Correlation Measures

Why This Matters

Understanding correlation measures is fundamental to making valid inferences from data—and that's exactly what you're being tested on in this course. These aren't just formulas to memorize; they represent different tools for different situations. The AP exam will challenge you to recognize when to use each measure based on your data type, distribution assumptions, and research question. Choosing the wrong correlation measure can lead to misleading conclusions and flawed decisions.

The key concepts here revolve around linearity assumptions, data types (continuous vs. ordinal vs. categorical), robustness to violations, and controlling for confounding variables. Each correlation measure makes specific assumptions about your data—violate those assumptions, and your results become meaningless. Don't just memorize the formulas; know what type of data each measure requires, what assumptions it makes, and when it outperforms alternatives.

---

Measuring Linear Relationships in Continuous Data

When both variables are continuous and you expect a straight-line relationship, these measures quantify how tightly your data points cluster around that line. The key assumption is linearity—if your scatter plot curves, these measures will underestimate the true relationship.

Pearson Correlation Coefficient

• Measures linear association between two continuous variables—the gold standard when assumptions are met, ranging from $-1$ to $+1$
• Assumes normality and linearity, meaning your data should be roughly bell-shaped and follow a straight-line pattern
• Highly sensitive to outliers, which can dramatically inflate or deflate your correlation value

Partial Correlation

• Controls for confounding variables while measuring the relationship between two variables of interest—essential for isolating direct effects
• Ranges from $-1$ to $+1$ like Pearson, but reveals the unique association after removing shared variance with control variables
• Critical for causal inference, helping distinguish genuine relationships from spurious correlations driven by lurking variables

Multiple Correlation

• Quantifies how well multiple predictors collectively explain a single outcome—ranges from $0$ to $1$ (always positive)
• Denoted as $R$ and its square ($R^2$) tells you the proportion of variance explained by your predictors combined
• Foundation of regression analysis, commonly tested in contexts asking about model fit and predictive power

---

Rank-Based Measures for Non-Normal Data

When your data violates normality assumptions or involves ordinal rankings, these measures assess monotonic relationships—whether variables consistently increase or decrease together, even if not in a straight line.

Spearman Rank Correlation

• Converts data to ranks before calculating correlation—perfect for ordinal data or when outliers would distort Pearson results
• Ranges from $-1$ to $+1$ and measures monotonic (not necessarily linear) relationships
• More robust to outliers than Pearson, making it the go-to choice when your data has extreme values

Kendall's Tau

• Compares concordant and discordant pairs of observations—more interpretable as the probability that rankings agree
• Handles ties better than Spearman, making it preferable when many observations share the same rank
• More conservative estimates with smaller sample sizes, often producing lower values than Spearman for the same data

---

Correlations Involving Categorical Variables

When one or both variables are categorical (binary or nominal), standard correlation measures don't apply. These specialized measures handle the unique properties of categorical data.

Point-Biserial Correlation

• Special case of Pearson for one continuous and one binary variable—mathematically equivalent to comparing group means
• Ranges from $-1$ to $+1$, indicating whether the continuous variable differs between the two categories
• Assumes normality within each group, making it appropriate when your continuous variable is roughly bell-shaped for both categories

Phi Coefficient

• Measures association between two binary variables—calculated from a $2 \times 2$ contingency table
• Ranges from $-1$ to $+1$, with the sign indicating whether the variables move together or opposite
• Related to chi-square: $\phi = \sqrt{\chi^2/n}$, connecting it to hypothesis testing for independence

---

Measuring Agreement and Reliability

Sometimes you're not measuring relationships between different constructs—you're assessing whether multiple measurements of the same thing agree with each other. This is fundamentally different from association.

Intraclass Correlation (ICC)

• Quantifies consistency among multiple raters or repeated measurements—ranges from $0$ to $1$, where higher values indicate better agreement
• Accounts for both correlation and systematic differences, unlike Pearson which only captures correlation
• Essential for reliability studies, determining whether your measurement instrument produces consistent results across observers or time points

---

Advanced Multivariate Relationships

When you're dealing with complex relationships involving multiple variables on both sides of the equation, or when linear measures fail to capture the full picture, these advanced techniques apply.

Canonical Correlation

• Finds linear combinations of two variable sets that maximize correlation—like running multiple regressions simultaneously
• Ranges from $0$ to $1$ for each canonical correlation, with the first capturing the strongest relationship
• Useful for multivariate research questions, such as relating a set of predictors to a set of outcomes

Distance Correlation

• Captures both linear and non-linear associations—ranges from $0$ to $1$, where $0$ means statistical independence
• Makes no distributional assumptions, working with any data type and detecting relationships Pearson would miss
• Equals zero only when variables are truly independent, unlike Pearson which can be zero even when non-linear relationships exist

---

Quick Reference Table

---

Self-Check Questions

1. You have continuous data with several extreme outliers. Which two correlation measures would be more appropriate than Pearson, and why do they share this advantage?

2. A researcher wants to know if test anxiety predicts exam performance after controlling for hours studied. Which correlation measure should they use, and what does it reveal that Pearson alone cannot?

3. Compare and contrast Spearman and Kendall's tau: when would you choose one over the other, and what do their different values for the same data tell you?

4. You're analyzing the relationship between a binary treatment variable (drug vs. placebo) and a continuous outcome (blood pressure). Which correlation measure applies, and what assumption must your data meet?

5. FRQ-style: A scatter plot shows a clear U-shaped relationship between two variables, but the Pearson correlation is near zero. Explain why this occurs and identify which correlation measure would better capture this relationship.