How to get the hazard ratio for a categorical variable according to a determined covariate

[delao victor]

Hi everyone,

I'm doing a Cox regression with a categorical variable "a" (5 categories) and I'm adjusting it for different covariates "b1, b2, b ..." continuous and categorical. In this case, I want to know the Hazard Ratio (HR) for each category (1 - 5) according to the variable "BMI" (dichotomous variable "<25" and "> = 25") and all the adjustment covariates that add to the model .

With this, I would like to obtain the HR for each category (1 - 5) of my variable "a" if the covariate BMI is "<25", and the HR for each category (1 - 5) of my variable "a" if the covariate BMI is "> = 25" adjusted for all the covariates that you have put in the model. Do you know the syntax?

I want something like this (but with HR):





Thanks a million!!

[Clyde Schechter]

The easiest way to do this is if you first run your Cox regression using factor variable notation. Then -margins- will easily give what you want.

gen byte oo = (bmi >= 25) if !missing(bmi)
label define oo 0 "Normal Weight" 1 "Overweight/Obese"
label values oo oo

stcox i.a##i.oo other_covariates_here
margins a#oo

[delao victor]

Thanks so much Clyde! The last question: after applying the command, the results are these:

So, if I'm interpreting correctly, the HR's for "bmi_r" for the cat ">=25" are the number in red. Is this correct?
And then I want to calculate the P for trend and the P-interaction but I get a value for each category (number) and I want a single value, so I used this syntax to give me just one value for each "P":

P for trend:
.testparm i.bmi_r#"continuous variable p for trend"

P-interact:
.levelsof i."a" if e(sample), local(levels)
.testparm i.bmi_r#i(`levels').i"a"

After running the syntax, both values of P are the same. I think I'm doing something wrong.
A lot of thanks!

[Clyde Schechter]

No. I don't think you did what I suggested. What you show in #3 is the stcox output. You won't find the hazard ratios you are looking for there. The hazard ratios you want will be in the output of the -margins- command.

To get a single interaction p-value, you would run -testparm i.variable_a#i.bmi_r- (but replace variable_a by the actual name of that variable.

It does not make sense to ask for a p for trend here because you have two different series of hazard ratios, one for bmi < 25 and the other for bmi >= 25. Trend p-values only make sense for a single ordered series.

[delao victor]

Okey! Thank you Clyde!