Predicting the Direction of Omitted-Variable Bias

Why the Direction of Bias Matters

A regression coefficient can change when a relevant predictor is left out of the model. The change is not automatically upward or downward: it depends on how the omitted predictor relates to the predictor that remains and how it relates to the outcome after accounting for that included predictor. Knowing the direction helps interpret a reduced model, but it does not by itself establish that either coefficient is a causal effect.

For example, suppose an analyst estimates the association between a treatment and a health outcome but does not include a baseline characteristic related to both treatment assignment and the outcome. The treatment coefficient from the reduced model may combine the treatment’s conditional association with the outcome and part of the omitted characteristic’s association. The sign of that added component can be reasoned through systematically.

The Sign Rule and Its Conditions

Let $X$ be the predictor whose coefficient is of interest, $Z$ a predictor that might be omitted, and $Y$ the outcome. Begin with the following additive population regression model:

Y=\beta_0+\beta_1X+\beta_2Z+\varepsilon

Assume the conditional mean of the error is zero given both predictors, $E(\varepsilon\mid X,Z)=0$, and that the outcome relationship is adequately represented by this additive linear model. In particular, the model has no $X Z$ interaction. Compare the coefficient $\beta_1$ from the model including both predictors with the population least-squares slope $\gamma_1$ from a model that regresses $Y$ on $X$ alone.

Result: Under the stated additive model and error condition, the reduced-model slope satisfies $$ \gamma_1-\beta_1=\beta_2\frac{\operatorname{Cov}(X,Z)}{\operatorname{Var}(X)} $$ provided $\operatorname{Var}(X)>0$. Because the variance in the denominator is positive, the direction of the coefficient change is the sign of the product $\beta_2\operatorname{Cov}(X,Z)$. A positive product means upward bias relative to $\beta_1$; a negative product means downward bias; a zero product means no bias from omitting this $Z$ under this model.

Here, $\beta_2$ describes the linear relationship between $Z$ and $Y$ at a fixed value of $X$. The covariance describes how $Z$ and $X$ vary together in the population. Equivalently, $\operatorname{Cov}(X,Z)/\operatorname{Var}(X)$ is the slope for predicting $Z$ from $X$ by linear regression. Its sign tells whether larger values of $X$ tend to accompany larger or smaller values of $Z$.

The coefficient $\beta_2$ is a conditional association, not automatically a causal effect. Likewise, the sign rule concerns a change between regression coefficients under the specified model. Causal interpretation requires additional design and causal assumptions.

How to Predict the Direction

Use the following sequence. First identify the coefficient being compared: the full-model coefficient $\beta_1$ or the reduced-model coefficient $\gamma_1$. Next determine the sign of the relationship between $X$ and $Z$, then the sign of $Z$'s conditional relationship with $Y$ given $X$. Multiply those signs. Do not use only the marginal association between $Z$ and $Y$: it may reflect the relationship of $Z$ with $X$ as well.

Relationship of $X$ with $Z$	Conditional relationship of $Z$ with $Y$	Change in the coefficient of $X$ when $Z$ is omitted
Positive	Positive	Upward
Positive	Negative	Downward
Negative	Positive	Downward
Negative	Negative	Upward
Zero covariance	Either sign	No change from this omitted predictor

Worked Example: An Omitted Predictor Produces Upward Bias

Suppose a hypothetical population follows the additive model $Y=\beta_0+\beta_1X+\beta_2Z+\varepsilon$, with $E(\varepsilon\mid X,Z)=0$. Let $X$ be weekly exercise time in hours, $Z$ be a baseline fitness score in points, and $Y$ be a health score in points. Assume $\beta_1=1.20$ health-score points per exercise hour and $\beta_2=0.50$ health-score points per fitness point. In this hypothetical population, let $\operatorname{Cov}(X,Z)=6.0$ hour-points and $\operatorname{Var}(X)=12.0$ hours squared. These summaries satisfy the model conditions and have positive $X$-$Z$ covariance.

The omitted-variable-bias component is

\gamma_1-\beta_1 =0.50\frac{6.0}{12.0} =0.25

Thus $\gamma_1=1.20+0.25=1.45$ health-score points per exercise hour. Omitting baseline fitness shifts the coefficient upward by $0.25$ points per hour in this hypothetical setting: fitter people tend to exercise more, and fitness is positively related to the outcome at a fixed exercise level. This calculation describes the regression coefficients; it does not prove that exercise causes a change in health score.

Worked Example: An Omitted Predictor Produces Downward Bias

Consider a different hypothetical population, with $X$ denoting treatment intensity in units, $Z$ denoting baseline illness severity in points, and $Y$ denoting a recovery score in points, where higher scores represent better recovery. Assume the same additive model and conditional mean-zero error condition. Suppose $\beta_1=2.0$ recovery points per treatment unit, $\beta_2=-0.80$ recovery points per severity point, $\operatorname{Cov}(X,Z)=4.0$ treatment-severity unit-points, and $\operatorname{Var}(X)=10.0$ treatment units squared.

Treatment intensity and severity have positive covariance: more severely ill individuals tend to receive more intensive treatment. At the same treatment intensity, greater severity is associated with lower recovery scores, so $\beta_2$ is negative. The calculation is

\gamma_1-\beta_1 =(-0.80)\frac{4.0}{10.0} =-0.32

The reduced-model slope is $\gamma_1=2.0-0.32=1.68$ recovery points per treatment unit, which is below the full-model coefficient of $2.0$. This is downward bias relative to $\beta_1$ under the stated model. It would be incorrect to conclude that treatment harms recovery: the direction of this omitted-variable bias does not settle the causal effect of treatment.

Worked Example: A Negative Relationship Can Also Produce Upward Bias

Suppose a hypothetical additive model has $X$ as a study-support score, $Z$ as hours spent at a paid job each week, and $Y$ as an exam score in points. Assume $E(\varepsilon\mid X,Z)=0$, with $\beta_1=1.5$ exam points per support-score point and $\beta_2=-0.60$ exam points per work hour. Let $\operatorname{Cov}(X,Z)=-3.0$ support-score-points times hours, and $\operatorname{Var}(X)=6.0$ support-score-points squared.

More study support tends to accompany fewer paid-work hours, so the covariance is negative. At a fixed support score, more work hours are associated with a lower exam score, so $\beta_2$ is also negative. Their product is positive:

\gamma_1-\beta_1 =(-0.60)\frac{-3.0}{6.0} =0.30

The reduced-model coefficient is $\gamma_1=1.5+0.30=1.80$ exam points per support-score point. Omitting work hours therefore biases the support coefficient upward in this hypothetical population. The two negative relationships combine to give a positive bias.

Why the Sign Rule Works

The result can be understood through the part of $Z$ that systematically varies with $X$. In the linear projection of $Z$ on $X$, the slope is

\delta_1=\frac{\operatorname{Cov}(X,Z)}{\operatorname{Var}(X)}

So, when $X$ increases by one unit, the linearly predicted value of $Z$ changes by $\delta_1$ units. In the full outcome model, a change of $\delta_1$ units in $Z$ corresponds to a change of $\beta_2\delta_1$ units in the outcome. That component is incorporated into the $X$-only slope when $Z$ is left out. Its direction is therefore the sign of $\beta_2\delta_1$, the same as the sign of $\beta_2\operatorname{Cov}(X,Z)$, because $\operatorname{Var}(X)>0$.

This argument explains the direction, not the full derivation of the coefficient identity. It also shows why knowing just one relationship is insufficient: if $Z$ predicts $Y$ but does not covary with $X$, there is no coefficient shift from this source; if $Z$ covaries with $X$ but has no conditional relationship with $Y$, there is likewise no shift under the model.

Limits and Common Pitfalls

Using a marginal relationship in place of a conditional one. The relevant outcome relationship is $\beta_2$, the relationship between $Z$ and $Y$ holding $X$ fixed. A marginal $Z$-$Y$ association can have a different sign.
Confusing upward bias with a positive coefficient. Upward bias means $\gamma_1>\beta_1$. If $\beta_1$ is negative, an upward shift could make the reduced coefficient less negative, not necessarily positive.
Assuming a sign proves causation. The rule compares population regression slopes under stated assumptions. It does not establish that changing $X$ would change $Y$.
Applying the single-omitted-predictor rule without checking the model. With several omitted predictors, their bias components can reinforce or oppose one another. Interactions, nonlinear relationships, or a violation of the error condition also mean the simple rule may not apply as written.

As a quick sign check, write the two signs before doing any arithmetic. If $Z$ is positively related to $X$ and negatively related to $Y$ conditional on $X$, the product is negative, so omission shifts the coefficient of $X$ downward. If the two signs match, the product is positive and the shift is upward. When the covariance or conditional relationship is zero, the corresponding product is zero.

Key takeaway: Under an additive linear model with a conditionally mean-zero error, the direction of bias from omitting one predictor is determined by the product of (1) its conditional outcome coefficient and (2) its covariance with the included predictor. A positive product shifts the included predictor's coefficient upward; a negative product shifts it downward.

Check Your Understanding

For each question, state the direction of the change relative to the full-model coefficient, or explain why the information is insufficient.

If $Z$ is positively associated with $X$ and has a negative conditional relationship with $Y$, which direction does omitting $Z$ shift the coefficient of $X$?
Suppose $\beta_2=0.40$, $\operatorname{Cov}(X,Z)=-5.0$, and $\operatorname{Var}(X)=10.0$. Calculate the bias component and its direction.
Why is the marginal association between $Z$ and $Y$ not necessarily enough to apply the sign rule?
If $Z$ is related to $X$ but has no conditional relationship with $Y$ given $X$, what does the rule predict under the stated model?
Does an upward shift in a treatment coefficient establish that treatment has a positive causal effect? Explain briefly.