Zero Conditional Mean and Exogeneity

Why the Error Mean Must Be Checked at Each Predictor Value

A regression line describes how a response changes with an explanatory variable. To interpret that line as the mean response at each predictor value, it is not enough for errors to average to zero across the whole population. Errors must average to zero within each predictor condition. This is the zero conditional mean assumption, also called mean exogeneity.

For example, a hypothetical insurer might use a policy’s annual mileage to predict its annual number of roadside-assistance calls. A model could capture the average call count for each mileage level only if the errors do not systematically tend to be positive at some mileage levels and negative at others. A positive and negative error pattern can cancel overall while leaving the line wrong for particular mileage values.

Definition: In the simple linear regression model $$ Y=\beta_0+\beta_1X+\varepsilon, $$ $Y$ is the response, $X$ is the explanatory variable, $\beta_0$ and $\beta_1$ are fixed model parameters, and $\varepsilon$ is the error. The zero conditional mean assumption is $$ E(\varepsilon\mid X)=0 $$ for every relevant value of $X$, or, for a continuous $X$, almost surely. The conditional expectation and the error are assumed to exist.

When $X$ is random, the condition says that among cases with the same predictor value, the average error is zero. It does not say that each individual error is zero. When the predictor values are treated as fixed by design, the corresponding assumption is that the error has mean zero conditional on the design: $E(\varepsilon_i\mid X_1,\ldots,X_n)=0$ for each observation $i$. In either setting, the condition concerns the error’s mean, not its particular realized value.

Why the Regression Line Becomes the Conditional Mean

The purpose of the assumption is clear by taking the conditional expectation of the model. Since $\beta_0$, $\beta_1$, and $X$ are known once we condition on $X$, linearity of conditional expectation gives the following result.

Theorem: If $Y=\beta_0+\beta_1X+\varepsilon$ and $E(\varepsilon\mid X)=0$, then $$ E(Y\mid X)=\beta_0+\beta_1X. $$ Conversely, if the error is defined by $\varepsilon=Y-(\beta_0+\beta_1X)$, then $E(Y\mid X)=\beta_0+\beta_1X$ implies $E(\varepsilon\mid X)=0$.

For a particular predictor value $x$, the result reads $E(Y\mid X=x)=\beta_0+\beta_1x$, wherever that conditional mean is defined. Thus the intercept is the conditional mean at $X=0$, and the slope describes the change in conditional mean for a one-unit increase in $X$. This interpretation requires the zero conditional mean assumption; it is not guaranteed merely by writing down a linear equation with an error term.

Condition on the predictor.
Start from $Y=\beta_0+\beta_1X+\varepsilon$ and take $E(\cdot\mid X)$ on both sides.

Take out quantities fixed by the condition.
The parameters are constants, and $X$ is known given $X$, so $E(Y\mid X)=\beta_0+\beta_1X+E(\varepsilon\mid X)$.

Apply the assumption.
Substituting $E(\varepsilon\mid X)=0$ yields $E(Y\mid X)=\beta_0+\beta_1X$.

Check the converse.
If $\varepsilon=Y-(\beta_0+\beta_1X)$, then $E(\varepsilon\mid X)=E(Y\mid X)-\beta_0-\beta_1X$. A conditional mean equal to the line therefore makes this expression zero.

The iterated-expectation rule also shows that zero conditional mean implies zero unconditional mean: $$ E(\varepsilon)=E\{E(\varepsilon\mid X)\}=E(0)=0. $$ The reverse implication does not hold. An overall average can be zero because positive conditional means at some predictor values cancel negative conditional means elsewhere. The condition is therefore stronger than $E(\varepsilon)=0$.

Worked Examples

Worked Example: Checking Conditional Means in a Hypothetical Prediction Model

A hypothetical roadside-assistance study uses $X$, annual mileage measured in units of 10,000 miles, to model $Y$, annual service calls per policy. Suppose $$ Y=12+4X+\varepsilon. $$ The random predictor has $P(X=1)=0.4$ and $P(X=3)=0.6$. Conditional on either predictor value, $\varepsilon$ has two equally likely values: at $X=1$, they are $-2$ and $2$; at $X=3$, they are $-1$ and $1$. Find the conditional error means, the conditional response means, and the unconditional error mean. All distributions and probabilities here are hypothetical; no interest or mortality assumption is relevant.

Method. Compute the error mean separately at each value of $X$, then apply the model and average over the distribution of $X$.

At $X=1$, $$ E(\varepsilon\mid X=1)=\tfrac12(-2)+\tfrac12(2)=-1+1=0. $$ At $X=3$, $$ E(\varepsilon\mid X=3)=\tfrac12(-1)+\tfrac12(1)=-0.5+0.5=0. $$ Thus the zero conditional mean condition holds at both possible predictor values.

At $X=1$, the model’s conditional response mean is $$ E(Y\mid X=1)=12+4(1)+0=16. $$ At $X=3$, it is $$ E(Y\mid X=3)=12+4(3)+0=24. $$ Finally, averaging the conditional error means using the probabilities of $X$, $$ E(\varepsilon)=0.4(0)+0.6(0)=0. $$ The unconditional mean is zero as required by the conditional assumption. Check: the conditional response means, 16 and 24, also equal $12+4x$ at $x=1$ and $x=3$. The common mistake this calculation avoids is checking only the overall error mean instead of checking the mean within each predictor condition.

Worked Example: An Unconditional Zero Mean That Does Not Give the Claimed Line

In a hypothetical study of claim-review workload, let $X$ be a workload category coded as 0, 1, or 2, with each value having probability $1/3$. Suppose the proposed model is $$ Y=5+3X+\varepsilon, $$ and the error is deterministic conditional on $X$: it equals 2 when $X=0$, 0 when $X=1$, and $-2$ when $X=2$. Determine whether the zero conditional mean assumption holds and find $E(Y\mid X=x)$ at each value. The stated finite distribution is hypothetical; there are no interest or mortality assumptions.

Method. Because the error is fixed at each predictor value, its conditional mean is just that value. Hence $$ E(\varepsilon\mid X=0)=2,\qquad E(\varepsilon\mid X=1)=0,\qquad E(\varepsilon\mid X=2)=-2. $$ The condition fails at $X=0$ and $X=2$.

The unconditional error mean nevertheless is $$ E(\varepsilon)=\tfrac13(2)+\tfrac13(0)+\tfrac13(-2) =\tfrac23+0-\tfrac23=0. $$ Now calculate the response mean at each predictor value using the model and its conditional error mean: $$ E(Y\mid X=0)=5+3(0)+2=7, $$ $$ E(Y\mid X=1)=5+3(1)+0=8, $$ $$ E(Y\mid X=2)=5+3(2)-2=9. $$ The claimed line gives 5, 8, and 11 at those same predictor values, so it does not give the conditional response means at $X=0$ or $X=2$. Check: the actual conditional means have overall mean $(7+8+9)/3=8$, while $E(5+3X)=5+3(1)=8$. Matching overall means does not make the conditional means match. The mistake to avoid is treating $E(\varepsilon)=0$ as if it established $E(\varepsilon\mid X)=0$.

Worked Example: Zero Conditional Mean Does Not Require Independence

Consider a hypothetical model for annual adjustment time, $Y=10+2X+\varepsilon$, where $X$ is a binary workload indicator with $P(X=0)=P(X=2)=1/2$. Conditional on $X=0$, let $\varepsilon$ be $-1$ or $1$, each with probability $1/2$. Conditional on $X=2$, let $\varepsilon$ be $-3$ or $3$, each with probability $1/2$. Check the conditional mean assumption and determine whether $\varepsilon$ is independent of $X$. These are hypothetical distributions; no interest or mortality assumption is involved.

Method. First calculate the error mean within each predictor group: $$ E(\varepsilon\mid X=0)=\tfrac12(-1)+\tfrac12(1)=0, $$ $$ E(\varepsilon\mid X=2)=\tfrac12(-3)+\tfrac12(3)=0. $$ Thus $E(\varepsilon\mid X)=0$ for every possible $X$, and the conditional response means are $$ E(Y\mid X=0)=10,\qquad E(Y\mid X=2)=10+2(2)=14. $$

The error is not independent of $X$. For example, $P(\varepsilon=1\mid X=0)=1/2$, whereas $P(\varepsilon=1\mid X=2)=0$. Independence would require the conditional distribution of $\varepsilon$ to be the same for both predictor values. It is not. Check: the error mean is zero in both groups even though the conditional error distributions differ. The common mistake is to interpret zero conditional mean as independence; the assumption restricts conditional means, not entire distributions.

What the Assumption Does and Does Not Say

The word exogeneity is sometimes used with different meanings in different settings. Here, mean exogeneity refers specifically to $E(\varepsilon\mid X)=0$. It is stronger than requiring only that the error have overall mean zero. Under appropriate finite-moment conditions, it also implies that the error is uncorrelated with functions of $X$: for a suitable function $g$, $$ E\{\varepsilon g(X)\}=E\{g(X)E(\varepsilon\mid X)\}=0. $$ But zero covariance alone is generally weaker than zero conditional mean, and neither statement by itself establishes independence.

The condition is a model assumption about the population or data-generating process, not an identity that must hold for observed errors. It also cannot be confirmed just because an estimated regression with an intercept has residuals that sum to zero in the observed sample. Residuals are calculated from the fitted model and data; the assumption concerns the conditional means of the underlying errors.

Exam trap: A solution reports $E(\varepsilon)=0$ and concludes that the regression line is $E(Y\mid X)$. That conclusion is not justified. The required step is to establish $E(\varepsilon\mid X)=0$; otherwise, the conditional mean is $\beta_0+\beta_1X+E(\varepsilon\mid X)$, which may differ from the proposed line.

A reliable check is to write the error mean conditionally before simplifying the regression equation. If the model is intended to describe the conditional mean, verify that the average error is zero at every relevant predictor value (or almost surely for a continuous predictor). This distinction is aligned with the SOA VEE Mathematical Statistics syllabus.

Check Your Understanding

For each problem, distinguish the mean error overall from its mean at each predictor condition.

A random predictor takes values 0 and 2 with probabilities 0.25 and 0.75. Given $X=0$, the error is equally likely to be $-2$ or 2. Given $X=2$, it is equally likely to be $-1$ or 1. Find both conditional error means and the unconditional error mean. (Answer: Both conditional means are 0; the unconditional mean is 0.)
A predictor takes values $-1$ and 1 with equal probabilities. The conditional error means are 1 when $X=-1$ and $-1$ when $X=1$. Find the unconditional error mean and decide whether zero conditional mean holds. (Answer: $E(\varepsilon)=0$; zero conditional mean does not hold.)
Suppose $Y=7+2X+\varepsilon$, $E(\varepsilon\mid X=1)=3$, and $E(\varepsilon\mid X=4)=-1$. Find the two conditional response means and decide whether the stated regression line is the conditional mean. (Answer: $E(Y\mid X=1)=12$, $E(Y\mid X=4)=14$; the line is not the conditional mean at either value.)
Let $P(X=0)=P(X=1)=1/2$. Given $X=0$, the error is $-2$ or 2 with equal probabilities; given $X=1$, it is $-4$ or 4 with equal probabilities. Does zero conditional mean hold, and are $X$ and $\varepsilon$ independent? (Answer: Zero conditional mean holds; $X$ and $\varepsilon$ are not independent.)
For a model $Y=\beta_0+\beta_1X+\varepsilon$, a report establishes only $E(\varepsilon)=0$. What additional condition justifies interpreting the line as $E(Y\mid X)$? (Answer: $E(\varepsilon\mid X)=0$ almost surely.)

What You'll Learn