 |
[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]
Re: st: survival
Sebastián
Your set-up implies that people were randomly selected within
strata. In most complex surveys the first unit selected is a
primary sampling unit which contains many people. If there is such a
variable in the data (e.g. psu_var), then the -svyset- command would go:
svyset psu_var [pw =exp] strata(estrato)
If indeed people were randomly selected at the first stage, then I
would expect, as you found, that the standard errors for -logit- and
for -svy: logit- would be very close.
-Steven
On Jan 2, 2009, at 7:52 AM, Sebastián Daza wrote:
Hi Steven,
In fact, I created a new observation for each person and period. I
have used svy: logit and plain logit, but there aren't differences
between results. Sampling desing is stratified and in person-level
data (not expanded) each person have a weight (inverse of the
probability that the observation is included in strata). I keep this
factor with the expanded person-period dataset.
svyset _n [pw=exp], strata(estrato)
svy: logit nevent d4-d18 sex, nocons
(running logit on estimation sample)
Survey: Logistic regression
Number of strata = 34 Number of obs
= 16098
Number of PSUs = 16098 Population size
= 50866.46
Design df
= 16064
F( 16, 16049)
= 210.37
Prob > F
= 0.0000
----------------------------------------------------------------------
--------
| Linearized
nevent | Coef. Std. Err. t P>|t| [95%
Conf. Interval]
-------------
+----------------------------------------------------------------
d4 | -8.328679 1.000458 -8.32 0.000
-10.28969 -6.367669
d5 | -7.196868 .5784475 -12.44 0.000
-8.33069 -6.063046
d6 | -6.457124 .8443208 -7.65 0.000
-8.112087 -4.80216
d7 | -5.112672 .3778669 -13.53 0.000
-5.853334 -4.372011
d8 | -4.454505 .2654388 -16.78 0.000
-4.974795 -3.934216
d9 | -4.36023 .2578381 -16.91 0.000
-4.865621 -3.854838
d10 | -2.906085 .1389361 -20.92 0.000
-3.178415 -2.633755
d11 | -3.376975 .1718178 -19.65 0.000
-3.713757 -3.040192
d12 | -2.344076 .1135556 -20.64 0.000
-2.566658 -2.121495
d13 | -2.003841 .1034748 -19.37 0.000
-2.206663 -1.801019
d14 | -1.533507 .1007537 -15.22 0.000
-1.730995 -1.336018
d15 | -1.20536 .1039184 -11.60 0.000
-1.409051 -1.001668
d16 | -1.09949 .1303548 -8.43 0.000
-1.355 -.8439799
d17 | -1.283726 .2046521 -6.27 0.000
-1.684867 -.8825846
d18 | -3.116538 1.02575 -3.04 0.002
-5.127122 -1.105953
sex | -.6876329 .1434859 -4.79 0.000 -.
9688814 -.4063845
----------------------------------------------------------------------
--------
plain logit
. logit nevent d4-d18 sex [pw=exp], nolog nocons
(sum of wgt is 5.0866e+04)
Logistic regression Number of obs
= 16098
Wald chi2(16)
= 3351.56
Log pseudolikelihood = -2734.8293 Prob > chi2
= 0.0000
----------------------------------------------------------------------
--------
| Robust
nevent | Coef. Std. Err. z P>|z| [95%
Conf. Interval]
-------------
+----------------------------------------------------------------
d4 | -8.328679 1.000511 -8.32 0.000
-10.28964 -6.367714
d5 | -7.196868 .5785008 -12.44 0.000
-8.330709 -6.063027
d6 | -6.457124 .8442876 -7.65 0.000
-8.111897 -4.80235
d7 | -5.112672 .3779941 -13.53 0.000
-5.853527 -4.371817
d8 | -4.454505 .2654323 -16.78 0.000
-4.974743 -3.934268
d9 | -4.36023 .2578938 -16.91 0.000
-4.865692 -3.854767
d10 | -2.906085 .1389697 -20.91 0.000
-3.178461 -2.63371
d11 | -3.376975 .1718226 -19.65 0.000
-3.713741 -3.040208
d12 | -2.344076 .1135966 -20.64 0.000
-2.566721 -2.121431
d13 | -2.003841 .1035396 -19.35 0.000
-2.206775 -1.800907
d14 | -1.533507 .1007861 -15.22 0.000
-1.731044 -1.335969
d15 | -1.20536 .1040183 -11.59 0.000
-1.409232 -1.001487
d16 | -1.09949 .1303473 -8.44 0.000
-1.354966 -.844014
d17 | -1.283726 .2046078 -6.27 0.000
-1.68475 -.8827017
d18 | -3.116538 1.025502 -3.04 0.002
-5.126485 -1.10659
sex | -.6876329 .1436219 -4.79 0.000 -.
9691267 -.4061391
----------------------------------------------------------------------
--------
When I computed plain logit I can compute deviance, BIC and AIC
without problem, it don't happen when I use svy logit. In this case,
are this methods equivalent?
Thanks for all you responses in advance.
Regards, Sebastián
Sebastián,
Without more information, I cannot tell whether your survival
setup is correct. With logistic regression, one has to create a
new observation for each person and period. I think that this is
what you have done. However, with a complex sample design you
should -svyset- your data and use -svy: logistic-, not plain -
logistic-. This will compute standard errors appropriate to the
design. You might also try -gllamm-, which will accept PSU's and -
pweights- and will fit a model with added heterogeneity.
With the expanded person-period setup, -svy: cloglog- will fit a
grouped proportional hazards model. Download Stephen Jenkins's -
hshaz- from SSC and see the references in the -help-.
-Steven
On Dec 31, 2008, at 1:13 PM, Sebastián Daza wrote:
I'm working with a discrete-time survival model, and I have doubts
about how to use weight (pweights) in a person-period data. I did
the
following:
logit nevent d4-d18 [pw=exp], nolog nocons
Is it correct?
exp is sampling weight, inverse of the probability that the
observation is included because of the sampling design. When I use
pweights I have problem with lrtest too, because I have to
"force" the
process.
*
* For searches and help try:
* http://www.stata.com/help.cgi?search
* http://www.stata.com/support/statalist/faq
* http://www.ats.ucla.edu/stat/stata/
--
Sebastián Daza Aranzaes
(56 2) 8 921 04 60 / (56 2) 28 307 45
[email protected]
--
Sebastián Daza Aranzaes
Sociólogo UC
[email protected]
*
* For searches and help try:
* http://www.stata.com/help.cgi?search
* http://www.stata.com/support/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/statalist/faq
* http://www.ats.ucla.edu/stat/stata/