Why Correlated Predictors Can Shift a Slope

Why Two Slopes Can Differ

A simple regression slope summarizes how the outcome varies with one predictor, without accounting for other predictors. A multiple regression slope describes how the outcome varies with that predictor when the other included predictor is held fixed. If the two predictors are correlated, these slopes need not be the same: changes in one predictor tend to come along with changes in the other, and the other predictor may also be related to the outcome.

The difference is not automatically a sign that either calculation is wrong. Each slope answers a different descriptive question. To understand how they differ, consider a two-predictor model and keep track of two ingredients: how strongly the predictors move together, and how the second predictor is related to the outcome after accounting for the first.

Use the additive model

Here, \(\beta_1\) and \(\beta_2\) are the adjusted slopes in the population model, and \(\varepsilon\) represents the part of the outcome not accounted for by the predictors. The additive form means that each predictor has a constant slope across values of the other predictor. It does not include a product term such as \(X_1X_2\).

For the population relationship below, assume the conditional mean of the error is zero: \(E(\varepsilon\mid X_1,X_2)=0\). Also assume the predictors are not perfectly linearly related, so the two separate adjusted slopes can be identified. These are model conditions, not guarantees that the model is a good description of any particular data set. Usual regression inference may require additional conditions, such as independent observations.

How the Predictor Relationship Enters the Slope

The marginal slope in the population is \(\operatorname{Cov}(X_1,Y)/\operatorname{Var}(X_1)\), when \(X_1\) has nonzero variance. Substitute the two-predictor model into the covariance in the numerator. Covariance distributes over sums, and a constant has zero covariance with \(X_1\):

Under the stated conditional-mean assumption, \(\operatorname{Cov}(X_1,\varepsilon)=0\). Dividing by \(\operatorname{Var}(X_1)\) gives the slope relationship:

The covariance-to-variance ratio is the slope of the linear regression of \(X_2\) on \(X_1\). Thus, the marginal slope combines the adjusted slope for \(X_1\) with a contribution from \(X_2\): that contribution depends on how \(X_2\) changes, on average, with \(X_1\), and on the adjusted slope for \(X_2\).

For a fitted sample, the corresponding relationship is exact when all three slopes come from ordinary least squares on the same observations, with an intercept in each model. If \(b_1\) and \(b_2\) are the fitted adjusted slopes and \(c\) is the sample slope from regressing \(X_2\) on \(X_1\), then

This sample identity follows because the multiple-regression residuals have sample covariance zero with each included predictor. Therefore, in the sample covariance of \(X_1\) with \(Y\), the residual contributes zero; the fitted part contributes \(b_1\operatorname{Var}(X_1)+b_2\operatorname{Cov}(X_1,X_2)\). Dividing by the sample variance of \(X_1\) yields the stated identity.

Reading the Direction and Size of the Shift

The formula gives a direct way to reason about the difference. The marginal slope equals the adjusted slope plus a product. If the predictors are positively correlated, the slope of \(X_2\) on \(X_1\) is positive; if they are negatively correlated, it is negative. The adjusted slope for \(X_2\) determines how that predictor’s co-movement contributes to the outcome trend. Both ingredients matter.

What Correlation Does—and Does Not—Tell You

Correlation alone is not enough to predict a slope shift. If \(X_1\) and \(X_2\) are uncorrelated in the population, the covariance term is zero, so the marginal and adjusted slopes for \(X_1\) agree under the stated model conditions. They also agree if \(\beta_2=0\), even when the predictors are correlated: in that case, the model assigns no linear outcome contribution to \(X_2\) after accounting for \(X_1\). In other cases, the shift depends on the product of the predictor relationship and the adjusted outcome slope for \(X_2\).

The model conditions matter. If the conditional mean is not adequately described by the additive form, or if the error is related to the predictors, the population formula may not describe the situation as stated. In practice, data can reveal patterns that raise concerns about linearity or unusual observations, but a slope calculation alone does not verify the model assumptions. Nor does holding a measured predictor fixed remove every possible source of difference between groups.

Common Mistake

A common mistake is to say, “The predictors are correlated, so the marginal slope must differ from the adjusted slope.” Correlation by itself does not guarantee a difference. If the second predictor has adjusted slope zero, its contribution is zero; if the predictors have zero covariance, their co-movement contributes zero under the model. The correction is to examine both factors in the formula, not correlation alone.

Another mistake is to interpret the shift as proof that one coefficient is biased or that adjustment has uncovered a causal effect. The slopes describe different associations under a specified model. Whether a causal interpretation is justified depends on the study design and assumptions beyond this arithmetic.