Right now, the MultipleImputer does not use the initialized imputed dataframe, but rather selects only complete rows in the imputation. This occurs when each SingleImputer within a MultipleImputer calls the _get_observed method. While this is a common practice when strict MCAR or MAR conditions are met, using observed only can lead to selection bias when these conditions break down.
@gjvd provided this note (pg 10). In our implementation, the user should have the option to use observed only or use initialized DataFrame. This can be implemented as a formal argument with default value set.
kearnz
commented
4 years ago
Right now, the
MultipleImputerdoes not use the initialized imputed dataframe, but rather selects only complete rows in the imputation. This occurs when eachSingleImputerwithin aMultipleImputercalls the_get_observedmethod. While this is a common practice when strict MCAR or MAR conditions are met, using observed only can lead to selection bias when these conditions break down.@gjvd provided this note (pg 10). In our implementation, the user should have the option to use observed only or use initialized DataFrame. This can be implemented as a formal argument with default value set.