Re: st: comparing 25th percentile survival time between two race groups

[Steve Samuels <sjsamuels@gmail.com>]

I should clarify.  When survival in one group is "better" that might not mean at all percentiles.  Think of two survival curves are similar up to 80% survival (i.e. 20-th percentile for failure) and then one curve is higher than the other thereafter. 

Steve

About -ranksum- :

1. It is not for use with censored survival data, so you should not have
used it. I apologize for not catching this mistake earlier.

2. The corresponding test for censored survival data is: -sts test- with
a "wilcoxon option". (The full name for the original rank sum test is
"Wilcoxon two-sample rank sum test")

3. I previously said that a rank-sum test "tests for any difference
where survival in one group is better, not just differences in medians."

That means that if one distribution is shifted to the right relative to
another, all percentiles, including 10th, 25th 40th, etc. will also be
shifted. There is no specific test for a 25th percentile. You will see
that the formulas for the Wilcoxon test for censored or
uncensored data do not single out a specific percentile.

4. You appear unfamiliar with elementary survival concepts. Before
going any further, I suggest that you consult the Cleves 
book that I referenced earlier or another introductory text.


Steve



On Sep 28, 2012, at 7:04 PM, Haena Lee wrote:

Thank you Steve, That's really good point to look at the graph
especially for the failures.

According to ranksum, I was not able to find an option for 25th percentiles.
When I was comparing the differences of median survival time between
African American and White across 10 different regions in US, I used
this following command;

. sort UNOS
. by UNOS: ranksum d_enrollment, by(race_all)

In this case, I didn't need to let Stata know it is median survival
time because I guess median was default. However my question is how I
can particularly tell Stata that it is to test 25th percentiles this
time. Any help? Thanks much.

On Fri, Sep 28, 2012 at 4:26 PM, Steve Samuels <sjsamuels@gmail.com> wrote:
> Dear Haena:
>> 
>> On Sep 28, 2012, at 2:43 PM, Haena Lee wrote:
>> 
>> Hi Listers,
>> I have tried to produce the median survival time and its 95% CI with
>> renal transplant data. "Stci" command, however, did not produce any
>> median values. So I generated the graph of median survival time by
>> race (African American vs. White) and the graph showed it only reached
>> to the length of 25% in the course of follow-up, indicating both of
>> groups have died not even near 50% (median), but mostly near the
>> length of 25% tile in the course of follow-up.
>> 
> What you observe can be better phrased as saying that followup was not
> long enough to observe 50% of failures.
> 
>> 1. Given this, I attempted to produce 25th percentile survival time.
>> However I am struggling with comparing the difference of 25th
>> percentile survival time between African American and White across 10
>> regions in US. The command that I used:
>> sort UNOS
>> by UNOS: stci , by(race_all) p(25)
> 
>> 
>> I was trying to use "ranksum" but it doesn't have an option for
>> testing differences of other percentiles besides median. Which command
>> should I use?
> 
> That's a correct command. It tests for any difference where survival
> in one group is better, not just differences in medians.
>> 
>> 2. In comparing the median to the 25 percentile, there are some
>> regions where the length to 25% tile is significantly longer than
>> median. In this case, what would you guys do? If you would choose to
>> use median survival time, then what else could I do in order not only
>> to produce the median survival time and 95% CI, but to test the
>> difference of median survival time between groups instead stci (since
>> it didn't produce in previous analysis)?
> 
> Your observation is impossible in theory; to persuade us, you'd need to
> follow the advice in Statalist FAQ (3.3) and show us the commands and
> results that lead you to this observation.
> 
> If I had to guess, I'd say that you looked at the Kaplan-Meier survival
> curve, not at -stci- results. A survival curve shows the proportion of
> people who haven't had the event at a time, not the proportion who have
> failed. Thus the point corresponding to the 25% mark on the y axis is where
> 25% of people haven't failed and 75% have. In other words it is the 75th
> percentile for failure. To see the failure time percentiles, the command
> is:
> 
> *******************
> sts graph, failure
> *******************
> 
> You might benefit from reading a good text on survival data, e.g. Cleves
> et al. (2010).
> 
> Reference: Cleves, Mario, William Gould, Roberto Gutierrez, and Yulia
> Marchenko. 2010. An Introduction to Survival Analysis Using Stata, Third
> Edition. College Station, Tex: Stata Press.
> 
> Steve
> 
>> Thank you so much!
> *
> *   For searches and help try:
> *   http://www.stata.com/help.cgi?search
> *   http://www.stata.com/support/faqs/resources/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/faqs/resources/statalist-faq/
> *   http://www.ats.ucla.edu/stat/stata/



-- 
=====================
Haena Lee
Ph.D Student
Sociology Department
The University of Chicago
312 - 405 - 3223
*
*   For searches and help try:
*   http://www.stata.com/help.cgi?search
*   http://www.stata.com/support/faqs/resources/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/faqs/resources/statalist-faq/
*   http://www.ats.ucla.edu/stat/stata/