4

I used the coxph() function from the survival package in multiply imputed dataset and encountered a warning when trying to pool the results. The warning message states: "In mice.df(m, lambda, dfcom, method) : Large sample assumed.

A reproducible example is below (with publically available data, without worrying to much about the appropriateness of using both mice() and coxph() with these data):

library(mice)
library(survival)

#load publically available data
data(pbc)

#select variables for the reproducable example
pbc.select <- pbc[pbc$status %in% c(0,1) , c("id", "time", "status", "trt")]

imp <- mice(pbc.select) #impute trt
fit <- with(imp, coxph(Surv(time, status) ~ trt)) #fit coxph in each imp
pool(fit) #pool the models; get's the error

This warning seems to result from the pool() function trying to require dfcom from

dfcom <- df.residual(object)

where df.residual() is not applicable to the object referred to in this context, which is of class coxph

class(fit) # "mira" "matrix" 
class(fit$analyses[[1]]) "coxph"

My questions are

  1. am I using the correct syntax for purpose
  2. if so, is there a way to supply pool() with the appropriate information?
  3. how does this assumption influence the results?
lovalery
  • 4,524
  • 3
  • 14
  • 28
Jeroen Hoogland
  • 133
  • 1
  • 9
  • Hi Jeroen - did you figure this out ? I have the same question using clogit (which is Cox under the hood). Simply wondering if this warning is something to worry about or not – user2498193 Jun 24 '16 at 21:02
  • Has a solution been found for this, by either poster asking? – AP30 Jan 30 '18 at 14:50
  • Hi, as of yet, I think the syntax is ok, but small sample methods (using adjusted degrees of freedom) have not been implemented for the multiparameter case in mice (e.g. methods by Reiter (2007)), whereas they were implemented for the single parameter (e.g. in pool.scalar which implements methods by Barnard and Rubin (1999)). Therefore, inference is based on classical Rubin's rules assuming complete-data df are infinite. Whether this is problematic depends on the sample size, as the warning already noted. Perhaps someone more knowledgeable on the topic can provide a more complete answer. – Jeroen Hoogland Feb 01 '18 at 15:41

0 Answers0