stcox and stcrreg

[Jiong Yao]

Hi guys:
I have a question regarding stcox and stcrreg. i try to find how does the independent variable, bullying strategy (dummy), influence different kinds of ways border claims ended. Therefore, the competing risk model is an ideal choice. also, i am interesting in finding the probabilities of certain outcomes within some periods, which leads to the choice of stcrreg, as i want to generate cif for individual outcomes. I found, however, using stcox and stcrreg produces different results. Below is the example to illustrate this point.
there are in total of 9 kinds of outcomes:

  resolved2 |      Freq.     Percent        Cum.
------------+-----------------------------------
          0 |     20,685       99.66       99.66
          1 |         10        0.05       99.71
          4 |         17        0.08       99.79
          5 |          2        0.01       99.80
          6 |          2        0.01       99.81
          7 |         14        0.07       99.87
         11 |          1        0.00       99.88
         12 |          5        0.02       99.90
         13 |         11        0.05       99.96
         14 |          9        0.04      100.00
------------+-----------------------------------
      Total |     20,756      100.00

i set resolved2 == 7 as the event of interest

stset claimserialend, id(claimdy) fail(resolved2==7) origin(time claimserialstart)  enter(enterdate)
                     scale(30)

then i first used stcox

stcox i.bullying i.bdout_3 viol icowsal cumu10mid avgterg avgterv i.special, nohr nolog efron robust

and get these results:

 failure _d:  resolved2 == 7
   analysis time _t:  (claimserialend-origin)/30
             origin:  time claimserialstart
  enter on or after:  time enterdate
                 id:  claimdy

Cox regression -- Efron method for ties

No. of subjects      =           87             Number of obs    =      20,756
No. of failures      =           14
Time at risk         =  20705.33333
                                                Wald chi2(8)     =       33.02
Log pseudolikelihood =   -30.568174             Prob > chi2      =      0.0001

                               (Std. Err. adjusted for 87 clusters in claimdy)
------------------------------------------------------------------------------
             |               Robust
          _t |      Coef.   Std. Err.      z    P>|z|     [95% Conf. Interval]
-------------+----------------------------------------------------------------
  1.bullying |   1.554178   .6693342     2.32   0.020     .2423068    2.866049
   1.bdout_3 |   1.212035   .7166583     1.69   0.091    -.1925897    2.616659
        viol |   .6764656   .3219261     2.10   0.036     .0455019    1.307429
     icowsal |   .2297885   .1818425     1.26   0.206    -.1266163    .5861933
   cumu10mid |  -.5565466   .1833163    -3.04   0.002    -.9158399   -.1972533
     avgterg |  -1.671215   .7281596    -2.30   0.022    -3.098381   -.2440479
     avgterv |   2.208843   1.918719     1.15   0.250    -1.551777    5.969463
   1.special |   .2368191   .7650886     0.31   0.757    -1.262727    1.736365
------------------------------------------------------------------------------

the first independent variable, bullying strategy dummy, is significant.
when i tried stcrreg, however, things are different:

 stcrreg i.bullying i.bdout_2 i.bdout_3 viol icowsal cumu10mid avgterg avgterv i.special,
              compete(resolved2 == 1 4 5 6 11 12 13 14) nohr nolog robust

         failure _d:  resolved2 == 7
   analysis time _t:  (claimserialend-origin)/30
             origin:  time claimserialstart
  enter on or after:  time enterdate
                 id:  claimdy

Competing-risks regression                       No. of obs       =     20,756
                                                 No. of subjects  =         87
Failure event   : resolved2 == 7                 No. failed       =         14
Competing events: (1)                            No. competing    =         57
                                                 No. censored     =         16

                                                 Wald chi2(9)     =    1620.55
Log pseudolikelihood = -45.009945                Prob > chi2      =     0.0000

                               (Std. Err. adjusted for 87 clusters in claimdy)
------------------------------------------------------------------------------
             |               Robust
          _t |      Coef.   Std. Err.      z    P>|z|     [95% Conf. Interval]
-------------+----------------------------------------------------------------
  1.bullying |   1.167243   1.018178     1.15   0.252    -.8283493    3.162836
   1.bdout_2 |   -16.4309   1.002408   -16.39   0.000    -18.39559   -14.46622
   1.bdout_3 |   .3128215   .7729912     0.40   0.686    -1.202213    1.827856
        viol |   .5064316   .3091432     1.64   0.101    -.0994779    1.112341
     icowsal |   .1148364   .1666615     0.69   0.491     -.211814    .4414869
   cumu10mid |  -.4802709   .2456184    -1.96   0.051    -.9616742    .0011324
     avgterg |  -1.438623   .7947055    -1.81   0.070    -2.996217    .1189714
     avgterv |   3.667996   1.440916     2.55   0.011     .8438523    6.492141
   1.special |   .6250076   .8436583     0.74   0.459    -1.028532    2.278548
------------------------------------------------------------------------------
(1) resolved2 == 1 4 5 6 11 12 13 14

so in the second model, the bullying strategy dummy becomes insignificant. How to explain this difference?
but based on the second model, i can still generate cif for bullying strategy dummy variable:

Thanks in advance!
Best
Jiong

[Clyde Schechter]

Apart from the general principle that the difference between statistically significant and not statistically significant is, itself, not statistically significant, you are trying to compare apples and oranges here. This is one of those situations where two things get the same label in different models but refer to different things.

The hazard ratio calculated in the Cox model is estimated on the assumption that all non-"death" (i.e. other than resolved2 = 7) terminations are censored, and that censoring is independent of survival (time to event) outcome. But in the competing risks model, the assumption is that the other terminations specified in the compete() option are not independent of resolved2 = 7. So this is a different event and there is no reason to expect that the hazard ratio associated with any predictor will be the same as in a Cox model.

[Jiong Yao]

Apart from the general principle that the difference between statistically significant and not statistically significant is, itself, not statistically significant, you are trying to compare apples and oranges here. This is one of those situations where two things get the same label in different models but refer to different things.

The hazard ratio calculated in the Cox model is estimated on the assumption that all non-"death" (i.e. other than resolved2 = 7) terminations are censored, and that censoring is independent of survival (time to event) outcome. But in the competing risks model, the assumption is that the other terminations specified in the compete() option are not independent of resolved2 = 7. So this is a different event and there is no reason to expect that the hazard ratio associated with any predictor will be the same as in a Cox model.

Hi Clyde. thank you very much for the prompt answer. i get your point and thanks for the clarification!

[Elena Nardi]

Hi, I'm new in this community.
My problem is related with stcurve command after having estimated a Cox model.
I have plotted the Survival funtcion after a Cox model, now I'm wondering if it is possible to show the number at risk table below a the graph as for the sts graph command.
I'm also interested in showing in the plot the value of the failure function at a specif point in time, is it possible?
Thanks in advance!
Elena