age2 age3 age4 age5 GPA1 GPA2 GPA3
GPA4 GPA5 GPA6
/PARTIALPLOT ALL
/SCATTERPLOT=(*ZRESID ,*ZPRED )
/RESIDUALS HIST(ZRESID) NORM(ZRESID)
/SAVE RESID .
examine

variables=res_1
/plot boxplot stemleaf histogram npplot.
The stepwise procdure lists 2 models in the output. Are the residuals and output charts based on the last model (the model that includes all the prdictor variables chosen in stepwise procedure)?
I am getting normality tests significant(Kolmogrov-Smirnov, Shapiro-Wilk) but from the stem and leaf plot and the noraml q-q plot of unstandardized residuals the data appear to be normal. Can I still trust the results?
----- Original Message ----
From: Christian Deindl
To: [hidden email]
Sent: Monday, June 2, 2008 2:52:36 AM
Subject: Re: Ordinal Regression in SPSS

I would suggest two practical approaches:

1) try if you can fit a linear regression model
one find it quite often in satisfaction research (with a 11 point scale)

2) but more important: build your model im small steps an enter one
independent variable after the other
then you might be able to find your problem

am I understanding your model correctly that "ethnic" is a categorical
variable?
Then you have to build dummy-variables

christian

Keval Khichadia schrieb:

Ornelas, Fermin-2

Re: Ordinal Regression in SPSS

In reply to this post by Keval Khichadia

For the number of variables that you have there is no need to let stepwise give you a final model. Instead you could have all the variables in the model then use modeling diagnostics to arrive at a sound model. Residual Plots, Lillifor's test for normality, Levine tests for constant variance, and collinear diagnostics should help you in arriving at the final model specification.

-----Original Message-----
From: SPSSX(r) Discussion [mailto:[hidden email]] On Behalf Of Keval Khichadia
Sent: Monday, June 02, 2008 3:37 PM
To: [hidden email]
Subject: Re: Ordinal Regression in SPSS

Hi,

Thank you for the responses. I have decided to use linear regression for the student satisfaction survey.

I have as my dependent variable (SAT- So far, how has your college experience met your expectations).

I coded most of the predictor variables as dummy (0,1) and for the categorical variables with many levels such as age and gpa I created AGE1-AGE5, GPA1-GPA5 and coded them (1,0) depending if they fell into that group.

I ran the stepwise regression below, and choice2, gpa4 were identified as significant. REGRESSION
/MISSING LISTWISE
/STATISTICS COEFF OUTS R ANOVA
/CRITERIA=PIN(.05) POUT(.10)
/NOORIGIN
/DEPENDENT SAT99
/METHOD=STEPWISE Gender2 Ethnic2 CurEnr2 CurLoad2 ClassLev2 employ2
currres2 resclass2 disab2 choice2 age1 age2 age3 age4 age5 GPA1 GPA2 GPA3
GPA4 GPA5 GPA6
/PARTIALPLOT ALL
/SCATTERPLOT=(*ZRESID ,*ZPRED )
/RESIDUALS HIST(ZRESID) NORM(ZRESID)
/SAVE RESID .
examine

variables=res_1
/plot boxplot stemleaf histogram npplot.
The stepwise procdure lists 2 models in the output. Are the residuals and output charts based on the last model (the model that includes all the prdictor variables chosen in stepwise procedure)?
I am getting normality tests significant(Kolmogrov-Smirnov, Shapiro-Wilk) but from the stem and leaf plot and the noraml q-q plot of unstandardized residuals the data appear to be normal. Can I still trust the results?
----- Original Message ----
From: Christian Deindl <[hidden email]>
To: [hidden email]
Sent: Monday, June 2, 2008 2:52:36 AM
Subject: Re: Ordinal Regression in SPSS

I would suggest two practical approaches:

1) try if you can fit a linear regression model
one find it quite often in satisfaction research (with a 11 point scale)

2) but more important: build your model im small steps an enter one
independent variable after the other
then you might be able to find your problem

am I understanding your model correctly that "ethnic" is a categorical
variable?
Then you have to build dummy-variables

christian

Keval Khichadia schrieb:

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

NOTICE: This e-mail (and any attachments) may contain PRIVILEGED OR CONFIDENTIAL information and is intended only for the use of the specific individual(s) to whom it is addressed. It may contain information that is privileged and confidential under state and federal law. This information may be used or disclosed only in accordance with law, and you may be subject to penalties under law for improper use or further disclosure of the information in this e-mail and its attachments. If you have received this e-mail in error, please immediately notify the person named above by reply e-mail, and then delete the original e-mail. Thank you.

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