[讨论]Cox Regression: Revisited

2618

收藏 2006-05-17

I have been thinking about our conversation about testing the proportional hazards assumption in a Cox regression. As we discussed, if we have *one* factor in our model with values x=1 and x=0 then entering this factor as a 'stratum' in a Cox regression produces two log(-log S(t)) curves on the same plot, one for each level of x. [The values of S(t) which SPSS uses are the Kaplan Meier estimates when the data is divided into 2 groups, one group for x=1 and one group for x=0].
I have a question, if we had, say, *two* binary variables in the model with levels x=1, x=0, y=1, y=0 in order to assess the proportional hazards assumption for each variable we would
need 4 divisions of the data corresponding to:
x=0, y=0
x=0, y=1
x=1, y=0
x=1, y=1

Do you know of a way in the Cox regression procedure whereby I can derive the four associated log(-log S(t)) curves on the same plot? If not, it will just be a case of deriving Kaplan Meier estimates (S(t)'s) for these 4 groups and then doing subsequent calculations by hand to derive the associated 4 log(-log(S(t)) plots.

扫码加我拉你入群

请注明：姓名-公司-职位

以便审核进群资格，未注明则拒绝

全部回复

SPSSCHEN

2006-5-17 00:34:00

I think the proportional hazards assumption is tested for each categorical variable separately: you should check that variable x follows the assumption (adjusting for the rest of the variables in the final model), then you check variable y (adjusting again for the rest of the model)...

For quantitative predictors, you can use a time dependent covariable to check the asumption, and, in this case, more than one variable can be tested at the same time:

TIME PROGRAM.
COMPUTE T_COV_ = T_*x .
COMPUTE T_COV2_= T_*y
COXREG
time /STATUS=status(1)
/METHOD=BSTEP(LR) T_COV_ T_COV2_ x y
/PRINT=CI(95)
/CRITERIA=PIN(1) POUT(1) ITERATE(20) .

HTH,
Marta

扫码加我拉你入群

请注明：姓名-公司-职位

以便审核进群资格，未注明则拒绝

SPSSCHEN

2006-5-17 00:37:00

Marta,

Sharing is always a good idea, in my view. Private exchanges are better for exchanging attachments, which is not good in mailing lists, but as you say questions and answers are likely to be useful also for others.

Regarding strata for two or more factors in Cox Regression, Kay may create a
new variable combining two dichotomies, as in:

COMPUTE X3=10*x1+x2.

This variable would have four values, 0, 1, 10, and 11, and would produce four strata if used in CoxReg. What is tested is the effect of X1, X2 and their interaction, all in one.

Another (and in my opinion inferior) possibility is temporarily selecting cases where, say, X1=0, and use x2 for stratum, and then doing the same with cases where X1=1. Underline TEMPORARILY, of course, and use the TEMPORARY command, to avoid losing data:

TEMPORARY.

SELECT IF X1 =0.

COXREG ....... /STRATUM X2.

TEMPORARY.

SELECT IF X1 =1.

COXREG ....... /STRATUM X2.

However I much prefer the first approach with X3, which also covers the possible interaction of X1 and X2 (if there is no interaction, the relative effect of X2 should be the same in the two X1 strata, so interaction is somehow perceived even in this SELECT IF approach).

Hector

扫码加我拉你入群

请注明：姓名-公司-职位

以便审核进群资格，未注明则拒绝

yiyucathy

2008-3-26 22:30:00

全英文啊？？

厉害

扫码加我拉你入群

请注明：姓名-公司-职位

以便审核进群资格，未注明则拒绝

扫码加我 拉你入群

扫码加我 拉你入群

扫码加我 拉你入群

扫码加我 拉你入群

分享

扫码加好友，拉您进群

扫码加我拉你入群

扫码加我拉你入群

扫码加我拉你入群

扫码加我拉你入群