Robust normality transformation and outlier detection in diverse distributions, with application to functional neuroimaging data
summary
The gist
Automatic detection of statistical outliers is facilitated through knowledge of the source distribution of regular observations <ref:2505.11806#pg2>.
In short
This research develops a novel robust transformation called SHASH to normalize data, effectively handling skewness and non-Gaussian tails better than existing methods. It proposes two initialization strategies—robust z-scoring and an isolation forest approach—to detect outliers reliably even when data is heavily contaminated (up to 30%). The method shows strong performance in functional neuroimaging analysis.
Key concepts
- SHASH Transformation
- A flexible distribution family used to transform raw data into a more normal distribution. It has four parameters controlling mean, variance, skewness, and tailweight. This transformation is superior because it can handle complex shapes like heavy tails and skewness that other methods struggle with.
- Robust Outlier Initialization
- The process of labeling outliers before the main transformation. The paper tests two methods: robust z-scoring using median and MAD, and an anomaly detection method based on an isolation forest. Combining these initialization steps helps ensure that outliers are correctly identified and not hidden during the normalization process.
- Functional Neuroimaging Data
- The specific application area for this study, involving fMRI data. The researchers used robust Principal Component Analysis (PCA) on ICA time courses to reduce dimensionality while preserving 80% of the variance. They applied the SHASH method to detect contaminated brain volumes in this complex dataset.
Terminology used across episodes
This episode discusses
- Robust normality transformation and outlier detection in diverse distributions, with application to functional neuroimaging data · Paper Radio
The paper
Robust normality transformation and outlier detection in diverse distributions, with application to functional neuroimaging data · Read on arXiv
Fatma Parlak, Saranjeet Singh Saluja, Amanda Mejia
Department of Statistics, Indiana University Bloomington · Department of Mathematics, Ankara Yıldırım Beyazıt University
Transcript
Introduction to the show: ident: Statistics. The day in statistics: methods, theory and the inference under them.
Tom: We're looking at the paper: "Robust normality transformation and outlier detection in diverse distributions, with application to functional neuroimaging data".
Jane: Here's the gist of the paper. Automatic detection of statistical outliers is facilitated through knowledge of the source distribution of regular observations <ref:2505.11806#pg2>.
Tom: Here, a novel robust method for transformation to Normality based on the highly flexible sinh-arcsinh (SHASH) family of distributions is developed to accommodate skew, non-Gaussian tail weights, and combinations of both <ref:2505.11806#pg2>.
Jane: This SHASH transformation outperforms existing methods by exhibiting high sensitivity to outliers even at heavy contamination levels (20-30%) <ref:2505.11806#pg2>.
Tom: The sinh-arcsinh (SHASH) family of distributions is characterized by four parameters controlling the mean (µ), variance (σ), skewness (ν), and tailweight (τ) <ref:2505.11806#pg7>.
Title and authors: Tom: So let's talk about who wrote this. Fatma Parlak, Saranjeet Singh Saluja, and Amanda Mejia are from Indiana University Bloomington and Ankara Yıldırım Beyazıt University in Turkey. They contributed equally to this work.
Jane: Yeah, what that means is they’re the ones who developed this robust method for transforming data to Normality using the SHASH distribution family.
Lu: The title itself points to the main things: robust normality transformation and outlier detection in diverse distributions with application to functional neuroimaging data. It tells you exactly what they did and where they tested it.
Meng: For me, what that implies is that if we have complex data, like fMRI time courses, this method might actually work where other methods fail because they don't handle those tail weights well.
Lalam: Exactly. They are trying to move away from transformations like Box-Cox or YeoJohnson which are limited to just skewness and see how much more robust the detection becomes with SHASH.
Tom: It’s about making sure that when we look for outliers in a distribution that isn't Gaussian, we aren't accidentally masking those outliers because the transformation itself gets messed up by them.
Jane: Right, so it’s not just about getting things normal; it’s about doing it in a way that respects the presence of outliers.
Lu: They are specifically focusing on avoiding the influence of outliers themselves in the detection process, which is a big challenge in outlier detection methodology Rousseeuw and Hubert wrote about.
The paper's summary: Tom: So what’s the actual mechanism they propose? The paper says that automatic detection of statistical outliers can be facilitated by knowing the source distribution of the regular observations.
Jane: Because we often don't know the true population distribution, so one approach is to apply a transformation to make it normal. But as we said, standard transformations are limited.
Lu: They develop this novel robust method based on the highly flexible sinh-arcsinh SHASH family of distributions specifically designed to accommodate skewness, non-Gaussian tail weights, and combinations of both.
Meng: So they’re building a distribution that is more flexible than what we typically use when we try to force data into a Gaussian shape.
Lalam: They define the transformation to standard normality by an equation where Z equals S nu tau of X equals sinh tau sinh minus one of X minus nu.
Tom: That equation means they’re using four parameters controlling the mean, the variance, the skewness, and that tailweight. That's a lot of control over how much distortion we allow.
Jane: And this four-parameter family incorporates location and scale parameters mu and sigma which leads to something called SHASH o two distribution.
Lu: The probability density function for this SHASH o two distribution is given by a complex expression involving hyperbolic identities. It’s mathematically heavy, but it captures the complexity of the required shape.
The paper's improvements: Tom: So what are the actual improvements they claim over existing methods? They say their SHASH transformation outperforms existing methods by exhibiting high sensitivity to outliers even at heavy contamination levels, specifically at twenty to thirty percent contamination.
Jane: That’s a big deal because most methods start struggling when you introduce a lot of outliers. It suggests it’s much better than what we have now for skewed data.
Lu: They tested this against robust z-scoring and the robust Box-Cox or YeoJohnson transformation across various distribution shapes and outlier contamination levels from light, one percent up to heavy, thirty percent.
Meng: I remember seeing some results suggesting that SHASH-i performs quite well even at thirty percent outlier contamination, which is a strong claim for real-world data.
Lalam: They also mentioned that the robust Box-Cox or YeoJohnson transformation seems to consistently fail even at five to ten percent outlier contamination for all three skewed distributions because of poor outlier initialization.
Tom: So the key improvement they highlight is this high sensitivity across a wide range of contamination levels, which is what makes it useful for messy data.
Jane: It’s like they’ve made the process much less sensitive to the initial noise in your dataset, which is a major hurdle for any outlier detection system.
Conclusion: Tom: So where does this leave us? The study suggests that combining the outlier initialization techniques underlying SHASH z and SHASH i might be an effective strategy to further improve sensitivity and avoid masking outliers.
Jane: It points toward a combined approach, using both methods for initializing labels to get the best results. They also noted that relying only on robust z-scoring can deteriorate rapidly as outlier contamination goes beyond one percent.
Lu: They pointed out that there is a need for alternative initialization methods like univariate minimum covariance determinant estimators because the robust z-scoring method has these limitations when things get really bad.
Meng: From a production standpoint, if you're shipping models, you need to know where this method breaks down—and they’re telling us that beyond one percent contamination, the simple robust z-scoring isn't reliable enough.
Lalam: They also suggest that future work should focus on embedding this SHASH transformation within a multivariate outlier detection procedure. That would be a big step for making it practical for complex datasets like functional neuroimaging data.
Tom: So, in summary, the paper "Robust normality transformation and outlier detection in diverse distributions, with application to functional neuroimaging data" shows that SHASH is more robust than standard methods when dealing with heavy contamination and complex shapes.
Jane: It’s a method that tries hard to keep things normal while acknowledging the weirdness of real-world data, which could make outlier detection much more reliable in medical imaging applications.
More episodes
- Daily Summary for 2026-11-19
- 2310.07973-Statistical Performance Guarantee for Subgroup Identification with Generic Machine Learning
- 2408.09155-Learning Robust Treatment Rules for Censored Data
- Daily Summary for 2026-11-19
- 2603.23397-Kinetic Langevin Splitting Schemes for Constrained Sampling
- 2606.09906-An information-geometric framework for mapping maximum potential biodiversity
- 2603.27718-Induced replication and the assessment of models
- 1307.0898-Entropy of a Zipfian Distributed Lexicon
- 2406.19974-Generalizing self-normalized importance sampling with couplings
- 2412.10213-Collaborative Optimal Design of Controlled Experiments with Subject Covariates