Hello!
The main idea is that subjects identified as outliers in multivariate analyses are generally expected to also appear as outliers in univariate analyses, or at least be detectable in one or more univariate dimensions.
The Outliers report identifies potential outliers using a univariate detection algorithm applied to individual tests within the findings domains. Robust scores, based on the methodology from Rousseeuw and Hubert (2011), are computed for each finding. Observations with robust scores greater than 2.5 or less than -2.5 are flagged as potential outliers.
Referring to Rousseeuw and Hubert (2011), the robust scores are computed as (x_i - median)/MAD, which is similar to z-scores' format. The value 2.5 is a suggested reference value. We picked 2.5 as scores beyond ±2.5 correspond roughly to data points more extreme than about 98–99% of a normally-distributed dataset.
These robust scores are summarized by study site in Figure 2, titled "Univariate Outliers by Study Site." This visualization allows users to quickly identify outliers at each site. Additional details can be accessed by hovering over individual points in the graph or by selecting points and displaying the associated tables (show tables).
The current boxplots displays robust scores per category, visually emphasizing the distribution, median, and outliers within each group. Unlike a volcano plot, which focuses primarily on magnitude versus significance (e.g., fold-change vs. p-value), the boxplot directly visualizes category-level variation and highlights individual outliers explicitly, making it easier and more intuitive for robust outlier detection and comparison across categories.
Please let us know if you have more questions! We are happy to discuss!
These videos provide JMP essentials for new users and those who want a refresher. Overviews Getting Started With JMP JMP Tips and...