Re: multivariate Cox model p values

arrayprofile@xxxxxxxxx wrote:
sorry, what I meant was: I can imagine that a variable may appear
significant in unvariable analysis and the significance disappeared in
multivariable analysis because the effect can be taken away by other
correlated variables

In addition to that, estimates of hazard ratios are biased towards 1.0 if important prognostic variables are omitted. Adding other variables can make the whole model fit better, and in models with no sigma parameter this translates to less bias in regression coefficients.

Frank Harrell

arrayprofile@xxxxxxxxx wrote:

Hi, I encountered a situation when I ran Cox proportional regression
with 6 independent variables. If I ran the regression with each
independent variable separately (i.e. univariable analysis), only 1
variable is significant. when I ran the Cox regression with all
variables together in one model (i.e. multivariable analysis), not only
the variable significant in the univariable analysis is still
significant, but 2 of the other variables showed significant as well.
How to explain this?

I can imagine that a variable may appear insignificant in unvariable
analysis and disappeared in multivariable analysis because the effect
can be taken away by other correlated variables, but can't think of a
reason why the opposite can happen. Can anybody explain?

Thanks

John


.