Table 1 Multivariate outlier removal via PyOD’s algorithms cannot reliably identify data rows that contain a single univariate outlier. Since all of the scalar outliers were simulated to be more extreme than the most extreme non-outlying scalar value across all features, most univariate outlier removal algorithms would have a 100% TPR in this test because they transform scalars to outliership scores monotonically

From: STAR_outliers: a python package that separates univariate outliers from non-normal distributions