Ordered logsitic regression

[nina]

Dear all,

I am a bit afraid that my question is not necessarily completely SPSS-related, but would greatly appreciate any comments/ideas/references regarding the following problem:

In a regression model,  my dependent variable is an attitudinal-item measured on a 4-point scale. Thus far, for my analyses I treated this variable as continouos. Now a reviewer challenges this view and recommends to conduct an ordered logistic regression analyses, e.g. via the PLUM Procedure in SPSS. OK, this seems doable but in my view complicates the analyses and interpretation a bit.

So I wonder what form of bias actually might result from modelling a categorical variable (with e.g. just 4 response options) as metric? Are there any papers available dealing with this topic?

Many thanks !
Nina

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

[bdates]

Nina,

There are actually a number of papers.  I'm sending a link to one of the papers which looks like a thesis, but does an ordered logit analysis of five areas, each using a 5-point Likert scale. Her finding is that for the most part Likert scales are ordinal, not interval.  She does say that the interval/ordinal dispute seems to be scale specific, which would mean that some kind of initial analysis should occur before applying a statistic with a particular scale type, akin to data modeling.  Some articles draw the line between ordinal and interval at the 5-point line, with anything less than five response categories needing treatment at the ordinal level. Another issue is whether a total/sum score is being used, or individual items are being used. IRT enthusiasts believe that all ordered data are ordinal until scaled. I'm typically in the minority, but for the most part treat Likert-type data as ordinal. A thousand years ago I had a course with Rensis Likert, and what most people call Likert scaling is not at all; it's simply attaching 5 or more numbers to response categories over a series of statements. I definitely draw the line there because no scaling technique was really ever used.  Good luck with this.  You'll get some views from the other side.

http://www.olin.wustl.edu/docs/CRES/MarkdaSilva.pdf 

Brian


-----Original Message-----
From: SPSSX(r) Discussion [mailto:[hidden email]] On Behalf Of Nina Lasek
Sent: Monday, September 15, 2014 11:36 AM
To: [hidden email]
Subject: Ordered logsitic regression

Dear all,

I am a bit afraid that my question is not necessarily completely SPSS-related, but would greatly appreciate any comments/ideas/references regarding the following problem:

In a regression model,  my dependent variable is an attitudinal-item measured on a 4-point scale. Thus far, for my analyses I treated this variable as continouos. Now a reviewer challenges this view and recommends to conduct an ordered logistic regression analyses, e.g. via the PLUM Procedure in SPSS. OK, this seems doable but in my view complicates the analyses and interpretation a bit.

So I wonder what form of bias actually might result from modelling a categorical variable (with e.g. just 4 response options) as metric? Are there any papers available dealing with this topic?

Many thanks !
Nina

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

[Rich Ulrich]

All in all -- That is not a very good dissertation.  And what it tests is
whether her own (probably non-Likert) examples are "ordinal" in some absolute
sense, as judged by the failure to appear "interval as analyzed by logistic."
The author rejects PLUM; the author does not test anything in comparison, to
say that PLUM works worse.

In introduction, she states that "most Likert scale data is severely skewed."
I would state, in partial contradiction, that any item which is severely
skewed has disqualified itself being properly called Likert, even loosely.

IMO, an understanding of "interval" versus "ordinal" requires the statement of
the criterion.... I don't know if the textbooks are that wise, but they should be.
 - She gives "degrees Fahrenheit" as the textbook example of equal intervals;
but that is true for many homely examples, but it is not useful (say) for
thermonuclear reactions, which are modeled more successfully with some
power law.  A similar argument demonstrates that degrees F may be, *indeed*,
a ratio measurement, with "absolute zero" at 91.5 (the temperature of skin that
allows heat loss for cooling), if you are modeling a human-subjective heat-index.

Since there is no reason that I recall for her criterion to be considered as forming
"equal intervals" in any substantial, a-priori sense, it is hard to be impressed with
the evidence that her scale data failed to match those arbitrary intervals.

PS.  It sounds to me like an overly-officious reviewer. 

Suggestion:  Try PLUM.  Hopefully, you can then say to the editor
(or in a footnote), "PLUM gave similar results and did not shed any
additional light on the question; and we find the ordinary regression
to be easier to understand and present."

--
Rich Ulrich

> Date: Mon, 15 Sep 2014 16:35:57 +0000

> From: [hidden email]
> Subject: Re: Ordered logsitic regression
> To: [hidden email]
>
> Nina,
>
> There are actually a number of papers. I'm sending a link to one of the papers which looks like a thesis, but does an ordered logit analysis of five areas, each using a 5-point Likert scale. Her finding is that for the most part Likert scales are ordinal, not interval. She does say that the interval/ordinal dispute seems to be scale specific, which would mean that some kind of initial analysis should occur before applying a statistic with a particular scale type, akin to data modeling. Some articles draw the line between ordinal and interval at the 5-point line, with anything less than five response categories needing treatment at the ordinal level. Another issue is whether a total/sum score is being used, or individual items are being used. IRT enthusiasts believe that all ordered data are ordinal until scaled. I'm typically in the minority, but for the most part treat Likert-type data as ordinal. A thousand years ago I had a course with Rensis Likert, and what most people call Likert scaling is not at all; it's simply attaching 5 or more numbers to response categories over a series of statements. I definitely draw the line there because no scaling technique was really ever used. Good luck with this. You'll get some views from the other side.
>
> http://www.olin.wustl.edu/docs/CRES/MarkdaSilva.pdf
>
> Brian
>
>
> -----Original Message-----
> From: SPSSX(r) Discussion [mailto:[hidden email]] On Behalf Of Nina Lasek
> Sent: Monday, September 15, 2014 11:36 AM
> To: [hidden email]
> Subject: Ordered logsitic regression
>
> Dear all,
>
> I am a bit afraid that my question is not necessarily completely SPSS-related, but would greatly appreciate any comments/ideas/references regarding the following problem:
>
> In a regression model, my dependent variable is an attitudinal-item measured on a 4-point scale. Thus far, for my analyses I treated this variable as continouos. Now a reviewer challenges this view and recommends to conduct an ordered logistic regression analyses, e.g. via the PLUM Procedure in SPSS. OK, this seems doable but in my view complicates the analyses and interpretation a bit.
>
> So I wonder what form of bias actually might result from modelling a categorical variable (with e.g. just 4 response options) as metric? Are there any papers available dealing with this topic?
>
> Many thanks !
> Nina

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