Scatterplots and Relationships

Scatterplots answer how two quantitative measures vary together. Position carries the primary evidence. Color, radius, a fitted line, or chronological connections should add one clearly stated dimension rather than compete with that relationship.

Choose the comparison#

Do not infer causation from proximity or a fitted trend. Show the model and preparation only when they answer the stated question.

Add a prepared regression#

A fitted line is derived data. Calculate its coefficients in the application, then render the raw observations and fitted endpoints as independent layers.

This separation lets the dot layer preserve every observation while the line layer consumes only the model output. See Marks and Layering for composition and Scales and D3 for the application-owned statistics boundary.

Connect observations only when order matters#

A connected scatterplot turns sequence into a path through two-dimensional measure space. Chronological labels and direction arrows make that additional ordering visible.

Without an explicit order, connecting points invents a relationship. Keep the path, arrow, selected labels, and points as separate layers so each can use the same scales without sharing renderer-specific state.

Compare each observation with its predecessor#

A lag plot moves time out of the axis and into data preparation. Each point pairs a current value with the previous value; an identity rule shows where those values would be equal.

Make the lag length explicit and decide how the first observation is handled. The chart should receive the resulting pairs rather than conceal the shift inside a mark.

Use spatial focus for dense points#

Nearest-point interaction should use two-dimensional distance when both axes matter. A spatial index can make that lookup efficient without changing the dot grammar or storing tooltip state inside the mark.

Tooltips and Focus defines the focus and formatting model. Use Interactions and Selections when focused data also changes application state.

Production checks#

• Use quantitative scales with intentional domains on both axes. Use a logarithmic scale only when multiplicative distance is the intended reading; see Scales and D3.
• Map magnitude through an area-preserving radial scale when point size carries a third quantitative value.
• Control opacity or aggregate spatially before thousands of overlapping dots obscure the distribution. See Large Data.
• Keep regression, lag pairs, and spatial indexes in data preparation or interaction capabilities rather than the dot mark.
• Provide keyboard-equivalent focus and a textual value path, as described in Accessibility.

The channel and styling contracts for points are in Dot and Hexagon Marks.