A partial correlation measures the strength of the relationship between two variables after removing the influence of one or more other variables. It answers a specific question: if we strip away the effect of a third factor, do these two things still relate to each other? The result is a single number, ranging from -1 to +1, just like a standard correlation coefficient.
Why Standard Correlation Can Be Misleading
A standard (zero-order) correlation between two variables captures everything that connects them, including indirect links through other variables. That can be a problem. Imagine you find a positive correlation between the number of hours people work and their exercise frequency. That might seem surprising, but both could be driven by age: younger adults tend to work more hours and exercise more. The correlation you see between work and exercise might partly, or entirely, reflect the fact that both relate to age.
Partial correlation handles this by statistically removing age from the picture. It asks: among people of the same age, is there still a relationship between work hours and exercise? If the partial correlation drops to near zero, age was doing most of the work. If it stays strong, the relationship between work and exercise is genuine, independent of age.
How It Works Conceptually
Think of three overlapping circles in a Venn diagram. Each circle represents one variable’s variance. The overlap between any two circles represents their shared variance, which corresponds to their squared correlation. When you calculate a partial correlation between variables X and Y while controlling for Z, you’re looking only at the overlap between X and Y that doesn’t also overlap with Z. You’re isolating the unique connection between X and Y.
Mathematically, the process boils down to the regular correlations among all three variables. You need the standard correlation between X and Y, the correlation between X and the control variable, and the correlation between Y and the control variable. The partial correlation formula combines these three numbers to produce the adjusted relationship. You don’t need to run a regression or do anything complicated by hand. If you already have the three standard correlations, the partial correlation is a direct function of them.
A Concrete Example
Researchers at SUNY Oswego examined whether verbal SAT scores predict college GPA after accounting for high school GPA. The standard correlation between verbal SAT and college GPA was moderate. But here’s the key question: is SAT score actually contributing something, or is it just picking up on the same academic ability already captured by high school GPA?
After partialling out high school GPA, the correlation between verbal SAT and college GPA dropped to .5751 and was no longer statistically significant. The takeaway: verbal SAT scores appeared unrelated to college GPA once high school performance was accounted for. High school GPA was doing the heavy lifting all along. This is exactly the kind of insight partial correlation is designed to reveal.
Partial vs. Semi-Partial Correlation
These two get confused constantly, and the difference matters. A partial correlation removes the control variable from both variables of interest. A semi-partial (also called “part”) correlation removes the control variable from only one of them.
Say you’re looking at how X1 relates to Y while controlling for X2. The partial correlation strips X2’s influence from both X1 and Y, then correlates the leftovers. The semi-partial correlation strips X2’s influence from X1 only, then correlates that cleaned-up version of X1 with the full, original Y.
This distinction changes what the squared value tells you. The squared semi-partial correlation tells you how much total variance in Y is uniquely explained by X1. The squared partial correlation tells you how much of Y’s remaining unexplained variance (after other variables are accounted for) is explained by X1. Because the partial correlation uses a smaller denominator (only the leftover variance), partial correlations will always be equal to or larger than the corresponding semi-partial correlations.
In practice, use the semi-partial when you want to know a variable’s unique contribution to the total picture. Use the partial when you want to know the strength of a relationship in a world where confounders don’t exist.
What the Data Needs to Look Like
Partial correlation shares the same assumptions as standard Pearson correlation, with one addition. All variables, including the ones you’re controlling for, need to be continuous. The relationships between all pairs of variables should be linear. And the data should be roughly normally distributed, though partial correlation is fairly robust to mild violations of this with larger sample sizes.
If your variables are ranked or ordinal rather than continuous, you’d use a Spearman-based approach instead. And if the relationship between your variables is curved rather than straight, partial correlation will underestimate or miss the true connection.
Statistical Significance and Degrees of Freedom
Testing whether a partial correlation is statistically significant works similarly to testing a standard correlation, with one adjustment. The degrees of freedom change. For a standard correlation, degrees of freedom equal n – 2, where n is your sample size. For a partial correlation, you subtract an additional number for each variable you control: n – 2 – c, where c is the number of control variables.
This means controlling for more variables costs you statistical power. Each additional control variable removes one degree of freedom, making it harder to reach significance. With a small sample, controlling for several variables at once can leave you with too few degrees of freedom to detect a real relationship. This is a practical constraint worth keeping in mind: partial correlation works best when your sample is reasonably large relative to the number of variables you’re adjusting for.
When Partial Correlation Is Most Useful
Partial correlation shines in observational research where you can’t run an experiment. In health research, you might want to know whether a dietary habit relates to blood pressure independent of body weight. In psychology, you might ask whether anxiety and academic performance are linked after removing the effect of sleep quality. In economics, you might examine whether education level correlates with income after controlling for years of work experience.
It’s also a useful diagnostic tool when building regression models. If a variable has a strong zero-order correlation with your outcome but a weak partial correlation after controlling for other predictors, that variable probably isn’t contributing much unique information. It’s redundant with what the other predictors already capture.
One important limitation: partial correlation controls for confounders statistically, not experimentally. It can only remove the linear influence of the control variable. If the confounding relationship is nonlinear, or if there are unmeasured confounders you haven’t controlled for, the partial correlation may still be misleading. It reduces confounding, but it doesn’t eliminate it the way a randomized experiment would.