Pooling of cox regression models going wrong

Question

I have saved the coefficients from a cox regression model (using the survival package) for 54 genetic variants. I have done this for 7 countries separately and now want to pool them together using the rma function in the metafor package. The cox regression went fine, and I am using the exp(coef) and the robust standard errors for the pooling. This is part of my script

library(metafor)   
HR.test <- c(1.0473445, 0.9463692, 0.8108665, 0.9981969, 0.9463692, 0.9858747, 0.9089416)
se.test <- c(0.06669897, 0.05569544, 0.1618536, 0.06813053, 0.05569544,
             0.05663756, 0.08702840)

pool.test <- rma(HR.test, sei=se.test, method="FE")
summary(pool.test)

The problem is, that the P values are impossibly small, especially considering that the confidence intervals go through 1, so the results are not significant. Is there something wrong with what I'm typing? I've used this package before to pool linear regression results and it worked fine. Is it different if you use hazard ratios (the values are the exponentials of the hazard ratios)?

If you want CI's you need to first calculate the log-est CI's and then transform. I doubt that exp(se)'s are particularly meaningful. The metafor RMA documentation says you are supposed to be working with the log-relative risk and either se's or variances, so I would think that also means you should NOT use exp(coef) but rather just coef and se. — IRTFM

Wolfgang Wolfgang · Accepted Answer · 2016-04-25T08:36:24

I doubt that the code as provided was actually used. First of all, beta.test is not defined. I assume that these are supposed to be the log hazard ratios extracted from the Cox regression models. Also, methods="FE" will give you an error. It should be method="FE".

Indeed, if you run rma(HR.test, sei=se.test, method="FE") then you get a very small p-value, but as @42 has already indicated in the comments, this is not correct. The se.test values are the estimated standard errors of the log hazard ratios. So, beta.test <- log(HR.test) and then rma(beta.test, sei=se.test, method="FE") is what you should do and this gives a p-value of about .20 for the combined estimate.

Pooling of cox regression models going wrong

1 Answers