Detecting outliers: Do not use standard deviation around the mean, use absolute deviation around the median

By C. Leys, Christophe Ley, Olivier Klein, P. Bernard, Laurent Licata. 2013

3,802 citations recorded by the source corpus.

View Paper: https://www.semanticscholar.org/paper/f34a372df4c449f914ce8273a096fe1e743e46b5

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Statistics and Probability

Citations in the corpus

Forecasting Error Calculation with Mean Absolute Deviation and Mean Absolute Percentage Error (2017, neutral): “MAD is useful when measuring prediction errors in the same unit as the original series[5][6].”
Generalized Out-of-Distribution Detection: A Survey (2021, neutral): “Density-Based Methods OD methods include Gaussian distribution (Altman & Bland, 2005; Leys et al., 2013), Mahalanobis distance (De Maesschalck et al., 2000), Gaussian mixtures (Yang et al., 2009), and Local outlier factor (LOF) (Breunig et al., 2000).”
NeuronInspect: Detecting Backdoors in Neural Networks via Output Explanations (2019, methodology): “We detect the out-liers based on the median absolute value (Leys et al. 2013).”
Experimental design and analysis and their reporting II: updated and simplified guidance for authors and peer reviewers (2018, background): “Guidance on ethical animal experimentation is published elsewhere (McGrath and Lilley 2015) and will be updated separately in 2018.”
The mechanical response of cellular materials with spinodal topologies (2019, methodology): “Outliers in κ1 and κ2 are determined based on a median absolute deviation method (Leys et al., 2013), and any (κ1, κ2) pair that has an outlier is removed.”
Learning One Class Representations for Face Presentation Attack Detection Using Multi-Channel Convolutional Neural Networks (2020, methodology): “Also, normalization using Mean Absolute Deviation (MAD) [44] is performed to convert the raw 16-bit values to the 8-bit range.”
DECAF: MEG-Based Multimodal Database for Decoding Affective Physiological Responses (2015, methodology): “These annotations were processed to arrive at the final set of 36 clips as follows: 1) To ensure that the annotations are comparable, we transformed all V and A annotations using the z-score normalization.”
Tangled up in BLEU: Reevaluating the Evaluation of Automatic Machine Translation Evaluation Metrics (2020, methodology): “Instead, we use the median and the Median Absolute Deviation (MAD) which are more robust (Iglewicz and Hoaglin, 1993; Rousseeuw and Hubert, 2011; Leys et al., 2013).”
Improving fluid intelligence with training on working memory: a meta-analysis (2014, neutral): “Outliers were categorized as data points that were 3 median absolute deviations away from the overall median (Leys, Ley, Klein, Bernard, & Licata, 2013).”
Recall length and measurement error in agricultural surveys (2020, methodology): “We apply the double Median Absolute Deviation method for outlier detection and correction to all constructed continuous variables in this analysis, that is, harvest quantity, storage and sales quantity, total, family, and hired labor person-days, organic and inorganic fertilizer quantity applied (Le”
TABOR: A Highly Accurate Approach to Inspecting and Restoring Trojan Backdoors in AI Systems (2019, methodology): “In the pioneering work [48], researchers have computed the L 1 norm for each local optimum and then utilized an outlier detection method (MAD [25]) to filter out local optima indicating incorrect triggers and false alarms.”
Results of the WMT20 Metrics Shared Task (2020, methodology): “This method, recommended by the statistics literature (Iglewicz and Hoaglin, 1993; Rousseeuw and Hubert, 2011; Leys et al., 2013) depends on the median and the median absolute deviation (MAD) which is the median of the absolute difference between each point and the median.”
Products as Pals: Engaging with Anthropomorphic Products Mitigates the Effects of Social Exclusion (2017, neutral): “Other research has also shown that a need to belong heightens preferences for nostalgic products (Loveland et al. 2010), which could (arguably) be one reason why participants felt more positively about a common childhood toy like a teddy bear.”
ECOD: Unsupervised Outlier Detection Using Empirical Cumulative Distribution Functions (2022, methodology): “This has motivated commonly used heuristic OD approaches such as the “three-sigma” rule that declares points more than three standard deviations from the mean to be outliers ([27], [28]; there’s also a robust variant called the “1.5 IQR” rule [29]).”