You can do this easily. Run your model
.logistic dead crib
.predict phat ----> this gives you the probability of "dead"
.predict xb, xb ----> this is the linear predictor, e.g., the solution of
your regression equation
.sort phat ----> this sorts the dataset from lowest to highest phat
values, to correspond with the H-L table
.list crib xb phat
The H-L table divides the dataset into percentiles of pr(dead) from lowest
to highest. The -list- command will associate each "crib" value with its
linear predictor, xb and its predicted probability, phat. The H-L cutpoints
will correspond to the phat listing.
paul
Paul F. Visintainer, PhD
School of Public Health
New York Medical College
Valhalla, NY 10595
(914) 594-4804 (phone)
(914) 594-4292 (fax)
-----Original Message-----
From: Kaaresen Per Ivar [mailto:[email protected]]
Sent: Wednesday, June 12, 2002 9:08 AM
To: '[email protected]'
Subject: st: Groups (qantiles) in Hosmer-Lemeshows goodness of fit test
Ikke sensitiv - ignore message - due to in-house security
Dear friends
Could anybody help me with this (again) probably rather simple problem.
I'm working with this CRIB (clinical risk index for babies) ability to
predict hospital death in a population (not individual) level. To test the
calibration of the score I've done a logistic regression and then a
Hosmer-Lemesow goodness of fit test. The variables are dead and crib (range
0-21, integers only).
. logit dead crib
Iteration 0: log likelihood = -206.16559
Iteration 1: log likelihood = -155.91351
Iteration 2: log likelihood = -150.13801
Iteration 3: log likelihood = -150.00793
Iteration 4: log likelihood = -150.00765
Logit estimates Number of obs =
443
LR chi2(1) =
112.32
Prob > chi2 =
0.0000
Log likelihood = -150.00765 Pseudo R2 =
0.2724
----------------------------------------------------------------------------
--
dead | Coef. Std. Err. z P>|z| [95% Conf.
Interval]
-------------+--------------------------------------------------------------
--
crib | .3516896 .0414307 8.49 0.000 .2704869
.4328923
_cons | -4.086216 .3701967 -11.04 0.000 -4.811788
-3.360644
----------------------------------------------------------------------------
--
. lfit,group(10) table
Logistic model for dead, goodness-of-fit test
(Table collapsed on quantiles of estimated probabilities)
note: because of ties, there are only 9 distinct quantiles
_Group _Prob _Obs_1 _Exp_1 _Obs_0 _Exp_0 _Total
1 0.0233 2 1.8 74 74.2 76
3 0.0328 2 2.2 65 64.8 67
4 0.0642 3 3.7 60 59.3 63
5 0.0888 2 3.2 34 32.8 36
6 0.1217 4 2.9 20 21.1 24
7 0.2188 15 13.7 56 57.3 71
8 0.2848 10 9.7 24 24.3 34
9 0.4458 12 12.4 20 19.6 32
10 0.9644 28 28.4 12 11.6 40
number of observations = 443
number of groups = 9
Hosmer-Lemeshow chi2(7) = 1.34
Prob > chi2 = 0.9874
My question is (yes, I see the possible problem with small numbers in 5 of
the cells - but that's not the question now): How can I find the intervals
in CRIB score which the different _Group (qantiles) represent? I would like
to present the results something like:
CRIB _Prob _Obs_1 _Exp_1 _Obs_0 _Exp_0 _Total
0-1 0.0233 2 1.8 74 74.2 76
2-4 0.0328 2 2.2 65 64.8 67
5-7 0.0642 3 3.7 60 59.3 63
. 0.0888 2 3.2 34 32.8 36
. 0.1217 4 2.9 20 21.1 24
. 0.2188 15 13.7 56 57.3 71
. 0.2848 10 9.7 24 24.3 34
. 0.4458 12 12.4 20 19.6 32
. 0.9644 28 28.4 12 11.6 40
I've seen this kind of presentation in different papers - and would be very
thankful if somebody could explain how I can do this in Stata.
Regards
Per Ivar Kaaresen
MD
University Hospital Nothern Norway
*
* 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/
*
* 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/
