The links from the help for -logit- in
your version of Stata show what StataCorp
has implemented.
Again, I wouldn't read too much into that.
On occasion, the existence of well-written
and documented user programs is a good argument for
StataCorp _not_ to implement something. That way
they save on writing the program "properly",
testing it hard with certification scripts,
writing help files to current standards,
reading up the literature,
writing manual documentation, technical support,
and future maintenance as other things change.
Asymptotically, all Stata programs will be
written by users, just as those who write
Fortran compilers don't write most Fortran programs.
More seriously, what's really important in Stata
development is whatever is quite beyond users.
Nick
[email protected]
Kallimanis, Bellinda
Thank you, Nick for your response. I checked the agreement between the 2
probabilities and found the probabilities not to be as in agreement as
what I thought at first by looking at the pairwise correlation. I think
I should then stick with cloglog or do some more reading on -relogit-.
When I asked about -relogit- being last updated in 1999 I simply was
wondering whether stata had developed something of it's own since then
for analyzing rare events.
n j cox
Your results suggest very high correlation between
predicted probabilities from complementary log log
and logit. Just check that there is also very high
agreement. For example, 10x and x are correlated +1
but only at 0 do they agree.
I'm not sure what you infer from the fact that Gary King's program
dates from 1999. Gary was a keen Gauss user, is now a
keen R user. Some of the people in his group have used
Stata heavily. I think that's the main reason -relogit-
has not been updated recently.
Nick
[email protected]
Kallimanis, Bellinda
I have data where my outcome is a rare event, it occurs in 0.97% of my
sample (n =11,618). So I was reading about complimentary log log
regression and thought it may be appropriate, but when I compared the
predicted probabilities of the complimentary log log model and a logit
model I get a pariwise correlation of 0.9991 which suggests to me that
the complimentary log log model isn't doing a better job of predicting
than the logit model. The coefficients are reasonably close to each
other, see output below.
Does this mean I should stick with a logit model and perhaps just alter
the cutoff value? Also I came across the work of Gary King and his
-relogit- command found at http://gking.harvard.edu/stats.shtml#relogit
though I see this was last updated in 1999 so I'm not sure how relevant
it is. Any thoughts would be greatly appreciated.
*
* For searches and help try:
* http://www.stata.com/support/faqs/res/findit.html
* http://www.stata.com/support/statalist/faq
* http://www.ats.ucla.edu/stat/stata/
