Mean Response Versus Future Observation

Two Different Targets at the Same Predictor Value

A regression line can answer more than one question at a given predictor value. It can estimate the average response among cases with that predictor value, or it can help predict the response of one particular future case. Those targets are related, but they are not the same. Choosing an interval before identifying the target can produce a mathematically correct calculation that answers the wrong question.

For example, a hypothetical insurer might study the relationship between a policy's age and its annual service cost. At a policy age of five years, the estimated average cost across policies of that age is one target. The actual cost for one new five-year-old policy is another. Even when both use the same fitted regression line, they require different interpretations and different intervals.

Model and assumptions: In simple linear regression, write $Y_i=\beta_0+\beta_1x_i+\varepsilon_i$, where $x_i$ is the predictor for case $i$, and $Y_i$ is its response. The errors have mean zero and common variance $\sigma^2$, and are independent across cases. For the usual exact $t$-based regression intervals, assume in addition that the errors are normally distributed. The predictor values are treated as fixed for the standard regression calculations.

At a specified predictor value $x_0$, the conditional mean response is $\mu_{Y\mid x_0}=E[Y\mid X=x_0]=\beta_0+\beta_1x_0$. It describes the population average response at that value of the predictor. Because the coefficients are unknown, the fitted line estimates this mean with $\hat y_0=b_0+b_1x_0$, where $b_0$ and $b_1$ are the estimated intercept and slope.

A future response at $x_0$, denoted $Y_{\text{new}}$, is one outcome for one new case. Under the model, $Y_{\text{new}}=\beta_0+\beta_1x_0+\varepsilon_{\text{new}}$. Its expected value is the conditional mean, but the realized response also contains the new case's error. Therefore, the estimated mean and the future observation share a center, $\hat y_0$, but the future observation has an additional source of variation.

Which Interval Matches the Question?

A confidence interval for the mean response at $x_0$ targets the unknown average $\mu_{Y\mid x_0}$. Its uncertainty comes from estimating the regression line using a sample. A prediction interval for a future response at $x_0$ targets the particular random outcome $Y_{\text{new}}$. It must account both for uncertainty in the fitted mean and for the natural case-to-case variation around that mean.

Question wording	Target	Interval to use
“What is the average response for cases with predictor $x_0$?”	$\mu_{Y\mid x_0}$, the conditional mean	Confidence interval for the mean response
“What response might one new case with predictor $x_0$ have?”	$Y_{\text{new}}$, one future outcome	Prediction interval for a future response
“Estimate the fitted response at $x_0$.”	$\hat y_0$, a point estimate	No interval is specified unless uncertainty is requested

For reference, in the normal simple linear regression model with $n$ observations, the usual two-sided $100(1-\alpha)\%$ intervals are centered at $\hat y_0$. Let $\bar{x}$ be the sample mean of the predictor, $S_{xx}=\sum_{i=1}^n(x_i-\bar{x})^2$, and $s$ the residual standard error. The intervals have the following forms:

\text{Mean-response CI: }\quad \hat y_0\ \pm\ t_{1-\alpha/2,\,n-2}\, s\sqrt{\frac{1}{n}+\frac{(x_0-\bar{x})^2}{S_{xx}}}

\text{Future-response PI: }\quad \hat y_0\ \pm\ t_{1-\alpha/2,\,n-2}\, s\sqrt{1+\frac{1}{n}+\frac{(x_0-\bar{x})^2}{S_{xx}}}

These formulas are included to make the distinction visible, not to replace identifying the target. The extra $1$ in the prediction interval is the contribution from the new response's own error. Under the stated assumptions, the prediction interval is wider than the mean-response confidence interval at the same $x_0$, confidence level, and fitted model. The formulas rely on the specified regression assumptions; if errors are not independent or have nonconstant variance, the usual intervals may not have their stated properties.

Why the Future-Response Interval Has Extra Uncertainty

The difference follows from what each interval is trying to cover. A mean-response interval concerns the unknown value $\mu_{Y\mid x_0}$. A future-response interval concerns $\mu_{Y\mid x_0}+\varepsilon_{\text{new}}$, which includes a new random error in addition to uncertainty about the mean.

For a short justification, let $\hat y_0$ estimate the conditional mean and consider the prediction error $Y_{\text{new}}-\hat y_0$. It can be written as the sum of the error in estimating the mean and the new case's error:

Y_{\text{new}}-\hat y_0 = (\mu_{Y\mid x_0}-\hat y_0)+\varepsilon_{\text{new}}.

The new error is independent of the sample used to fit the line, has mean zero, and has variance $\sigma^2$. Consequently, the variance of the prediction error is the variance of the mean-estimation error plus $\sigma^2$. Estimating that variance adds the $1$ inside the prediction interval's square root. The mean-response interval does not include this new-case term because it is not trying to cover a realized individual response.

Worked Examples

Worked Example: Average Processing Time or One New Case?

A hypothetical claims team fits a simple linear regression of processing time $Y$, in days, on document count $x$. The fitted line is $\hat y=4.5+1.2x$. At $x_0=8$ documents, the team asks for (a) the estimated average processing time for all comparable cases and (b) a range for the processing time of one new case with eight documents. Identify each target, calculate the fitted value, and select the appropriate interval type.

Method. First calculate the fitted response at the specified predictor value. Then distinguish an average across comparable cases from an individual outcome.

Step 1: Calculate the fitted response.

\hat y_0=4.5+1.2(8)=4.5+9.6=14.1\text{ days}.

Step 2: Match the first question to its target. Part (a) asks for the average processing time among comparable cases with eight documents. That target is $\mu_{Y\mid 8}$, so an interval for it is a confidence interval for the mean response. The point estimate is 14.1 days.

Step 3: Match the second question to its target. Part (b) asks about one new case. Its response includes that case's own deviation from the average, so the appropriate interval is a prediction interval, also centered at the fitted value of 14.1 days.

Check. Both targets have the same fitted center, 14.1 days, but they are different quantities: an average versus a single outcome. No interval endpoints can be calculated from the fitted line alone; additional information about the regression's residual variation and sample design is needed.

Common mistake avoided. Using the mean-response confidence interval for the new case would address an average, not the requested individual processing time.

Worked Example: Calculate the Appropriate Interval from Supplied Width Information

A hypothetical service provider models annual repair cost $Y$, in hundreds of dollars, using equipment age $x$, in years. At $x_0=6$, the fitted value is $\hat y_0=18$. A regression calculation supplies a 95% mean-response confidence interval half-width of 2.4 and a 95% future-response prediction interval half-width of 7.8. Find the interval for the average cost at age six and the interval for one new unit at age six. Report the results in dollars.

Method. Use the supplied half-width that matches each target. Since the response is measured in hundreds of dollars, convert each endpoint from hundreds of dollars to dollars by multiplying by 100.

Step 1: Average response. The target is the average repair cost for all comparable six-year-old units, so use the mean-response confidence interval:

18\pm2.4=[18-2.4,\ 18+2.4]=[15.6,\ 20.4]

In dollars, this is $[15.6(100),20.4(100)]=[\$1{,}560,\$2{,}040]$.

Step 2: One future unit. The target is an individual future repair cost, so use the prediction interval:

18\pm7.8=[18-7.8,\ 18+7.8]=[10.2,\ 25.8]

In dollars, this is $[10.2(100),25.8(100)]=[\$1{,}020,\$2{,}580]$.

Check. Both intervals are centered at $18$ hundreds of dollars, or $\$1{,}800$. The prediction interval is wider: its total width is $25.8-10.2=15.6$ hundreds of dollars, compared with $20.4-15.6=4.8$ hundreds for the mean-response interval.

Common mistake avoided. Applying the smaller half-width to an individual unit would understate the variation relevant to that target. The interval type is determined by the question, not by which interval looks more plausible.

Worked Example: Separate a Group Average from an Individual Prediction

A hypothetical laboratory models assay time $Y$, in minutes, as a function of the number of preparation steps $x$. At $x_0=5$, the fitted mean is 32 minutes. At the 90% level, the supplied mean-response confidence interval is $[30,34]$ minutes, and the supplied prediction interval for one new assay is $[24,40]$ minutes. A supervisor asks whether the average assay time at five steps is below 35 minutes and whether a particular new assay can be guaranteed to finish below 35 minutes. Answer each question using the interval information.

Method. For the average question, compare the entire confidence interval for the mean with 35. For the individual question, use the prediction interval and distinguish an interval estimate from a guarantee.

Step 1: Evaluate the average-response question. The entire mean-response interval lies below 35 because its upper endpoint is 34 and $34<35$. Thus, at the stated confidence level, the interval supports the conclusion that the mean response at five preparation steps is below 35 minutes.

Step 2: Evaluate the future-assay question. The prediction interval runs from 24 to 40 minutes. Since $40>35$, this interval does not place all plausible future responses below 35 minutes. In particular, it does not support a claim that one new assay will finish below 35 minutes.

Step 3: State the limits of the conclusion. A 90% prediction interval is not a guarantee about a specific outcome. Under the model assumptions and repeated use of the same procedure, such intervals are constructed to cover future responses at the stated rate. The interval does not say that every future response must fall between 24 and 40 minutes.

Check. The fitted value, 32 minutes, is inside both intervals. The mean question uses $[30,34]$; the individual question uses $[24,40]$. The different conclusions follow from the different targets, not from different fitted centers.

Common mistake avoided. Treating the mean-response interval as a range for every individual assay confuses an average with a case-level outcome. Also, a prediction interval is not a deterministic guarantee.

Common Mistakes and Exam Traps

The most frequent error is choosing an interval from a familiar phrase without checking what is being covered. “At predictor value $x_0$” only specifies where the model is evaluated; it does not identify whether the target is a mean or an individual response. Read for words such as average, mean, typical across cases, one new case, and future observation.

Wrong target: A question asks for one future response, but the solver uses a confidence interval for the mean. The correction is to use a prediction interval because an individual response includes new-case variation.
Wrong interpretation: A confidence interval for the mean is described as containing a specified percentage of individual outcomes. It concerns the unknown conditional mean, not a proportion of individual observations.
False guarantee: A prediction interval is described as a promise that the next response will lie inside it. It is a procedure with stated model-based coverage, not certainty for one case.
Confusing point estimate and interval: The fitted value $\hat y_0$ is a point estimate shared by both interval types. The fact that it is the same does not make the intervals interchangeable.

Key takeaway: Identify the target before selecting a regression interval. An average response at $x_0$ calls for a confidence interval for the mean response; one future response at $x_0$ calls for a prediction interval. Both are centered on the fitted value, but only the prediction interval includes variation from the new case.

Check Your Understanding

For each question, identify the target and select the relevant interval or conclusion. Assume the stated regression model is appropriate wherever intervals are supplied.

A hypothetical call center fits a model for call duration from the number of issues reported. At seven issues, the fitted value is 18 minutes. The question asks for the average duration among all calls with seven issues. Which interval type is appropriate, and what is the point estimate? (Answer: Confidence interval for the mean response; 18 minutes.)
A model estimates the mean delivery delay at a route length of 40 miles as 12 minutes. A dispatcher asks about the delay of one new delivery on that route. Which target is requested? (Answer: One future response; use a prediction interval centered at 12 minutes.)
At $x_0=3$, a fitted model gives $\hat y_0=9.5$, in thousands of dollars. The supplied 95% mean-response half-width is 0.8 thousand dollars. Find the mean-response confidence interval in dollars. (Answer: $[8.7,10.3]$ thousand dollars, or $[\$8{,}700,\$10{,}300]$.)
For one new component with a specified predictor value, a 90% prediction interval is $[41,59]$ hours. Can the result be stated as a guarantee that this component will last between 41 and 59 hours? (Answer: No; it is a model-based prediction interval, not a guarantee for an individual outcome.)
At the same predictor value and confidence level, a mean-response interval has half-width 1.6 and a prediction interval has half-width 5.2. A question asks whether the population average response is below 20. Which interval should be compared with 20, and why? (Answer: Compare the mean-response confidence interval with 20 because the target is the conditional average, not an individual future response.)

Question wording	Target	Interval to use
“What is the average response for cases with predictor \(x_0\)?”	\(\mu_{Y\mid x_0}\), the conditional mean	Confidence interval for the mean response
“What response might one new case with predictor \(x_0\) have?”	\(Y_{\text{new}}\), one future outcome	Prediction interval for a future response
“Estimate the fitted response at \(x_0\).”	\(\hat y_0\), a point estimate	No interval is specified unless uncertainty is requested

What You'll Learn