Recognizing Clusters in Scatterplots

Why Look for Clusters?

A scatterplot can show more than a single overall relationship. Its points may form two or more relatively concentrated clouds with visible gaps between them. Those clusters can be clues that the observations come from different sites, cohorts, or other subgroups. If we describe only the full collection of points, we may miss how its pieces are arranged.

Recognizing clusters is a visual observation, not proof that the data contain distinct populations or that a particular subgroup caused the pattern. The plot shows where observations lie; labels and knowledge of how the data were collected help determine whether the separation has a meaningful explanation.

Reading the Shape and Separation

To assess a possible cluster, inspect the plotted points across both axes. Ask whether several observations lie near one another and whether a visible gap separates them from another concentration of points. Describe where each cloud lies—for example, lower or higher on the predictor axis, the response axis, or both. Naming the axes and units makes the description specific.

A cluster need not be round, equally sized, or equally dense. One cloud may be wider than another, and the groups may be separated mainly horizontally, mainly vertically, or in both directions. A small gap in a sparse plot may also be less convincing than a clear separation between dense clouds. These judgments depend on the display, so axis scales and limits must not distort the apparent distances.

Before interpreting separation, check that the plot uses the intended observations and that the axes, units, and measurement definitions are comparable across them. A plotting or data-coding error can create an apparent split. If group labels are available, use them to see whether points in different clouds share a recorded feature such as site or cohort. A label can help explain the layout, but it does not by itself establish why the groups differ.

Clusters and Pooled Summaries

When the observations have known group labels, a pooled summary combines observations across those groups. For example, the pooled mean of a response is the sum of all response values divided by the total number of observations. If each group has a different mean, the pooled mean is a weighted average: larger groups contribute more because they contain more observations. It need not describe either group particularly well.

For two groups, let their sample sizes be \(n_A\) and \(n_B\), and their response means be \(\bar{y}_A\) and \(\bar{y}_B\). The pooled mean \(\bar{y}\) is

This follows by adding the response totals. Group A's total is \(n_A\bar{y}_A\), and group B's total is \(n_B\bar{y}_B\); dividing their combined total by the combined sample size gives the formula. When the group means differ, the pooled mean lies closer to the mean of the larger group. If the group means are equal, the pooled mean equals both means.

What the Overall View Can Conceal

A pooled scatterplot places every observation together, which is useful for seeing the full data set. But it can hide the fact that points belong to separate clouds. Likewise, one pooled numerical summary cannot show the separate locations of those clouds. When a split is visible, describe it before summarizing the full set, and check whether recorded group information helps characterize it.

For instance, suppose points in one cloud come from one site and points in another come from a second site. A description of all points together might make the data sound like one uniform collection, even though observations are concentrated in two distinct regions. It is safer to state what is visible—such as “two separated clouds, with the first lower on both axes”—and then report what is known about group membership separately.

Checking Whether a Cluster Has an Explanation

Once a possible cluster is identified, look for context that could account for the separation. Relevant information might include site, cohort, measurement method, or a feature used to select participants. If group labels were recorded, color or shape can make them easier to compare with the visible clouds. The plot remains descriptive: matching a label to a cloud indicates an association between the label and the plotted location, not the reason for that association.

Also consider whether the apparent split is stable under an honest view of the data. A narrow axis range, misleading scale, or coding mistake may exaggerate or create separation. Conversely, a plot with many overlapping points can make concentrations harder to see. Clustering is a visual judgment, so report the observed pattern clearly rather than presenting a subjective boundary as a precise statistical fact.

Common Mistakes

• Calling every isolated point a cluster. A cluster is a cloud of multiple nearby observations, not simply a point that lies far from the rest.
• Assuming a cluster's cause. A separated cloud may line up with a site or cohort label, but the plot alone does not establish why it does.
• Describing only the pooled pattern. If the points form distinct clouds, state that structure rather than implying all observations are distributed uniformly.
• Misstating the pooled mean. With unequal group sizes, the pooled mean is closer to the larger group's mean only when the group means differ. If the group means are equal, the pooled mean equals both.
• Overstating a visual boundary. Clusters are recognized by visible concentrations and separation; there is no universal gap size that defines them.