|Date: ||Sun, 8 Mar 2009 09:26:53 -0700|
|Sender: ||"SAS(r) Discussion" <SAS-L@LISTSERV.UGA.EDU>|
|From: ||"Dennis G. Fisher, Ph.D." <dfisher@CSULB.EDU>|
|Organization: ||California State University, Long Beach|
|Subject: ||Re: ordinal data|
|Content-Type: ||text/plain; charset="utf-8"|
What exactly did the reviewer say who did not like the analysis? If the reviewer was very strong on the issue, then you may not get past that reviewer and be accepted by the journal. On rare occasions, you might be successful, but usually you will not be. How important is it, that you use the original analysis? If it is important, then you might want to submit to a different journal and justify your analysis better in a re-write of the paper. If not, then analyze with ordinal logistic etc. FWIW
Dennis G. Fisher, Ph.D.
Professor and Director
Center for Behavioral Research and Services
California State University, Long Beach
1090 Atlantic Avenue
Long Beach, CA 90813
tel: 562-495-2330 x121
From: SAS(r) Discussion [mailto:SAS-L@LISTSERV.UGA.EDU] On Behalf Of Peter Flom
Sent: Sunday, March 08, 2009 8:55 AM
Subject: Re: ordinal data
Arthur Tabachneck <art297@NETSCAPE.NET> wrote
>I haven't kept current in the area, but still wouldn't be as quick as
>Peter to agree.
>Take a look at: http://en.wikipedia.org/wiki/Level_of_measurement
>Can an underlying theoretical distribution be postulated that one could
>use as an argument that the scale is more than just ordinal? I would
>think that the critical issue will be the properties of the numbers themselves.
>I.e., if 3 isn't three times greater than 1, than who can one justify
>using multiplication or division on those numbers?
This gets into tricky areas.
The levels of measurement have become enshrined, although this makes no more sense than the enshrinement of .05 as THE level of significance.
Technically, with (say) a five point ordinal scale, one could use
1 1.1 1.2 1.3 1000000
just as much as
1 2 3 4 5
that is, after all, what ordinal means.
But, in the real world, we often have a sense, even if we cannot prove it, that the scale is *roughly* interval. And fairly few things are really ratio scaled .... even something like income, which looks to be perfectly interval ($10,000 is ten times $1,000, right?) might better be treated differently, as the difference in (say) $20,000 annual income and $10,000 is not the same as between $100,000 and $110,000.
Further, if one looks into the details of many techniques proposed for ordinal scale really depend on more than ordinality.
Ordinal logistic regression at least allows a test of the proportional odds assumption, although the test isn't very good, and violations, unless they are extreme, seem not to make a big difference.
So, what to do?
Rather than get hung up on technical details, I sometimes do the analysis several ways, and compare results. If the predicted values from the different models are similar, then I go with the simplest model.
Peter L. Flom, PhD
www DOT peterflomconsulting DOT com