An outlier is a measurement unusually far from the other observations. In a two-parameter response-surface experiment, one markout-profit measurement at threshold 0.5 and order_size 3.5 was much lower than the rest, making a simple quadratic model unreliable. section 4.2.2
Treat outliers as a model-and-experiment diagnostic:
- Do not rely on a universal cutoff; available rules require a threshold chosen by the analyst. section 4.2.2
- Inspect measurements visually when using a manual response-surface method, while recognizing that automatic bad-point detection is possible but deferred. section 4.2.2
- Shrink the parameter ranges around the current center and iterate when an outlier suggests that the fitted function is too complex. section 4.2.2
