Scatter Plot
A scatter plot represents the relationship between two numerical variables by plotting each observation as a single point, with one variable's value determining its horizontal position and the other its vertical position — advertising spend against revenue, hours studied against exam score, height against weight. It's the standard chart for visually assessing whether and how two numerical variables relate to each other, revealing not just whether a relationship exists but its shape (linear or curved), its strength (tightly clustered points versus a loose, scattered cloud), and its direction (positive, negative, or none) all in a single view.
A scatter plot is the single most important diagnostic tool for catching exactly the kind of structure a correlation coefficient alone can hide, which is precisely the lesson Anscombe's quartet — four datasets sharing nearly identical correlation coefficients despite looking completely different when actually plotted — is most commonly used to teach: a correlation coefficient is a single number that assumes a linear relationship, and it can report a moderate or even near-zero correlation on data that actually has a strong, obvious, but curved (non-linear) relationship, a pattern invisible in the number but immediately visible in a scatter plot. This is why checking a scatter plot before trusting a reported correlation value is treated as a required step, not an optional one, in careful statistical practice — the correlation coefficient and the scatter plot answer related but genuinely different questions, and relying on the number alone risks missing exactly the kind of real relationship the number isn't built to detect.
A single outlier's influence on a scatter plot's apparent relationship is often visually obvious in a way it wouldn't be from summary statistics alone — a single extreme point positioned far from the main cluster can visibly pull a fitted regression line toward itself, and spotting that single influential point by eye is frequently the first and fastest way an analyst catches a data quality problem or an unusually influential observation before it has a chance to distort a formal regression model built on the same data. This is also where a scatter plot reveals distinct subgroups mixed within what initially looks like a single dataset: two visually separate clusters of points, each following its own internal pattern, can combine to produce a single, misleading overall trend line when treated as one dataset — exactly the kind of structure Simpson's paradox describes, and a scatter plot is often the fastest way to actually see it happening rather than just infer it from aggregated statistics.
Overplotting — where a large number of data points overlap so heavily that individual points become indistinguishable, obscuring the true density of the data — becomes a genuine practical problem as dataset size grows into the thousands or millions of points, addressed through techniques like transparency (making points semi-transparent so overlapping regions appear visually darker, revealing density), jittering (adding small random offsets to separate points that would otherwise land on exactly the same position, common with discrete or rounded data), or switching entirely to a density-based visualization like a 2D histogram or hexbin plot once the point count makes a traditional scatter plot too cluttered to usefully interpret.
A fitted trend line (commonly a linear regression line) is often overlaid directly on a scatter plot specifically to make the overall direction and strength of a relationship more explicit, but adding one prematurely — before actually looking at the raw scatter of points — risks imposing an assumed linear shape onto a relationship that might genuinely be curved, which is exactly the trap a scatter plot, examined on its own first, is meant to help an analyst avoid.
Read a scatter plot as the primary tool for seeing a relationship between two numerical variables rather than just computing a number that summarizes it: Anscombe's quartet remains the standard demonstration of why a correlation coefficient alone is an insufficient substitute for actually looking at the data, outliers and hidden subgroups are often visually obvious in a scatter plot well before they'd be caught through summary statistics, and techniques like transparency and jittering exist specifically to keep the chart interpretable once the sheer number of data points threatens to overwhelm its basic visual logic.