How does TargetDiscovery handle missing data?
When estimating the variable importance of a given target variable, TargetDiscovery discards any observations that have missing values in the outcome, target variable, or subgroup variable (if specified). Missing values in any of the confounders are imputed as follows. The missing values are replaced by a typical value for that variable among the non-missing observations, and an indicator variable is added to the list of confounders that records whether a given value represents an observed or an imputed measurement. For continuous variables, the mean of the non-missing values is used as the typical value; for discrete variables, the mode of the non-missing values is used. This approach ensures that the information content of the original data is preserved since the imputation step could be reversed (zero loss of information). Conventional variable importance algorithms approach missing data in one of two ways: either observations with missing values for any of the variables are discard
