When an Adjusted Coefficient Changes
A regression coefficient can change substantially when another predictor is included in the model. It may become larger, smaller, or even switch signs. These changes are important: an unadjusted association and an adjusted association answer different questions. But a coefficient change, by itself, does not identify why the change occurred. In particular, it does not establish whether a variable is a suppressor, a common cause, or something else in the process that generated the data.
This tutorial uses standardized regression coefficients to make those changes easy to compare. A standardized coefficient describes the change in the predicted outcome, in standard deviation units, for a one-standard-deviation increase in a predictor, holding the other included predictors fixed. The numbers below are hypothetical and illustrate regression patterns; they are not results from real studies.
A common cause can make an unadjusted coefficient differ from the association conditional on that cause. Suppression can also produce a large coefficient change, including a sign reversal, without the correlations alone telling us that a common cause is responsible. To distinguish the explanations, regression calculations must be combined with subject-matter knowledge about how variables relate.
A Two-Predictor Calculation
Let \(X\), \(Y\), and \(Z\) be standardized variables, each with mean 0 and standard deviation 1. Write \(r_{XY}\) for the correlation between \(X\) and \(Y\), with the other correlations defined similarly. In a simple linear regression of \(Y\) on \(X\), the slope is \(r_{XY}\). In a multiple regression of \(Y\) on both \(X\) and \(Z\), the standardized coefficient of \(X\) is
The condition \(|r_{XZ}|<1\) means that \(X\) and \(Z\) are not perfectly correlated, so their separate regression coefficients can be calculated. The numerator subtracts the part of the \(X\)-\(Y\) association that is aligned with \(Z\); the denominator accounts for the remaining variation in \(X\) after accounting for \(Z\). This formula is an algebraic result for the fitted linear regression with these correlations. A causal interpretation requires additional justification.
Why Adjustment Can Increase a Coefficient
The formula also gives a useful justification for how suppression can occur. Suppose \(Z\) is positively correlated with \(X\), but has very little correlation with \(Y\). Then the product \(r_{XZ}r_{ZY}\) may be small, while \(Z\) accounts for some variation in \(X\) that is not useful for predicting \(Y\). The adjusted coefficient can then be larger in magnitude than the simple-regression slope.
More directly, the multiple-regression normal equations for standardized variables are
The first equation says that the marginal \(X\)-\(Y\) correlation combines the coefficient for \(X\) with the contribution associated with \(Z\), which is correlated with \(X\). Solving these two equations for \(\beta_X\) gives
Thus the adjusted coefficient can differ in either direction from \(r_{XY}\). This algebra explains the change in coefficients; it does not tell us whether the adjustment is appropriate for a causal question. That depends on the role of \(Z\) and the assumptions behind the analysis.
Worked Example: A Suppression Pattern Increases the Coefficient
Suppose hypothetical standardized measurements have \(r_{XY}=0.30\), \(r_{XZ}=0.60\), and \(r_{ZY}=0.05\). First, the simple-regression slope of \(Y\) on \(X\) is \(0.30\). Calculate the coefficient of \(X\) after adjusting for \(Z\).
Substitute into the formula:
The coefficient increases from \(0.30\) to about \(0.422\). In context, the fitted mean of \(Y\), measured in standard deviation units, is about \(0.422\) standard deviations higher for a one-standard-deviation increase in \(X\), comparing observations with the same \(Z\). The adjusted association is stronger than the unadjusted one.
Here \(Z\) is strongly associated with \(X\) (\(r_{XZ}=0.60\)) but only weakly associated with \(Y\) (\(r_{ZY}=0.05\)). Those facts are consistent with a suppression pattern: adjusting for \(Z\) removes variation in \(X\) that does not align much with variation in \(Y\). They do not establish what caused \(X\), \(Y\), or \(Z\). A causal explanation would need information beyond these correlations.
Common-Cause Confounding Can Reverse a Sign
When \(Z\) causes both \(X\) and \(Y\), part of the unadjusted association between \(X\) and \(Y\) may reflect their shared relationship with \(Z\). If the association between \(X\) and \(Y\) conditional on \(Z\) points in the opposite direction, the marginal and adjusted coefficients can have opposite signs. A sign reversal is possible, but it is not guaranteed whenever confounding is present.
Worked Example: A Common Cause Masks a Negative Conditional Association
Consider hypothetical standardized variables with \(r_{XY}=0.40\), \(r_{XZ}=0.70\), and \(r_{ZY}=0.70\). Suppose subject-matter knowledge supports the following structure: \(Z\) influences both \(X\) and \(Y\), and \(X\) also has a negative direct association with \(Y\). For illustration, a compatible set of linear equations is
where \(Z\), \(U\), and \(V\) are uncorrelated, with variances 1, 0.51, and approximately 0.4941, respectively. These choices make the variables standardized. For example, \(\operatorname{Var}(X)=0.70^2+0.51=1\). The resulting correlations are approximately \(r_{XZ}=0.70\), \(r_{XY}=0.40\), and \(r_{ZY}=0.70\).
The unadjusted slope of \(Y\) on \(X\) is \(r_{XY}=0.40\), which is positive. After adjusting for \(Z\):
The adjusted coefficient is negative, while the marginal coefficient is positive. In the illustrative equations, \(Z\) has a positive influence on both \(X\) and \(Y\), making \(X\) and \(Y\) positively associated overall. Once \(Z\) is held fixed, the negative association represented by the coefficient of \(X\) is visible. This is a common-cause confounding pattern because the causal structure was specified as part of the example, not because the sign reversed. Without the subject-matter information, the three correlations alone would not establish that structure.
The coefficient is in standard deviation units: conditional on \(Z\), a one-standard-deviation increase in \(X\) corresponds to a fitted mean of \(Y\) about \(0.1765\) standard deviations lower. Calling this a causal effect would require the relevant causal assumptions to be justified; the regression calculation alone is not enough.
A Sign Reversal Is Not a Diagnosis
A regression sign reversal can also arise as a suppression pattern. The next hypothetical correlations show such a change. No causal diagram or mechanism is specified, so the example supports a statistical description but not a conclusion about common-cause confounding.
Worked Example: A Suppression Pattern Reverses the Sign
Suppose \(r_{XY}=-0.20\), \(r_{XZ}=0.60\), and \(r_{ZY}=-0.60\). The simple-regression slope of \(Y\) on \(X\) is \(-0.20\). The adjusted coefficient is
The coefficient changes from \(-0.20\) to \(0.25\): the marginal association is negative, while the adjusted association is positive. In standardized units, at the same value of \(Z\), a one-standard-deviation increase in \(X\) corresponds to a fitted mean of \(Y\) that is \(0.25\) standard deviations higher.
This is a sign-reversal pattern in the regression results. Calling it a suppression pattern is a description of how adjustment changes the association; the correlations alone do not establish why that pattern arose. In particular, they cannot determine whether \(Z\) is a common cause, a consequence of another variable, or part of a different causal structure. The structural explanation must be argued separately.
Comparing the Interpretations
| Feature | Suppression pattern | Common-cause confounding pattern |
|---|---|---|
| What is being described? | A change in coefficient magnitude or direction after adjustment | A shared cause of the predictor and outcome that contributes to their marginal association |
| What can regression coefficients show? | The marginal and adjusted associations, including whether one grows or changes sign | A marginal and adjusted association that may differ |
| What else is needed? | For a causal explanation, knowledge of the variable relationships | Subject-matter support for the common-cause structure and the assumptions needed for the causal interpretation |
| Does a sign reversal prove the explanation? | No | No |
The terms need not be treated as rival labels for every data set. A common cause can contribute to a coefficient change that also resembles suppression in the regression arithmetic. The key distinction is that suppression describes the pattern of coefficients, whereas confounding is a claim about causal relationships among variables. Regression output supplies the first kind of information; causal knowledge and assumptions are needed for the second.
Common Mistake
Wrong reasoning: “The coefficient changed sign after adjustment, so \(Z\) must be a confounder and the adjusted coefficient must be the causal effect.”
Correction: A sign reversal shows that the marginal and conditional linear associations differ. It does not identify the causal role of \(Z\), prove that the chosen adjustment is appropriate, or establish that the adjusted coefficient is causal. Check the causal structure and the assumptions relevant to the question before giving the coefficient a causal interpretation.
The same caution applies when an adjusted coefficient gets larger. That change is consistent with suppression, but its size alone does not prove that \(Z\) is a suppressor in a causal sense or that adjustment improves the answer to every research question. The useful habit is to report which association is being described, state what was adjusted for, and separate the regression result from its causal interpretation.
Check Your Understanding
Use the formula and interpretations in this tutorial to answer the following questions.
- If \(r_{XY}=0.25\), \(r_{XZ}=0.50\), and \(r_{ZY}=0.10\), calculate the standardized coefficient of \(X\) after adjusting for \(Z\). Is its magnitude larger or smaller than the simple-regression slope?
- In the common-cause example, why can the marginal coefficient be positive while the adjusted coefficient is negative?
- What does a sign reversal establish about the regression associations, and what does it not establish about causation?
- In the third worked example, interpret the adjusted coefficient of \(0.25\) in standardized units.
- What additional kind of information is needed to support the claim that a variable is a common cause?