Correlation measures the strength and direction of the linear relationship between two numerical variables. When one variable changes, does the other tend to change in a predictable way? The Pearson correlation coefficient (r) quantifies this relationship on a scale from −1 to +1.
Examples of correlated variables:
The Pearson r coefficient answers: given one variable's value, how well can we predict the other? A value near +1 means strong positive relationship (both increase together). Near −1 means strong negative relationship (one increases as the other decreases). Near 0 means little to no linear relationship.
Important caveat: correlation does not imply causation. Two variables can be strongly correlated because they're both caused by a third variable, because the relationship is coincidental, or because of confounding factors in how the data was collected. Always interpret correlations in context before drawing causal conclusions.
The Pearson correlation coefficient is calculated as the covariance of two variables divided by the product of their standard deviations:
r = Σ[(xi − x̄)(yi − ȳ)] / √[Σ(xi − x̄)² × Σ(yi − ȳ)²]
Equivalently, using raw scores:
r = [nΣxy − ΣxΣy] / √{[nΣx² − (Σx)²][nΣy² − (Σy)²]}
Where n is the number of paired observations.
R² (coefficient of determination) = r² tells you the proportion of variance in Y that is explained by X. If r = 0.80, then R² = 0.64, meaning 64% of the variation in Y is accounted for by the linear relationship with X.
Worked example: Weekly training miles (X) and 5K time in minutes (Y) for 5 runners:
| Runner | Miles/week (X) | 5K time (Y) |
|---|---|---|
| A | 20 | 28 |
| B | 35 | 24 |
| C | 50 | 21 |
| D | 65 | 19 |
| E | 80 | 17 |
x̄ = 50, ȳ = 21.8. After computing the formula: r ≈ −0.99. As miles per week increase, 5K time decreases — an extremely strong negative correlation. R² ≈ 0.98, meaning training volume explains 98% of the variance in race times for these runners.
| r value | Strength | Direction | Example |
|---|---|---|---|
| +0.90 to +1.00 | Very strong | Positive | Height vs weight in adults |
| +0.70 to +0.89 | Strong | Positive | Study hours vs test score |
| +0.50 to +0.69 | Moderate | Positive | Income vs education level |
| +0.30 to +0.49 | Weak | Positive | Shoe size vs IQ score |
| +0.10 to +0.29 | Very weak | Positive | Birth month vs sports success |
| −0.10 to +0.10 | Negligible | None | TV color preference vs salary |
| −0.30 to −0.10 | Very weak | Negative | Hours slept vs reaction time |
| −0.50 to −0.30 | Weak | Negative | Distance from work vs satisfaction |
| −0.70 to −0.50 | Moderate | Negative | Training load vs injury risk |
| −0.90 to −0.70 | Strong | Negative | Body fat % vs VO2 max |
| −1.00 to −0.90 | Very strong | Negative | Speed vs race time |
Note: These cutoffs (0.30, 0.50, 0.70) are Cohen's guidelines and are conventions, not absolute rules. In social science with many confounding variables, r = 0.30 may be meaningful. In physics where theory predicts r = 1.00, r = 0.90 might indicate a measurement problem. Always interpret r in the context of your field and research question.
A correlation coefficient by itself doesn't tell you whether the relationship is statistically significant — that depends on both r and your sample size. With a small sample, even a moderately high r could be due to chance. With a large sample, even a small r can be statistically significant (though possibly not practically meaningful).
The t-test for correlation significance:
t = r × √(n−2) / √(1−r²), df = n−2
Minimum r needed for significance at p < 0.05 (two-tailed):
| Sample Size (n) | Min |r| for p < 0.05 | Min |r| for p < 0.01 |
|---|---|---|
| 5 | 0.878 | 0.959 |
| 10 | 0.632 | 0.765 |
| 20 | 0.444 | 0.561 |
| 30 | 0.361 | 0.463 |
| 50 | 0.279 | 0.361 |
| 100 | 0.197 | 0.256 |
| 200 | 0.139 | 0.182 |
| 500 | 0.088 | 0.115 |
Key takeaway: with n = 10, you need r > 0.63 to be confident the correlation isn't due to chance. With n = 500, r = 0.09 is statistically significant but practically meaningless — X explains less than 1% of Y's variance. Always report both r and its significance, and consider practical significance alongside statistical significance.
Before calculating r, visually inspecting a scatter plot reveals patterns that r alone may miss:
Anscombe's Quartet (1973) famously demonstrated four datasets with identical r = 0.816 but completely different scatter patterns: one linear, one curved, one with an outlier, one perfectly linear except for one anomalous point. This underscores why visualizing data before interpreting correlations is essential.
Pearson r is not the only correlation coefficient. Different situations call for different measures:
| Method | Data Type | Assumption | Best For |
|---|---|---|---|
| Pearson r | Continuous | Linear relationship, normal distribution | Height/weight, temperature/sales |
| Spearman ρ | Ordinal or continuous | Monotonic relationship (ranks) | Rankings, survey scales, skewed data |
| Kendall τ | Ordinal or small samples | Monotonic relationship (pairs) | Small n, many ties in ranking |
| Point-biserial | One binary, one continuous | Normal distribution of continuous var | Gender vs test score, pass/fail vs study hours |
When to prefer Spearman over Pearson: your data is ordinal (e.g., survey responses rated 1–5), your data is heavily skewed or has outliers, or you're testing for any monotonic relationship (not just linear). Spearman converts values to ranks before computing r, making it robust to outliers and non-normality. Our calculator computes Pearson r; for Spearman, rank your data first and then use the same formula.
Correlation analysis is fundamental to sports science research. Understanding which training variables correlate with performance outcomes helps coaches and athletes make evidence-based decisions.
Training load and performance correlations: Studies consistently find strong negative correlations (r = −0.7 to −0.9) between weekly mileage and race times among recreational runners — more miles, faster times. However, correlation doesn't mean more is always better: beyond ~80–100 miles/week, injury risk increases and the performance-mileage correlation weakens or reverses.
Heart rate and pace: At steady-state aerobic effort, heart rate and running pace show strong positive correlation (r ≈ 0.85–0.95) for individual runners in controlled conditions. This forms the basis of heart rate training zones and aerobic threshold testing.
VO2 max and race performance: Maximal oxygen uptake correlates strongly with endurance performance across athletes (r = −0.70 to −0.85 with race times). However, within a group of elite runners with similarly high VO2 max, the correlation weakens — other factors like lactate threshold and running economy then distinguish performance.
Sleep and recovery metrics: Heart rate variability (HRV) correlates positively with perceived recovery status (r ≈ 0.5–0.7) in athletes. This moderately strong correlation has driven widespread adoption of HRV monitoring in elite sport, though the individual variability means population-level correlations don't perfectly predict individual responses.
Pitfalls in sports correlation research: Publication bias favors high correlations, so the literature may overstate relationships. Cross-sectional studies (measuring both variables at one time point) can't establish causation. Aggregating across athletes with different training levels (beginners vs. elites) can create artificially strong correlations that don't apply within any subgroup. Always consider the specific population when generalizing correlation findings.
It depends on the field. In psychology and social science, r > 0.50 is considered strong. In physics or engineering, r > 0.99 might be expected. For predicting athletic performance with a single training variable, r = 0.70–0.85 is typical and considered meaningful. Always interpret r in context — what matters is whether the relationship is strong enough for your specific application.
R² is the proportion of variance in Y explained by X. If r = 0.80, then R² = 0.64, meaning 64% of the variation in Y is accounted for by the linear relationship with X. The remaining 36% is due to other factors or random variation. R² ranges from 0 (no linear relationship) to 1 (perfect linear relationship).
Correlation tells you about the strength of a relationship, but for prediction use linear regression. Regression gives you the equation of the best-fit line (Y = a + bX), enabling actual predictions. The correlation r is the square root of R² in simple linear regression. If you have a strong correlation, regression will give you a reliable prediction equation.
Correlation without causation can arise from: (1) Third-variable causation — both X and Y are caused by Z (ice cream sales and drowning both increase in summer due to hot weather, not because of each other); (2) Reverse causation — Y causes X, not X causes Y; (3) Coincidence — spurious correlations in small samples or time-series data. Establishing causation requires experimental design (randomized controlled trials) or rigorous causal inference methods.
A minimum of n = 30 pairs is a common rule of thumb for reliable Pearson r estimates. For detecting small correlations (r ≈ 0.2), you may need n = 150–200. For high-stakes research, power analysis should determine the exact n based on expected effect size and desired significance level. With n < 10, correlation estimates are highly unstable and should be interpreted with extreme caution.