Pearson’s Correlation Coefficient: The Most Common Way to Measure Linear Relationships

Whenever researchers want to know whether two things move together, one of the first tools they reach for is correlation. Does more study time go with higher exam scores? Does taller height go with greater weight? Pearson’s product-moment correlation coefficient is the most widely used way to answer questions like these. It turns a vague sense that “these two variables seem related” into a single precise number that captures both how strong the relationship is and which direction it runs. This post walks through what that number means, how to calculate it by hand, and the important situations where it can quietly mislead you.

What is Pearson’s correlation? 🔗

Pearson’s correlation coefficient, usually written as a lowercase r, is a measure of the strength and direction of the linear relationship between two continuous variables. It was formalised by the British mathematician Karl Pearson in the 1890s, building on earlier ideas from Francis Galton. The coefficient is defined as the ratio of the covariance of two variables to the product of their standard deviations, which is just a technical way of saying it standardises how the two variables vary together so the result always lands on a fixed scale.

That scale runs from −1 to +1. A value of +1 means a perfect positive linear relationship, −1 means a perfect negative one, and 0 means no linear association at all. The sign tells you the direction. When r is positive, the two variables rise together. When r is negative, one rises as the other falls. The size of the number, ignoring the sign, tells you the strength.

How to read the size of r 🔗

There is no single official cut-off for calling a correlation “strong” or “weak,” but a common rough guide treats values around 0.1 as small, 0.3 as moderate, and 0.5 and above as large. Many textbooks describe values between roughly 0.5 and 0.8 as moderate and anything above 0.8 as strong. The right interpretation depends on your field. In physics a correlation of 0.9 might be unremarkable, while in social research a value of 0.4 can be a meaningful finding.

It also helps to separate two ideas that students often confuse: the strength of a correlation and its statistical significance. A correlation can be tiny yet still come out “significant” simply because the sample is huge. The classic illustration is that an r of 0.31 and an r of 0.64 can share the same p-value, yet one is weak and the other is moderate to strong. The p-value tells you how likely the result is to be a fluke; the value of r tells you how big the relationship actually is. Always report and read both.

The coefficient of determination 🔗

If you square the correlation coefficient, you get r², called the coefficient of determination. This number has a very concrete meaning: it is the proportion of the variation in one variable that can be explained by the other. An r of 0.5 gives an r² of 0.25, which means only 25% of the variability is shared, leaving 75% to be explained by other factors. An r of 0.9 gives an r² of 0.81, meaning 81% of the variation is accounted for. Squaring is a useful reality check because it stops people from overstating modest correlations. A relationship that “sounds” decent at r = 0.5 is actually explaining only a quarter of what is going on.

Example calculation using height and weight 🔗

The cleanest way to understand r is to compute it once by hand. Suppose we measure the height (in centimetres) and weight (in kilograms) of six people:

Heights (X): 150, 155, 160, 165, 170, 175
Weights (Y): 50, 56, 58, 62, 67, 72

We will use the computational version of the formula, which avoids working with decimals:

r = [ n·ΣXY − (ΣX)(ΣY) ] ÷ √{ [ n·ΣX² − (ΣX)² ] · [ n·ΣY² − (ΣY)² ] }

Here n is the number of pairs, which is 6. We need five sums. Working through each column:

ΣX (sum of heights) = 150 + 155 + 160 + 165 + 170 + 175 = 975
ΣY (sum of weights) = 50 + 56 + 58 + 62 + 67 + 72 = 365
ΣXY (sum of each height times its weight) = (150×50) + (155×56) + (160×58) + (165×62) + (170×67) + (175×72) = 59,680
ΣX² (sum of squared heights) = 158,875
ΣY² (sum of squared weights) = 22,517

Now we slot these into the formula. First the numerator:

Numerator = (6 × 59,680) − (975 × 365) = 358,080 − 355,875 = 2,205

Next the two pieces of the denominator:

6 × 158,875 − (975)² = 953,250 − 950,625 = 2,625
6 × 22,517 − (365)² = 135,102 − 133,225 = 1,877

Denominator = √(2,625 × 1,877) = √4,927,125 ≈ 2,219.71

Finally:

r = 2,205 ÷ 2,219.71 ≈ 0.99

The result is about 0.99, which is very close to +1. This tells us there is an extremely strong positive linear relationship between height and weight in this small group: taller people are heavier in a way that follows an almost perfectly straight line. Squaring it gives r² ≈ 0.99, meaning roughly 99% of the variation in weight here is explained by height. That figure is unusually high because this is a tidy, made-up dataset. Real human data, with all its messiness, almost never lines up this neatly. A study using actual measurements would more realistically report something moderate, and one such analysis of height and weight returned a correlation of around 0.69.

Limitations of Pearson’s r 🔗

Pearson’s r is powerful, but it is also, in the words of one statistics resource, “much abused.” Knowing where it breaks down is just as important as knowing how to calculate it.

It only detects linear relationships 🔗

The biggest constraint is built into the definition. Pearson’s r measures how well the data fit a straight line. If two variables have a strong but curved relationship, r can be small or even zero, suggesting “no relationship” when in fact there is a clear one. For instance, anxiety and performance often follow an inverted-U shape: a little pressure helps, too much hurts. Pearson’s coefficient would underestimate that association badly. The coefficient being zero only rules out a linear relationship, not every kind of relationship. Plotting your data in a scatterplot before trusting r is essential.

It is highly sensitive to outliers 🔗

Because the calculation is based on means and squared deviations, a single extreme point can drag the result up or down dramatically. One worked example shows the coefficient jumping from 0.54 to 0.71 once a lone outlier was removed, which also changed the explained variance from under a third to over half. Whether to keep or drop such a point is a judgement call: outliers that genuinely belong to the population should usually stay, while those caused by measurement error may be removed, but never simply because they are inconvenient.

It assumes certain conditions about the data 🔗

Pearson’s r works best when several assumptions hold: the variables are measured on a continuous scale, the data come in related pairs, there are no major outliers, and the relationship is linear. Many texts also mention an assumption that the data follow a roughly normal distribution, although the coefficient is fairly robust to moderate departures from normality. When your data are ranks or are heavily skewed, a non-parametric alternative such as Spearman’s rank correlation is usually a safer choice, because it relies on the order of the values rather than their exact size and so resists the pull of outliers.

Correlation is not causation 🔗

This is the limitation that trips up even experienced readers. A high value of r tells you two variables move together; it says nothing about why. A third, hidden variable may be driving both. The textbook example is that ice cream sales and crime rates rise together not because one causes the other, but because hot weather drives both. Establishing genuine cause and effect generally requires a controlled experiment, not a correlation calculated from observational data. Treat a strong r as the start of an investigation, not the conclusion.

When Pearson’s r is the right tool 🔗

Used within its limits, Pearson’s correlation remains one of the most useful summaries in all of statistics. It is ideal when you have two continuous variables, you have plotted them and seen a roughly straight-line pattern, you have checked for distorting outliers, and you are careful to describe the result as an association rather than a cause. Pair the r value with its r², report the sample size, and run a significance test so readers can judge whether the relationship is likely to hold in the wider population. Handled this way, a single number can carry a remarkable amount of honest information.

What do you think? If you found a strong correlation between students’ attendance and their exam marks, what steps would you take before claiming that attending class actually causes better results? And can you think of two variables in everyday life that probably move together only because a third hidden factor links them?

References

1. https://en.wikipedia.org/wiki/Pearson_correlation_coefficient
2. https://statistics.laerd.com/statistical-guides/pearson-correlation-coefficient-statistical-guide.php
3. https://mathspace.co/textbooks/syllabuses/Syllabus-1059/topics/Topic-20583/subtopics/Subtopic-268663/
4. https://pmc.ncbi.nlm.nih.gov/articles/PMC6107969/
5. https://www.scribbr.com/statistics/coefficient-of-determination/
6. https://statisticsbyjim.com/basics/correlations/
7. https://explorable.com/pearson-product-moment-correlation
8. https://medium.com/@anthony.demeusy/pearson-correlation-methodology-limitations-alternatives-part-2-limitations-63c20b21e53b
9. https://www.statisticssolutions.com/pearson-product-moment-correlation/
10. https://academic.oup.com/ckj/article/14/11/2332/6262634
11. https://fastercapital.com/topics/interpreting-pearson-correlation-results.html