Survival Analysis: sts test, stcox which output should be used

[aozkan]

Dear listers,

I am working on a data set that is about renal cell carcinoma survival.

I have some categorical variables like tumor histology, tumor grade, T stage etc.

For tumor histology I have 3 different p values. sts test Histology results p<0.05 , stcox Histology results p>0.05 and stcox i.Histology p>0.05.


How can I properly read these results. Below you can find detailed stata output. Also I put a pdf file to show my results in a better shape.

Appreciate your help.
Thank you

Alp


. sts test Histology

failure _d: CSS == 1
analysis time _t: followuptimemonths


Log-rank test for equality of survivor functions

| Events Events
Histology | observed expected
-----------------------+-------------------------
Clear Cell | 43 40.44
Papillary | 6 6.47
Chromophobe | 2 4.99
Collecting duct | 1 0.09
Unclassified | 0 0.01
------------------------+-------------------------
Total | 52 52.00

chi2(4) = 10.92
Pr>chi2 = 0.0275



. stcox Histology

failure _d: CSS == 1
analysis time _t: followuptimemonths



Cox regression -- Breslow method for ties

No. of subjects = 1233 Number of obs = 1233
No. of failures = 52
Time at risk = 42660.6
LR chi2(1) = 0.70
Log likelihood = -325.80498 Prob > chi2 = 0.4034

------------------------------------------------------------------------------
_t | Haz. Ratio Std. Err. z P>|z| [95% Conf. Interval]
-------------+----------------------------------------------------------------
Histology | .8248506 .1981739 -0.80 0.423 .5150751 1.320931
------------------------------------------------------------------------------




. stcox i.Histology

failure _d: CSS == 1
analysis time _t: followuptimemonths


Cox regression -- Breslow method for ties

No. of subjects = 1233 Number of obs = 1233
No. of failures = 52
Time at risk = 42660.6
LR chi2(4) = 5.51
Log likelihood = -323.39883 Prob > chi2 = 0.2388

----------------------------------------------------------------------------------
_t | Haz. Ratio Std. Err. z P>|z| [95% Conf. Interval]
-----------------+----------------------------------------------------------------------
Histology |
Papillary | .8724689 .3802907 -0.31 0.754 .3713037 2.050079
Chromophobe | .3766941 .2725427 -1.35 0.177 .0912293 1.555404
Collecting duct | 10.38974 10.63399 2.29 0.022 1.397631 77.2355
Unclassified | 1.83e-13 3.49e-06 -0.00 1.000 0 .
----------------------------------------------------------------------------------

[Svend Juul]

In the following I assume that the histology categories are coded 1 to 5. (I prefer using the numlabel command to make the codes behind value labels visible in tables).

Your second suggestion is definitely invalid; it tests whether there is a linear trend, using the histology codes, and that is obviously nonsense.

Your first and third suggestion both make sense, but due to two very thin strata, there is a lot of noise in the results, and the P-values are quite different. The noise is illustrated by the very wide confidence intervals in your third suggestion.

With the sparse data, you ask for too much. You could compare clear-cell-histology vs. the rest, and if you do that I guess that your method 1 and 3 yield P-values that are not very different. Whether it is interesting, is your business.

Hope this helps,
Svend

[aozkan]

Svend,

Your suggestions make it clear for me. I was annoyed of that huge CI's grouping them in two categories helped me. p values are now closer and CI is not that wide. I am dealing with not just histology I have 13 variables need to handled. I wish I had a better data set but this is what I have

Thank you for help.