Cook's Distance
You'll learn how to measure a point's overall influence on your regression model by combining leverage and residual size.
What Is Cook's Distance?
Cook's Distance (Cook's D) is a single metric that answers: "If I remove this one observation and refit the model, how much will my predictions change everywhere?"
You've already learned about leverage (how extreme a point's X-value is) and residuals (how badly the model predicts a point). Cook's D combines both: a point is influential if it's far from the mean in X-space and poorly predicted.
The Intuition
Think of it like identifying troublemakers in a tug-of-war:
- A person pulling from the far end of the rope (high leverage) has mechanical advantage
- If they're also pulling in the wrong direction (large residual), they throw everyone off balance
- Cook's D measures both their position advantage and how misaligned their pull is
A high Cook's D means: "This point is yanking the regression line away from where it would naturally go."
Why It Matters
Rather than checking leverage and residuals separately, Cook's D gives you a single diagnostic number per observation. Points with Cook's D > 0.5 or 1.0 (common thresholds) deserve investigation—they might be data entry errors, outliers, or legitimate extreme cases that need special handling.
Unlike just looking at residuals, Cook's D tells you which points are changing your model's coefficients and predictions throughout the entire dataset, not just fitting poorly themselves.
Key Takeaway: Cook's Distance combines leverage and residual magnitude into one number that identifies observations exerting disproportionate influence on your regression model—high Cook's D points are "pulling" the model away from the pattern formed by other data.