Correlation Coefficients

Why This Matters

Correlation is one of the most useful tools in AP Statistics for understanding how two quantitative variables relate to each other. The correlation coefficient $r$ tells you both the direction (positive or negative) and strength (weak, moderate, or strong) of a linear relationship. This concept connects directly to scatterplot interpretation, least-squares regression, residual analysis, and the coefficient of determination $r^2$, all of which are heavily tested on the AP exam.

What you're really being tested on: Can you interpret what $r$ means in context? Can you recognize when correlation is misleading (outliers, nonlinear patterns, lurking variables)? Can you connect $r$ to the regression slope formula $b = r \cdot \frac{s_y}{s_x}$? Don't just memorize that $r$ ranges from -1 to 1. Know why certain values matter, what affects them, and what correlation can and cannot tell you about causation.

---

The Pearson Correlation Coefficient: Your Primary Tool

The Pearson correlation coefficient $r$ is the standard measure of linear association on the AP exam. It quantifies how closely data points cluster around a straight line by comparing the standardized values (z-scores) of both variables.

Pearson Correlation Coefficient ($r$)

• Measures the strength and direction of linear relationships between two quantitative variables
• Formula uses z-scores: $r = \frac{1}{n-1} \sum \left(\frac{x_i - \bar{x}}{s_x}\right)\left(\frac{y_i - \bar{y}}{s_y}\right)$, though you'll typically use technology to compute it
• Unit-free and bounded: ranges from $-1$ to $1$, making it comparable across different contexts regardless of the original units

Because $r$ is calculated from z-scores, changing the units of either variable (say, converting inches to centimeters) has no effect on its value. This is a common multiple-choice trap.

Coefficient of Determination ($r^2$)

• Represents the proportion of variation explained. If $r = 0.8$, then $r^2 = 0.64$, meaning 64% of the variation in $y$ is explained by the linear relationship with $x$.
• Always between 0 and 1: since it's a squared value, $r^2$ is never negative and gives you an intuitive percentage interpretation.
• Key for FRQs: when asked to interpret a regression, a standard response is "$r^2$ percent of the variability in [response variable] is explained by the linear relationship with [explanatory variable]."

---

Interpreting Strength and Direction

Understanding what correlation values actually mean is essential for multiple-choice questions and FRQ interpretations. The magnitude (absolute value) indicates strength, while the sign indicates direction.

Strength Categories

• Strong correlation: $|r| \geq 0.8$ indicates data points cluster tightly around the regression line with little scatter
• Moderate correlation: $|r|$ between approximately 0.5 and 0.8 shows a clear trend but with noticeable spread
• Weak correlation: $|r| < 0.5$ means the linear pattern is hard to see and data points are widely scattered

These cutoffs are rough guidelines, not rigid rules. Context matters. In some fields (like psychology), $r = 0.5$ is considered quite strong, while in physics you might expect $r$ values very close to 1.

Direction of Association

• Positive correlation ($r > 0$): as $x$ increases, $y$ tends to increase. The scatterplot slopes upward left to right.
• Negative correlation ($r < 0$): as $x$ increases, $y$ tends to decrease. The scatterplot slopes downward left to right.
• No linear correlation ($r \approx 0$): no consistent linear pattern exists, though a nonlinear relationship might still be present.

Perfect Correlation

• $r = 1$ or $r = -1$: all data points fall exactly on the regression line with zero residuals. This is rare in real data.
• Indicates a deterministic relationship: knowing $x$ perfectly predicts $y$ with no error.
• Exam context: perfect correlations typically appear in theoretical questions or as benchmarks for comparison.

---

Connecting Correlation to Regression

The correlation coefficient isn't just a standalone statistic. It's mathematically linked to the least-squares regression line, and understanding this connection helps you move between correlation and regression problems.

Slope Formula Using $r$

• The slope formula $b = r \cdot \frac{s_y}{s_x}$ directly connects correlation to the regression line. Know this relationship cold.
• The sign of the slope always matches the sign of $r$: a positive correlation produces a positive slope, and vice versa. (This makes sense because $s_y$ and $s_x$ are always positive.)
• Standardized interpretation: when both variables are converted to z-scores, the slope of the regression line equals $r$.

Regression Line Properties

• Always passes through $(\bar{x}, \bar{y})$: the point of averages lies exactly on the least-squares regression line.
• Minimizes the sum of squared residuals: this is the "least squares" criterion that defines the LSRL.
• Y-intercept formula: $a = \bar{y} - b\bar{x}$, derived from the fact that the line passes through the mean point.

Residual Connection

• Residuals sum to zero: $\sum(y_i - \hat{y}_i) = 0$ for any least-squares regression line.
• Residual plots reveal fit quality: a patternless scatter of residuals indicates the linear model is appropriate; curves suggest nonlinearity.
• $r^2$ relates to residual variation: higher $r^2$ means smaller residuals relative to total variation in $y$.

---

What Affects Correlation: Outliers and Influential Points

One of the most tested concepts is how outliers and influential points can distort correlation. A single unusual observation can dramatically change $r$, which is why visual inspection of scatterplots is so important.

Outliers in Bivariate Data

• Can inflate or deflate $r$ depending on their position. An outlier far from the overall pattern typically weakens correlation, while an outlier that happens to fall along the trend can strengthen it.
• Identified through scatterplots and residual plots: look for points with unusually large residuals or unusual $x$-values.
• Always investigate before removing: outliers may represent data errors, but they might also be legitimate and important observations.

High-Leverage Points

• Have substantially larger or smaller $x$-values than other observations. They "pull" the regression line toward themselves.
• May or may not be influential: a high-leverage point that follows the existing pattern has little effect on the line; one that deviates from the pattern is highly influential.
• Critical for interpretation: if removing a point substantially changes $r$, slope, or intercept, that point is influential.

Influential Points

• Change the regression results substantially when removed, including changes to slope, intercept, and/or correlation.
• Often are both outliers and high-leverage: the combination of unusual $x$ and unusual $y$ creates maximum influence.
• Exam strategy: if asked about the effect of removing a point, consider whether it's pulling the line toward or away from the overall pattern of the remaining data.

---

Limitations and Common Misconceptions

Understanding what correlation cannot tell you is just as important as knowing what it measures. These limitations appear frequently in multiple-choice questions designed to test conceptual understanding.

Correlation Does Not Imply Causation

This is perhaps the most important statistical principle to internalize. Two variables can be strongly correlated without one causing the other.

• Lurking (confounding) variables may drive the relationship between both variables, creating a spurious correlation. For example, ice cream sales and drowning deaths are correlated because hot weather (the lurking variable) increases both.
• Establishing causation requires experiments: only randomized controlled experiments can demonstrate cause-and-effect relationships. Observational studies, no matter how strong the correlation, cannot establish causation on their own.

Nonlinear Relationships

• $r$ only measures linear association: a perfect curved relationship (like a parabola) can have $r \approx 0$ because Pearson correlation doesn't detect nonlinear patterns.
• Always examine scatterplots: before interpreting $r$, verify that a linear model is appropriate for the data.
• Transformations may help: if data show curvature, applying log, square root, or power transformations might linearize the relationship so that $r$ becomes meaningful.

Other Assumptions and Limitations

• Requires quantitative, paired data: both variables must be numerical and measured on the same individuals or cases.
• Sensitive to outliers: a single extreme point can shift $r$ dramatically, as discussed above.
• Switching $x$ and $y$ doesn't change $r$: correlation is symmetric, but regression is not. The regression of $y$ on $x$ gives a different line than the regression of $x$ on $y$.

---

Visualizing Correlation with Scatterplots

Scatterplots are your primary tool for assessing relationships before calculating correlation. Visual inspection reveals patterns, outliers, and potential problems that numbers alone cannot show.

Scatterplot Interpretation

When describing a scatterplot, cover these four features in order:

1. Direction: positive, negative, or neither
2. Form: linear, curved, or no clear pattern
3. Strength: how tightly the points follow the form
4. Unusual features: outliers, clusters, or gaps

The explanatory variable goes on the x-axis and the response on the y-axis. This convention matters for regression, but not for correlation itself (since $r$ is symmetric).

Residual Plots

• Plot residuals vs. fitted values (or vs. $x$): this diagnostic tool reveals whether a linear model is appropriate.
• Random scatter indicates good fit: if residuals show no pattern, the linear model captures the relationship well.
• Curved patterns indicate nonlinearity: a U-shape or other systematic pattern in the residual plot means the linear model is inadequate, even if $r$ looks decent.

Using Technology

• Graphing calculators compute $r$ automatically: on the TI-84, use LinReg(ax+b) after entering data in lists. Make sure "DiagnosticOn" is enabled, or $r$ and $r^2$ won't display.
• Always plot first: technology will calculate $r$ for any paired data, even when correlation is meaningless (nonlinear data, categorical data miscoded as numbers).
• Report $r$ in context: stating "$r = 0.85$" alone is incomplete. Interpret what this means for the specific variables in the problem.

---

Quick Reference Table

---

Self-Check Questions

1. If $r = -0.92$ for the relationship between hours of TV watched and GPA, what does this tell you about the strength and direction of the relationship? What does $r^2$ tell you that $r$ alone doesn't?

2. Two datasets both have $r^2 = 0.64$. One has $r = 0.8$ and the other has $r = -0.8$. How do these relationships differ, and how are they similar?

3. A scatterplot shows a clear U-shaped pattern, but the calculated correlation is $r = 0.05$. Explain why this happens and what it reveals about the limitations of correlation.

4. Compare how an outlier in the middle of the $x$-range versus an outlier at an extreme $x$-value would affect the correlation coefficient and regression line.

5. A study finds a strong positive correlation ($r = 0.78$) between ice cream sales and sunburn rates. A newspaper headline claims "Eating ice cream causes sunburns." Using the concept of lurking variables, explain why this causal claim is flawed and what study design would be needed to establish causation.