Is a scatterplot matrix the same as a correlation matrix?

No. One shows paired observations visually; the other usually shows numerical correlation summaries.

How many unique pairs are there for k variables?

There are k(k−1)/2 distinct unordered pairs. A full matrix displays each pair twice outside the diagonal.

Can categorical variables be included?

They can be encoded or handled with other plot types, but ordinary scatterplots may be less informative than displays designed for categories.

AI Glossary / Mathematics and optimization / Beginner

Scatterplot matrix

Definition

A scatterplot matrix is a grid of plots showing pairwise relationships among several variables. Each off-diagonal panel compares one variable with another, allowing a reader to inspect patterns, outliers and groups across a dataset in one coordinated display.

What it means

A single scatterplot shows two variables. When a dataset has several numerical variables, examining them one pair at a time can become cumbersome. A scatterplot matrix arranges those comparisons systematically, with a variable assigned to each row and column.

In a common convention, the row variable is on the vertical axis and the column variable is on the horizontal axis. Panels on opposite sides of the diagonal contain the same pair with the axes reversed. Some displays keep only one triangle to avoid duplication.

The diagonal often contains variable names, histograms or density plots instead of a variable plotted against itself. This combines views of individual distributions with views of relationships. The display is exploratory: a visible association suggests a question to investigate, not a causal conclusion.

How it works

Choose a manageable set of variables measured on corresponding records. For each pair, plot the two values from each record as one point. Keep labels, units and axis conventions clear so readers can follow a variable across its row or column.

Look for straight trends, curves, clusters, changes in spread and isolated points. A relationship can be nonlinear even when a linear correlation coefficient is near zero. Coloring by a known group can reveal differences hidden in the combined data, provided the color legend is meaningful.

Large datasets create overplotting. Transparency, smaller markers, sampling or density-based panels can help, but each changes what the reader sees. Record any sampling or transformation. Missing values also require care: pairwise deletion can cause different panels to use different subsets of records.

Interactive linking highlights the same records across panels. Brushing lets a reader select a region in one plot and see where those records appear elsewhere. This supports investigation of a suspicious group without treating its visual distinctiveness as an explanation.

Three practical examples

A maintenance analyst compares temperature, vibration and load to identify nonlinear patterns or unusual operating records.
A marketing research team explores relationships among numerical campaign measurements before choosing a regression specification.
A data-quality reviewer highlights suspicious observations in one panel and traces them across other variable pairs.

Think of it this way

Think of viewing the same collection of objects from several paired perspectives on a contact sheet. One view may reveal a pattern that another hides.

The analogy stops short of a complete three-dimensional reconstruction. Pairwise views can miss relationships that depend on several variables together, and the arrangement of dots does not explain why a relationship exists.

Worked example

A curve hidden by zero correlation

Create five teaching records with X values −2, −1, 0, 1 and 2. Let Y=X², giving Y values 4,1,0,1,4. Let Z=X+10, giving Z values 8,9,10,11,12.

A three-variable scatterplot matrix has nine cells. Six off-diagonal cells represent three unique pairs, each displayed in both axis orders. The diagonal can show the three individual distributions.

The Y-versus-X panel forms a U-shaped pattern. The mean of X is zero, the mean of Y is two, and the sum of products X(Y−2) is −4+1+0−1+4=0. Their sample covariance and Pearson correlation are therefore zero, even though Y is completely determined by X in these records.

The Z-versus-X panel is a straight line with slope one, while Y versus Z has the same quadratic relationship shifted horizontally. The display reveals structure that a table containing only linear correlations would obscure. These five constructed points illustrate reading the display; they are not sufficient evidence for a general real-world model.

Illustrative case study

A fictional equipment supplier receives a dataset from a pilot installation. A correlation table shows little linear association between a centered control setting and a quality measurement, so an early review proposes dropping the setting.

An analyst builds a scatterplot matrix and notices a curved relationship. Coloring by machine identifier also exposes a small cluster that merits investigation. The team checks the raw measurements, operating ranges and preprocessing before fitting candidate models.

The display changes the questions the team asks. It does not by itself prove that changing the control setting will improve quality or that the highlighted machine is faulty.

Common misconceptions and mistakes

A scattered cloud proves independence. Pairwise plots can miss complex or higher-order dependence.
Zero correlation means no relationship. The worked quadratic example disproves this.
Every apparent cluster is a meaningful population. Rounding, filtering and plotting choices can create patterns.
More variables always improve the display. The number of panels grows quickly and can overwhelm readers.
A trend establishes causation. Confounding, selection effects and other mechanisms may explain it.

Business use

Scatterplot matrices are useful early in analysis, when a team needs to understand data structure before selecting a model or making strong claims. They help generate targeted checks for outliers, transformations and possible interactions.

Start with variables relevant to the decision, then inspect important panels more closely. Report axis units, transformations, group encodings and the number of records used. Where privacy or confidentiality matters, use appropriate aggregation or access controls; simply shrinking points on a page does not remove the underlying records' sensitivity.

Check your understanding

A four-variable scatterplot matrix has 16 cells. How many unique off-diagonal variable pairs does it represent?

Answer: Six, because 4×3/2=6. A full display shows each pair twice with reversed axes, plus four diagonal cells.

Related terms

Learn next

Read correlation and exploratory data analysis. Plot the five worked records by hand, then explain why a single straight-line correlation coefficient cannot describe the U-shaped panel adequately.

Sources

NIST/SEMATECH e-Handbook, Scatterplot Matrix: https://www.itl.nist.gov/div898/handbook/eda/section3/scatterb.htm. Official treatment of panel layout, exploratory questions, linking and brushing. The five-record example, covariance check and fictional equipment scenario are original teaching material.

Previous: Scatter plot Next: Scatterplot smoothing