Analyse a variable as continuous and as categorical

noob

New Member
#1
Hello everybody

I performed a univariate analysis of a set of predictors. My plan was to include the significant predictors in a multivariate cox regression

There are good reasons to perform the analysis of these predictors as continuous as well as categorical variables (in 3 or 2 categories). . Now, you lose information when using a variable as categorical instead of continuous, so I would expect if a variable is not significant when analyzed as continuous it would most likely not be significant as categorical.

I'm planning 2 seperate multivariate analyses to build 2 models. One with the categorical variabels and one with the continuous predictors. However it's important to first exclude the same predictors after univariate analyses so I ll get the same dataset (after deleting the rows with missing values) for both analyses.

My results of the univariate analysis:
Well I have 1 significant predictor when analyzed continuously but not categorically. However I also have 3 predictors not significant continuously but significant categorically. Which predictors to include in my multivariate cox regression?
I did not find any information about how to deal with this situation?

Any experience with this?
 
Last edited:

hlsmith

Omega Contributor
#2
I big part of this may be where you are setting the cutoffs. Some groups may have a different effect or small/large samples. It pretty much comes down to you just justifying your decisions and how you may use these data/results.
 
#3
There is no reason why this shouldn't happen: Your two analyses ask different questions, so they get different answers.

Which analysis should you run? That depends on what your question is.

But 1) Bivariate screening of variables prior to entering them in a model is a bad idea. 2) Categorizing continuous variables is nearly always a bad idea as well.
 

Karabiner

TS Contributor
#4
My plan was to include the significant predictors in a multivariate cox regression
As said before, generally this is not a good idea.
You might get biased and nongeneralizable results
from your final analysis.
Your Cox regression doesn't know that it deals
with a thoroughly pre-tested and pre-selected
set of variables.
How big is your sample size, by the way?
There are good reasons to perform the analysis of these predictors as as continuous as well categorical variables (in 3 or 2 categories).
Could you describe why there are good reasons to do
so either way? This sounds very uncommon.

With kind regards

K.