Outlier detection using multiple statistical methods from the easystats performance package. This module provides comprehensive outlier detection through univariate methods (Z-scores, IQR, confidence intervals), multivariate methods (Mahalanobis distance, MCD, OPTICS, LOF), and composite scoring across multiple algorithms. Complements existing data quality assessment modules with state-of-the-art outlier detection capabilities. Perfect for clinical research data quality control and preprocessing.
Usage
outlierdetection(
data,
vars = NULL,
method_category = "composite",
univariate_methods = "zscore_robust",
multivariate_methods = "mahalanobis",
composite_threshold = 0.5,
zscore_threshold = 3.29,
iqr_multiplier = 1.7,
confidence_level = 0.999,
show_outlier_table = TRUE,
show_method_comparison = FALSE,
show_exclusion_summary = FALSE,
show_visualization = FALSE,
show_interpretation = FALSE,
sampleThreshold = 10000,
sampleSize = 5000,
seed = 123
)Arguments
- data
The data as a data frame.
- vars
Continuous variables to analyze for outliers. The module will detect outliers based on the selected variables using the chosen detection methods.
- method_category
Category of outlier detection methods to use. CLINICAL GUIDANCE: Univariate methods analyze each variable separately (ideal for lab values, vital signs). Multivariate methods consider relationships between variables (useful for correlated biomarkers). Composite combines multiple approaches for robust detection (recommended for most clinical data). Examples: Use univariate for hemoglobin levels, multivariate for complete blood count panels.
- univariate_methods
Specific univariate method for outlier detection when univariate category is selected.
- multivariate_methods
Specific multivariate method for outlier detection when multivariate category is selected.
- composite_threshold
Threshold for composite outlier score (0.1-1.0). Default 0.5 means observations classified as outliers by at least half of the methods are considered outliers.
- zscore_threshold
Threshold for Z-score based methods. CLINICAL EXAMPLES: 3.0 = 99.7 percent confidence (standard screening for most lab values), 3.29 = 99.9 percent confidence (stringent, for critical values like cardiac enzymes), 2.5 = 98.8 percent confidence (sensitive detection for research). Recommended: 3.29 for clinical quality control.
- iqr_multiplier
Multiplier for IQR-based outlier detection. CLINICAL EXAMPLES: 1.5 = Tukey's standard (sensitive, may flag ~0.7 percent of normal data), 1.7 = conservative (recommended for clinical screening), 2.0 = very conservative (for critical biomarkers). Useful for non-normal distributions common in clinical data.
- confidence_level
Confidence level for interval-based methods (ETI, HDI). Default 99.9 percent identifies the most extreme observations.
- show_outlier_table
Display a comprehensive table of outlier detection results including outlier scores, distances, and classification for each observation.
- show_method_comparison
Compare results across different detection methods when using composite approach.
- show_exclusion_summary
Provide summary of observations recommended for exclusion and impact analysis.
- show_visualization
Generate plots showing outlier detection results and distribution of outlier scores.
- show_interpretation
Display detailed interpretation of outlier detection results and methodological notes.
- sampleThreshold
Row count above which a random subsample is analysed instead of the full dataset. Datasets at or below this size are always analysed in full. Raise it to analyse a large dataset completely at the cost of speed.
- sampleSize
Number of rows retained when subsampling. Outliers among the rows that are not sampled cannot be detected, so the reported count is a lower bound; larger values recover more of them at the cost of speed.
- seed
Random seed used for the reproducible random subsample that is drawn when the dataset exceeds the subsampling threshold. Change this value to draw a different sample; the default of 123 reproduces the previous fixed behaviour.