I ran multiple imputation in R using mice. Only one categorical variable had missingness and I specified the imputation model to imputate it using polyreg.
After imputation, I run the Cox model below (minimal example):
cox_simple <- with(
data_imputed,
coxph(Surv(start_time, stop_time) ~
gender +
age +
pspline(hemoglobin)
)
)
# pool the estimates
summary(pool(cox_simple))
The problem I run into are with summary(pool(cox_simple)). It returns NA for the spline estimates. When I run the model with pspline without imputed data, everything works fine and model diagnostics look good. When I run the model with imputed data, but replace the penalised spline by a natural spline ns(hemoglobin, df = 3) everything works fine as well. However, for multiple reasons I would vastly prefer using a penalised spline for this.
Is there a way to obtain pooled model estimates despite using it on imputed data?
