Please add an option to skip remove.lindep()

[shangzhi-hong]

Please add option to skip the call to remove.lindep() in the sampler.univ() function.
I understand it is necessary for most models, but for methods like random-forest-based ones, it may not be necessary.
It causes errors when I do simulations, and I found it impossible to skip it.

Thanks.

[stefvanbuuren]

Did you already try something like mice(data, maxcor = 1.0, eps = 0)?

This will set the maxcor and eps arguments of remove.lindep() to its extremes, and hence all variables will pass.

This is a dangerous option. The setting could crash mice() due to numerical difficulties if collinearity is not properly handled.

[gerkovink]

For many such numerical difficulties, the estimation will fall back to ridge regression for at least the default methods. If this happens, the logged events will notify the users about this expected behaviour.

[shangzhi-hong]

Did you already try something like mice(data, maxcor = 1.0, eps = 0)?

This will set the maxcor and eps arguments of remove.lindep() to its extremes, and hence all variables will pass.

This is a dangerous option. The setting could crash mice() due to numerical difficulties if collinearity is not properly handled.

Thanks, I am not aware of this combination of paramters (did not see the ... in the function call), sorry
I am aware of the consequences, but I am doing experiments with machine-learning-based methods (like CART, RF, etc.), I believe that the are robust against collinearity

[shangzhi-hong]

It seems that the remove.lindep function has following lines:

  if (eps <= 0)
    stop("\n Argument 'eps' must be positive.")

that eps = 0 would not pass this check

[shangzhi-hong]

I am wondering whether eps = .Machine$double.xmin or eps = .Machine$double.eps would do the trick

[stefvanbuuren]

mice 3.8.4 tweaks remove.lindep() as follows.

  # setting eps = 0 bypasses remove.lindep()
  if (eps == 0) 
    return(rep.int(TRUE, ncol(x)))
  if (eps < 0)
    stop("\n Argument 'eps' must be positive."

You should now be able to bypass remove.lindep() by setting mice(..., eps = 0).

[shangzhi-hong]

Thanks, I will change my code accordingly