BUG - Fixes response leakage

[mayer79]

Fixes a major issue #77: the model response was used as covariate, which does not make sense.

[teecrow]

@mayer79 This may be worth revising. Counterintuitively, the response/outcome variable should sometimes be used for imputation. Here's a paper that explains this in much more detail. The TLDR is in the abstract:

We mathematically demonstrate that including the outcome variable in imputation models is not just a recommendation but a requirement to achieve unbiased results when using stochastic imputation methods. Moreover, we dispel common misconceptions about deterministic imputation models and demonstrate why the outcome should not be included in these models.

So it depends on the person's use case - so I would argue there should be an option to include or not include the response variable in the imputation. My vote would be the default is to not include it, given my assumption that most people using your package will be from the data science world where deterministic imputation is more common, rather than e.g., the social science world, where multiple imputation is more common.

[mayer79]

Thanks for digging into this.

1. When your task is to build a regression model with a response, and your covariates have missing values (at random), then it can make sense to use the regression model response in the imputation models (for the features).
2. The bug in missRanger(), however, used each variable y both as response and covariate in the same model. This does not make sense. It is not so bad as it seems though, because any missing value in y had been imputed by the other variables in previous iterations. Still, I expect imputations to become better by fixing the bug.

[teecrow]

Ah I see - I misunderstood the nature of the bug. My apologies and thanks for a great package!