Correlation and scatterplots: inspect shape and hidden groups

MCAT content library / Research & data / Correlation and scatterplots: inspect shape and hidden groups

Interpret paired data by its shape, groups, and sampled range before summarizing it with a correlation coefficient.

About 14 minutes, including the worked examples and checks.

Useful foundations

• Reading graphs: axes, scales, and the comparison that matters
• Observational and experimental studies: what can we conclude?

On this page

• Concept: inspect the pairs before summarizing the relationship
• Application: a pooled pattern reverses within groups
• Pattern: zero linear correlation can hide a perfect curve
• Rule: distinguish strength from slope and influence from error
• Write the claim the plot actually supports
• Check your reasoning
• Continue learning
• Sources

Concept: inspect the pairs before summarizing the relationship

A scatterplot places one paired observation at each point. Before interpreting its shape, identify what each point represents, the units on both axes, and the range sampled. Then ask about direction, form, scatter, unusual points, and distinct groups. A cloud rising from left to right suggests a positive association over the observed range; a falling cloud suggests a negative one. Curvature can make a single straight-line description misleading.¹

The Pearson correlation coefficient, r, summarizes the direction and strength of a linear association. Its range is −1 to +1. Values close to either endpoint indicate a strong straight-line pattern; the sign distinguishes direction. A value near zero does not rule out a strong curved relationship. The coefficient is dimensionless, while a regression slope retains units. Neither a large correlation nor a visually persuasive trend by itself identifies a causal mechanism.²

Application: a pooled pattern reverses within groups

All datasets and scenarios here are original teaching examples. Six participants report practice sessions per week and complete a task scored in points. Three have daytime schedules and three have evening schedules. The simplified data make the comparison easy to inspect; they are not estimates of a real scheduling effect.

Schedule	Sessions per week	Task score (points)
Day	1	60
Day	2	65
Day	3	70
Evening	4	30
Evening	5	35
Evening	6	40

Looking only at all six points together, participants with more reported practice tend to have lower scores. The daytime group averages two sessions and 65 points; the evening group averages five sessions and 35 points. That between-group separation dominates the pooled pattern.

Within each schedule group, however, the observed relationship is positive: each additional session corresponds to five more score points in these constructed observations. The pooled negative association therefore cannot be treated as a description of the relationship within either group. The figure illustrates how aggregation can hide or reverse a subgroup pattern.

Do these data show that practice raises scores within a schedule group? No. Practice was reported rather than assigned, and the group-specific patterns still lack a design that rules out alternative explanations. Prior preparation, available resources, or reverse causation could matter. Do the data show that an evening schedule lowers scores? That also exceeds the design. Schedule labels distinguish groups; they do not identify the cause of the gap.

The appropriate next step depends on the question. To describe the observed associations, report both the pooled and group-specific patterns. To estimate a causal effect of practice, use the study design and background knowledge to determine which variables and comparisons can address confounding. Automatically adjusting for every measured characteristic can also mislead if a variable lies downstream of the exposure or selection process. Stratifying a plot is an informative diagnostic, not an automatic causal solution.⁵

Pattern: zero linear correlation can hide a perfect curve

Consider five pairs with x values −2, −1, 0, 1, and 2 and y values 4, 1, 0, 1, and 4. Every point satisfies y = x². The negative and positive x values have symmetric y values, so the upward and downward contributions to the linear covariance cancel. Pearson r is zero even though y is completely determined by x in this constructed dataset.

A straight line loses the central feature: y decreases toward x = 0 and then increases. Inspecting the scatterplot or the paired table makes the relationship visible immediately. A model with a squared term can represent the shape, but the observed relationship still supports prediction only within an appropriate range and context. The pattern does not justify extrapolating indefinitely or assigning a causal interpretation. NIST uses curved scatterplot examples to show why form must be inspected before selecting a linear model.³

Rule: distinguish strength from slope and influence from error

A steep line is not necessarily a stronger association than a shallow one. Changing an x-axis from hours to minutes makes the numerical slope smaller without changing how tightly the points follow their line. For a positive linear unit conversion, Pearson r is unchanged. By contrast, a negative rescaling reverses direction, so its sign changes. Keep the definition of the transformed variable explicit.

An unusual point can strongly influence a fitted line or a correlation, especially when its x value lies far from the rest. Influence does not prove that the point is a mistake. Check the measurement record and whether the observation belongs to the intended population. Compare results with and without it when useful, disclose the comparison, and apply a justified rule rather than deleting a point because the conclusion becomes more attractive.

Restriction of range creates another interpretive limit. A relationship seen across a broad range may look weak when only a narrow subset is sampled. For example, selecting participants with nearly identical practice frequency reduces the observed x variation. The resulting correlation describes that selected range; it does not establish that practice and performance are unrelated everywhere.

Write the claim the plot actually supports

A defensible interpretation identifies the variables, direction, form, groups, and sampled range. For the practice example: “The pooled sample has a negative association, while both observed schedule groups have positive within-group associations.” That is more informative than “practice hurts performance,” which introduces a causal conclusion and ignores the grouping structure.

When an answer choice says one variable causes another, look beyond the plot to the assignment, timing, measurement, and alternative explanations. When it says no relationship exists because r is small, inspect the shape. When it uses a group average to describe every individual, recover the actual paired observations. These checks turn a visual impression into a bounded inference.

Check your reasoning

Choose an answer before opening its feedback. Each option explains the reasoning behind that choice.

Continue learning

Related lessons

• Mean, median, and spread: describing a dataset
• Sampling and generalizability: trace who enters the estimate

Browse research & data · MCAT content library

MCAT preparation and roadmap · Aim High, Score Higher

Sources

1. NIST/SEMATECH — Scatter Plot

2. OpenStax Statistics — The Regression Equation

3. NIST/SEMATECH — Scatter Plot: Quadratic Relationship

4. AAMC — What is on the MCAT exam?

5. Rohrer (2018) — Thinking Clearly About Correlations and Causation