Partial Correlation (linear, controlling for covariates)
VerifiedAdvanced & specialized
Run this test straight away on a free built-in teaching dataset — no data of your own needed — or bring your own. Either opens the guided workspace: variable setup, assumption diagnostics, results with effect sizes and confidence intervals, figures, and APA-ready reporting.
Loading teaching datasets…
Or use your own dataset
Loading your datasets…
Measures the LINEAR association between two continuous variables x and y after linearly removing the effect of one or more covariates z from both.
Answers the question: 'What is the relationship between x and y for cases equated on z?' Reports partial r, Fisher-z 95% CI, t, p, partial r², and the zero-order (unadjusted) r for comparison — when partial and zero-order r diverge substantially, the data exhibit a confounding (or suppression) pattern.
Worked example
Do ice-cream sales relate to drownings once temperature is controlled?
Across 100 days, ice-cream sales, drownings and temperature were recorded. A partial correlation removes temperature to test the direct sales–drownings link.
The zero-order correlation (r = .58) vanished once temperature was partialled out, r = .03, p = .78 — the association was entirely confounded by heat.
After controlling for temperature, the partial correlation between ice-cream sales and drownings was non-significant, r = .03, p = .78 (zero-order r = .58).
Try it yourself: Load this ready-made sample and follow the run above.
When to use it
- Confounder removal — controlling for a known third variableAn economist observes a strong positive Pearson r between ice-cream sales and drowning deaths across US cities.
- Multiple-covariate adjustmentA psychologist measures impulsivity, risk-taking, and controls for age, sex, and IQ in 250 adolescents.
- Suppressor variable detectionA workplace researcher finds zero zero-order r between training hours and job performance.
When NOT to — use instead
- Categorical covariateLinear partialling on a nominal covariate isn't meaningful — use ANCOVA which dummy-codes the factor properly. → One-Way ANCOVA (covariate-adjusted)
- Multiple predictors of interest (not just covariates to remove)When you have multiple SUBSTANTIVE predictors of Y, fit a multiple regression — partial r is for one X–Y question at a time. → Linear Regression \u2014 OLS (continuous y + predictors)
- Ranked / ordinal dataPearson partial assumes interval x and y; for ordinal data use Spearman with the same residualisation idea. → Spearman's Rank-Order Correlation (monotonic / nonparametric)
- Strong non-linear X-Y relationshipPartial correlation removes LINEAR effects of the covariate. → GAM \u2014 Generalized Additive Model (non-linear spline regression)
Hypotheses
Parameter tested: population partial correlation ρ_{xy·z}
Assumptions (and what to do if they fail)
Check: Shapiro-Wilk on e_x and e_y.
If violated: Non-normal residuals make p-values / CI for small-n partial r unreliable.
Check: Breusch-Pagan p on e_y ~ e_x.
If violated: Heteroscedasticity inflates SE of the partial r; CI coverage and p-value accuracy suffer.
Check: Count Mahalanobis-extreme cases in the residual space.
If violated: A single influential case can dramatically inflate or attenuate the partial r.
Ready to run a Partial Correlation (linear, controlling for covariates) on your own data?
Guided setup, automatic assumption checks, effect sizes, figures and an APA write-up.
Run this test →