Absolute risk from a Cox-regression model

data analysis
,
1

I am currently working on a survival-analysis project. My clinical supervisor asked me to compute “absolute risks at year 2”. How can I do this? I thought that Hazard ratios are the way to go to describe a Cox-model, but he keeps asking for an “absloute risk” in categories.

As an idea although it is a totally different field, he showed me the Table3 of this paper:
https://doi.org/10.1093/eurheartj/ehae557
where absolute risks are defined for different subgroups.
I do not really understand, how these are computed. As well as I dont found informations how they computed the Hazard ratios for these “subgroups”. Did the splitted the data and recomputed the model? Or do they just computed contrasts with mean/median of the subgroup value?

2

Essentially, you use the Cox model to calculate estimated survival at a certain time threshold (in this case year 2). The survest function of the rms package can do this.

Then, “absolute risk at year 2” is just 1 - estimated survival at year 2.

1 Like
3

Thanks, that sounds good. The later the timepoint the higher the absolute risk, but yeah.

1 Like
4

Yes and these are estimates of cumulative incidence of an outcome. See this and this.

5

Can survest be used on the output from fit.mult.impute and get correct estimate acounting for missing values from aregImpute?
Asks Fredrik L

6

When the model is parametric, yes. For a semiparametric model where different bootstrap samples have different distinct failure times, I’m uncertain.

7

The model I have in mind is a cph and time of death and censoring have no missing data. Missing data are only among the independent variables. If so, is survest correct applied after fit.mult.impute and aregImpute?

8

If you don’t get an error message it’s probably correct. I would need to check the code in detail to be sure. Even though Y is complete in your case, certain bootstrap samples will exclude some distinct failure times, creating a problem. Ultimately we need to use full Bayesian modeling which does not use Y to impute X.