Set Cox hazard ratio to 1 for stratified variable

[Edsel Ing]

Two Questions:

1) Is there a way of setting the hazard ratio to 1 for the normal range of a stratified variable?
I saw the offset command but it seems to be for the coefficient, not the hazard ratio.

2) To avoid collinearity in the analysis where one variable is mathematically derived from another, do I just run stcox separately for each variable, but with the same set of covariates?

Thank you.
Edsel Ing

=================
Details of problem

I am trying to calculate hazard ratios for stroke according to different levels of blood pressure.
I will have four blood pressure categories: low, normal, high and very high
and four blood pressure indices systolic, diastolic, pulse pressure and mean blood pressure

I would like to set the normal blood pressure risk e.g. systolic 115 -125 mmHg at hazard ratio of 1.0

My covariates include age, gender, cholesterol, smoking, bmi etc.

[Carlo Lazzaro]

As far as question 1) is concerned, I suspect you can't go that way, as the remaining covariates obviously play a role in determining the hazard ratio for stroke. However, you have different options to create a reference group/category in your independent variables. I would recommend you to take a compehensive tour through the entries of Stata 13.1 .pdf manual -help fvvarlist- points you out to.
I am not absolutely clear with your question # 2): what is mathematically derived from another? The dependent variable or one or more independent variables?

Kind regards,
Carlo

[David Fisher]

Do you mean that you wish to estimate the hazard of the other blood pressure categories relative to the hazard of the "normal" category (adjusted for other covariates)? If so, you need to look at setting the base level of your category variable. See "help help fvvarlist##bases". The hazard ratio within the base level will then, by definition, be 1.

For your second question: I'm also not entirely clear, but I guess it depends whether you wish your final model to only contain one or the other. If so, your suggested approach sounds reasonable.

[Carlo Lazzaro]

Expanding on my previous reply, I would add that your reference group also depends on the number of indicators variables included in the right-hand side of the Cox regression, as from the following oversimplified Stata example::

use http://www.stata-press.com/data/r13/hip3,clear
stcox protect i.male, nohr ///data are already in the -stset- form
di exp((-2.028436*0) + (-1.377045 *0))///the reference group: female///with
no protection
di exp((-2.028436*0) + (-1.377045 *1))
di exp((-2.028436*1) + (-1.377045 *0))
di exp((-2.028436*1) + (-1.377045 *1))
predict HR, hr
tab HR///just to compare the results obtained manually

Back to your query, set aside continuous predictors, the reference group also depends on your decision about categorizing (say) diastolic blood presure and not sistolic only (provided that this approach is reasonable under a clinical framework).

As a side light, in more general terms, whenever it comes to categorizing continuous predictors, my warning chime rings and points me to: Royston P, Altman DG, Sauerbrei W. Dichotomizing continuous predictors in multiple regression: a bad idea. Statistics in Medicine 2006; 25: 127-141.

Kind regards,
Carlo