Bivariate Quantitative Data

Bivariate quantitative data consists of paired numerical measurements

(x_{i}, y_{i})

on the same subjects, analyzed to detect and quantify relationships between two variables. The key result: a scatterplot reveals the form, direction, and strength of association, and the correlation coefficient

r

summarizes linear association while the least-squares regression line predicts one variable from the other. Used everywhere two measurements covary — economics, biology, medicine, engineering. Foundations laid by Galton and Pearson (1880s–1900s).

Bivariate quantitative data refers to a collection of ordered pairs of numerical values, each pair measured on a single individual or experimental unit, such as a person's height and weight or a car's speed and stopping distance. It matters because most interesting questions in science and society are about relationships — how does one quantity change as another changes? — and bivariate data is the simplest setting in which such relationships can be measured rigorously. The central tools are the scatterplot (a visual display), the correlation coefficient $r$ (a single number measuring linear strength and direction), and the least-squares regression line (an equation for prediction). A surprising fact: two datasets can have identical means, standard deviations, correlation, and regression lines yet look completely different when plotted — a lesson dramatized by Anscombe's quartet, proving that summary statistics never replace the picture.

The Core Insight

When you measure two numbers on each subject, plot them as points. The pattern of the cloud of points — its shape (form), its tilt (direction), and its tightness (strength) — encodes the relationship. A straight-line trend is the simplest and most important pattern, and we can compress it into two numbers: a correlation that says how tightly the points hug a line, and a slope/intercept pair that says which line. Both come from a single quantity: how $x$ and $y$ vary together.

Formal Statement

Bivariate Quantitative Data

A bivariate quantitative dataset is a set of $n$ ordered pairs

{(x_{1}, y_{1}), (x_{2}, y_{2}), \dots, (x_{n}, y_{n})}

⟳This is a definition, not a derivation — each pair records two numerical measurements on the same unit

i

where $x_{i}$ and $y_{i}$ are real-valued (quantitative) measurements on the same unit $i$ . The variable plotted horizontally ( $x$ ) is the explanatory (independent) variable; the variable plotted vertically ( $y$ ) is the response (dependent) variable. Both must be quantitative — measured on a numerical scale where arithmetic is meaningful.

Covariance

The sample covariance measures how the two variables vary together:

s_{x y} = \frac{1}{n - 1} i = 1 \sum n (x_{i} - \overset{x}{ˉ}) (y_{i} - \overset{y}{ˉ})

⟳Average the product of paired deviations from each mean; positive when both deviate in the same direction, negative when opposite.

where $\overset{x}{ˉ} = \frac{1}{n} \sum x_{i}$ and $\overset{y}{ˉ} = \frac{1}{n} \sum y_{i}$ are the sample means. Covariance has units of $x$ times $y$ , which makes it hard to interpret directly.

Pearson Correlation Coefficient

The correlation coefficient $r$ standardizes covariance by the two standard deviations, producing a unitless number:

r = \frac{s _{x y}}{s _{x} s _{y}} = \frac{\sum _{i = 1}^{n} ( x _{i} - x ˉ ) ( y _{i} - y ˉ )}{\sum _{i = 1}^{n} ( x _{i} - x ˉ ) ^{2} \sum _{i = 1}^{n} ( y _{i} - y ˉ ) ^{2}}

⟳Divide covariance [2] by

s_{x} s_{y}

; the

(n - 1)

factors cancel, leaving the ratio of summed cross-products to the geometric mean of summed squares.

It can be written compactly using standardized scores ( $z$ -scores):

r = \frac{1}{n - 1} i = 1 \sum n (\frac{x _{i} - x ˉ}{s _{x}}) (\frac{y _{i} - y ˉ}{s _{y}})

⟳Substitute

z_{x_{i}} = (x_{i} - \overset{x}{ˉ}) / s_{x}

and

z_{y_{i}} = (y_{i} - \overset{y}{ˉ}) / s_{y}

into [3];

r

is the average product of standardized scores.

Properties of the Correlation Coefficient

For any bivariate quantitative dataset:

$- 1 \leq r \leq 1$ always.
$r = + 1$ exactly when all points lie on a line of positive slope; $r = - 1$ when all lie on a line of negative slope.
$r$ is unitless and unchanged by changes of scale or origin (e.g. converting cm to inches).
$r$ is symmetric: the correlation of $x$ with $y$ equals that of $y$ with $x$ .
$r$ measures only linear association — it can be near $0$ for strong nonlinear relationships.

Least-Squares Regression Line

The least-squares regression line $\overset{y}{^} = a + b x$ is the line minimizing the sum of squared vertical residuals. Its slope and intercept are

b = \frac{s _{x y}}{s _{x}^{2}} = r \frac{s _{y}}{s _{x}}, a = \overset{y}{ˉ} - b \overset{x}{ˉ}

⟳Minimize

\sum (y_{i} - a - b x_{i})^{2}

by setting partial derivatives w.r.t.

a

and

b

to zero; solving the two normal equations gives these closed forms.

The predicted value $\overset{y}{^}_{i} = a + b x_{i}$ is the fitted value; the residual is $e_{i} = y_{i} - \overset{y}{^}_{i}$ .

Coefficient of Determination

The fraction of variation in $y$ explained by the linear regression on $x$ equals $r^{2}$ :

r^{2} = 1 - \frac{\sum _{i = 1}^{n} ( y _{i} - y ^ _{i} ) ^{2}}{\sum _{i = 1}^{n} ( y _{i} - y ˉ ) ^{2}} = \frac{explained variation}{total variation}

⟳Decompose total sum of squares

\sum (y_{i} - \overset{y}{ˉ})^{2}

into explained

\sum (\overset{y}{^}_{i} - \overset{y}{ˉ})^{2}

plus residual

\sum (y_{i} - \overset{y}{^}_{i})^{2}

; the explained fraction simplifies to

r^{2}

Why It Holds / Derivation

Deriving the Least-Squares Slope and Intercept

We seek $a, b$ minimizing the total squared vertical error

S (a, b) = i = 1 \sum n (y_{i} - a - b x_{i})^{2}

⟳This is the objective function — squared residuals summed over all points.

Step 1 — Intercept condition. Differentiate with respect to

a

and set to zero:

\frac{\partial S}{\partial a} = - 2 i = 1 \sum n (y_{i} - a - b x_{i}) = 0 \Rightarrow \sum y_{i} = na + b \sum x_{i}

⟳Chain rule on each squared term; the sum of residuals must be zero. Divide by

n

to get

\overset{y}{ˉ} = a + b \overset{x}{ˉ}

This already gives $a = \overset{y}{ˉ} - b \overset{x}{ˉ}$ , so the line passes through the point of averages $(\overset{x}{ˉ}, \overset{y}{ˉ})$ .

Step 2 — Slope condition. Differentiate with respect to

b

and set to zero:

\frac{\partial S}{\partial b} = - 2 i = 1 \sum n x_{i} (y_{i} - a - b x_{i}) = 0 \Rightarrow \sum x_{i} y_{i} = a \sum x_{i} + b \sum x_{i}^{2}

⟳The residuals must be uncorrelated with the

x

-values; this is the second normal equation.

Step 3 — Solve. Substitute

a = \overset{y}{ˉ} - b \overset{x}{ˉ}

into Step 2 and simplify using

\sum (x_{i} - \overset{x}{ˉ}) (y_{i} - \overset{y}{ˉ})

and

\sum (x_{i} - \overset{x}{ˉ})^{2}

b = \frac{\sum ( x _{i} - x ˉ ) ( y _{i} - y ˉ )}{\sum ( x _{i} - x ˉ ) ^{2}} = \frac{s _{x y}}{s _{x}^{2}}

⟳Algebraic rearrangement; the centered sums emerge once

\overset{x}{ˉ}, \overset{y}{ˉ}

are subtracted. Dividing numerator and denominator by

(n - 1)

gives the covariance/variance form.

Finally, since $s_{x y} = r s_{x} s_{y}$ , we get $b = r s_{x} s_{y} / s_{x}^{2} = r s_{y} / s_{x}$ , completing equation [5]. $■$

Francis Galton | 1822–1911

The English polymath and cousin of Charles Darwin, Galton invented the concepts of correlation and regression while studying heredity. Plotting parents' heights against children's heights, he noticed that tall parents tended to have tall — but on average less tall — children, a phenomenon he named "regression toward mediocrity" (now regression toward the mean). His 1886 paper introduced the regression line, and his student Karl Pearson formalized the correlation coefficient $r$ — still called the Pearson correlation — in the 1890s. Galton's statistical insights were genuine and lasting, even as his views on eugenics were later, rightly, discredited.

Building Intuition

Think of correlation as a measure of how much knowing $x$ helps you guess $y$ . If $r = 0$ , knowing a person's $x$ tells you nothing about their $y$ — your best guess is always $\overset{y}{ˉ}$ . If $r = \pm 1$ , knowing $x$ tells you $y$ exactly. In between, $r^{2}$ tells you what fraction of the "spread" in $y$ you can account for: $r = 0.8$ means $r^{2} = 0.64$ , so $64%$ of $y$ 's variation is explained by the line.

The slope formula $b = r (s_{y} / s_{x})$ also explains regression to the mean: when $x$ is $1$ standard deviation above average, the predicted $y$ is only $r$ standard deviations above average — and since $∣ r ∣ < 1$ in real data, $y$ is pulled back toward its mean. Outstanding parents tend to have merely good children; record-breaking seasons tend to be followed by ordinary ones. This is not a force of nature; it is pure arithmetic.

Scatterplot with least-squares regression line and residuals

<svg viewBox="0 0 400 300" xmlns="http://www.w3.org/2000/svg">
<rect width="400" height="300" fill="#0a0e17"/>
<!-- axes -->
<line x1="50" y1="250" x2="370" y2="250" stroke="#94a3b8" stroke-width="1.5"/>
<line x1="50" y1="250" x2="50" y2="30" stroke="#94a3b8" stroke-width="1.5"/>
<text x="360" y="270" fill="#e2e8f0" font-size="12">x</text>
<text x="30" y="40" fill="#e2e8f0" font-size="12">y</text>
<!-- regression line y-hat -->
<line x1="60" y1="220" x2="360" y2="70" stroke="#d4a853" stroke-width="2.5"/>
<text x="300" y="80" fill="#d4a853" font-size="12">ŷ = a + bx</text>
<!-- data points and residuals -->
<!-- point 1 -->
<line x1="100" y1="200" x2="100" y2="180" stroke="#64748b" stroke-width="1" stroke-dasharray="3"/>
<circle cx="100" cy="180" r="4" fill="#e2e8f0"/>
<!-- point 2 -->
<line x1="150" y1="170" x2="150" y2="190" stroke="#64748b" stroke-width="1" stroke-dasharray="3"/>
<circle cx="150" cy="190" r="4" fill="#e2e8f0"/>
<!-- point 3 -->
<line x1="200" y1="145" x2="200" y2="125" stroke="#64748b" stroke-width="1" stroke-dasharray="3"/>
<circle cx="200" cy="125" r="4" fill="#e2e8f0"/>
<!-- point 4 -->
<line x1="250" y1="120" x2="250" y2="140" stroke="#64748b" stroke-width="1" stroke-dasharray="3"/>
<circle cx="250" cy="140" r="4" fill="#e2e8f0"/>
<!-- point 5 -->
<line x1="310" y1="92" x2="310" y2="78" stroke="#64748b" stroke-width="1" stroke-dasharray="3"/>
<circle cx="310" cy="78" r="4" fill="#e2e8f0"/>
<!-- residual label -->
<text x="120" y="195" fill="#94a3b8" font-size="10">residual e = y − ŷ</text>
<!-- mean point -->
<circle cx="205" cy="145" r="3" fill="#d4a853"/>
<text x="210" y="160" fill="#94a3b8" font-size="10">(x̄, ȳ)</text>
</svg>

Worked Examples

Computing r and the Regression Line by Hand

Given: Five students' hours studied (

x

) and exam scores (

y

| $x$ | 1 | 2 | 3 | 4 | 5 | |---|---|---|---|---|---| | $y$ | 52 | 60 | 67 | 73 | 78 |

Step 1 — Means.

\overset{x}{ˉ} = 3

\overset{y}{ˉ} = (52 + 60 + 67 + 73 + 78) /5 = 330/5 = 66

. Step 2 — Deviations and products.

| $x_{i} - \overset{x}{ˉ}$ | $y_{i} - \overset{y}{ˉ}$ | product | $(x_{i} - \overset{x}{ˉ})^{2}$ | $(y_{i} - \overset{y}{ˉ})^{2}$ | |---|---|---|---|---| | −2 | −14 | 28 | 4 | 196 | | −1 | −6 | 6 | 1 | 36 | | 0 | 1 | 0 | 0 | 1 | | 1 | 7 | 7 | 1 | 49 | | 2 | 12 | 24 | 4 | 144 |

Sums: $\sum (x - \overset{x}{ˉ}) (y - \overset{y}{ˉ}) = 65$ , $\sum (x - \overset{x}{ˉ})^{2} = 10$ , $\sum (y - \overset{y}{ˉ})^{2} = 426$ .

Step 3 — Correlation [3].

r = \frac{65}{10 \cdot 426} = \frac{65}{4260} = \frac{65}{65.27} = 0.9959

Step 4 — Slope and intercept [5].

b = \frac{65}{10} = 6.5, a = 66 - 6.5 (3) = 66 - 19.5 = 46.5

Step 5 — Coefficient of determination [6].

r^{2} = 0.995 9^{2} = 0.992

\overset{y}{^} = 46.5 + 6.5 x, r = 0.996, r^{2} = 0.992 (99.2% of score variation explained)

Each extra hour of study predicts

6.5

more points.

Prediction and Residual

Given: Using the line

\overset{y}{^} = 46.5 + 6.5 x

from the previous example, predict the score for a student who studies

x = 3

hours, and find the residual since the actual score was

67

. Step 1 — Fitted value.

\overset{y}{^} = 46.5 + 6.5 (3) = 46.5 + 19.5 = 66 points

Step 2 — Residual.

e = y - \overset{y}{^} = 67 - 66 = + 1 point

Step 3 — Interpretation. The student scored

1

point above the line's prediction; the positive residual means they outperformed the model. Predicting for

x = 10

hours would be extrapolation — outside the observed range

1

–

5

— and is unreliable.

\overset{y}{^} = 66 points, e = + 1 point

Common Misconceptions

Correlation Is Not Causation

A high $∣ r ∣$ means two variables move together — it does not mean one causes the other. Ice cream sales and drowning deaths are strongly correlated, but neither causes the other; both are driven by a lurking variable, hot weather. Establishing causation requires a controlled experiment or careful confounding adjustment, never correlation alone.

r Only Measures Linear Association

A correlation near zero does not mean "no relationship." It means no linear relationship. Data tracing a perfect parabola $y = x^{2}$ symmetric about $x = 0$ has $r = 0$ exactly, yet $y$ is perfectly determined by $x$ . Always plot the data first.

Anscombe's Quartet

In 1973 statistician Francis Anscombe constructed four datasets, each with $11$ points, that share identical summary statistics: $\overset{x}{ˉ} = 9$ , $\overset{y}{ˉ} = 7.5$ , $r = 0.816$ , regression line $\overset{y}{^} = 3 + 0.5 x$ , and $r^{2} = 0.67$ . Yet plotted, they are utterly different: one is a clean linear trend, one is a perfect curve (nonlinear), one is a line dragged off course by a single outlier, and one is a vertical strip held together by one extreme point. The lesson: summary statistics never substitute for a scatterplot.

Don't Extrapolate or Confuse the Roles of x and y

The regression line of $y$ on $x$ is generally different from the line of $x$ on $y$ — the two slopes are $r s_{y} / s_{x}$ and $r s_{x} / s_{y}$ , which are equal only when $∣ r ∣ = 1$ . Choose the explanatory variable deliberately. And never use the line to predict far outside the data's $x$ -range: a study-hours model fitted on $1$ – $5$ hours cannot validly predict the score for $100$ hours.

Connections

Bivariate quantitative analysis sits at the heart of a web of statistical ideas. The covariance and correlation in [2]–[3] generalize the single-variable Standard Deviation to two variables. The least-squares method [5] is the simplest case of Linear Regression and extends to many predictors in multiple regression. The decomposition behind $r^{2}$ in [6] is an instance of Analysis of Variance. When both variables are random, correlation connects to the Bivariate Normal Distribution and the concept of Statistical Independence (independence implies $r = 0$ , but not conversely). For categorical pairs, the analogous tool is the Two-Way Table and the chi-squared test rather than $r$ .

Real-World Applications

Medicine: Correlating blood pressure with stroke risk, or drug dosage with response, to build predictive clinical models.
Economics: The Phillips curve relates unemployment to inflation; regression estimates how consumer spending responds to income (the marginal propensity to consume).
Finance: The beta of a stock is literally a regression slope of the stock's returns on the market's returns, measuring systematic risk.
Sports analytics: Relating a baseball team's run differential to wins, or a sprinter's stride length to race time, to optimize training.
Agriculture & climate: Yield versus rainfall or fertilizer; global temperature versus atmospheric CO₂ concentration over decades.
Engineering & quality control: Calibration curves relating a sensor's reading to the true measured quantity, fitted by least squares.
Psychometrics: Validating a test by correlating its scores with an external criterion (e.g. SAT scores against college GPA).

Historical Development

1795/1805: Carl Friedrich Gauss (used ~1795) and Adrien-Marie Legendre (published 1805) independently develop the method of least squares to fit astronomical orbits.
1809: Gauss links least squares to the normal distribution of errors in Theoria Motus.
1846: Auguste Bravais derives mathematical expressions resembling correlation while studying errors in two dimensions.
1877–1886: Francis Galton introduces regression and the regression line through studies of heredity (sweet peas, then human heights), coining "regression toward mediocrity."
1888: Galton publishes "Co-relations and Their Measurement," giving the first explicit index of correlation.
1896: Karl Pearson formalizes the product-moment correlation coefficient $r$ in its modern algebraic form.
1908: William Sealy Gosset ("Student") develops small-sample inference, enabling significance tests for $r$ and regression slopes.
1922–1925: Ronald A. Fisher places regression and correlation on rigorous inferential footing and develops the analysis of variance.
1973: Francis Anscombe publishes his quartet, dramatizing why graphs must accompany statistics.

Subtleties & Limits

Several deep cautions surround bivariate analysis. Simpson's paradox shows that a correlation present in every subgroup can reverse direction when groups are pooled — so the "right" grouping is genuinely ambiguous and context-dependent. Ecological correlation (computed from group averages, e.g. state-level data) is typically far stronger than the individual-level correlation, and mistaking one for the other is the ecological fallacy. Establishing causal direction from observational bivariate data remains a frontier: modern causal-inference frameworks (Pearl's do-calculus, instrumental variables, randomized trials) exist precisely because $r$ cannot do it. Finally, $r$ is highly sensitive to outliers and to the range of $x$ sampled — restricting the range (range restriction) attenuates $r$ even when the underlying relationship is unchanged, a persistent problem in test validation.

Bivariate Quantitative Data

The Core Insight

Formal Statement

Bivariate Quantitative Data

Covariance

Pearson Correlation Coefficient

Properties of the Correlation Coefficient

Least-Squares Regression Line

Coefficient of Determination

Why It Holds / Derivation

Deriving the Least-Squares Slope and Intercept

Francis Galton | 1822–1911

Building Intuition

Scatterplot with least-squares regression line and residuals

Worked Examples

Computing r and the Regression Line by Hand

Prediction and Residual

Common Misconceptions

Correlation Is Not Causation

r Only Measures Linear Association

Anscombe's Quartet

Don't Extrapolate or Confuse the Roles of x and y

Connections

Real-World Applications

Historical Development

Subtleties & Limits

See Also

Before this

Where this leads