Statistics Glossary → Minimum covariance determinant (MCD) and robust distance
Minimum covariance determinant (MCD) and robust distanceconcept
A robust estimator for multivariate data that uses the 'tightest' subset of observations to obtain location and covariance estimates unaffected by outliers; the Mahalanobis distance computed from it is called the robust distance.
MCD, proposed by Rousseeuw (1984), searches for the subset of h out of n observations (usually between about half and three quarters) whose covariance matrix has the smallest determinant. The determinant is related to the volume of the data cloud, so the selected subset is the densest, most consistent core of the data. The mean and (consistency-scaled) covariance matrix of this subset serve as the robust estimates. The FAST-MCD algorithm is widely used to compute it.
The ordinary Mahalanobis distance can hide outliers (masking) because the mean and covariance are themselves distorted by those outliers. The MCD-based robust distance largely removes this problem; outliers are flagged by comparing the squared distance with a high percentile (e.g. 0.975) of the chi-square distribution with degrees of freedom equal to the number of variables. In robust regression, MCD is used to identify leverage points in the space of the predictors. The breakdown point depends on h: the smaller h, the more outliers the estimator withstands, but the lower its efficiency. The method becomes unstable as the number of variables approaches the number of observations.
An everyday example
A bank examines customers' monthly spending and number of transactions together. A few fraudulent accounts make a small number of very large transactions. Because the ordinary mean and covariance are pulled by these accounts and stretch the data cloud, the accounts can look 'normal'. MCD determines the true shape of the cloud from the most consistent majority of customers, and the robust distance then exposes these accounts as clear outliers.
In MerQur
It is used in Robust Regression (M-estimator) and Anomaly Detection to identify multivariate outliers and leverage points.
Analyses where this concept appears
Robust Regression (M-estimator)Anomaly Detection
Related concepts
Mahalanobis distanceRobust regression estimators (MM, S, LTS) and the breakdown point