Understanding Pearson Correlation: The Five Essential Assumptions

The Pearson correlation coefficient (PCC), often formally known as the product-moment correlation coefficient, stands as a cornerstone in statistical analysis. Its primary function is to rigorously quantify the linear strength and direction of the relationship observed between two distinct continuous variables.

The coefficient itself is constrained to yield a value strictly bounded between -1 and 1. The magnitude of this value directly indicates the strength of the linear association, while the sign specifies the direction of the relationship:

• -1 indicates a perfectly inverse, or negative, linear relationship between the two variables. As one variable increases, the other decreases consistently.
• 0 indicates the complete absence of any linear relationship between the variables under observation.
• 1 indicates a perfectly direct, or positive, linear relationship between the variables. As one variable increases, the other increases consistently.

To ensure that the calculated Pearson coefficient is a reliable and valid measure of association, five foundational statistical assumptions must be met by the data. Failure to satisfy these prerequisites may lead to inaccurate conclusions regarding the variables’ true relationship.

These five critical assumptions are summarized below:

1. Level of Measurement: Both variables must possess a continuous scale of measurement, specifically either interval or ratio.

2. Linear Relationship: The underlying connection between the variables must be linear, meaning it can be approximated by a straight line.

3. Normality: The distribution of data for both variables should closely approximate a normally distributed pattern.

4. Related Pairs: Each subject or observation within the sample must be composed of a corresponding, matched pair of values for the two variables being examined.

5. No Outliers: The dataset must be free from extreme outliers, which possess the potential to disproportionately distort the correlation results.

The subsequent sections provide an in-depth exploration of each assumption, along with practical guidelines for utilizing both visual inspection and formal tests to determine if your data satisfies these vital statistical requirements.

Assumption 1: Requirements for Level of Measurement

A fundamental prerequisite for the legitimate application of the Pearson correlation coefficient is that the data for both variables under analysis must be measured at either the interval or ratio level. These two levels fall under the category of continuous data, which is essential because the mathematical calculation of the Pearson coefficient relies heavily on calculating the precise differences between individual data points. This operation is statistically meaningless when applied to categorical data.

The illustration below provides a concise summary of the four primary levels at which variables can be measured in statistical research, clarifying the distinction between continuous and categorical data types.

Variables categorized by the interval scale are characterized by having meaningful and equal distances between values. Crucially, however, they lack a true, non-arbitrary zero point, meaning ratios are not valid comparisons. Examples of interval variables include:

• Temperature: Standard measurements expressed in units such as Fahrenheit or Celsius.
• Credit Scores: Standardized, scaled measurements typically ranging from 300 to 850 points.
• Standardized Test Scores (e.g., SAT): Scores that range from 400 to 1,600 where zero does not imply the absence of the trait being measured.

Variables measured on a ratio scale satisfy all the characteristics of an interval scale but also incorporate a true zero point. This absolute zero signifies the complete absence of the quantity being measured, thereby allowing for statistically valid ratio comparisons (e.g., 2 meters is twice as long as 1 meter). Examples include:

• Height: Measured in metric or imperial units (e.g., centimeters, inches).
• Weight: Measured in units such as kilograms or pounds.
• Length: Measured distance in units like meters or feet.

If your data is measured at the ordinal level (data that can be ranked but where the difference between ranks is not necessarily equal), the Pearson method is inappropriate. In such instances, researchers must turn to non-parametric alternatives, most commonly the Spearman’s rank correlation coefficient, which is specifically designed for ranked data.

Related:

Assumption 2: Establishing a Linear Relationship

The Pearson correlation coefficient is mathematically engineered to quantify the strength of a linear relationship exclusively. Consequently, a fundamental assumption is that the association between the two continuous variables must be capable of being accurately modeled by a straight line. If the true underlying relationship is curvilinear—for example, following a quadratic or exponential curve—the calculated Pearson coefficient will fail to capture the comprehensive and genuine association between the variables.

The most accessible and reliable method for visually assessing the assumption of linearity is the generation of a scatter plot, which graphically represents the bivariate relationship. If the data points displayed on the plot exhibit a noticeable clustering around a single, implied straight line, as demonstrated in the figure below, the assumption of linearity is considered adequately fulfilled:

In contrast, if the data points are either randomly dispersed across the plot or clearly define a curved trajectory (such as the distinct quadratic pattern illustrated in the subsequent figure), the assumption of a linear relationship has been violated.

When faced with a distinctly non-linear relationship, applying the Pearson correlation coefficient will inevitably result in a value that substantially underestimates or misrepresents the true connection. In these statistical scenarios, researchers are advised to explore data transformation techniques (to linearize the relationship) or to utilize alternative statistical models that are specifically designed for non-linear associations.

Assumption 3: Assessment of Data Normality

The third assumption for valid Pearson correlation analysis dictates that the data for both variables should be approximately normally distributed. While the correlation coefficient itself is generally robust to minor departures from perfect normality, substantial non-normality can compromise the accuracy of statistical inference, especially concerning the calculation of confidence intervals and the associated p-values used to test the significance of the coefficient.

Verification of this assumption should ideally involve a combination of visual data inspection and objective formal statistical testing.

1. Visual Check using Histograms

A histogram offers a quick, clear visual overview of the frequency distribution of a dataset. If the histogram displays a symmetrical, bell-shaped curve that tapers smoothly on both sides, it strongly suggests that the data is reasonably normally distributed.

2. Visual Check using Q-Q Plots

The Quantile-Quantile (Q-Q) plot serves as a more precise visual tool by comparing the actual distribution of your sample data against a theoretical normal distribution. In this plot, the theoretical quantiles (representing perfect normality) are plotted on the x-axis, and the sample quantiles (your actual data points) are plotted on the y-axis.

If the entire collection of data points aligns tightly along the central 45-degree reference line, the variable is deemed to be adequately normally distributed. Deviations from this line, particularly at the tails, signal potential non-normality.

3. Formal Statistical Tests for Normality

To complement visual assessments, formal statistical tests provide an objective measure of deviation from normality. For these tests, if the resulting p-value falls below a predetermined significance level (e.g., the standard α = 0.05), the researcher must reject the null hypothesis, concluding that the data is significantly non-normal.

Key statistical tests frequently employed for assessing the normality assumption include:

1. The Jarque-Bera Test

• This test bases its calculation on the skewness and kurtosis of the data distribution.
• It is often the preferred choice when analyzing large sample sizes.

2. The Shapiro-Wilk Test

• The Shapiro-Wilk test is widely regarded as one of the most statistically powerful normality tests available.
• It is highly recommended for use with smaller to moderate sample sizes (typically when the sample size, n, is less than 50).

3. The Kolmogorov-Smirnov Test

• This test operates by comparing the cumulative distribution function of the sample data directly against the theoretical normal distribution function.
• Although generally less powerful than the Shapiro-Wilk test, it remains a common and standard option in many statistical packages.

Assumption 4: Requirement of Matched and Related Pairs

The core methodology of the Pearson correlation coefficient is fundamentally dependent on the assumption that every single observation in the dataset consists of a related pair of values. In practical terms, this necessitates that for every data point recorded for variable X (e.g., individual height), there must exist a corresponding, accurately matched data point for variable Y (e.g., that same individual’s weight).

Verifying this assumption is primarily an exercise in careful data management and hygiene. If, for instance, a researcher intends to correlate the number of hours studied with final exam scores, they must meticulously ensure that every student in the sample has both a measured study duration and a corresponding exam score. While dealing with legitimately missing data (which often results in list-wise deletion) can reduce statistical power, any instance where data points are mismatched, misaligned, or improperly aggregated will render the correlation calculation mathematically invalid and meaningless.

Assumption 5: Sensitivity and Absence of Extreme Outliers

The final, and perhaps most practically relevant, assumption involves ensuring the dataset is free of extreme outliers. The Pearson correlation is intrinsically vulnerable to the undue influence of outliers because its calculation is based on the mean and standard deviation of the variables. Since these metrics are not robust to extreme values, a single data point that deviates significantly from the rest of the sample can drastically inflate, deflate, or even reverse the resulting correlation coefficient.

To clearly illustrate this dramatic effect, consider the small, hypothetical dataset presented below:

For this initial, tightly clustered dataset, the calculated Pearson correlation coefficient between variables X and Y is a very strong 0.949, indicating an almost perfect positive linear association.

Now, observe the consequence of introducing just one single extreme outlier (the point 10, 2) into the otherwise clean dataset, as shown here:

With the incorporation of this sole extreme data point, the Pearson correlation coefficient between X and Y drops significantly from 0.949 to 0.711.

This stark numerical comparison underscores how critical it is to identify and address outliers, as they possess the power to substantially obscure or distort the true underlying correlation between variables. Researchers should systematically use visual aids (such as box plots or scatter plots) to detect outliers and then determine the most appropriate course of action, whether that involves removing the outlier (if justified) or opting for a more robust, non-parametric alternative like Spearman’s rho, which is less sensitive to distributional extremes.

Related:

Additional Resources for Correlation Analysis

For researchers seeking further detailed guidance, the following resources provide comprehensive tutorials regarding the calculation, interpretation, and proper application of the Pearson correlation coefficient and related statistical methods: