How Additional Controls Change Partial Correlation

Why the Effect of Adding Controls Is Not Predictable From Their Number

A partial correlation describes the association between two variables after linear adjustment for a specified set of controls. Adding another control changes both the residual covariance between the target variables and their residual variances. Consequently, a partial correlation can shrink in magnitude, reverse sign, or remain close to its original value. The number of controls alone does not determine which outcome occurs.

The useful calculation is a covariance-block update. It separates the part of the target association removed by the added control from the changes in residual variation that also affect the correlation’s denominator. All calculations below concern population covariances and linear residuals; no Gaussian assumption is required.

The Covariance-Block Update

First remove the controls already in the model. Denote the resulting residuals for \(X\) and \(Y\) by \(R_X\) and \(R_Y\). Suppose a new vector of controls \(W\) is to be added. Residualize \(W\) on the existing controls as well, and call that residual vector \(R_W\). Write

and let \(u=\operatorname{Cov}(R_X,R_W)\) be a row vector and \(v=\operatorname{Cov}(R_W,R_Y)\) a column vector. Assume \(Q\) is positive definite and the residual variances after adding \(W\) are positive.

Proof. By construction, \(R_X\), \(R_Y\), and \(R_W\) are each uncorrelated with the existing controls. Therefore, projecting \(X\) on the existing controls and then adding \(W\) is equivalent to projecting \(R_X\) on \(R_W\); the corresponding statement holds for \(Y\). The least-squares coefficient row for projecting \(R_X\) on \(R_W\) is \(uQ^{-1}\), because its covariance with the projection error must be zero. Thus the new residual is \(R_X-uQ^{-1}R_W\). Likewise, the new residual for \(Y\) is \(R_Y-v^{\mathsf T}Q^{-1}R_W\).

Taking the covariance of these two residuals and using \(\operatorname{Var}(R_W)=Q\) gives

The two middle terms are scalars and are equal, so one is canceled by the final term. Taking each residual’s variance in the same way gives \(v_X-uQ^{-1}u^{\mathsf T}\) and \(v_Y-v^{\mathsf T}Q^{-1}v\). Dividing the updated covariance by the product of the updated residual standard deviations proves the correlation formula. \(\square\)

For one added control, write \(q=\operatorname{Var}(R_W)>0\), \(u=\operatorname{Cov}(R_X,R_W)\), and \(v=\operatorname{Cov}(R_W,R_Y)\). The formulas simplify to

The covariance adjustment is the product of the two target-control covariances divided by the control variance. It can be positive, negative, or zero. Meanwhile, each residual variance is reduced or unchanged. These two features jointly determine the new correlation.

When Does the Sign or Magnitude Change?

Justification. A correlation has the same sign as its covariance because its denominator is positive. The old covariance is \(c\), and the new one is \(c-uv/q\); their signs differ precisely when their product is negative. For the magnitude condition, the old and new absolute correlations are nonnegative, so comparing them is equivalent to comparing their squares. Squaring the two correlation formulas gives the displayed inequality, with positive denominators by assumption. Equality means the magnitude is unchanged; reversing the inequality means it increases. \(\square\)

This result highlights a key distinction: a reduction in residual covariance does not by itself establish attenuation of the correlation. The denominator changes too. In particular, if the adjustment makes both residual variances much smaller, a smaller covariance can still correspond to a similar or larger absolute correlation.

How to Interpret the Change

The block calculation is useful because it identifies what the added controls contribute beyond the controls already present. In the scalar case, \(u\) and \(v\) are the associations of the added control with the two target residuals. If they have the same sign, their product is positive and the target residual covariance decreases. If they have opposite signs, the product is negative and that covariance increases. Neither observation alone settles the change in partial correlation, since the residual variances also change.

With several added controls, the term \(uQ^{-1}v\) accounts for their joint covariance structure. It is not generally correct to calculate each control’s effect in isolation and simply add those effects: correlated controls are combined through \(Q^{-1}\). The block update applies to the whole set at once and also makes clear which quantities must be evaluated after the existing controls are removed.

Partial correlation is a description of linear association after adjustment, not a causal effect. A change after adding controls means that the adjusted linear association differs for the two control sets; it does not, by itself, show that a control caused the change or that one adjustment is preferable. The choice of controls must come from the scientific question and study design.

Common Mistake

Wrong reasoning: “Adding a control removes variation, so the partial correlation must get smaller.” This confuses residual variances with correlation magnitude and overlooks the covariance adjustment. The update reduces each residual variance, but the correlation divides the updated covariance by the square root of their product. Depending on \(u\), \(v\), and \(Q\), the result can attenuate, reverse sign, increase in magnitude, or change very little. Calculate the covariance and both residual variances before interpreting the updated correlation.