Predictor Design and Coefficient Interval Width

Why Predictor Placement Matters

Two studies can use the same number of observations and have similar residual variability, yet give different precision for an estimated regression slope. One reason is that the predictor values may be spread out differently. If the observations cover a wider range of predictor values, changes in the response across that range can generally provide more information about the slope.

Placement also matters when the goal is to estimate the mean response at a particular predictor value. Estimates near the center of the observed predictor values are generally more precise than estimates far from that center. These are related effects of design, but they concern different quantities: the slope describes a change per predictor unit, while a fitted mean response describes an expected response at one predictor value.

The discussion below concerns simple linear regression with a quantitative predictor. Treat the observed predictor values as fixed, and suppose the response follows a linear mean model with independent errors, mean zero, and common variance \(\sigma^2\). Normal errors support the usual exact \(t\)-based confidence intervals. For these intervals, \(s\) estimates the residual standard deviation and the degrees of freedom are \(n-2\).

Predictor Spread and the Slope Interval

In simple linear regression, the estimated slope is \(b_1\). Its estimated standard error is \(SE(b_1)=s/\sqrt{S_{xx}}\). Thus, if sample size and residual standard deviation are held fixed, increasing \(S_{xx}\) decreases the slope standard error. A two-sided confidence interval for the population slope \(\beta_1\), at confidence level \(1-\alpha\), has the form \(b_1\) plus or minus a \(t\) critical value times that standard error.

The total interval width is \(2t^*_{n-2}s/\sqrt{S_{xx}}\). With the same \(n\), confidence level, and \(s\), a design with greater predictor spread produces a narrower slope interval. The estimated slope \(b_1\) itself is not made larger by greater spread; rather, the estimate is more precise. In actual studies, designs can also differ in residual variability, so predictor spread alone does not guarantee narrower intervals unless the other factors are accounted for.

Distribution Matters, Not Just the Endpoints

The minimum and maximum predictor values give a quick sense of range, but they do not determine \(S_{xx}\). The number and placement of observations between the endpoints also matter. Because \(S_{xx}\) adds the squared distances of all predictor values from \(\bar{x}\), it summarizes the whole design rather than just its endpoints.

Uncertainty at a Particular Predictor Value

A different design question arises when estimating the mean response at a specified predictor value \(x_0\). The fitted mean is \(\hat{y}(x_0)=b_0+b_1x_0\), where \(b_0\) is the estimated intercept. Its estimated standard error is \(s\sqrt{1/n+(x_0-\bar{x})^2/S_{xx}}\). The first term reflects the contribution of having a finite sample; the second increases as \(x_0\) moves farther from the predictor mean. The confidence interval for the population mean response at \(x_0\) uses this standard error and the \(t\) critical value with \(n-2\) degrees of freedom.

This interval concerns the mean response at \(x_0\), not the response of an individual. The predictor placement affects its width through both \(S_{xx}\) and the distance \(x_0-\bar{x}\). Therefore, a design that is helpful for estimating the slope does not make every fitted mean equally precise: the target value’s location relative to the design also matters.

Why the Mean-Response Formula Depends on Placement

The relationship can be justified by writing the fitted mean at \(x_0\) as a weighted combination of the observed responses. In simple linear regression, \(b_1=\sum (x_i-\bar{x})y_i/S_{xx}\) and \(b_0=\bar{y}-b_1\bar{x}\). Substituting these into \(b_0+b_1x_0\) gives \(\hat{y}(x_0)=\bar{y}+(x_0-\bar{x})b_1\).

Under the stated error model, \(\bar{y}\) has variance \(\sigma^2/n\), \(b_1\) has variance \(\sigma^2/S_{xx}\), and their covariance is zero. The covariance is zero because the centered predictor values sum to zero. Consequently, the variance of their indicated combination is \(\sigma^2/n+(x_0-\bar{x})^2\sigma^2/S_{xx}\). Replacing \(\sigma\) with \(s\) gives the estimated standard error in the interval formula.

At \(x_0=\bar{x}\), the distance term is zero, so the estimated standard error of the fitted mean is \(s/\sqrt{n}\). Moving away from \(\bar{x}\) adds uncertainty. Greater \(S_{xx}\) reduces that added term for a fixed distance, but it does not erase the effect of being far from the center.

Using Design Comparisons Carefully

The formulas help separate design effects that are easy to conflate. For a slope interval, width depends on the residual standard deviation, the critical value, and \(S_{xx}\). For a mean-response interval at \(x_0\), width also depends on how far \(x_0\) lies from \(\bar{x}\). A useful comparison therefore identifies which factors are held fixed before attributing a difference in interval width to predictor placement.

• With the same \(n\), \(s\), confidence level, and \(S_{xx}\), the slope standard errors and interval widths are the same.
• With \(n\), \(s\), and confidence level fixed, larger \(S_{xx}\) means a smaller slope standard error.
• For estimating a mean response, a target value closer to \(\bar{x}\) has a smaller distance contribution to its standard error.
• The same minimum and maximum predictor values do not imply the same \(S_{xx}\).

These formulas describe precision under the regression model; they do not establish that a linear relationship is appropriate or that an estimate at a distant predictor value is scientifically useful. In particular, an interval for a mean response should not be treated as evidence that predictions far beyond the observed predictor values are reliable.

Common Mistake

A common mistake is to say, “The predictor range doubles, so the slope interval must be half as wide.” That conclusion does not follow from range alone. The standard error depends on \(\sqrt{S_{xx}}\), and \(S_{xx}\) depends on every observation’s distance from \(\bar{x}\). Also, in a new study, \(s\) or the degrees of freedom may differ. The correction is to compare the complete width formula, checking \(s\), \(S_{xx}\), and the critical value rather than relying on an informal description of range.

Another mistake is to use the slope standard error to describe uncertainty in the mean response at \(x_0\). The slope standard error is \(s/\sqrt{S_{xx}}\); the fitted-mean standard error includes the additional \(1/n\) term and the distance from \(x_0\) to \(\bar{x}\). Choose the formula for the quantity being estimated.