0
$\begingroup$

I performed univariate analysis to identify predictors of survival in a cohort. In a second step I performed multiple cox regression of the factors significant in univariate analysis. One variable is categorical with 4 categories. I am trying to decide whether I should consider this variable as significant in cox regression or not. So kind of a global p-value for this variable is needed. SPSS shows four p-values: one for the reference value (marked black) and one for each of the other values (vs. the reference value).

Is this marked p value kind of a global p-value for this variable? Or do I have to perform some other test ?

The variables significant on cox-regression will be incorporated in a predictive score, so a decision needs to be made whether the "whole" variable is considered significant.

x

$\endgroup$

1 Answer 1

1
$\begingroup$

The key here is noting the number of degrees of freedom for the Wald test that provided the p-value highlighted in black. Your 4-level categorical predictor has 3 degrees of freedom, so that presumably represents the multi-parameter Wald test that evaluates together all levels of the predictor. The p-values, regression coefficients (B) and hazard ratios (Exp(B)) are for each of the other levels against the reference level.

I would caution you, however, with respect to your approach:

I performed univariate analysis to identify predictors of survival in a cohort. In a second step I performed multiple cox regression of the factors significant in univariate analysis.

That can be a misleading way to set up a Cox model. Any type of automated predictor selection is generally not a good idea. Furthermore, a Cox model can behave similarly to logistic regression with omitted-variable bias, so that omitting any predictor associated with outcome can bias downward the magnitudes of the predictors retained in the model. You might thus have missed a predictor that might be important when other predictors are taken into account.

There are much better ways to set up such a model. You generally want to to use your knowledge of the subject matter to select or potentially combine predictors rather than use single-predictor p-values to drive the process. If your interest is in prediction, you generally want to have as complete a model as you can without overfitting, or use penalization methods to include more predictors while minimizing overfitting. See for example Frank Harrell's course notes on regression modeling strategies.

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.