Hello,
when we are exploring certain predictors for bad outcome (mortality), we usually run univariate, followed by multivariate models (including all univariate analysis with p < 0.20) with backward elimination, where we report outcomes of this final model with specific OR including CI and p values.
Now recently , when I explored proportional hazards models, as we do have for all patients time to event (death) or discharge home - I was suprised to see different outcomes than in my multivariate models, where I used the same variables.
Is it generally better if you have a time to event , exploring those variables with proportional hazards than with univariate / multivariate analysis ? Do I first run better a univariate analysis and based on those identified with p < 0.2 (?) , which I enter into my proportional hazard model ? Or do I enter all potential variables ? Do I have to explore and exclude colinearity in proportional hazard first ?
What would be the advantage / disadvantage for either multivariate or proportional hazard ?
Thanks a lot, Marc
I am not sure what you specifically mean by "we usually run univariate, followed by multivariate models (including all univariate analysis with p < 0.20) with backward elimination" but if you mean a linear model with regression, then there are likely some important differences for your data. The main differences are (1) censored life data and (2) non-normal distributions of life data. Because of these differences, the proportional hazards model or the parametric survival model tends to give better and more realistic estimates and tests.
I am not sure what you specifically mean by "we usually run univariate, followed by multivariate models (including all univariate analysis with p < 0.20) with backward elimination" but if you mean a linear model with regression, then there are likely some important differences for your data. The main differences are (1) censored life data and (2) non-normal distributions of life data. Because of these differences, the proportional hazards model or the parametric survival model tends to give better and more realistic estimates and tests.