Partial Correlation Calculator
Calculate the partial correlation between two variables while controlling for a third. Enter three paired data columns or three pairwise Pearson correlations, and get the partial correlation coefficient with interpretation.
Find the Relationship After Controlling for a Third Variable
Partial Correlation Formula
Where:
- \(r_{XY \cdot Z}\) — partial correlation of X and Y, controlling for Z
- \(r_{XY}\) — Pearson correlation between X and Y
- \(r_{XZ}\) — Pearson correlation between X and Z
- \(r_{YZ}\) — Pearson correlation between Y and Z
The formula removes the part of the X–Y relationship that is explained by their separate linear associations with Z. The result lies between −1 and 1, just like an ordinary correlation coefficient.
Partial Correlation – Calculation Example
Example: eight observations with X = 1; 2; 3; 4; 1; 2; 3; 4, Y = 3; 3; 5; 9; 4; 8; 10; 10 and a binary control variable Z = 0; 0; 0; 0; 1; 1; 1; 1 (the calculator's default values).
First compute the three pairwise Pearson correlations:
\[ r_{XY} = 0.7785, \qquad r_{XZ} = 0.0000, \qquad r_{YZ} = 0.5222 \]
Then apply the partial correlation formula:
\[ r_{XY \cdot Z} = \frac{0.7785 - (0)(0.5222)}{\sqrt{(1 - 0^{2})(1 - 0.5222^{2})}} = \frac{0.7785}{0.8528} \approx 0.9129 \]
Result: the partial correlation is approximately 0.9129. Controlling for Z strengthens the X–Y association from 0.78 to 0.91, because Z is related to Y but not to X in this dataset.
How to Use the Calculator
- Choose the Input method: “Three paired data columns” if you have raw measurements, or “Three pairwise correlations” if you already know the three Pearson coefficients.
- With raw data, paste the values into Variable X, Variable Y and Control variable Z. Separate values with spaces, semicolons or new lines; decimal commas are accepted. Keep all three columns in the same observation order and of equal length (4 to 2,000 values).
- With coefficients, enter Correlation of X and Y, Correlation of X and Z and Correlation of Y and Z, each between −1 and 1.
- Read the result: the partial correlation of X and Y controlling for Z, the three pairwise correlations, the number of paired observations (data mode), and a plain-language interpretation.
What Is Partial Correlation?
A partial correlation measures the strength of the linear relationship between two variables after removing the influence of one or more control variables. In the one-control-variable case handled by this calculator, it answers the question: how strongly are X and Y related once the variation each of them shares with Z has been factored out?
This matters because an ordinary correlation can be misleading. Two variables may appear correlated simply because both are driven by a third variable — a classic confounding situation. For example, ice cream sales and drowning incidents correlate over the year, but both are driven by temperature. Partial correlation helps you see whether a direct association remains after accounting for the shared cause.
The result is interpreted like any correlation coefficient: values near +1 or −1 indicate a strong positive or negative linear association, values near 0 indicate little or no association. A partial correlation can be larger or smaller in magnitude than the raw correlation. It becomes larger when Z masks part of the true relationship (as in the example above), and smaller when the raw correlation was mostly an artifact of Z.
Note that partial correlation, like any correlation, measures association, not causation. Removing Z's linear influence reduces one possible source of spurious association, but it cannot establish the direction of causality or rule out other confounders. For modelling several predictors jointly, the multiple linear regression calculator is a natural next step.
Assumptions, Limitations and Tips
- Linearity. The formula is based on Pearson correlation, so it captures only linear relationships. If associations are monotonic but curved, consider rank-based measures such as the Spearman rank correlation.
- One control variable. This calculator controls for a single variable Z. With several controls, the same idea generalizes via the precision (inverse correlation) matrix, but the arithmetic becomes matrix-based.
- Consistent coefficients. In coefficient mode, the three correlations must be mutually compatible — the calculator rejects combinations that cannot form a valid correlation matrix.
- Undefined cases. The result is undefined when X or Y is perfectly (or nearly perfectly) correlated with Z, because the denominator approaches zero.
- Sample size. Small samples produce unstable estimates. With few observations, a partial correlation of moderate size can easily arise by chance; a statistical power check helps you judge whether your sample can detect the effect you expect.
- Complete pairs. Each row must contain a value for X, Y and Z. Missing values in one column break the pairing, so clean your data first.
Frequently Asked Questions
How do I calculate a partial correlation by hand?
First compute the three Pearson correlations \(r_{XY}\), \(r_{XZ}\) and \(r_{YZ}\) from your paired data. Then plug them into \( r_{XY \cdot Z} = \frac{r_{XY} - r_{XZ} r_{YZ}}{\sqrt{(1 - r_{XZ}^2)(1 - r_{YZ}^2)}} \). Multiply the two correlations with Z, subtract from \(r_{XY}\), and divide by the square root of the two remaining terms. The calculator does all steps at once and shows the pairwise coefficients it used.
Is my partial correlation statistically significant?
The calculator does not run a significance test, but you can do one yourself with \( t = r_{XY \cdot Z}\sqrt{\frac{n-3}{1 - r_{XY \cdot Z}^2}} \), where \(n\) is the number of complete observations. Compare \(t\) to a t-distribution with \(n - 3\) degrees of freedom. Roughly, with \(n = 50\), a partial correlation beyond about \(\pm 0.28\) is significant at the 5% level (two-tailed).
What counts as a strong or good partial correlation?
There is no universal "good" value — it depends on your field. As rough rules of thumb used in behavioral sciences, magnitudes near 0.1 are considered small, near 0.3 moderate, and 0.5 or above strong. In physics or engineering contexts, much weaker correlations can still matter. Judge the value against your sample size: a small coefficient from a large sample may be meaningful, while a large one from a handful of observations may be pure chance.
Can partial correlation be larger than the ordinary correlation, or even flip sign?
Yes to both. When Z suppresses part of the true X–Y relationship, the partial correlation comes out larger in magnitude than \(r_{XY}\). A sign flip (a negative partial correlation from a positive raw one) signals classical suppression — the control variable reveals a relationship opposite to what the raw correlation suggested. Both are legitimate results, not errors, though they usually deserve a careful interpretation of how Z relates to X and Y.
What is the difference between partial and semi-partial (part) correlation?
Partial correlation removes the influence of Z from both X and Y, then correlates what remains. Semi-partial correlation removes Z from only one variable, say Y, while leaving X untouched. Use partial correlation when you ask "is X related to Y beyond Z for both?" and semi-partial when you ask "does X explain the part of Y that Z does not account for?" This calculator computes the partial version.
Why did the calculator reject my three correlation coefficients?
In coefficient mode, the three values must be mutually consistent: they must form a valid correlation matrix, meaning \( |r_{XY} - r_{XZ}\, r_{YZ}| \) cannot exceed \( \sqrt{(1 - r_{XZ}^2)(1 - r_{YZ}^2)} \). If it does, no real dataset could produce that combination. Common causes are mixing up which coefficient belongs to which pair, or taking correlations from different samples or studies that did not share the same observations.
Can I control for more than one variable?
This calculator handles a single control variable Z. With several controls, you need the precision matrix approach: invert the full correlation matrix of all variables, negate the off-diagonal cells, and rescale them back to correlations. In practice, most people instead run a multiple linear regression and look at the standardized coefficients or partial \(R^2\) values, which generalize the same idea and provide confidence intervals and p-values.