Outlier detection in multivariate analytical chemical data.

Area of Science:

• Statistics
• Chemometrics
• Data Analysis

Background:

• Traditional multivariate outlier detection methods (Mahalanobis distance, hat matrix leverage) are unreliable.
• Robust statistical methods for multivariate outlier detection are increasingly important in chemical literature.
• Existing robust methods (MVE, MVT, M-estimators, MCD) have programming difficulties and long processing times.

Purpose of the Study:

• To introduce and evaluate novel, robust methods for multivariate outlier detection.
• To compare the performance of proposed methods against established techniques like MCD and MVT.
• To address the limitations of existing outlier detection algorithms in multivariate data analysis.

Main Methods:

• Development of two new outlier detection methods: resampling by the half-means and smallest half-volume.
• Evaluation of proposed methods using selected datasets and simulations.
• Comparison with Minimum Covariance Determinant (MCD) and Multivariate Trimming (MVT).

Main Results:

• Multivariate Trimming (MVT) was found unreliable due to its reliance on Mahalanobis distance.
• The proposed resampling by the half-means and smallest half-volume methods demonstrated superior performance.
• Both new methods are simple, conceptually clear, and outperform MCD and MVT.

Conclusions:

• The proposed resampling by the half-means and smallest half-volume methods are recommended for detecting multiple outliers in multivariate data.
• These novel techniques offer a more reliable and efficient alternative to existing methods.
• Simplicity and conceptual clarity enhance the practical applicability of the new outlier detection strategies.