Scatterplots & Association — Scatterplots and Association (Direction, Form, Strength) - College Statistics

SECTION 1

Historical Context & Motivation

The desire to visualize relationships between two quantitative variables is deeply rooted in the history of statistics and data analysis. Long before digital computing allowed instantaneous plotting, scientists and social reformers recognized that pairing observations on two dimensions could reveal patterns invisible in univariate summaries alone. The scatterplot — a graph that places one variable on the horizontal axis and another on the vertical axis, plotting each case as a single point — became the foundational tool for exploring bivariate association. Understanding how this tool emerged helps us appreciate why characterizing association through direction, form, and strength remains one of the first skills taught in any statistics course.

1686 • First Known Scatterplot

Edmund Halley plotted barometric pressure against altitude, producing what many historians consider the earliest known scatterplot. His goal was to model the atmosphere, but the visual representation itself became a lasting contribution to data display.

1833 • Statistical Graphics for Social Science

André-Michel Guerry published scatterplot-like displays comparing crime rates and educational attainment across French departments, demonstrating that bivariate plots could illuminate social phenomena and challenge prevailing assumptions about moral statistics.

1886 • Galton's Regression Framework

Francis Galton used scatterplots of parents' and children's heights to develop the concept of regression toward the mean, laying the groundwork for the formal study of association, correlation, and linear modeling.

1896 • Pearson's Correlation Coefficient

Karl Pearson formalized the product-moment correlation coefficient r, giving statisticians a single numerical measure to quantify the strength and direction of a linear relationship — a quantity best interpreted alongside a scatterplot.

1977 • Tukey and Exploratory Data Analysis

John Tukey's seminal text Exploratory Data Analysis championed scatterplots as essential first steps in any analysis, arguing that visual inspection of form, direction, and strength should precede formal inference. Modern statistical software made this advice universally actionable.

The central question that scatterplots address is deceptively simple: When one variable changes, does the other tend to change as well — and if so, in what manner? Answering this question requires a vocabulary for describing the direction of the trend (positive or negative), the form of the pattern (linear, curved, clustered), and the strength with which the points adhere to that pattern. These three descriptors, together with the identification of outliers, constitute the complete framework for qualitative scatterplot interpretation that we develop in this lesson.

SECTION 2

Core Principles & Definitions

Before diving into visual or numerical analysis, it is essential to establish the conceptual vocabulary that statisticians use when discussing bivariate relationships. A scatterplot encodes each observational unit as a point (xᵢ, yᵢ) in Cartesian space, where x is conventionally the explanatory (independent) variable and y is the response (dependent) variable. When neither variable has a clear explanatory role, the assignment to axes is arbitrary, but consistency matters for communication. The four aspects of any scatterplot that must be described are captured in the following principles.

Direction

A positive association occurs when increases in x tend to accompany increases in y. A negative association occurs when increases in x tend to accompany decreases in y. If no discernible trend exists, the variables show no association.

Form

The overall shape of the cloud of points may be linear (points cluster around a straight line), curvilinear (points follow a curve such as a parabola or exponential), or display no clear functional form. Identifying form guides model selection.

Strength

Strength describes how tightly the points cluster around the underlying pattern. A strong association means little scatter about the trend; a weak association means the points deviate substantially. For linear relationships, Pearson's r quantifies this.

Outliers & Unusual Features

Individual points that deviate markedly from the overall pattern are outliers. Clusters, gaps, or subgroups in the data are also notable features. Outliers can heavily influence summary statistics like r and the least-squares regression line.

✦

KEY TAKEAWAY

Think of a scatterplot as a weather radar image for data. Direction tells you whether the 'storm' is moving left or right; form tells you whether it follows a straight highway or a winding mountain road; strength tells you how tightly the raindrops cluster along that road. Just as a meteorologist checks the radar before issuing a forecast, a statistician inspects the scatterplot before fitting any model — skipping this step is like forecasting weather without looking outside.

SECTION 3

Visual Explanation — Reading a Scatterplot

The diagram below presents three scatterplots side by side, each illustrating a different combination of direction, form, and strength. By comparing these panels, you can train your eye to classify association at a glance — a skill that becomes second nature with practice but requires deliberate attention when first learning.

Three panels illustrating contrasting associations. Left: strong positive linear relationship (r ≈ 0.98) — points tightly follow an upward-sloping line. Center: weak negative linear relationship (r ≈ −0.45) — a downward trend is visible but scatter is substantial. Right: strong nonlinear (quadratic) relationship with r ≈ 0.02 — Pearson's r is near zero despite a clear pattern, underscoring that r measures only linear association.

In the left panel, the direction is clearly positive because y increases as x increases, the form is linear because the points align along a straight trend, and the strength is strong because each point lies close to the dashed reference line. The center panel reverses direction and weakens considerably: although the downward slope is perceptible, points scatter widely, producing a moderate-to-weak association. The right panel delivers the most important lesson of this section — a near-zero Pearson's r does not mean the variables are unrelated. The quadratic form means that x and y are strongly associated, but Pearson's r, which captures only the linear component, fails to detect it. This is why examining the scatterplot visually is an indispensable complement to computing numerical summaries.

⚠

Always Plot First

Anscombe's Quartet (1973) demonstrated that four datasets can share nearly identical means, variances, correlations, and regression lines while exhibiting wildly different scatterplot patterns — including one that is nonlinear and another dominated by a single outlier. The scatterplot reveals what summary statistics hide.

SECTION 4

Mathematical Framework

While the qualitative descriptors of direction, form, and strength are powerful, quantifying the linear component of a relationship demands a formal measure. The most widely used statistic for this purpose is Pearson's product-moment correlation coefficient, denoted r. Understanding its derivation illuminates why it captures direction and strength but is restricted to linear form.

PEARSON CORRELATION COEFFICIENT

r = (1 / (n − 1)) × Σ [(xᵢ − x̄) / sₓ] × [(yᵢ − ȳ) / sᵧ]

Where n is the sample size, x̄ and ȳ are the sample means, sₓ and sᵧ are the sample standard deviations, and the summation runs from i = 1 to n. Each factor in the product is a z-score, so r is the average product of standardized deviations.

The formula expresses r as the mean cross-product of z-scores. When both variables lie above their respective means (positive z-scores) or both lie below (negative z-scores), the product is positive, pulling r toward +1. When one is above and the other below, the product is negative, pulling r toward −1. If positive and negative products cancel, r approaches 0. This algebraic structure explains why r detects linear direction and linear strength but is insensitive to curvature: in a symmetric U-shape, positive and negative products cancel even though the relationship is deterministic.

EQUIVALENT COMPUTATIONAL FORM

r = [n Σxᵢyᵢ − (Σxᵢ)(Σyᵢ)] / √{[n Σxᵢ² − (Σxᵢ)²][n Σyᵢ² − (Σyᵢ)²]}

This algebraically equivalent form avoids computing z-scores individually and is sometimes more convenient for hand calculation or understanding the relationship between r and the least-squares regression slope.

COEFFICIENT OF DETERMINATION

r² = 1 − (SSresid / SStotal)

r² (r-squared) represents the proportion of variance in y explained by the linear relationship with x. For instance, r = 0.80 yields r² = 0.64, meaning 64% of the variability in y is accounted for by the linear model.

ℹ

Properties of r

The correlation coefficient r is bounded: −1 ≤ r ≤ 1. It is unitless and symmetric (r(x, y) = r(y, x)). Adding a constant or multiplying by a positive constant does not change r, because z-scores absorb location and scale. Multiplying by a negative constant reverses the sign. These invariance properties make r a pure measure of linear association, free from the units of the original variables.

SECTION 5

Classifying Association — A Detailed Breakdown

In practice, one of the most important skills is translating the visual impression of a scatterplot into precise descriptive language and then connecting that description to the appropriate numerical summary. The diagram below organizes the most common association patterns into a taxonomy that distinguishes among direction, form, and strength, providing reference examples for each category.

A six-panel taxonomy of bivariate association. The top row covers linear associations: strong positive, moderate negative, and no association. The bottom row illustrates nonlinear forms (quadratic and exponential) and the effect of a single outlier on an otherwise strong linear trend.

Pattern	Direction	Form	Strength Indicator	r Reliable?
Strong positive linear	Positive	Linear	r ≈ +0.85 to +1.00	Yes
Moderate negative linear	Negative	Linear	r ≈ −0.50 to −0.70	Yes
No association	None	None	r ≈ 0.00	Yes (correctly reports none)
Quadratic	Mixed	Curvilinear	r near 0 despite strong pattern	No — misleading
Exponential	Positive	Curvilinear	r moderate but underestimates	Partially — log transform helps
Linear with outlier	Positive	Linear (+ outlier)	r distorted by outlier	Only after investigating outlier

Summary of association patterns and the reliability of Pearson's r for each

A common guideline for interpreting the magnitude of r in the social sciences (following Cohen, 1988) categorizes |r| < 0.30 as weak, 0.30 ≤ |r| < 0.70 as moderate, and |r| ≥ 0.70 as strong. However, these thresholds are discipline-dependent: in physics or engineering, r = 0.70 might signal a poor fit, whereas in psychology or education, it would be considered impressively strong. Context always matters when interpreting effect sizes.

SECTION 6

Worked Example — Computing and Interpreting r

Consider a dataset of n = 6 students recording the number of hours studied (x) and exam score (y). We will compute Pearson's r, interpret the scatterplot qualitatively, and connect the two perspectives.

Student	Hours (x)	Score (y)
A	2	55
B	4	62
C	5	70
D	7	78
E	9	85
F	10	92

Study hours and exam scores for six students

Computing Pearson's r Step by Step

1
Step 1 — Compute the means

x̄ = (2 + 4 + 5 + 7 + 9 + 10) / 6 = 37 / 6 ≈ 6.167. Similarly, ȳ = (55 + 62 + 70 + 78 + 85 + 92) / 6 = 442 / 6 ≈ 73.667.

x̄ ≈ 6.167, ȳ ≈ 73.667
2
Step 2 — Compute the standard deviations

sₓ = √[Σ(xᵢ − x̄)² / (n − 1)]. The squared deviations are (2 − 6.167)² = 17.36, (4 − 6.167)² = 4.69, (5 − 6.167)² = 1.36, (7 − 6.167)² = 0.69, (9 − 6.167)² = 8.03, (10 − 6.167)² = 14.69. Sum = 46.83, so sₓ = √(46.83/5) = √9.367 ≈ 3.061. By analogous calculation, sᵧ ≈ 13.707.

sₓ ≈ 3.061, sᵧ ≈ 13.707
3
Step 3 — Compute z-scores for each observation

For Student A: zₓ = (2 − 6.167) / 3.061 ≈ −1.361, $z_y$ = (55 − 73.667) / 13.707 ≈ −1.362. Product = 1.854. Repeat for all six students: products are 1.854, 0.551, 0.249, 0.164, 0.649, 1.548.

Cross-products: 1.854, 0.551, 0.249, 0.164, 0.649, 1.548
4
Step 4 — Average the cross-products

r = (1 / (n − 1)) × Σ(zₓ × $z_y$ ) = (1/5) × (1.854 + 0.551 + 0.249 + 0.164 + 0.649 + 1.548) = (1/5) × 5.015 ≈ 1.003. Due to rounding at intermediate steps, the exact computation yields r ≈ 0.993.

r ≈ 0.993
5
Step 5 — Interpret

With r ≈ 0.993, we describe the association as a strong positive linear relationship. Direction: positive (more hours → higher score). Form: linear (points closely follow a straight trend). Strength: very strong (|r| > 0.99). The coefficient of determination r² ≈ 0.986 indicates that approximately 98.6% of the variance in exam scores is explained by the linear relationship with study hours in this sample.

Strong positive linear association; r² ≈ 0.986

SECTION 7

Strengths, Limitations, and Common Pitfalls

Scatterplots and Pearson's r are among the most widely used tools in descriptive statistics, but they carry important limitations that, if overlooked, lead to misinterpretation. The table below contrasts the strengths of this framework with its known weaknesses, and the key takeaway afterward synthesizes the most critical caveat.

Strengths	Limitations
Scatterplots give an immediate, intuitive overview of bivariate data — no statistical training needed to see patterns.	Overplotting in large datasets can obscure patterns; transparency, jittering, or binning may be needed.
Pearson's r provides a single, unitless, standardized measure of linear association that is comparable across studies.	r captures only linear relationships; strong nonlinear associations may produce r ≈ 0, giving a false sense of independence.
r² has a clear proportional-reduction-in-error interpretation, making it accessible to non-technical audiences.	r is highly sensitive to outliers; a single influential point can inflate or deflate r substantially.
Scatterplots reveal outliers, clusters, and heteroscedasticity that numerical summaries alone conceal.	Correlation does not imply causation. Lurking variables, confounders, or reverse causation cannot be ruled out from association alone.
Scatterplots work naturally with continuous data and extend to multiple groups via color or faceting.	Restricted range on one variable attenuates r, making associations appear weaker than they are in the broader population.

Strengths and limitations of scatterplot analysis and Pearson's r

⚠️

CORRELATION ≠ CAUSATION

Perhaps the most misunderstood result in all of statistics: a strong correlation between x and y does not prove that x causes y. Consider the famous example of ice cream sales and drowning deaths — both rise in summer because both are driven by a lurking variable (temperature). The scatterplot would show a strong positive linear association, and r would be high, yet eating ice cream does not cause drowning. In research design, establishing causation typically requires randomized controlled experiments or sophisticated causal-inference frameworks such as instrumental variables or difference-in-differences, not just correlation.

SECTION 8

Connection to Regression and Advanced Methods

Scatterplot analysis and Pearson's r form the descriptive foundation upon which simple linear regression is built. In regression, we move from merely describing association to modeling it: fitting the line ŷ = b₀ + b₁x that minimizes the sum of squared residuals. The connection between correlation and regression is tight — the slope of the least-squares line is b₁ = r × (sᵧ / sₓ), and the proportion of variance explained by the model is exactly r². Understanding direction, form, and strength at the scatterplot stage prepares you to evaluate whether a linear model is appropriate before fitting one.

Feature	Scatterplot + Correlation	Simple Linear Regression
Goal	Describe the association between x and y	Predict y from x and quantify the relationship
Key statistic	r (direction and strength)	b₁ (slope), b₀ (intercept), R² = r²
Assumption	None required for description	Linearity, independence, normality, equal variance (LINE)
Handles nonlinearity?	Visually, yes; numerically (r), no	Only after transforming variables or using polynomial terms
Causation?	No — association only	No — still association; causation requires design

Comparison of descriptive correlation and inferential regression

Beyond simple linear regression, the ideas introduced here extend to multiple regression (where partial correlations describe associations after controlling for other predictors), Spearman's rank correlation (a nonparametric alternative that measures monotonic rather than strictly linear association), and nonlinear regression models. The critical habit to develop now — always plotting your data, characterizing direction, form, and strength before computing — will serve you well regardless of the complexity of the models you eventually fit.

SECTION 9

Practice Problems

Question 1: CONCEPTUAL

A researcher finds that the Pearson correlation between two variables is r = −0.02. Can we conclude that the two variables are unrelated? Explain your reasoning, referencing at least one scenario in which this conclusion would be incorrect.

Show answer

No, we cannot conclude that the variables are unrelated. Pearson's r measures only the strength and direction of the linear component of association. If the true relationship is curvilinear — for example, a symmetric U-shape (quadratic) — positive and negative cross-products of z-scores will cancel, yielding r ≈ 0 even though a strong, deterministic pattern exists. The correct procedure is to inspect the scatterplot: if the points trace a curve or other systematic pattern, the variables are clearly associated despite a near-zero r. Another scenario involves subgroup structure (Simpson's paradox), where different groups may each show a positive trend, but the combined data appear flat.

Question 2: BASIC CALCULATION

For the following five data points — (1, 10), (2, 14), (3, 18), (4, 22), (5, 26) — compute Pearson's r by hand. Describe the direction, form, and strength of the association.

Show answer

First, note that y = 6 + 4x exactly for each point, so the data fall perfectly on a straight line. Means: x̄ = 3, ȳ = 18. Standard deviations: sₓ = √(10/4) = √2.5 ≈ 1.581, sᵧ = √(160/4) = √40 ≈ 6.325. Each z-score product equals the square of the x z-score because the relationship is perfectly linear with a positive slope. Summing the products: (−1.265)(−1.265) + (−0.632)(−0.632) + 0 + (0.632)(0.632) + (1.265)(1.265) = 1.601 + 0.400 + 0 + 0.400 + 1.601 = 4.002. r = 4.002 / (5 − 1) ≈ 1.000. Direction: positive. Form: perfectly linear. Strength: perfect (r = 1.00). All points lie exactly on the line, so r² = 1 and 100% of the variance in y is explained by x.

Question 3: INTERMEDIATE

A scatterplot of hours of television watched per week (x) versus GPA (y) for a sample of 200 college students shows a roughly linear cloud sloping downward, with moderate scatter. The computed correlation is r = −0.58. Compute r², interpret it in context, and explain why we cannot conclude that watching more television causes lower GPAs.

Show answer

r² = (−0.58)² = 0.3364, or approximately 33.6%. This means that about 33.6% of the variability in GPA among these 200 students can be accounted for by the linear relationship with weekly television hours. The remaining 66.4% of GPA variability is due to other factors not captured by this bivariate relationship. We cannot conclude causation because this is an observational study, not a randomized experiment. Several confounders could explain the association: students with heavier course loads may have less free time for both television and studying, motivation levels could simultaneously reduce TV consumption and raise GPA, or socioeconomic factors could influence both variables. To establish causation, one would need to randomly assign students to different television-watching regimes and observe the effect on GPA, which is typically impractical and ethically questionable.

Question 4: APPLIED

An environmental scientist measures the average daily temperature (°C) and dissolved oxygen concentration (mg/L) in a lake over 12 months. The scatterplot reveals a strong, roughly linear, downward-sloping pattern with one point (August: 32°C, 4.2 mg/L) sitting well below the trend. Without the outlier, r = −0.94; with it, r = −0.88. Discuss how the outlier affects r and what steps the scientist should take before removing it.

Show answer

The outlier in August attenuates the magnitude of r from −0.94 to −0.88 because it deviates from the overall linear pattern, introducing additional residual variance that weakens the observed linear fit. However, the change is modest (Δr = 0.06), suggesting the outlier is influential but not dominant. Before removing this point, the scientist should investigate its legitimacy: was there a measurement error (e.g., malfunctioning sensor), or does August represent a genuine physical phenomenon such as algal bloom-driven oxygen depletion that is distinct from the temperature-oxygen mechanism? If the data point is valid and reflects a real process, removing it would misrepresent the data. A better approach might be to report r with and without the outlier, note the August anomaly in the narrative, and consider whether a nonlinear model (e.g., incorporating algal dynamics) better captures the full system. This example underscores that outlier handling requires scientific judgment, not just statistical rules.

Question 5: CRITICAL THINKING

Prove algebraically that Pearson's r is invariant under positive linear transformations — that is, if we define x* = a + bx and y* = c + dy (with b > 0 and d > 0), then r(x*, y*) = r(x, y). Briefly explain why this property matters for practical data analysis.

Show answer

Since r is computed from z-scores, we need to show that the z-score of x* equals the z-score of x. The mean of x* is x̄* = a + bx̄, and the standard deviation is

s_{x*}

= |b| × sₓ = b × sₓ (since b > 0). Therefore,

z_{x*}

= (x* − x̄*) /

s_{x*}

= ((a + bxᵢ) − (a + bx̄)) / (b × sₓ) = b(xᵢ − x̄) / (b × sₓ) = (xᵢ − x̄) / sₓ = zₓ. By identical reasoning,

z_{y*}

z_y

when d > 0. Since r = (1/(n−1)) × Σ(zₓ ×

z_y

), and neither z-score has changed, r(x*, y*) = r(x, y). If b < 0 (or d < 0), the z-score's sign flips, reversing the sign of r — hence the requirement that b, d > 0 for invariance rather than merely a sign change. This property matters practically because it means r is unaffected by changes in measurement units (e.g., Celsius to Fahrenheit, dollars to euros) or by centering/scaling operations, making it a truly standardized measure of linear association that allows cross-study comparison.

SUMMARY

Lesson Summary

Summary

A scatterplot displays the relationship between two quantitative variables by plotting each observation as a point in the Cartesian plane. When interpreting a scatterplot, always describe three features: direction (positive, negative, or none), form (linear, curvilinear, or no clear form), and strength (how tightly the points cluster around the underlying pattern). Additionally, note any outliers or unusual features such as clusters or gaps.

The Pearson correlation coefficient r quantifies the direction and strength of a linear association, ranging from −1 (perfect negative) to +1 (perfect positive), with 0 indicating no linear relationship. Its square, r², represents the proportion of variance in y explained by the linear model. Crucially, r is sensitive to outliers, blind to nonlinear patterns, and does not establish causation. Always inspect the scatterplot before relying on any numerical summary — a practice that connects this descriptive analysis to the more advanced regression models you will encounter next.