Visualizing a correlation with a third variable

WHY IT MATTERS A correlation between two continuous measures is best understood visually before you make claims, and a strong bivariate relationship can be entirely reshaped by a hidden third variable.

WHAT IT TESTS Choosing the correct chart for two continuous variables and showing maturity about confounders and causality.

A GOOD ANSWER COVERS For two continuous variables the natural choice is a scatter plot, with load time on the x-axis and session duration on the y-axis. Each point is a session or a user. Add a fitted trend or regression line to make the negative slope explicit, and consider a density or hexbin overlay if there are too many points to read. To probe whether network type drives the relationship, introduce it as a third encoding: color the points by WiFi versus Cellular and fit a separate trend line per group, or use faceting with small multiples, one scatter per network type, so slopes can be compared side by side. A bubble chart could add a fourth variable through point size if needed. The key insight is that comparing within-group slopes can reveal Simpson's paradox, where the pooled trend masks or even reverses the within-segment trend, for example if cellular users simply have both slower loads and shorter sessions for unrelated reasons.

WHEN IT MATTERS Use color or faceting whenever a categorical covariate might confound a continuous relationship; reach for it before asserting any causal story.

COMMON WRONG ANSWERS A bar chart, which collapses the point-level relationship. Asserting load time causes shorter sessions without controlling for confounds. Cramming too many encodings into one unreadable chart. Forgetting to add a trend line, leaving the relationship to the eye.

ONE CONCRETE EXAMPLE A scatter shows session duration falling as load time rises. Coloring by network reveals two clouds: within both WiFi and cellular the slope is mild, but cellular sessions cluster at high load and low duration, so the steep pooled slope was partly driven by network type, prompting a deeper, segment-aware analysis rather than a premature causal claim.