A New Outlier Identification Test for Method Comparison Studies Based on Robust Regression

Abstract

The identification of outliers in method comparison studies (MCS) is an important part of data analysis, as outliers can indicate serious errors in the measurement process. Common outlier tests proposed in the literature usually require a homogeneous sample distribution and homoscedastic random error variances. However, datasets in MCS usually do not meet these assumptions. In this work, a new outlier test based on robust linear regression is proposed to overcome these special problems. The LORELIA (local reliability) residual test is based on a local, robust residual variance estimator, given as a weighted sum of the observed residuals. The new test is compared to a standard test proposed in the literature by a Monte Carlo simulation. Its performance is illustrated in examples.

Key Words:

• Heteroscedasticity
• Local variance estimator
• Method comparison studies
• Outliers

Notes

¹Number in parentheses: TP in % of the number of simulated outliers.

²Number in parentheses: FP in % of the number of non-outlying normal data.

^hHomogeneous outlier distribution.

^clClustered outliers.

¹Number in parentheses: TP in % of the number of simulated outliers.

²Number in parentheses: FP in % of the number of non-outlying normal data.

^hHomogeneous outlier distribution.

^clClustered outliers.

¹Number in parentheses: TP in % of the number of simulated outliers.

²Number in parentheses: FP in % of the number of non-outlying normal data.

^hHomogeneous outlier distribution.

^clClustered outliers.