Outlier Test: A Scientific Judgment of Abnormal Data
The Principle of Outlier Testing
Statistical methods determine whether abnormal values exist in the data: Grubbs' test is based on deviation from the mean (assuming normality), Dixon's test on ratios of ranges (small samples, no normality assumption), and the IQR rule on the interquartile range (1.5 x IQR) for a quick screen. The tool outputs test statistics and p-values, allowing a scientific decision rather than deleting data by feel. The methods complement each other across normal and non-normal data scenarios, and because statistics and p-values are both reported, conclusions are quantitative and verifiable.
When to Use It
Use it to screen for data-entry errors; as a data-quality check before capability analysis, control charts and regression; to judge whether an unusual point is worth investigating; and as pre-processing before merging multiple data sets. Outlier testing is a quality gate before data analysis, preventing unusual points from distorting downstream statistical conclusions. It upgrades outlier judgment from experience-based to test-based, and it is recommended to make outlier testing a standard step in the data-analysis workflow.
How to Read the Results
A p-value below 0.05 (the default alpha) identifies the point as an outlier; the tool flags each suspicious point with its test statistic. Before removing a flagged point, verify its source (data-entry error, special cause) and then decide. AI interpretation explains the deviation level of each suspicious point. Decide to keep or remove only after verification, so that arbitrary deletion does not change the analysis conclusions.
Precautions
Outlier tests assume the data (apart from the outliers) come from a single population; multiple testing inflates false positives, so be cautious with iterative removal; a significant outlier is not a mandate to delete it, so find the cause first; Grubbs' test is unreliable for heavily skewed data, so use the IQR rule or transform the data first. After removing outliers, re-analyze and record the treatment process in the report; prefer robust methods for skewed data and control the overall false-positive risk in multiple testing.