Dear All, I am contemplating centering the covariates in my Cox model to reduce multicollinearity between the predictors and the interaction term and to render a more meaningful interpretation of the regression coefficient. Suppose I have two indicator variables, x1 and x2 which represent age categories (x1 is patients less than 16 while x2 is for patients older than 65). If I use the following Cox model, is there anyway I can centre the variables? Do I have to do it before I fit them into the model and if so, how? fit2=coxph(Surv(rem.Remtime,rem.Rcens)~x1(partial)+x2(partial),data=partial,method="breslow") Thank you, Laura
Centring variables in Cox Proportional Hazards Model
2 messages · Laura Bonnett, Frank E Harrell Jr
Laura Bonnett wrote:
Dear All, I am contemplating centering the covariates in my Cox model to reduce multicollinearity between the predictors and the interaction term and to render a more meaningful interpretation of the regression coefficient. Suppose I have two indicator variables, x1 and x2 which represent age categories (x1 is patients less than 16 while x2 is for patients older than 65). If I use the following Cox model, is there anyway I can centre the variables? Do I have to do it before I fit them into the model and if so, how? fit2=coxph(Surv(rem.Remtime,rem.Rcens)~x1(partial)+x2(partial),data=partial,method="breslow") Thank you, Laura
There is no need to center the variables. Also, you are going to find a huge lack of fit for the shape of the age effect you are using, as opposed to using a smooth function in continuous age. Your notation is treating x1 and x2 as functions of data frame which is strange. With data=partial you would ordinarily just have something like x1+x2 in the model. Frank
Frank E Harrell Jr Professor and Chair School of Medicine
Department of Biostatistics Vanderbilt University