Scatterplots suit the relation between two continuous variables

What it is

A scatterplot maps two continuous variables measured on the same observational unit to horizontal and vertical position, preserving one paired record per point. It supports questions about direction, strength, shape, outliers, and clusters—not merely the value of a linear correlation coefficient. Meaningful pairing is the first requirement: one participant's completion time and error count may form a point, whereas aggregates from different people or time windows do not become a valid pair because they share a row number.

Why it happens

Two position channels share a Cartesian frame, allowing readers to inspect individual records and ensemble structure. Position does not force the relation to be linear, so curvature, heteroscedasticity, multiple subgroups, and high-leverage points can remain visible. Axis limits, scales, and transformations nevertheless change distances and cloud shape. A logarithmic transform may straighten a multiplicative relation and stabilize spread, but it also changes what a unit difference means. A plotted position may reflect underlying differences, process or sampling variation, and measurement error on either axis; error magnitude and correlation can alter cloud shape and fitted relations, so a narrow band is not evidence of error-free measurement.

Studying it

Use data with known generating processes to manipulate correlation, nonlinearity, outliers, clusters, sample size, and measurement error, then measure accuracy, response time, and confidence for direction, shape, and anomaly judgments. Anscombe presented four datasets with nearly identical common summaries and linear regressions but very different point-cloud structures, showing why graphical inspection should complement rather than replace statistical summaries. In realistic tasks, define the observational unit, join key, time window, missing-data treatment, and transformation before comparing chart judgments with numerical summaries, model diagnostics, and domain interpretation. When testing linear against log scales, test whether readers understand the units as well as whether one display looks cleaner.

Where it stops holding

If one axis is nominal, strips or grouped comparisons may better match the task; if temporal sequence and path are central, a line chart may fit better. With few observations, accidental points dominate apparent shape; with many, overplotting erases density. Adding further variables through color, shape, and size increases decoding cost. Association also cannot identify causal direction. A screen reader cannot reconstruct paired values from plotted pixels, so an image plus hover-only tooltips is not equivalent access.

Applying it

• State the question, observational unit, and pairing key before assigning two meaningful continuous measures to axes with units; disclose missingness, filters, and transformations.
• Inspect the original scale and any theoretically justified transformed scale. Label log scales explicitly, and never silently alter zero or negative values to make them fit.
• Keep the roles of raw points, trend summaries, and uncertainty distinct. If a fit is shown, identify its model form and interval without hiding the raw structure.
• Provide a keyboard-accessible data table, sample size, ranges, and a text summary of direction, shape, clusters, outliers, and known limitations.
• Test whether intended readers find the target patterns and can explain what each point represents. A failed pairing explanation blocks any interpretation of correlation.

Related

• Same group: U2.05.2 Many points overplot; use transparency or binning · U2.05.3 Correlation is not causation
• Nearby: U2.06.1 Bubbles carry a third variable on the least precise common channel · U2.14.2 Only adjacent-axis relations read directly
• Search terms: scatterplot · bivariate relationship · paired observations · measurement error · log transform