Date: Tue, 9 Sep 2008 17:48:06 -0300
Reply-To: Hector Maletta <hmaletta@fibertel.com.ar>
Sender: "SPSSX(r) Discussion" <SPSSX-L@LISTSERV.UGA.EDU>
From: Hector Maletta <hmaletta@fibertel.com.ar>
Subject: Re: Logistic Regression - Threshold effect sizes?
In-Reply-To: <cd3517fa0809091319w1a06d699g70bd05667775fff@mail.gmail.com>
Content-Type: text/plain; charset="us-ascii"
Of course, Jon, my formulation of the dummies is algebraically equivalent to
Scott's.
Hector
_____
From: Jon Bernard [mailto:jon563@gmail.com]
Sent: 09 September 2008 17:19
To: Hector Maletta; SPSSX-L@listserv.uga.edu
Subject: Re: Logistic Regression - Threshold effect sizes?
Thank you Hector.
Both of the approaches you recommended produced useful results. I am still
working out the interpretation - see my email to Scott from a few minutes
ago - but I'm getting there.
As you might expect, the first type of dummy coding you demonstrated
("COMPUTE DUMMYCAT3=(yourvar=3).") produced results identical to those from
the categorical contrasts method Scott recommended.
Thanks again.
Jon
On Fri, Sep 5, 2008 at 7:31 PM, Hector Maletta <hmaletta@fibertel.com.ar>
wrote:
Dummy variables are only partially useful in the case of Jon, because they
do not necessarily generate ordered results (as expected).
Jon: to create dummy variables in general you simply create one dummy with
values (0,1) for each category minus one. The omitted category is omitted
because it can be deduced from the rest. General syntax for one category:
COMPUTE DUMMYCAT3=(yourvar=3).
The new variable, DUMMYCAT3, will equal 1 whenever your original variable
YOURVAR equals 3, and will equal 0 otherwise. Repeat this for all categories
except one (for instance, you may choose the first or last category for
omission).
However, if you have represented your 5-category variables by four dummies
each, you are not guaranteed to obtain monotonic results, i.e. that the
effect increases or decreases in a monotonic way (by equal or different
amounts) for each increase in the ordinal response from 1 to 5. You may well
find that the effect increases for category 1 and 3, but decreases for
category 2 and 4, which could be incomprehensible from a theoretical point
of view and may be a random effect of your sample.
You can create "incremental dummies" by assigning the value 1 to cases
having one value OR LESS:
COMPUTE UPTOONE=(YOURVAR=1).
COMPUTE UPTOTWO=(YOURVAR LE 2).
COMPUTE UPTOTHREE=(YOURVAR LE 3).
COMPUTE UPTOFOUR=(YOURVAR LE 4).
(it is not necessary to create UPTO FIVE, because people choosing category
five can be deduced as the complement of UPTOFOUR).
The increase (or decrease) in the effect between UPTOTWO and UPTOTHREE would
be the specific effect of choosing 3. This ASSUMES the underlying variable
represented by your ordinal IV is monotonic. If you are not sure about this,
you may use simple dummies as in my first example with DUMMYCAT1 to
DUMMYCAT4. Or you may try both and look at the results.
In any case, try to avoid dummies where very few cases have 0 or 1 (i.e.
dummies with very low or very high frequencies in each value): results could
be unstable and unreliable. If, for instance, very few people chose category
5, you may group them with category 5 and define the variable in question as
a four-category ordinal variable. Of course this way you sacrifice
information (difference between 4 and 5) but you avoid the unpleasant
consequences of a very small sample of cases in category 5.
Hope this helps.
Hector
-----Original Message-----
From: SPSSX(r) Discussion [mailto:SPSSX-L@LISTSERV.UGA.EDU] On Behalf Of SR
Millis
Sent: 05 September 2008 18:19
To: SPSSX-L@LISTSERV.UGA.EDU
Subject: Fw: Re: Logistic Regression - Threshold effect sizes?
Jon,
If you're uncomfortable treating your ordinal variable
as though it were interval, you can always "dummy
code" the ordinal variable---such that the 5-category
variable become 4 separate variables---and direct
comparisons can be made. This can work pretty well if you
don't have a lot of ordinal vairables and if you're
sample size is sufficiently large.
Scott Millis
> --- On Fri, 9/5/08, Justin Black
> <justin.black@gmail.com> wrote:
>
> > From: Justin Black <justin.black@gmail.com>
> > Subject: Re: Logistic Regression - Threshold effect
> sizes?
> > To: "SR Millis" <srmillis@yahoo.com>
> > Date: Friday, September 5, 2008, 6:13 PM
> > Scott, thank you for following up.
> >
> > What's puzzling to me is the seemingly omnibus
> nature
> > of the effect. So, I
> > could say that a 1-point increase in IV1 is associated
> with
> > a, e.g., 40%
> > increase in the likelihood of the event occurring.
> But
> > that assumes that a
> > 1-point increase in the IV has the same effect on the
> DV
> > regardless of the
> > baseline level of the IV. I don't think that
> > assumption is accurate in
> > these particular data.
> >
> > I really feel like I'm missing something here,
> just
> > can't figure out what.
> > A look back at Hosmer & Lemeshow didn't help
> any.
> >
> > Thank you,
> >
> > Jon
> >
> > On Fri, Sep 5, 2008 at 6:00 PM, SR Millis
> > <srmillis@yahoo.com> wrote:
> >
> > > Jon,
> > >
> > > What is puzzling in the results?
> > >
> > > Have you examined the degree of collinearity
> among the
> > > predictors/covariates?
> > >
> > >
> > >
> > > Scott R Millis, PhD, MEd, ABPP (CN,CL,RP), CStat
> > > Professor & Director of Research
> > > Dept of Physical Medicine & Rehabilitation
> > > Wayne State University School of Medicine
> > > 261 Mack Blvd
> > > Detroit, MI 48201
> > > Email: smillis@med.wayne.edu
> > > Tel: 313-993-8085
> > > Fax: 313-966-7682
> > >
> > >
> > > --- On Fri, 9/5/08, Jon Bernard
> > <jon563@gmail.com> wrote:
> > >
> > > > From: Jon Bernard <jon563@gmail.com>
> > > > Subject: Logistic Regression - Threshold
> effect
> > sizes?
> > > > To: SPSSX-L@LISTSERV.UGA.EDU
> > > > Date: Friday, September 5, 2008, 5:42 PM
> > > > Fellow SPSSers,
> > > >
> > > > I am struggling with a logistic regression
> issue
> > and
> > > > thought I'd put it out
> > > > to the list for some clarity.
> > > >
> > > > The dependent variable of interest is a
> binary
> > event (0 =
> > > > Did not happen, 1
> > > > = Did happen). The independent variables of
> > interest are
> > > > ordinal
> > > > attitudinal survey items, with responses on
> a
> > 5-point scale
> > > > (1 = Strongly
> > > > Disagree, 5 = Strongly Agree). I have been
> using
> > logistic
> > > > regression for
> > > > the analysis, but either I'm missing
> > something in the
> > > > results output or I'm
> > > > using the wrong statistical technique. I
> have an
> > inkling
> > > > that the intervals
> > > > between categories of the independent
> variables
> > are not all
> > > > equal. In other
> > > > words, I think that the impact on the
> dependent
> > variable of
> > > > an independent
> > > > variable score of 3 vs. one of 2 is greater
> than
> > that of a
> > > > score of 5 vs.
> > > > one of 4. Is that clear? If so, what would
> you
> > recommend
> > > > in order to test
> > > > that hypothesis? Is there a class of
> techniques
> > designed
> > > > particularly for
> > > > this kind of test?
> > > >
> > > > Very many thanks in advance for your
> assistance
> > with this
> > > > matter.
> > > >
> > > > Kind regards,
> > > >
> > > > Jon
> > > >
> > > > =====================
> > > > To manage your subscription to SPSSX-L, send
> a
> > message to
> > > > LISTSERV@LISTSERV.UGA.EDU (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
LISTSERV@LISTSERV.UGA.EDU (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
LISTSERV@LISTSERV.UGA.EDU (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
LISTSERV@LISTSERV.UGA.EDU (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
| 
