Why do I get different sign of the coefficients of covariates when I run the semi-parametric proportional hazard model (coxph) compared to the parametric proportional hazard model (survreg)? Anyone with experience in extracting information form survreg to make predictions are free to contact me. Cheers, Ullrika [[alternative HTML version deleted]]
Ullrika Sahlin <ullrika.sahlin <at> ekol.lu.se> writes:> > Why do I get different sign of the coefficients of covariates when I run the > semi-parametric proportional hazard model (coxph) compared to the parametric > proportional hazard model (survreg)? > > Anyone with experience in extracting information from survreg to make > predictions are free to contact me. > > Cheers, > > Ullrika >coxph fits a proportional hazards model. survreg fits an accelerated failure time model. These are parametrized differently. In the first, a higher linear predictor means greater hazard (i.e., shorter lifetime). In the second, a higher linear predictor means greater expected (log) lifetime. Consult an introductory survival text such as Klein and Moeschberger for more on these types of model. The Design library may help you with predictions. Chris
> Why do I get different sign of the coefficients of covariates when I run the > semi-parametric proportional hazard model (coxph) compared to the parametric > proportional hazard model (survreg)?Survreg models the time till death, a postive coefficient --> a longer time. Coxph models the death rate, a positive coefficient --> higher death rate. So in the first a positive is good, in the second it is bad.> Anyone with experience in extracting information form survreg to make > predictions are free to contact me.Commonly one would use predict(fit) to get predictions. There are several options. Terry Therneau