This method is an extension of method SDID because it accounts for the
“outlyingness” of each observations. This is a quite natural approach
since outliers do have a higher risk of re-identification and therefore
these outliers should have larger disclosure risk intervals as observations
in the center of the data cloud.
The algorithm works as follows:
1. Robust Mahalanobis distances are estimated in order to get a robust
multivariate distance for each observation.
2. Intervals are estimated for each observation around every data point of
the original data points where the length of the interval is
defined/weighted by the squared robust Mahalanobis distance and the
parameter $k$. The higher the RMD of an observation the larger the
interval.
3. Check if the corresponding masked values fall into the intervals around
the original values or not. If the value of the corresponding observation
is within such an interval the whole observation is considered unsafe. So,
we get a whole vector indicating which observation is save or not, and we
are finished already when using method RMDID1).
4. For method RMDID1w: we return the weighted (via RMD) vector of disclosure
risk.
5. For method RMDID2: whenever an observation is considered unsafe it is
checked if $m$ other observations from the masked data are very close
(defined by a parameter $k2$ for the length of the intervals as for SDID or
RSDID) to such an unsafe observation from the masked data, using Euclidean
distances. If more than $m$ points are in such a small interval, we
conclude that this observation is ``save''.