gFormulaImpute with custom method fails on mice 3.18

[martindanka]

Firstly, I wanted to thank you for all your fantastic work on g-methods and multiple imputation!

Following the recent mice package dependency update from 3.17.0 to 3.18.0, gFormulaImpute() is no longer compatible with mice when a user supplies a custom method.

Reproducing the issue

The following code executes correctly with mice 3.17.0, but throws an error with mice 3.18.0:

# Packages --------------------------------------------------------------------------------------------------------
library(mice)          
library(gFormulaMI)


# Reproduce the bug -----------------------------------------------------------------------------------------------
set.seed(7626)
imp1 <- gFormulaImpute(
  data = simDataFullyObs,
  M = 1,
  trtVars = c("a0", "a1", "a2"),
  trtRegimes = list(c(0,0,0), c(1,1,1))
)
meth <- imp1$method
pred <- imp1$predictorMatrix

# This will throw an error even for mice v. 3.17.0 because the method vector includes 'regime', which is not accepted by the function. 
try(
  gFormulaImpute(
    data = simDataFullyObs,
    M = 1,
    trtVars = c("a0", "a1", "a2"),
    trtRegimes = list(c(0,0,0), c(1,1,1)),
    method = meth,
    predictorMatrix = pred
  )
)

# This was a workaround with mice 3.17.0, which no longer works with mice 3.18.0.
# Remove 'regime' from method
meth2 <- meth[!grepl("regime", names(meth))]

# This will throw an error for mice v. 3.18.0 but not earlier versions
try(
  gFormulaImpute(
    data = simDataFullyObs,
    M = 1,
    trtVars = c("a0", "a1", "a2"),
    trtRegimes = list(c(0,0,0), c(1,1,1)),
    method = meth2,
    predictorMatrix = pred
  )
)

Output with mice 3.18.0:

[1] "Input data is a regular data frame."
Error : All elements of 'method' must be named and match blocks.

A potential fix

I believe I may have found the source of this issue. From mice 3.18.0 onward, mice() insists that its internal blocks vector has the same names and order as the column names in the data frame.gFormulaImpute() creates a new column called regime and adds an empty element to the user-supplied method vector intended for 'regime', but this element is not named.

Below is a potential solution that also allows gFormulaImpute() to accept a method vector both with or without the regime element. I am not sure this covers every scenario, so I have opened this as an Issue and not a Pull request. The fix is for lines 192-195 and 242-245 of the original gFormulaImpute() function, so this change needs to be applied at two places. This change should also keep the function compatible with mice 3.17.

-      #add on an empty imputation method for the new variable regime
-      method <- c(method,"")
+      if (!"regime" %in% names(method)) {
+        #add on an empty imputation method for the new variable regime
+        method <- c(method, regime = "")
+      }

Let me know if this works and if you'd like me to open a pull request.

[jwb133]

Thanks Martin for this!

For the first call which throws an error, I get:

Error: Length of method differs from number of blocks

but I think that is what should happen, because meth is length 8 now (including regime) while there are 7 columns in the input data frame.

But the second call you have should be ok, but doesn't now work I think for the reason you have identified (thank you!). I think for fixing it we can just have:

method <- c(method,regime = "")

in these two places?

[martindanka]

Thanks for checking this so promptly! That's definitely a simpler solution and it preserves the error on the first call as you mentioned. I did find this error a bit unintuitive, because this is inconsistent with how the predictor matrix is handled, since that is accepted whether or not it includes regime. It also means gFormulaImpute does not accept its own method vector if passed back in unchanged. But I can also see that the error is consistent with how mice itself treats the methods vector (and, more importantly, either solution works!).

[jwb133]

Thanks for looking quickly. I've committed the modified line in my previous message and modified some of the tests that tested functionality when the user supplies their own method argument. I've added the documentation to the function to say that if you supply a method vector, it should be named.

[martindanka]

Thanks, Jonathan! I've installed the updated version and it works on my end too.