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]
FW: st: RE: RE: RE: Competing risk calulations based on stratified Cox model
From
Christel Häggström <[email protected]>
To
"[email protected]" <[email protected]>
Subject
FW: st: RE: RE: RE: Competing risk calulations based on stratified Cox model
Date
Mon, 23 Sep 2013 11:22:26 +0200
Dear Statausers,
In my group we have now decided to analyse the full dataset taking competing risk into account and hence calculate cumulative incidence curves based on a stratified Cox model, which will look similar to this.
xi:stcox cause#(BMI_categories ), nolog strata(cause birth_cohort_categories sub_cohort_categories )
I have not found an option within stcompadj how I can handle the covariates needed to be stratified within the Cox model. Any suggestions? Or any alternative commands to use?
Thank you for the consideration.
Christel
-----Original Message-----
From: [email protected] [mailto:[email protected]] On Behalf Of Christel Häggström
Sent: den 19 september 2013 10:36
To: [email protected]
Subject: RE: st: RE: RE: RE: Competing risk calulations based on stratified Cox model
Hi Kieran,
Thanks for your advice. I have mostly relied on p-values, but it looks like most of the covariates are not proportional using graphs as well. Just did a second look using your suggested method and decided that smoking does not need to be stratified for.
I am currently looking at prostate cancer incidence, and it's have been an increase in detection during the last decades (due to PSA testing), and this looks like it have shifted how BMI is associated to the disease. This introduction of PSA is at different point in calendar times in different sub cohorts, which is the explanation to why there is a need to stratify to sub cohort and birth cohort categories. An "easy" workaround the problem would be to only use parts of the data or to divide the data in before and after this increase (after deciding a point in calendar time of diagnosis for this), but my first attempt was to solve this with the full dataset, and to calculate cumulative incidence of the disease over timescale (attained age). I am happy to hear other suggestions and comments to this.
Best regards,
Christel
-----Original Message-----
From: [email protected] [mailto:[email protected]] On Behalf Of Kieran McCaul
Sent: den 18 september 2013 22:28
To: [email protected]
Subject: RE: st: RE: RE: RE: Competing risk calulations based on stratified Cox model
...
I'm not aware of any literature specifically related to the PH test. I'm simply relying on the fact that a small p-value does not, in and of itself, indicate an important effect: you have to look at the size of the effect itself and within the context of the study determine if it is important or not.
When I was doing my PhD, I had a cohort of nearly 500,000 individuals with over 14,000 events observed over time. Nearly everything I looked at was "statistically significant". In Cox models when I tested the PH assumption, most variables produced small p-values, but I was able to dismiss all but one of these when I looked at the Scaled Schoenfeld residual plots. The one remaining variable was cancer stage, but I could explain how and why proportionality would be expected to vary over time for this variable..
This underlies an issue with testing the PH assumption in Cox models that I think is often ignored: lack of proportionality is an epidemiological feature that should be explainable.
-----Original Message-----
From: [email protected] [mailto:[email protected]] On Behalf Of Adam Olszewski
Sent: Thursday, 19 September 2013 3:49 AM
To: [email protected]
Subject: Re: st: RE: RE: RE: Competing risk calulations based on stratified Cox model
Hi,
I am actually quite interested in the subject. Can you indicate any literature that dealt with the subject of "significance" of PH violation in very large data sets and for alternative assessments of the assumption?
Best,
AO
On Wed, Sep 18, 2013 at 3:36 PM, Kieran McCaul <[email protected]> wrote:
> ...
>
> Hi Christel,
>
> Before considering how to get the cumulative incidence curves, can you tell me how you've determined that you have a proportionality problem?
>
> Are you relying principally on p-values when testing the proportional hazards assumption?
>
> If you have 300,000 subjects then presumably you have many events and hence sufficient power to declare small effects "statistically significant".
>
> In such a setting, it's quite possible to find covariates that fail the proportional hazards test "statistically", but for which the variation in proportionality is small and could be ignored. In essence, the proportionality test is looking at a regression line and testing the null that the slope is zero: with a large sample, a slope that is very near zero could be " statistically significant".
>
> In the output from the proportionality test you need to look at the rho values. The p-values are testing rho=0.
> Small rho values don't indicate a problem with proportionality, regardless of the size of the p-value.
>
> By small, I mean something like |rho| <0.02 or <0.03, depending on my mood. I tend to rely more on graphical assessments.
>
> Use the -estat phtest- options to produce plots of the scaled Schoenfeld residuals against time overlayed with a lowess smoother: if the smoother line looks flat, there's little evidence of a problem with proportionality.
>
>
> Kieran
>
>
>
> -----Original Message-----
> From: [email protected]
> [mailto:[email protected]] On Behalf Of Christel
> Häggström
> Sent: Wednesday, 18 September 2013 9:04 PM
> To: [email protected]
> Subject: st: RE: RE: Competing risk calulations based on stratified
> Cox model
>
> Dear Kieran,
>
> Thank you for the help and quick reply.
>
> I can fit a combined Cox model as in those pages you referred to, if I have understand the documentation right, in my data it looks something similar to this:
> xi:stcox cause#(BMI_categories ), nolog strata(cause
> birth_cohort_categories smoking_categories sub_cohort_categories )
>
> There are different effect of the covariates I need to stratify for both of the endpoints, and none of these covariates satisfy the proportional hazard assumption. The main cause are a disease among the elderly, competing cause are all-cause death and my dataset contain approx. 300 000 subjects.
>
> What I want to do is to calculate the cumulative incidence based on this Cox model, and plot that cumulative incidence over the timescale, but I don't know how I should write the stcompadj function to do this possible. Suggestions?
>
> Best regards,
> Christel
>
>
>
>
>
> -----Original Message-----
> From: [email protected]
> [mailto:[email protected]] On Behalf Of Kieran
> McCaul
> Sent: den 18 september 2013 10:37
> To: [email protected]
> Subject: st: RE: Competing risk calulations based on stratified Cox
> model
>
> ...
>
> -stcompadj- is a user-written program. There are ancillary files that can be downloaded as well and one of these is a PDF file that outlines the general approach used.
>
> If you look at Pages 6 and 7 of this PDF file, you can see how the data is set up for analysis using standard cox models.
> Why not try doing this and test proportionality in each of the models? Then try the combined model (bottom of page 7) with the stratum variable needed for the competing risk analysis plus the variables that you need to stratify on.
>
>
>
> -----Original Message-----
> From: [email protected]
> [mailto:[email protected]] On Behalf Of Christel
> Häggström
> Sent: Wednesday, 18 September 2013 4:08 PM
> To: [email protected]
> Subject: st: Competing risk calulations based on stratified Cox model
>
> Dear statausers,
>
> I am going to perform a competing risk analysis, and plan to use stcompadj for the ability to adjust for covariates. I have previously calculated hazard ratios with the same material with stcox and there are some covariates not satisfying the proportional hazard assumption which has been stratified for within the Cox model (strata option).
>
> I planned to use the same model for cumulative incidence calculations, but have not found an option within stcompadj how I can handle the covariates not satisfying proportional hazards assumption. Any suggestions? Or any alternative commands to use?
>
> Thank you for the consideration.
> Christel
>
>
>
> *
> * 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/
>
> *
> * 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/
*
* 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/
*
* 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/