Updating Partial Correlations When Adding a Control

Why an Update Formula Is Useful

A partial correlation measures the association between two variables after their linear relationships with a specified set of controls have been removed. When a new variable joins that set, the partial correlation generally changes. The update formula makes that change calculable from four correlations already adjusted for the existing controls: the partial correlation of the target pair and each target variable’s partial correlation with the new control.

This recursive calculation is useful when controls are added sequentially or when one needs to compare partial correlations for nested control sets. It also makes an important requirement explicit: every input to a single update must be calculated using the same existing control set.

Partial Correlation and the Update Formula

Definition: Let $X,Y,W$ be random variables with finite second moments, and let $Z$ denote a set of control variables. Regress each variable linearly on the variables in $Z$, and let $R_{X\mid Z},R_{Y\mid Z},R_{W\mid Z}$ be the resulting residuals. The partial correlation of $X$ and $Y$ given $Z$, written $\rho_{XY\cdot Z}$, is the correlation of $R_{X\mid Z}$ and $R_{Y\mid Z}$, provided both residual variances are positive. If $Z$ is empty, this is the ordinary correlation.

Here linear regression means population least-squares projection onto the linear span of the controls (with an intercept when needed). No Gaussian assumption is required. For a nonempty $Z$, assume its covariance matrix is nonsingular so the projection coefficients are well-defined. The residuals have mean zero when an intercept is included.

Theorem: Suppose the residual variances needed to define the partial correlations below are positive. Then adding $W$ to the conditioning set gives $$ \rho_{XY\cdot (Z,W)} = \frac{\rho_{XY\cdot Z}-\rho_{XW\cdot Z}\rho_{YW\cdot Z}} {\sqrt{\bigl(1-\rho_{XW\cdot Z}^{\,2}\bigr) \bigl(1-\rho_{YW\cdot Z}^{\,2}\bigr)}}. $$ In particular, the denominator is positive when the residual of $W$ after controlling for $Z$ has positive variance and the residual variances of $X$ and $Y$ after controlling for $(Z,W)$ are positive.

Worked Example: Adding a Control With No Existing Controls

Suppose the correlations are $\rho_{XY}=0.60$, $\rho_{XW}=0.50$, and $\rho_{YW}=0.20$. Find the partial correlation of $X$ and $Y$ given $W$.

With $Z$ empty, the three inputs are ordinary correlations. The numerator is $0.60-(0.50)(0.20)=0.50$. The denominator is $$ \sqrt{(1-0.50^2)(1-0.20^2)} =\sqrt{(0.75)(0.96)} =\sqrt{0.72} \approx 0.8485. $$ Thus $$ \rho_{XY\cdot W}=\frac{0.50}{0.8485}\approx 0.5893. $$

The inputs are compatible with a nonsingular three-variable correlation matrix: its determinant is $1+2(0.60)(0.50)(0.20)-0.60^2-0.50^2-0.20^2=0.47>0$.

Derivation by Residualizing the New Control

The update follows from one more linear projection. First remove the effects of $Z$ from all three variables, and write the residuals as $U=R_{X\mid Z}$, $V=R_{Y\mid Z}$, and $T=R_{W\mid Z}$. Controlling for both $Z$ and $W$ is equivalent to controlling for $Z$ first and then removing the linear component associated with $T$. This follows because $T$ is the part of $W$ orthogonal to the linear span of $Z$.

The residuals after this additional projection are $U^*=U-\operatorname{Cov}(U,T)T/\operatorname{Var}(T)$ and $V^*=V-\operatorname{Cov}(V,T)T/\operatorname{Var}(T)$. Expanding their covariance gives $$ \operatorname{Cov}(U^*,V^*) = \operatorname{Cov}(U,V) -\frac{\operatorname{Cov}(U,T)\operatorname{Cov}(V,T)} {\operatorname{Var}(T)}. $$ For example, the two subtracted cross-terms each contribute $-\operatorname{Cov}(U,T)\operatorname{Cov}(V,T)/\operatorname{Var}(T)$, while the term involving $T^2$ contributes one copy with a positive sign. The net result is one negative copy.

For the variance of $U^*$, the same expansion gives $$ \operatorname{Var}(U^*) = \operatorname{Var}(U) -\frac{\operatorname{Cov}(U,T)^2}{\operatorname{Var}(T)}. $$ There is an identical expression for $\operatorname{Var}(V^*)$. To express these quantities in correlations, divide the covariance formula by $\sqrt{\operatorname{Var}(U)\operatorname{Var}(V)}$, and divide each variance formula by its original variance. By definition, $\rho_{XY\cdot Z}=\operatorname{Corr}(U,V)$, $\rho_{XW\cdot Z}=\operatorname{Corr}(U,T)$, and $\rho_{YW\cdot Z}=\operatorname{Corr}(V,T)$. Therefore $$ \frac{\operatorname{Cov}(U^*,V^*)} {\sqrt{\operatorname{Var}(U)\operatorname{Var}(V)}} = \rho_{XY\cdot Z}-\rho_{XW\cdot Z}\rho_{YW\cdot Z}, $$ and $$ \frac{\operatorname{Var}(U^*)}{\operatorname{Var}(U)} =1-\rho_{XW\cdot Z}^{\,2}, \qquad \frac{\operatorname{Var}(V^*)}{\operatorname{Var}(V)} =1-\rho_{YW\cdot Z}^{\,2}. $$ Dividing the standardized covariance by the square root of the two standardized variances gives the stated formula. Positive residual variances ensure that this division is valid. $\square$

Worked Example: Updating After Existing Controls

Suppose that after controlling for $Z$, the relevant partial correlations are $\rho_{XY\cdot Z}=0.40$, $\rho_{XW\cdot Z}=0.30$, and $\rho_{YW\cdot Z}=-0.20$. Find $\rho_{XY\cdot (Z,W)}$.

The numerator is $0.40-(0.30)(-0.20)=0.46$. The denominator is $$ \sqrt{(1-0.30^2)(1-(-0.20)^2)} = \sqrt{(0.91)(0.96)} = \sqrt{0.8736} \approx 0.9347. $$ Therefore $$ \rho_{XY\cdot (Z,W)} = \frac{0.46}{0.9347} \approx 0.4921. $$

The negative sign in $\rho_{YW\cdot Z}$ matters: subtracting the product of the two correlations adds $0.06$ to the numerator. A useful check on the inputs is that their $3\times3$ correlation matrix has determinant $1+2(0.40)(0.30)(-0.20)-0.40^2-0.30^2-(-0.20)^2=0.662>0$.

Applying the Formula Recursively

The formula can be applied repeatedly, one control at a time. To add a second control after a first, first calculate all three partial correlations needed for that next update conditional on the first control. The output of one update is not, by itself, enough: the two partial correlations involving the next control must also be available under the current conditioning set.

Worked Example: Adding Two Controls Sequentially

Consider four variables $X,Y,W_1,W_2$ with every pairwise correlation equal to $0.20$. Their correlation matrix is positive definite: its eigenvalues are $0.80$ (with multiplicity three) and $1.60$. First find the correlation of $X$ and $Y$ given $W_1$.

The update gives $$ \rho_{XY\cdot W_1} = \frac{0.20-(0.20)(0.20)} {\sqrt{(1-0.20^2)(1-0.20^2)}} = \frac{0.16}{0.96} = \frac{1}{6}. $$

To add $W_2$, the other two required inputs, also conditional on $W_1$, are $$ \rho_{XW_2\cdot W_1} = \rho_{YW_2\cdot W_1} = \frac{0.20-(0.20)(0.20)}{0.96} = \frac{1}{6}. $$ Hence $$ \rho_{XY\cdot (W_1,W_2)} = \frac{\frac16-\frac1{36}} {\sqrt{(1-\frac1{36})(1-\frac1{36})}} = \frac{\frac5{36}}{\frac{35}{36}} = \frac17 \approx 0.1429. $$

The calculation illustrates the recursion: the second update uses correlations already adjusted for $W_1$, not the original pairwise correlations.

What the Formula Does—and Does Not—Say

The numerator removes the component of the existing residual association that is accounted for by the new control. The denominator rescales by the residual standard deviations of $X$ and $Y$ after that control is removed. Both parts are necessary: subtracting the product alone gives a covariance on the old residual scale, not the new correlation.

The update can change the magnitude of the partial correlation, and it can change its sign. The formula does not say that adding a control must weaken an association, nor does it establish a causal interpretation. It describes a change in a linear residual correlation; interpretation beyond that requires assumptions specific to the study and model.

The same algebra applies to sample partial correlations when the quantities are computed from one common sample using consistent linear regressions and the relevant sample residual variances are positive. In that setting, rounding intermediate correlations too aggressively can produce small discrepancies between a direct calculation and successive updates.

Common Mistakes

Mixing conditioning sets. Using $\rho_{XY\cdot Z}$ with ordinary correlations $\rho_{XW}$ and $\rho_{YW}$ does not implement the update. All three inputs must be conditional on the same $Z$.
Dropping the denominator. The numerator is not the updated correlation by itself. The denominator standardizes the residual covariance after adding $W$.
Using the wrong sign. The numerator is $\rho_{XY\cdot Z}-\rho_{XW\cdot Z}\rho_{YW\cdot Z}$. If the product is negative, subtracting it increases the numerator.
Ignoring degeneracy. If a relevant residual variance is zero, a partial correlation involving that residual is undefined, and the update cannot be used as written.

Key takeaway: To add one control $W$, use the partial correlations of $X,Y$, $X,W$, and $Y,W$ all conditional on the existing controls $Z$. Subtract the product of the latter two, then divide by the square root of their two residual-variance factors.

Check Your Understanding

Use the update formula and show the substitutions in each calculation.

What three partial correlations are required to update $\rho_{XY\cdot Z}$ when adding $W$?
If $\rho_{XY\cdot Z}=0.50$, $\rho_{XW\cdot Z}=0.40$, and $\rho_{YW\cdot Z}=0.25$, calculate $\rho_{XY\cdot (Z,W)}$.
Why must the correlations involving $W$ use the same conditioning set $Z$ as the correlation of $X$ and $Y$?
Which residual variances must be positive for the updated partial correlation to be defined?
Explain why the numerator alone is not generally the updated partial correlation.