Error message in Fine Gray regression analysis in R (package cmprsk)

Question

I am working on competing risk analysis in R thanks to a Fine & Gray regression analysis. Here is my code with death as the competing risk:

fg.model <- crr(ftime,fstatus,cov,failcode=1,cencode=0)

ftime is a numerical variable ranging from 1 to 180 days that indicates the period of follow-up of patients until their death (fstatus==1). If they are still alive until the end of the follow-up, this variable is equal to 180 days and their status is equal to 0. In summary, If a person dies after 30 days of follow-up, the variable ftime will therefore be equal to 30 days and the variable fstatus to 1 and if a person is still alive after the end of the follow-up (max 180 days) and did not die during it, the variable ftime will therefore be equal to 180 and the variable fstatus to 0. fstatus is also a numerical variable.
The parameter "cov" is a dataframe with two covariates (age and sex converted in factors). Failcode is equal to 1 as death is the competing event and cencode is equal to 0 as survivors as considered as censored.

I have the following error message:

# NAs introduced by coercion Error in crr(ftime,fstatus,cov,failcode=1, : 
# NAs introduced by coercion NA/NaN/inf in foreign function call(arg4)

Since I have no missing data in my database, what can explain this error message and how can I solve it?

I already tried to use na.omit, complete.case, and other code to be sure that there is no missing data in my code. I also check the structure of the data but time and status are well numerical and cov converted in factors.

Here is a code reproducing my dataset and the error message:

# Set the sample size
n<- 8076
# CReate a variable for follow-up time
time<- c(rep(180,6533),sample(1:179,n-6533,replace = TRUE))
# Create a variable for status
status<-ifelse(time==180,0,1) # O= alive/censored
                              # 1 = death
# Create age
    age<-sample(18:90,n,replace = TRUE)
# Create gender
sex<-sample(c("male","female"),n,replace = TRUE)
# Combine
df<-data.frame(time,status,age,sex)
# Create cov
cov<-subset(df,select = c("age","sex"))
cov$sex<-as.factor(cov$sex)
# Run Fine Gray model
library(cmprsk)
fg.mod <-crr(df$time,df$status,cov,failcode = 1,cencode = 0) 
                                                                                                           

Answer 1

Pass your own design matrix to the model using model.matrix()

The solution to the question in your comment - what happens if you have a factor with several levels - is to do this:

# Create factor with three levels
cov$income <- factor(sample(c("high", "med", "low"), nrow(cov), replace = TRUE))

# Define factors for model spec
factors <- c("sex", "income")
model_spec <- reformulate(
    paste0("age+", paste(factors, collapse = "+"))
) # ~age + sex + income

covariates_matrix <- model.matrix(
    model_spec,
    data = cov,
    contrasts.arg = lapply(cov[factors], contrasts)
)[, -1] # first column is constant intercept (1)

head(covariates_matrix, 3)
#   age sexmale incomelow incomemed
# 1  60       1         0         1
# 2  39       0         1         0
# 3  23       0         0         0

crr(df$time, df$status, covariates_matrix, failcode = 1, cencode = 0)

# convergence:  TRUE
# coefficients:
#        age    sexmale  incomelow  incomemed
#  0.0002871  0.0506400 -0.0391500 -0.0098910
# standard errors:
# [1] 0.001193 0.050910 0.062340 0.062020
# two-sided p-values:
#       age   sexmale incomelow incomemed
#      0.81      0.32      0.53      0.87

As you can see, your factor variables now have a coefficient for each level, rather than being treated as continuous.

Why you need to do this

The difficulty you have is that the cmprsk::crr() function does not support model formulas. Instead it takes a matrix, which means that all covariates will be coerced to the same class. In your case, your factor variables lead to everything being coerced to a character, and NAs introduced by coercion. As the cmprsk docs state:

The model.matrix function can be used to generate suitable matrices of covariates from factors

In your original question, which had a binary factor, we could just do cov$sex <- cov$sex=="male", and get a binary covariate which is easy to interpret. However, with more levels, we need to use model.matrix(). If you want to change how the reference level is represented, see this question.