Reading Scatter Plots and Correlation

Read the scatter plot before the coefficient

A scatter plot displays one point for each paired observation. It lets you ask questions a single correlation coefficient cannot answer: Is the pattern roughly straight? Is it driven by one point? Are there clusters? Does the spread change across X?

This lesson focuses on graphical interpretation. For the calculation and inference, continue to Pearson correlation.

Four features to describe

1. Direction: as X increases, does Y generally increase, decrease, or show neither tendency?
2. Form: is the pattern roughly linear, curved, or separated into groups?
3. Strength: how tightly do points cluster around the relevant pattern?
4. Unusual observations: are some points far from the pattern or isolated along X?

Include the variables’ names and units. “Positive relationship between study hours and score” is more informative than “r is positive.”

A linear pattern

Pairs (1, 3), (2, 5), (3, 7), (4, 9), (5, 11) lie on Y = 1 + 2X. Their Pearson correlation is 1. That perfect coefficient says the observed points form an increasing straight line; it does not imply one variable causes the other or that a new observation must lie on the same line.

A few small deviations from the line reduce |r|. Compare the point cloud with the pattern that matters for your question, rather than applying universal “weak/moderate/strong” categories.

A nonlinear pattern with zero correlation

Pairs (−2, 4), (−1, 1), (0, 0), (1, 1), (2, 4) form a U shape. Their Pearson r is 0 because positive and negative linear tendencies cancel.

There is a perfect relationship, Y = X², but no linear trend across this symmetric range. “No correlation” is too broad; “no measured linear association” is the accurate statement.

An influential observation

Four pairs are (1, 1), (2, 2), (3, 3), (4, 4), with r = 1. Add (5, −10) and r becomes about −0.5547. One added observation changes both direction and magnitude.

Investigate it: data entry, a different population, a valid extreme case, or a mechanism the simple pattern missed? Show analyses with and without an observation only when the comparison is justified and documented. Do not delete a point to restore the relationship you expected.

Clusters and pooled data

Suppose one branch serves mostly small orders with short waits, while another serves large orders with long waits. A pooled scatter plot may show a strong order-size/wait relationship even if the relationship within each branch is weak.

Plot relevant groups with distinct labels or separate panels. Differences between groups and associations within groups answer different questions. This is one reason confounding and Simpson’s paradox matter.

Preserve meaningful scales and pairing

• Label axes with units and define what each point represents.
• Do not sort X and Y independently.
• Use transparency or jitter carefully when observations overlap, and disclose any jitter.
• Avoid connecting unrelated observations with lines; that suggests a sequence that may not exist.
• Use a log axis only when it serves the question, and label it clearly.

A tight-looking pattern can be affected by the axis ranges. A numeric coefficient and a plot provide complementary evidence.

Practice

A scatter plot curves upward, with Y increasing slowly at low X and quickly at high X. Pearson r is 0.90. Is a straight regression line automatically adequate?

Show the explanation

No. A high r can coexist with systematic curvature. Inspect a residual plot after fitting; a curved residual pattern suggests the mean relationship is not adequately represented by a straight line. Consider a model or transformation justified by the setting and validate it.

Next steps

Use the correlation calculator to compare the linear, U-shaped, and influential-point examples. Learn the Pearson calculation, then linear regression and regression diagnostics.