Course contentsShow
← Data Science
Lesson 481 of 2,14513. Correlation AnalysisFree lesson

Outliers and Their Impact on r

How extreme values can inflate or deflate correlation coefficients and methods to identify influential points.

Outliers and Their Impact on r

You'll learn how a single extreme data point can dramatically alter your correlation coefficient — and how to catch these troublemakers.

Why Outliers Matter for Correlation

Pearson's correlation coefficient (r) is calculated using every point in your dataset, and it's surprisingly sensitive to extreme values. Just one outlier can inflate r (making variables look more related than they truly are) or deflate it (hiding a real relationship).

Think of r like a seesaw: most of your data creates a stable balance, but one heavy person sitting far from the center can tip the entire thing. That heavy person is your outlier.

How Outliers Distort r

Inflating r: Imagine plotting height vs. weight for 20 adults, showing a moderate correlation. Now add one data point: a 7-foot basketball player weighing 280 pounds. This extreme point often pulls the correlation line toward itself, artificially strengthening r.

Deflating r: Conversely, an outlier that doesn't follow the general pattern (like an extremely tall person who weighs very little due to illness) can weaken an otherwise strong correlation by pulling the line away from the main cluster.

Identifying Influential Points

Since you've already learned about scatterplots, use them to spot outliers visually. Look for points that:

  • Sit far from the main cloud of data
  • Don't follow the general trend
  • Appear isolated in the plot's corners

Once spotted, ask: "Does this point represent a genuine observation or a data entry error?" Calculate r with and without the suspicious point. A dramatic change (say, r = 0.85 vs. r = 0.62) signals that point is highly influential.

Key Takeaway: Outliers can severely distort Pearson's r by disproportionately pulling the correlation coefficient up or down; always examine scatterplots and consider calculating r with and without suspicious points to assess their influence.