However, the point of the MICE algorithm is to not use only the complete dataset, but to use the previous round's imputations as placeholders in the variables that are not currently being imputed (if I am imputing column x i want to use imputed placeholders in columns y, and z before doing the predictive imputing of column x)
For the first round, missing values are replaced by simple placeholders (mean/mode imputation) - additionally the order in which the columns are imputed at each round should be customizable
In the current implementation, the algorihm simply runs the multiple imputer K times, and the result is equal to the K-1th run, regardless of the results of the previous runs
See for reference: Azur, Melissa J., et al. "Multiple imputation by chained equations: what is it and how does it work?." International journal of methods in psychiatric research 20.1 (2011): 40-49.
The series imputers used by the MICE Imputer, still call _get_observed (https://github.com/kearnz/autoimpute/blob/d1a4c3966ea4138cd52111d51ef22d2fb43648e2/autoimpute/imputations/dataframe/single_imputer.py#L185)
However, the point of the MICE algorithm is to not use only the complete dataset, but to use the previous round's imputations as placeholders in the variables that are not currently being imputed (if I am imputing column x i want to use imputed placeholders in columns y, and z before doing the predictive imputing of column x)
For the first round, missing values are replaced by simple placeholders (mean/mode imputation) - additionally the order in which the columns are imputed at each round should be customizable
In the current implementation, the algorihm simply runs the multiple imputer K times, and the result is equal to the K-1th run, regardless of the results of the previous runs
See for reference: Azur, Melissa J., et al. "Multiple imputation by chained equations: what is it and how does it work?." International journal of methods in psychiatric research 20.1 (2011): 40-49.