How to Perform a Correlation Test in R (With Examples)

The Foundation: The Pearson Correlation Coefficient

The most widely utilized metric for quantifying the linear relationship between two continuous variables is the Pearson correlation coefficient, often denoted by $r$. This measure, formally known as Pearson’s product-moment correlation coefficient, quantifies the extent to which two variables are linearly related and change together. It assumes that both variables are normally distributed and that the relationship, if present, can be described by a straight line.

The value of the Pearson coefficient is strictly bounded, always falling within the interval $[-1, 1]$. The magnitude of this value indicates the strength of the relationship, while the sign indicates the direction. A coefficient close to zero suggests a weak or non-existent linear relationship, whereas values approaching the extremes, $-1$ or $1$, signify a very strong relationship.

Interpreting the bounded values is straightforward and provides immediate insight into the bivariate relationship:

• $r = -1$: This indicates a perfectly negative linear correlation. As one variable increases, the other decreases consistently along a straight line.
• $r = 0$: This indicates no linear correlation between the two variables. Note that a zero correlation does not necessarily mean the variables are independent; non-linear relationships might still exist.
• $r = 1$: This indicates a perfectly positive linear correlation. As one variable increases, the other increases consistently along a straight line.

While the Pearson coefficient is powerful, it is crucial to first inspect the data visually, typically through a scatterplot, to confirm that a linear model is appropriate before relying solely on the numerical coefficient. Non-linear patterns or heteroscedasticity can result in misleadingly low Pearson coefficients, highlighting the need for visual data exploration.

Determining Statistical Significance: T-Scores and P-Values

Finding a correlation coefficient, $r$, is only the first step. The more critical question is whether this observed relationship is statistically significant, meaning it is unlikely to have occurred by random chance alone. To address this, we compare the sample correlation coefficient against the null hypothesis ($H_0$), which posits that the true population correlation ($rho$) is zero.

This test of significance is conducted by calculating a standardized test statistic, typically the T-score (or T-statistic). This T-score transforms the correlation coefficient into a value that follows a known distribution—the Student’s t-distribution—under the assumption that the null hypothesis is true. The magnitude of the T-score reflects how many standard errors the sample coefficient is away from the hypothesized population correlation of zero.

The calculation of the T-statistic is standardized based on the sample size ($n$) and the calculated correlation coefficient ($r$). The formula provided is essential for understanding the underlying mechanics of the test:

T = r * $sqrt{n-2}$ / $sqrt{1-r^2}$

Once the T-score is determined, it is used to calculate the corresponding P-value. The P-value represents the probability of observing a correlation as extreme or more extreme than the one calculated, assuming that no linear relationship exists in the population (i.e., $H_0$ is true).

This P-value calculation relies on the t-distribution with $n-2$ degrees of freedom. If the calculated P-value is less than a predetermined significance level (commonly $alpha = 0.05$), we reject the null hypothesis and conclude that the observed correlation is statistically significant. Conversely, a P-value greater than $alpha$ suggests insufficient evidence to claim a significant linear relationship.

Beyond Pearson: Non-Parametric Correlation Methods

While the Pearson coefficient is the default standard for continuous data, its successful application relies heavily on assumptions of normality, homoscedasticity, and strict linearity. When these underlying distributional assumptions are violated, or when dealing with ordinal data or data containing significant outliers that skew the mean and variance, alternative non-parametric methods are often more appropriate for measuring association.

The Spearman Rank Correlation Coefficient ($rho$ or $r_s$) is a non-parametric measure of the strength and direction of the monotonic relationship between two variables. Instead of using the raw data values, Spearman’s method ranks the data for each variable separately and then calculates the Pearson correlation coefficient on these ranks. This transformation makes the test robust to outliers and effective for assessing non-linear, but consistently monotonic, relationships.

The Kendall Rank Correlation Coefficient ($tau$) is another key non-parametric test available in R. It assesses the association between two variables based on the number of concordant and discordant pairs of observations. Kendall’s Tau is generally considered more robust and less sensitive to errors and variances than Spearman’s coefficient, especially when dealing with smaller sample sizes or data with many tied ranks. Although the interpretation of its magnitude is often less intuitive than that of Pearson or Spearman, it provides a highly reliable measure of directional agreement.

The cor.test() Function in R

To perform a comprehensive correlation test that includes both the coefficient estimation and the crucial test of statistical significance (T-score and P-value), the R environment provides the specialized function cor.test(). This function is far superior to the basic cor() function when formal hypothesis testing is required, as it automatically calculates the necessary statistics, handles the degrees of freedom, and reports confidence intervals.

The generalized syntax for utilizing this function allows for essential flexibility in the choice of correlation method, accommodating various data types and underlying distributions, making it a versatile tool for data analysts:

cor.test(x, y, method=c(“pearson”, “kendall”, “spearman”))

The parameters passed into the function precisely control which data are analyzed and which statistical method is employed:

• x, y: These are the required numeric vectors containing the data for the two variables being analyzed. It is imperative that these vectors contain an equal number of observations (i.e., they must be the same length).
• method: This optional argument specifies the type of correlation to be calculated. The default setting is “pearson”, but it can be explicitly set to “kendall” or “spearman” if non-parametric testing is preferred due to data characteristics such as non-normality or ordinal scaling.

Understanding the implications of these arguments is crucial for selecting the appropriate statistical tool for your specific research question and ensuring that the test’s assumptions align with the data’s characteristics. Choosing the wrong method can lead to inaccurate conclusions about the relationship between variables.

Practical Example: Data Preparation and Visualization

To illustrate the powerful capabilities and practical application of cor.test(), consider a scenario where we are analyzing the relationship between two hypothetical variables: X, representing the level of training hours completed, and Y, representing job performance scores. We first define these variables as numeric vectors in the R environment:

x <- c(2, 3, 3, 5, 6, 9, 14, 15, 19, 21, 22, 23)
y <- c(23, 24, 24, 23, 17, 28, 38, 34, 35, 39, 41, 43)

Before proceeding with the formal correlation test, it is universally recommended practice in data analysis to visualize the relationship between the two variables. A scatterplot is the ideal visualization for bivariate data, as it allows us to visually confirm linearity, identify potential outliers, and get an initial estimate of the direction and strength of the relationship.

We generate the scatterplot using the base R plotting function. This step helps us confirm that the relationship appears generally positive and reasonably linear, which, if confirmed, justifies the subsequent use of the Pearson correlation coefficient method:

#create scatterplot 
plot(x, y, pch=16)

As evidenced by the scatterplot, the data points tend to trend upward from left to right, suggesting a positive linear relationship. The points are clustered relatively tightly around an imaginary line, indicating a strong positive correlation. This visual confirmation allows us to proceed confidently to the formal statistical assessment to precisely quantify this observation.

Interpreting the cor.test() Output

With the data prepared and a visual assessment complete, we execute the primary correlation test using the default settings of the cor.test() function. Since we did not specify a method, R defaults to the Pearson correlation, performing a two-sided test against the null hypothesis that the true population correlation is zero.

The command generates a comprehensive output detailing the test results:

#perform correlation test between the two vectors
cor.test(x, y)

	Pearson's product-moment correlation

data:  x and y
t = 7.8756, df = 10, p-value = 1.35e-05
alternative hypothesis: true correlation is not equal to 0
95 percent confidence interval:
 0.7575203 0.9799783
sample estimates:
      cor 
0.9279869 

The output provides several crucial metrics necessary for drawing statistical conclusions. First, the sample estimate for the correlation coefficient (cor) is calculated as 0.9279869. This value confirms the initial visual assessment, indicating a very strong positive linear association between the two vectors.

Next, we examine the hypothesis testing components. The test statistic, t, is reported as 7.8756. This T-value is exceptionally high, suggesting the sample correlation is many standard errors away from zero. It is used in conjunction with the degrees of freedom (df = 10) to determine the probability of observing this result under the null hypothesis.

Drawing Statistical Conclusions

The resulting P-value is the most direct measure of significance, reported as 1.35e-05 (which translates to 0.0000135). When comparing this P-value to the standard significance level ($alpha = 0.05$), we clearly observe that $1.35e-05 < 0.05$. Therefore, we have overwhelming evidence to reject the null hypothesis that the true population correlation is zero. We confidently conclude that the correlation between X and Y is statistically significant.

Furthermore, the output provides the 95 percent confidence interval (CI), ranging from 0.7575203 to 0.9799783. This interval represents the range within which the true population correlation coefficient is expected to fall 95% of the time. Since this entire interval is positive and does not contain the value zero, it provides an additional, robust confirmation that the true correlation coefficient in the population is significantly non-zero, reinforcing the conclusion drawn from the P-value.

Conclusion and Next Steps

Performing a correlation test in R using the cor.test() function is an indispensable skill for rigorous quantitative analysis. It moves beyond merely calculating an association measure to providing a formal, statistically verifiable statement about the likelihood of that association existing in the broader population. By correctly interpreting the T-statistic, P-value, and confidence intervals, researchers can confidently communicate the significance and magnitude of their findings.

It is vital to remember the core statistical principle that correlation does not imply causation. Even a strong, statistically significant correlation suggests only that two variables move together predictably. Further experimental design, advanced regression techniques, and domain knowledge are always required to explore and establish causal relationships between variables.

For those wishing to deepen their understanding of correlational statistics, especially the nuances of different coefficient types and their underlying assumptions, the following resource provides valuable additional information:

An Introduction to the Pearson Correlation Coefficient