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: AW: AW: st: Competing Hazards with Multiple-Record-per-Subject Data (2)
From
Steven Samuels <[email protected]>
To
[email protected]
Subject
Re: AW: AW: st: Competing Hazards with Multiple-Record-per-Subject Data (2)
Date
Mon, 25 Jul 2011 11:25:12 -0400
Other thoughts: Use the more general Lum-McNeil stratum model and cluster on ID to get robust standard errors.
Other than these, I'm out of ideas, so perhaps others will chime in. And, please, as the FAQ request, show _all_ your code (including the -stsplit- and -stset- commands) and what Stata typed. That will save further misunderstandings and wasted time.
Reference:
Lunn, M., & McNeil, D. (1995). Applying Cox regression to competing risks. Biometrics, 51(2), 524-532.
Steve
[email protected]
On Jul 25, 2011, at 8:54 AM, Stefan Göke wrote:
Thanks,
just to be clear: the subepisode variable is the one generated by Stata when
you stsplit the data and consequently not used in the regression. I excluded
the key covariates of my model to ease the discussion. (obviously not
successfully). Yet, the data setup works perfectly fine with Cox and Fine
and Gray models. My concern is "only" the implementation of Lunn/McNeil 1005
augmentation technique with multiple-record-per-subject data. Here, breaking
the ties is I think already a step forward. But further comments are
welcome.
Thanks and best regards
Stefan
-----Ursprüngliche Nachricht-----
Von: [email protected]
[mailto:[email protected]] Im Auftrag von Steven Samuels
Gesendet: Montag, 25. Juli 2011 14:23
An: [email protected]
Betreff: Re: AW: st: Competing Hazards with Multiple-Record-per-Subject Data
(2)
Stefan,
I notice two things.
1) Your "subepisode" variable is perfectly collinear with time.
2) Your "risk type" variable looks like your outcome indicator (it's the
only variable that differs in the two parts of your listing). I don't have
Lunn/McNeill, but I don't see how you can use the outcome indicator as a
covariate for predicting that outcome.
I can suggest:
1) Make sure that your multiple failure model corresponds to one of those in
http://www.stata.com/support/faqs/stat/stmfail.html
2) Get that model working with a single outcome first: ordinary survival
analysis, treating the other outcome as censored.
Good luck
Steve
[email protected]
On Jul 25, 2011, at 4:46 AM, Stefan Göke wrote:
Hi Steve,
thanks for your reply! I do not use Fine and Gray because I seek to model
risk type as a covariate following Lunn/McNeil 1995.
Following your suggestion, I added to each ending time of one of the two
risk types the constant of 0.001.
After stset, the data looks like the following excerpt for ids “2” and “6”,
with
- id = id,
- supedno = subepisode after stsplit,
- risk = risk type either 0 or 1,
- ddes_risk = failure variable,
- dtime1_risk = ending time of the record
+-----------------------------------------------------------------------+
| id subepno risk ddes_r~k dtime1~k _st _d _t
_t0 |
|-----------------------------------------------------------------------
|
20. | 4 0 1 . 1 1 0 1
0 |
21. | 4 1 1 . 2 1 0 2
1.0001 |
22. | 4 2 1 . 3 1 0 3
2.0000999 |
23. | 4 3 1 1 4 1 1 4
3.0000999 |
+-----------------------------------------------------------------------+
|-----------------------------------------------------------------------
|
32. | 6 0 1 . 1 1 0 1
0 |
33. | 6 1 1 . 2 1 0 2
1.0001 |
34. | 6 2 1 . 3 1 0 3
2.0000999 |
35. | 6 3 1 . 4 1 0 4
3.0000999 |
36. | 6 4 1 . 5 1 0 5
4.0001001 |
37. | 6 5 1 . 6 1 0 6
5.0001001 |
38. | 6 6 1 . 7 1 0 7
6.0001001 |
39. | 6 7 1 0 8 1 0 8
7.0001001 |
+-----------------------------------------------------------------------+
|-----------------------------------------------------------------------
|-
|
2094. | 2 0 0 . 1.0001 1 0 1.0001
1 |
2095. | 2 1 0 . 2.0001 1 0 2.0000999
2 |
2096. | 2 2 0 1 3.0001 1 1 3.0000999
3 |
+------------------------------------------------------------------------+
|-----------------------------------------------------------------------
|-
|
2117. | 6 0 0 . 1.0001 1 0 1.0001
1 |
2118. | 6 1 0 . 2.0001 1 0 2.0000999
2 |
2119. | 6 2 0 . 3.0001 1 0 3.0000999
3 |
2120. | 6 3 0 . 4.0001 1 0 4.0001001
4 |
2121. | 6 4 0 . 5.0001 1 0 5.0001001
5 |
2122. | 6 5 0 . 6.0001 1 0 6.0001001
6 |
2123. | 6 6 0 . 7.0001 1 0 7.0001001
7 |
2124. | 6 7 0 0 8.0001 1 0 8.0001001
8 |
+------------------------------------------------------------------------+
Is this how you meant it? I am asking because my regression does not work
sound when integrating the “risk” variable into my stcox survival analysis –
for “risk” the p-value is 1.000 and no confidence interval is reported.
Moreover, no rho-value is reported in the Schoenfeld residuals test. I tried
clustering the data when using stcox now by id with the option vce(cluster
varname) to get robust standard errors given the replication of
observations. However, the rho-value still is not reported. How could this
be solved?
Many thanks!
Stefan
-----Ursprüngliche Nachricht-----
Von: [email protected]
[mailto:[email protected]] Im Auftrag von Steven Samuels
Gesendet: Sonntag, 24. Juli 2011 00:19
An: [email protected]
Betreff: Re: st: Competing Hazards with Multiple-Record-per-Subject Data (2)
Break the ties: add a small constant (e.g. 0.001) to the times for one of
the failure types.
Also -stcrreg- (which fits the Fine-Gray model) does not require the
augmentation trick, I believe.
-Steve
[email protected]
Dear statalisters,
in earlier posts implementation of data replication in Lunn/McNeil 1995
Applying Cox Regression to Competing Risks were discussed. Under the subject
line above, Alex Gelber (see below) asked how to implement this particularly
when using multiple-record-per-subject data – unfortunately with no reply.
I stsplitted my data by year to model a linear time trend and age covariates
and thus, I am facing the same problem: stsetting does not work when the
data is replicated along Lunn/McNeil 1995 since at every instant two records
are attached to the id variable. The question is how to solve this stset
problem?
Given that 5 years have passed by since Alex posted, please allow myself to
basically repost this question – maybe even Alex has the answer and is still
subscribed.
Many thanks for your consideration
Stefan
--
Stefan Göke
Ph.D. Candidate
Department of Management
University Paderborn
____________________________
In earlier Statalist posts, May Boggess of Statacorp explained how to
estimate a competing hazard model using Lunn and McNeil's methods A and B.
There seems to be a problem, however, with implementing Lunn and McNeil's
Method B when using multiple-record-per-subject data.
The problem is that to implement Lunn and McNeil Method B, you need to
duplicate each subject's record twice (in the case of two competing
hazards). In the case of multiple-record-per-subject data, when using the
stset command to stset the data, you need to specify the ID variable using
id(idname), where "idname" is the name of the ID variable. But then Stata
gives you an error message, because each ID has two records attached to it
at every instant, and the stset command with multiple-record-per-subject
data only works correctly when there is only one record for each id-time
combination.
For example, in my case, the time variable is "month," the failure variable
is "status," and the id variable is "id," and here is what happens when I
try to stset my data:
. stset month, failure(status) id(id)
id: id
failure event: status != 0 & status < .
obs. time interval: (month[_n-1], month] exit on or before: failure
----------------------------------------------------------------------------
--
2000 total obs.
2000 multiple records at same instant PROBABLE ERROR
(month[_n-1]==month)
----------------------------------------------------------------------------
--
0 obs. remaining, representing
0 subjects
0 failures in single failure-per-subject data
0 total analysis time at risk, at risk from t = 0 earliest observed entry t
= .
last observed exit t = .
Does anyone know how to address this problem? Any help would be much
appreciated.
Alex
--
Alexander Gelber
Ph.D. Candidate
Harvard University Department of Economics Littauer Center Cambridge, MA
02138
*
* 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/
*
* 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/
*
* 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/