Thanks! And a great article, as usual! -Don
On Thu, Sep 4, 2008 at 7:46 PM, Gary King <king(a)harvard.edu> wrote:
for this and any other likelihood-based model, you
can compute a pseudo-R2
as follows. Run your model and record the log-likelihood. Then run the
same model without any of your explanatory variables (i.e., with only a
constant term). Then compute the proportionate reduction in error from one
to the other.
A lot of people use these statistics, and so I can see why you'd want to
pay attention to them. But they can be highly misleading. Some time ago, I
explained these points in a AJPS article called "how not to lie with
statistics"; a copy is here
http://gking.harvard.edu/files/abs/mist-abs.shtml and a follow-up article
is here
http://gking.harvard.edu/files/abs/truth-abs.shtml
Gary
---
http://gking.harvard.edu
On 9/4/2008 7:17 PM, Donald Braman wrote:
You are probably right statistically speaking, but when presenting findings
to people who are unfamiliar with logits but who are familiar with standard
mvr, it helps them develop a rough understanding the results in terms they
are familiar with. Zelig helps by making results pretty easy to graph (like
the old Clarify in Stata), but I would love to have pseudo r2 to help
communicate findings.
On Thu, Sep 4, 2008 at 7:06 PM, Kosuke Imai <kimai(a)princeton.edu> wrote:
I don't know. I also don't think that
the pseudo R^2 is very useful.
Kosuke
On Sep 4, 2008, at 5:53 PM, Donald Braman wrote:
Interesting -- I see that polr doesn't provide much in the way of fit
statistics and that Zelig draws on polr. Do you
know of a way to calculate
the pseudo R2 from the Zelig results? I'm coming from Stata where it's
provided as part of the ologit command. I'd hate to go back to Stata just
for this.
On Thu, Sep 4, 2008 at 5:34 PM, Kosuke Imai <kimai(a)princeton.edu> wrote:
You can always take a mean of statistics across multiply imputed data
sets and then use that as an estimate.
Kosuke
On Sep 4, 2008, at 5:02 PM, Donald Braman wrote:
Is there a simple way to view estimated fit statistics in Zelig for
ologit analyses with multiply imputed data sets? If not, any workarounds?