What a Pointwise Interval Is About
A regression confidence interval for a mean response concerns the average response at a specified predictor value. It does not describe the range of individual outcomes, and its confidence level is not a probability assigned to the unknown mean after an interval has been calculated. Correct interpretation depends on keeping the target fixed and understanding what would vary across repeated samples.
For simple linear regression, write the model as \(Y_i=\beta_0+\beta_1x_i+\varepsilon_i\), where \(x_i\) is the predictor value for observation \(i\), \(\beta_0\) and \(\beta_1\) are fixed but unknown coefficients, and \(\varepsilon_i\) is the random error. At a specified value \(x_0\), the mean response is \(\mu(x_0)=\beta_0+\beta_1x_0\). This is a fixed, unknown quantity when the model and \(x_0\) are fixed.
For the usual exact small-sample interval, assume that the predictor values are fixed, the errors are independent and normally distributed with mean zero and common variance \(\sigma^2\), the predictor values are not all identical, and \(n>2\). Define \(\bar{x}=\sum_{i=1}^n x_i/n\), \(S_{xx}=\sum_{i=1}^n(x_i-\bar{x})^2\), and let \(s\) be the square root of the residual mean square from the fitted regression. The interval at \(x_0\) is
Here \(t_{1-\alpha/2,n-2}\) is the appropriate quantile of the \(t\) distribution with \(n-2\) degrees of freedom. The interval’s center and endpoints depend on the sample, so they are random before the data are observed. The target \(\mu(x_0)\), by contrast, is fixed under the model.
Why the Coverage Statement Is About Repetition
Under the stated model, for a fixed design and a fixed \(x_0\), the standardized difference between the estimated mean and the true mean has a \(t\) distribution with \(n-2\) degrees of freedom when the unknown error variance is estimated by \(s^2\). Thus, the central \(1-\alpha\) portion of that distribution gives the interval procedure its coverage.
Rearranging the two inequalities places \(\mu(x_0)\) between the random lower and upper endpoints of the confidence interval. Therefore, if the same fixed predictor design and target \(x_0\) were used repeatedly to collect new samples and calculate intervals in the same way, a proportion \(1-\alpha\) of those intervals would cover \(\mu(x_0)\), in the long run. For a 95% procedure, the coverage rate is 95% under the model assumptions.
This is a property of the interval-producing procedure over repeated samples. Once one particular interval has been calculated, its endpoints are fixed, and the mean response is still fixed. The frequentist interpretation does not say that there is a 95% probability that this already calculated interval contains the parameter. The interval either contains the parameter or it does not; the 95% describes the procedure’s performance across repetitions.
The degrees of freedom condition matters. With \(n=2\), even if the two predictor values differ, the fitted line uses both observations and there are no residual degrees of freedom to estimate the error variance. The stated \(t\)-interval is therefore not defined. Likewise, if all predictor values are identical, \(S_{xx}=0\), so the slope and the displayed standard error formula cannot be used as stated.
Worked Example: Interpreting One 95% Interval at the Design Center
A hypothetical study measures a response at 10 fixed predictor values. Assume the simple linear regression model has independent normal errors with mean zero and common variance, and that the predictor values are not all identical. The fitted mean response at \(x_0=5\) is 19.5, the residual standard deviation is \(s=4\), \(\bar{x}=5\), and \(S_{xx}=30\). Find the 95% interval and interpret its coverage.
Method. Use the pointwise mean-response interval with \(n=10\), so the degrees of freedom are \(10-2=8\). The 95% critical value is approximately \(t_{0.975,8}=2.306\).
Since \(x_0=\bar{x}=5\), the squared-distance term is zero. The estimated standard error is
Using the unrounded standard error, the margin of error is \(2.306(1.264911\ldots)=2.9169\ldots\). The endpoints are \(19.5-2.9169\ldots=16.5831\ldots\) and \(19.5+2.9169\ldots=22.4169\ldots\). To two decimal places, the interval is \((16.58,22.42)\).
Interpretation. The interval procedure, applied repeatedly to new samples under the same model and fixed design, would cover the fixed mean response at \(x=5\) in 95% of repetitions. It is not correct to say that the fixed mean has a 95% probability of lying in this particular calculated interval.
Check and common mistake. The interval is centered at 19.5, and each endpoint is about 2.92 from the center. A common mistake is to call 19.5 an individual future response or to treat the 95% as a probability about the parameter after seeing the endpoints. Here the target is the mean response at \(x=5\), and the coverage claim concerns repeated intervals.
Pointwise Means One Predictor Value at a Time
A pointwise interval targets one specified \(x_0\). Its coverage statement does not automatically extend to a collection of predictor values, or to every value in a range. If an analyst calculates separate 95% intervals at several values, each interval has its own pointwise coverage statement under the model. The collection of intervals does not thereby become a 95% simultaneous confidence band.
A value chosen in advance is especially straightforward to interpret. If the analyst selects \(x_0\) after inspecting the data or searches across many values and reports only a favorable interval, the simple pointwise interpretation may no longer match the selection process. A simultaneous procedure, designed to cover a set of mean responses together, is needed for a claim about an entire range or a group of values. That is a different inferential target.
Worked Example: Reading an Off-Center Pointwise Interval
A hypothetical regression study has \(n=12\) fixed predictor values, \(\bar{x}=4\), \(S_{xx}=44\), and residual standard deviation \(s=3.6\). At the prespecified value \(x_0=7\), the fitted mean response is 23.9. Assume independent normal errors with mean zero and common variance. Calculate the 95% interval and state precisely what it covers.
Method. There are \(12-2=10\) residual degrees of freedom, and \(t_{0.975,10}\) is approximately 2.228. The target is the mean response at \(x_0=7\), not a single observation at that predictor value.
The squared distance from the design center is \((7-4)^2=9\). The quantity under the square root is
Therefore, the standard error is
which is 1.9316 when displayed to four decimal places. Using the unrounded standard error, the margin is \(2.228(1.931556\ldots)=4.3035\ldots\). The endpoints are \(23.9-4.3035\ldots=19.5965\ldots\) and \(23.9+4.3035\ldots=28.2035\ldots\). To two decimal places, the interval is \((19.60,28.20)\).
Interpretation. For repeated samples with this same predictor design, the 95% interval procedure at the fixed value \(x=7\) would cover the fixed mean response \(\mu(7)\) in 95% of repetitions. It does not assert that 95% of individual responses at \(x=7\) fall between 19.60 and 28.20.
Check and common mistake. The standard error exceeds the center-of-design standard error because \(x_0=7\) is three units from \(\bar{x}=4\), so the added term \(9/44\) is positive. A common arithmetic error is to round the standard error before multiplying by the critical value; the endpoints above use the unrounded value.
Several Pointwise Intervals Are Not a Simultaneous Guarantee
The word “pointwise” becomes important when an analysis reports intervals at more than one predictor value. Consider two fixed values. Each 95% interval can be formed with a procedure that has 95% coverage for its own mean response. But the event that both intervals cover their respective means is a joint event. The two separate marginal coverage statements do not establish that this joint event has probability 95%.
Worked Example: Two Locations, Two Pointwise Statements
A hypothetical study has \(n=16\) fixed predictor values with \(\bar{x}=10\), \(S_{xx}=80\), and residual standard deviation \(s=5\). The fitted line is \(\hat{\mu}(x)=10+2x\). Assume independent normal errors with mean zero and common variance. Calculate 95% pointwise intervals at \(x=10\) and \(x=14\), then explain what may be claimed about their coverage.
Method. The degrees of freedom are \(16-2=14\), and \(t_{0.975,14}\) is approximately 2.145. At \(x=10\), the fitted mean is \(10+2(10)=30\). Its standard error is
The margin is \(2.145(1.25)=2.68125\). Thus the interval at \(x=10\) is \((30-2.68125,30+2.68125)=(27.31875,32.68125)\), or \((27.32,32.68)\) to two decimal places.
At \(x=14\), the fitted mean is \(10+2(14)=38\), and the squared distance from the design center is \((14-10)^2=16\). The standard error is
The margin, using the unrounded standard error, is \(2.145(2.561738\ldots)=5.4949\ldots\). The interval is \((38-5.4949\ldots,38+5.4949\ldots)=(32.5051\ldots,43.4949\ldots)\), or \((32.51,43.49)\) to two decimal places.
Interpretation. Under the stated model, each interval separately has 95% pointwise coverage for its corresponding fixed mean: \(\mu(10)\) for the first and \(\mu(14)\) for the second. These calculations alone do not show that the probability both intervals cover their targets is 95%.
Check and common mistake. At \(x=10\), the predictor is at the design center and the distance contribution is zero. At \(x=14\), the contribution is positive, so its standard error is larger. A common mistake is to report “the two intervals together have 95% coverage” just because each is a 95% interval. The pointwise results do not provide that joint guarantee.
Interpretation Checklist
When reporting a pointwise mean-response interval, state the predictor value and the target mean clearly. Then describe confidence as the long-run coverage rate of the method, conditional on the regression model and the fixed design. Keep the following distinctions explicit:
- Target: the fixed mean response \(\mu(x_0)\) at one stated predictor value, not an individual response.
- Random quantity: the interval endpoints, which vary from sample to sample.
- Confidence level: the repeated-sampling coverage rate of the procedure under its assumptions, not a posterior probability for the fixed parameter.
- Scope: pointwise coverage at \(x_0\), not automatic simultaneous coverage across multiple predictor values.
The exact \(t\)-based interpretation depends on the model conditions. If errors are not independent, do not have constant variance, or are not normally distributed in a small sample, the stated exact result may fail. Approximate methods may be appropriate under different conditions, but they require their own justification. The interval also requires \(S_{xx}>0\) and \(n>2\); a formula should not be interpreted when its required quantities cannot be estimated.
Check Your Understanding
For each question, distinguish the target from the interval procedure and retain the stated regression assumptions.
- A study has \(n=9\), residual standard deviation \(s=2.4\), and \(x_0=\bar{x}\). The fitted mean response is 14. Assume fixed, nonidentical predictor values and independent normal errors with mean zero and common variance. Using \(t_{0.975,7}=2.365\), calculate the 95% pointwise interval to two decimals and state its coverage meaning. (Answer: standard error \(0.80\), interval \((12.11,15.89)\); the procedure covers the fixed mean at this \(x_0\) in 95% of repeated samples.)
- An analyst reports a 95% interval \((6.2,8.7)\) for the mean response at a prespecified predictor value. Is it correct to say there is a 95% probability that the fixed mean lies in these observed endpoints? State the correct interpretation. (Answer: No; the interval either covers the fixed mean or does not. The procedure has 95% repeated-sampling coverage under its assumptions.)
- For \(n=20\), \(\bar{x}=4\), \(S_{xx}=50\), \(s=3\), and \(x_0=9\), the fitted mean is 27. Assuming the usual model and using \(t_{0.975,18}=2.101\), find the 95% interval to two decimals. (Answer: standard error \(3\sqrt{0.55}=2.2249\), margin \(4.6744\), interval \((22.33,31.67)\).)
- At each of four predictor values, an analyst calculates a separate 95% pointwise interval. Does this alone guarantee that all four intervals cover their respective means with probability 0.95? (Answer: No; each interval has its own pointwise 95% coverage statement, but those statements do not establish 95% joint coverage.)
- A proposed exact \(t\)-interval uses \(n=2\) observations with distinct predictor values and otherwise satisfies the normal-error assumptions. Can the usual residual-based interval be calculated? Explain the key condition. (Answer: No; \(n-2=0\) residual degrees of freedom, so the error variance cannot be estimated by the stated procedure. The exact interval requires \(n>2\).)