Margins after stcox with negative lower 95% CI - is there a fix?

[Martin Mueller]

Dear all

I have the following dataset with three binary variables, i.e. outcome (death after six months), older_65 (age above 65) and sex, as well as two continous variable, follow_up_days and age_years:

* Example generated by -dataex-. For more info, type help dataex
clear
input float(age_years follow_up_days) byte(outcome older_65 sex_female1)
72 358 0 1 0
72 363 0 1 0
76 357 0 1 0
87 209 1 1 0
66 363 0 1 0
80 359 0 1 0
61 360 0 0 0
73 358 0 1 1
81 359 1 1 0
end
label values outcome yes_no_lab
label values sex_female1 sex_lab
label values older_65 yes_no_lab
label def yes_no_lab 0 "No", modify
label def yes_no_lab 1 "Yes", modify
label def sex_lab 0 "Male", modify
label def sex_lab 1 "Female", modify

I declared the data to be survival-time data and ran a cox regression with an interaction between older age (older 65 years) and sex and age as a covariate.
I wanted to obtain the predicted hazard ratios according to the different strata adjusted for age, thus I used the - margins - respectively - marginsplot - command.

stset follow, failure(outcome)
stcox i.older_65##i.sex_female1 age_years
margins i.older_65#i.sex_female1
marginsplot

In the plot shown and the margins output (not shown) the 95% CI for the predicted Hazard ratios seem to be wrong as they are negative.



Steve Samuels wrote here on a similar topic https://www.statalist.org/forums/for...ns-after-stcox an explanation why this happens, but unfortunately I could not find a fix for it:

Margins uses the delta method to compute standard errors, then forms CIs by adding ± a multiple of the standard error. As a general purpose command, margins doesn't "know" that hazard ratios are supposed to be positive, so the CI extends below zero.

Is there a fix for it? How can I get the correct 95% CI?

Thank you in advance for your advise!

Best wishes
Martin

[Andrea Discacciati]

https://www.statalist.org/forums/for...ns-after-stcox

Useful? But you can't use marginsplot - you'll need to write your code to draw a plot.

[Martin Mueller]

Dear Andrea

Sorry for overseen your post. This was of great help!

Thank you.

Best wishes
Martin