The Standard Error of a Regression Intercept

Why the Intercept Has Its Own Standard Error

A regression intercept is an estimated response value at a particular predictor value: zero. Its standard error describes how much the estimated intercept would vary across repeated samples, conditional on the predictor values, under the regression model. It is not the same quantity as the residual standard deviation, which describes the typical scatter of individual responses around the regression line.

The intercept’s precision depends not only on residual scatter and sample size, but also on where the observed predictor values lie relative to zero. If zero is far from their center, estimating the response at zero can be less precise than estimating it near the center of the observed predictor values. This tutorial derives that result and shows how to calculate the standard error.

The Model and the Intercept Standard Error

Consider simple linear regression with one quantitative predictor. For observations $i=1,\ldots,n$, write the model as

Y_i=\beta_0+\beta_1x_i+\varepsilon_i.

Here, $Y_i$ is the response, $x_i$ is the predictor value, $\beta_0$ is the population intercept, and $\beta_1$ is the population slope. For the usual standard error formula, treat the observed predictor values as fixed, or condition on them. Assume the errors have mean zero, a common variance $\sigma^2$, and are uncorrelated across observations. In symbols, $E(\varepsilon_i\mid x)=0$, $\operatorname{Var}(\varepsilon_i\mid x)=\sigma^2$, and $\operatorname{Cov}(\varepsilon_i,\varepsilon_j\mid x)=0$ for $i\ne j$. These assumptions make the conditional mean response linear in $x_i$ and give the conventional variance formula. Normal errors are not required for this variance derivation; they are a further assumption used for exact small-sample inference.

Definition: Let $\hat{\beta}_0$ and $\hat{\beta}_1$ be the least-squares estimates of the intercept and slope, let $\bar{x}$ be the sample mean of the predictor, and define $S_{xx}=\sum_{i=1}^{n}(x_i-\bar{x})^2$. The estimated standard error of the intercept is $$ SE(\hat{\beta}_0)=s\sqrt{\frac{1}{n}+\frac{\bar{x}^{\,2}}{S_{xx}}}, $$ where $s=\sqrt{\mathrm{SSE}/(n-2)}$ estimates the residual standard deviation and $\mathrm{SSE}=\sum_{i=1}^{n}(Y_i-\hat{Y}_i)^2$.

The formula requires $n>2$ and $S_{xx}>0$: there must be enough observations to estimate residual variation, and the predictor must vary. If all predictor values are identical, the slope and intercept cannot both be estimated from the data. The $n-2$ in $s$ reflects the two fitted regression coefficients.

Why the Formula Has This Form

The least-squares intercept can be written as the sample mean response minus the estimated slope multiplied by the sample mean predictor:

\hat{\beta}_0=\bar{Y}-\hat{\beta}_1\bar{x}.

Under the model assumptions, the sample mean response has conditional variance $\sigma^2/n$, while the slope estimate has conditional variance $\sigma^2/S_{xx}$. The sample mean response and the slope estimate have conditional covariance zero. The following derivation shows why, and then combines these results.

Express the slope estimate using the errors.
Since the fitted slope is $\hat{\beta}_1=\sum (x_i-\bar{x})Y_i/S_{xx}$, substitution of the model gives $\hat{\beta}_1=\beta_1+\sum (x_i-\bar{x})\varepsilon_i/S_{xx}$. The term involving $\beta_0$ disappears because $\sum (x_i-\bar{x})=0$.

Find the slope variance and its covariance with $\bar{Y}$.
Independence in the covariance sense and common error variance give $\operatorname{Var}(\hat{\beta}_1\mid x)=\sigma^2/S_{xx}$. Also, $\bar{Y}=\beta_0+\beta_1\bar{x}+\bar{\varepsilon}$, so $\operatorname{Cov}(\bar{Y},\hat{\beta}_1\mid x)=\sigma^2\sum(x_i-\bar{x})/(nS_{xx})=0$.

Use the variance of a difference.
Because $\hat{\beta}_0=\bar{Y}-\bar{x}\hat{\beta}_1$, its conditional variance is $\operatorname{Var}(\bar{Y}\mid x)+\bar{x}^{\,2}\operatorname{Var}(\hat{\beta}_1\mid x)-2\bar{x}\operatorname{Cov}(\bar{Y},\hat{\beta}_1\mid x)$. The covariance term is zero, leaving the result below.

\operatorname{Var}(\hat{\beta}_0\mid x) =\sigma^2\left(\frac{1}{n}+\frac{\bar{x}^{\,2}}{S_{xx}}\right).

The standard deviation of the sampling distribution is therefore $\sigma\sqrt{1/n+\bar{x}^{\,2}/S_{xx}}$. Since $\sigma$ is unknown, replace it with the residual standard deviation estimate $s$ to obtain $SE(\hat{\beta}_0)$. This is an estimated standard deviation of the intercept estimator’s sampling distribution, not a statement that the intercept is likely to be within one standard error of its estimate.

Calculating the Standard Error

Worked Example: Calculating an Intercept Standard Error

In a hypothetical sample of six observations, the predictor values are $1,2,3,4,5,6$, and the response values are $13,12,17,19,18,23$. The fitted line is $\hat{Y}=10+2x$, with residuals $1,-2,1,1,-2,1$. Assume the linear mean model and error conditions above are reasonable for these data.

First calculate the predictor summaries. The mean is $\bar{x}=(1+2+3+4+5+6)/6=3.5$. The centered sum of squares is

S_{xx}=(-2.5)^2+(-1.5)^2+(-0.5)^2+(0.5)^2+(1.5)^2+(2.5)^2=17.5.

The residual sum of squares is $1^2+(-2)^2+1^2+1^2+(-2)^2+1^2=12$. Thus $s=\sqrt{12/(6-2)}=\sqrt{3}=1.732$ response units. Substituting all values,

SE(\hat{\beta}_0) =1.732\sqrt{\frac{1}{6}+\frac{3.5^2}{17.5}} =1.732\sqrt{0.1667+0.7000} =1.612.

The estimated intercept is 10 response units, and its standard error is about 1.612 response units. In repeated samples with these predictor values and the stated model conditions, the estimated intercept would have a sampling standard deviation estimated to be 1.612 response units. The residual standard deviation, 1.732 response units, describes a different kind of variation: scatter of individual outcomes around the fitted line.

How Predictor Location and Spread Matter

The term $1/n$ reflects the contribution from estimating the mean response. The term $\bar{x}^{\,2}/S_{xx}$ reflects the additional uncertainty involved in locating the fitted line at $x=0$. A predictor mean farther from zero can increase this contribution, but only when the other relevant quantities are held fixed. If the predictor spread $S_{xx}$ also changes, the ratio need not increase. The important comparison is the ratio of the squared predictor mean to the predictor’s centered sum of squares, not the mean alone.

For fixed $n$, $\bar{x}$, and $s$, a larger $S_{xx}$ reduces $\bar{x}^{\,2}/S_{xx}$ and therefore reduces the intercept standard error. For fixed $n$, $S_{xx}$, and $s$, increasing $|\bar{x}|$ increases the ratio and the standard error. Holding the predictor summaries and $s$ fixed, increasing $n$ reduces the $1/n$ contribution. These are controlled comparisons: if more than one quantity changes, compare the complete formula rather than assuming one change determines the result.

Worked Example: Why Distance from Zero Is Not Enough

Consider two hypothetical predictor designs, each with $n=10$ and residual standard deviation $s=1.5$ response units. Design A has $\bar{x}=1$ and $S_{xx}=1$; Design B has $\bar{x}=2$ and $S_{xx}=100$. The model conditions are assumed to be reasonable for both designs.

For Design A, the squared standard-error multiplier is

\frac{1}{10}+\frac{1^2}{1}=1.10, \qquad SE(\hat{\beta}_0)=1.5\sqrt{1.10}=1.573.

For Design B, it is

\frac{1}{10}+\frac{2^2}{100}=0.14, \qquad SE(\hat{\beta}_0)=1.5\sqrt{0.14}=0.5612.

Although Design B’s predictor mean is farther from zero, its much larger spread makes the ratio $\bar{x}^{\,2}/S_{xx}$ smaller: $0.04$ rather than $1$. Consequently, its calculated intercept standard error is smaller. This comparison demonstrates why it is incorrect to say that moving the predictor mean farther from zero necessarily increases the intercept standard error when the predictor spread may also change.

Worked Example: Centering the Predictor

Suppose a hypothetical regression has $n=12$, $\bar{x}=5$ hours, $S_{xx}=80$ hours squared, and $s=2.4$ response units. Assume the model conditions are reasonable. The intercept in the original predictor scale is the estimated response at $x=0$ hours. Its standard error is

SE(\hat{\beta}_0) =2.4\sqrt{\frac{1}{12}+\frac{5^2}{80}} =2.4\sqrt{0.3958} =1.510

response units. Now define a centered predictor $z=x-5$. Its sample mean is zero, and the centered sum of squares remains 80 hours squared. The intercept in a regression using $z$ is the estimated response at $z=0$, which corresponds to $x=5$ hours. Its standard error is

SE(\text{intercept at } z=0) =2.4\sqrt{\frac{1}{12}+\frac{0^2}{80}} =\frac{2.4}{\sqrt{12}} =0.6928

response units. The smaller standard error is for a different intercept: the estimated mean response at the predictor’s sample mean rather than at zero hours. Centering changes the reference point and the interpretation of the intercept; it does not change the fitted responses or remove uncertainty from the data.

Interpretation and Common Pitfalls

The intercept $\beta_0$ is the model’s mean response when $x=0$. Its practical meaning depends on whether zero is a meaningful predictor value and whether it is within or near the observed predictor range. If zero is far outside that range, the intercept can be an extrapolation. Its standard error quantifies sampling uncertainty under the model, but it does not establish that the model is appropriate at unobserved predictor values.

Confusing the two standard deviations. The residual standard deviation $s$ measures residual scatter in response units. $SE(\hat{\beta}_0)$ measures estimated sampling variation of the intercept estimate, also in response units.
Ignoring predictor location. A small residual standard deviation does not by itself guarantee a precise intercept. The multiplier also depends on $n$, $\bar{x}$, and $S_{xx}$.
Making an unqualified comparison based on $|\bar{x}|$. A larger distance from zero increases the intercept standard error only when $n$, $S_{xx}$, and $s$ are held fixed. If $S_{xx}$ changes, evaluate the full expression.
Forgetting what zero means. An intercept at zero hours, zero dose, or another zero value may be scientifically unhelpful or outside the data range. Centering can express the intercept at a more relevant predictor value, but that is a different parameter.

Key takeaway: In simple linear regression, $SE(\hat{\beta}_0)=s\sqrt{1/n+\bar{x}^{\,2}/S_{xx}}$. Intercept precision depends on residual variability, sample size, and the predictor mean relative to its spread. Interpret the intercept at its stated predictor value, especially when zero is not represented in the data.

Check Your Understanding

Use the formula and its assumptions to answer the following.

In the intercept standard-error formula, what does $s$ measure, and how does that differ from what $SE(\hat{\beta}_0)$ measures?
For fixed $n$, $\bar{x}$, and $s$, what happens to the intercept standard error as $S_{xx}$ increases? Explain using the formula.
Why is it not always valid to conclude that a predictor mean farther from zero leads to a larger intercept standard error?
A predictor is centered by subtracting its sample mean. At what original predictor value does the new intercept refer, and what happens to the $\bar{x}^{\,2}/S_{xx}$ contribution?
Why can a precisely calculated intercept still have limited practical meaning when zero lies far outside the observed predictor range?