Bookmark and Share

Notice: On April 23, 2014, Statalist moved from an email list to a forum, based at statalist.org.


[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: st: fit index for ordered logistic regression


From   Darcy Hannibal <dlhannibal@ucdavis.edu>
To   statalist@hsphsun2.harvard.edu
Subject   Re: st: fit index for ordered logistic regression
Date   Sun, 15 Dec 2013 22:44:34 -0800

While I prefer the Bayesian approach, if you haven't used it before, this might be more than you want to tackle for a manuscript that is pretty much done. You could just respond to the comment by the reviewer and to say you are not using Bayesian methodology. To do this right, you typically need many variables to build a useful model, not just a few variables to test a specific hypothesis.

On 12/14/2013 2:13 PM, Lucas wrote:
David Hoaglin makes a good point.  In addition, you might look at:

Raftery, Adrian E.  1995.  "Bayesian Model Selection in Social
Research." *Sociological Methodology* 25: 111-163.

He has a nice discussion of what you might regard as cut-offs for a
meaningful difference between models.  At the risk of undermining
incentive to read Raftery's very informative paper, check out, for
example, Tables 8 and 9 on page 141. The text further discusses what
counts as weak, strong, and very strong evidence in a comparison of
models.

Take care
Sam

On Sat, Dec 14, 2013 at 2:03 PM, David Hoaglin <dchoaglin@gmail.com> wrote:
Hi, Julie.

I'm not sure what you mean by "imbedded."

AIC is not an absolute measure.  If you have a set of models that
should be good models on substantive grounds, you can choose the model
that has the smallest AIC.  Those models do not need to be nested, and
they do not have to be based on the same family of distributions (as
long as the calculation of AIC includes all the constants in the
likelihood).  The usual definition of AIC applies to large samples.
If the ratio of the sample size (n) to the total number of parameters
(K) is not "large," say n/K < 40, it is better to use a modified
version of AIC.

Are you able to use deviance (and DIC) to compare each of your models
against the corresponding "saturated" model?

Also, Agresti (2010) discusses ways of assessing goodness of fit.

David Hoaglin

Agresti, A. (2010).  Analysis of Ordinal Categorical Data, 2nd ed.
John Wiley & Sons.


On Sat, Dec 14, 2013 at 2:48 PM, Julie Lamoureux <drjuliel@comcast.net> wrote:
Good day,

I am looking for "normed" goodness of fit indices for ordered logistic
regression.  We submitted a manuscript with the results of 2 OLR (on two
different unrelated outcomes) and the reviewer is asking how we assessed
model goodness of fit and its usefulness.  I know AIC and BIC let us compare
models that are "imbedded" but can we use AIC and BIC to determine how
useful the model is?  If so, what "cut-off" for those indices can be
considered "good"?  Is there something else I am not aware of to answer this
question?

Thank you for your time

Julie Lamoureux
*
*   For searches and help try:
*   http://www.stata.com/help.cgi?search
*   http://www.stata.com/support/faqs/resources/statalist-faq/
*   http://www.ats.ucla.edu/stat/stata/
*
*   For searches and help try:
*   http://www.stata.com/help.cgi?search
*   http://www.stata.com/support/faqs/resources/statalist-faq/
*   http://www.ats.ucla.edu/stat/stata/


---
This email is free from viruses and malware because avast! Antivirus protection is active.
http://www.avast.com

*
*   For searches and help try:
*   http://www.stata.com/help.cgi?search
*   http://www.stata.com/support/faqs/resources/statalist-faq/
*   http://www.ats.ucla.edu/stat/stata/


© Copyright 1996–2018 StataCorp LLC   |   Terms of use   |   Privacy   |   Contact us   |   Site index