Is the band always wider away from the predictor mean?

In the stated simple-regression formula, yes, because the squared distance from x̄ increases the mean-response standard error.

Does 95% confidence assign a 95% probability to the already fixed true line?

The frequentist statement concerns the long-run coverage of bands constructed by the procedure under its assumptions.

Can it justify forecasting far outside the data?

No. Model validity under extrapolation is a separate issue.

AI Glossary / Statistics and probability / Advanced

Working - Hotelling procedure

Definition

The Working - Hotelling procedure constructs a simultaneous confidence band for the mean response in a linear regression. In simple linear regression, it expands the usual pointwise uncertainty multiplier so that the band covers the whole true regression line with the stated confidence under the model assumptions. It describes uncertainty in the mean relationship, not individual future outcomes.

What it means

A confidence interval at one chosen input value answers a narrower question than a band covering the regression line across inputs. Looking at many ordinary pointwise intervals does not automatically give the advertised confidence that every interval covers its corresponding true mean simultaneously.

The Working - Hotelling construction addresses this simultaneous question. Its wider band reflects protection across the mean-response curve. That protection concerns sampling uncertainty within the specified linear model. It cannot protect against a wrong functional form, a changed population or an unsupported extrapolation beyond where the model is credible.

How it works

For simple linear regression, let n be the sample size, s the residual standard deviation, x̄ the mean predictor value and Sxx = Σ(xᵢ − x̄)². The estimated mean at x is ŷ(x), with standard error s√[1/n + (x − x̄)²/Sxx].

For confidence level 1 − α, multiply this standard error by W = √[2F₁₋α;₂,ₙ₋₂], where F denotes the stated cumulative F quantile. The band is ŷ(x) ± W times the standard error. The classical result assumes the correct linear mean structure, independent normal errors of constant variance and a nondegenerate predictor design. The band is narrowest at x̄ and widens away from it.

Three practical examples

1. Capacity planning: An analyst displays a fitted mean processing-time relationship over a supported workload range. A simultaneous band communicates uncertainty in the whole mean curve when managers examine several possible workloads.

2. Calibration: Engineers estimate mean instrument response against a reference input. The band helps assess uncertainty across the fitted response line, while uncertainty for an individual measurement requires a different calculation.

3. AI performance reporting: A team fits a simple mean-runtime model within a stable experimental setting. It distinguishes a simultaneous band for average runtime from the variability of individual model runs.

Think of it this way

Think of enclosing an entire drawn line inside a protective corridor rather than placing a guardrail at just one selected point. A corridor that protects the line everywhere needs a different allowance. It still surrounds the model's mean line; it does not claim to contain every individual observation scattered around that line.

Worked example

Consider an original fitted-summary illustration with n = 20, x̄ = 5, Sxx = 100 and residual standard deviation s = 2. Suppose the fitted line is ŷ(x) = 10 + 3x. For a 95% band, the 0.95 F quantile with 2 and 18 degrees of freedom is approximately 3.55456. Therefore W = √(2 × 3.55456) ≈ 2.66629.

At x = 5, the estimated mean is 25 and its standard error is 2√(1/20) ≈ 0.44721. The half-width is approximately 1.19240, giving a band slice from 23.8076 to 26.1924.

At x = 8, the estimated mean is 34. Its standard error is 2√[0.05 + 9/100] ≈ 0.74833. The half-width is approximately 1.99527, giving 32.0047 to 35.9953. These are two slices of one simultaneous band, calculated from illustrative fitted summaries. They are not intervals guaranteed to contain 95% of individual future observations.

Illustrative case study

In this fictional case, Birch Processing compares average completion times across job sizes. Its dashboard originally connects ordinary 95% pointwise limits and labels the result “95% confidence for the entire curve.” An analyst replaces that display with the appropriate simultaneous mean-response band and explains the distinction.

The team also adds a separate display of individual-run variation. When managers consider much larger jobs than those observed, the analyst identifies that as extrapolation and requests additional measurements. The wider statistical band is useful, but it is not treated as evidence that the linear relationship continues indefinitely.

Common misconceptions and mistakes

Do not substitute a prediction interval for a mean-response band or interpret the band as a service-level guarantee. Individual outcomes include residual variation beyond the uncertainty in the estimated mean.

The procedure is not simply a collection of ordinary t intervals relabeled as simultaneous. It also differs from Bonferroni intervals for a finite set of specified comparisons. Neither method is uniformly the narrowest for every question. Heteroskedastic or dependent errors need appropriate alternative treatment. The classical formula does not automatically remain valid after arbitrary model selection or repeated searching for a favorable regression specification.

Business use

The band is helpful when a report discusses a whole relationship rather than one isolated prediction point. It can support calibration review, operational planning and communication of regression uncertainty across a meaningful input range.

State the model, confidence level, predictor range and whether the display concerns a mean or individual outcomes. Retain residual diagnostics and the design information used in the calculation. For management decisions, combine statistical uncertainty with practical tolerances; a narrow confidence band can still surround a relationship that is operationally inadequate or irrelevant to the intended use.

Check your understanding

Why is the band wider at eight than at five in the example?

Answer: Eight is farther from the observed predictor mean, increasing estimation uncertainty.

Does the interval at five cover individual processing times with the stated confidence?

Answer: No. It is a slice of a simultaneous confidence band for the mean response.

What determines the residual degrees of freedom here?

Answer: Twenty observations minus two fitted regression coefficients gives eighteen.

Related terms

Prediction interval
F distribution
Simple linear regression
Simultaneous confidence band
pointwise confidence interval
mean response
Bonferroni procedure
extrapolation

Learn next

Next, compare pointwise intervals and a simultaneous band on the same fitted line. Add individual-response prediction intervals and explain why their interpretation differs. Then study how multiple regression and alternative error structures change simultaneous inference.

Multiple regression

Sources

Purdue STAT 525 lecture notes, “Statistical Inference,” derive mean-response uncertainty and introduce the Working–Hotelling multiplier: https://www.stat.purdue.edu/~bacraig/notes525/topic3.pdf

The fitted-summary calculation and Birch Processing case are original teaching illustrations. The numerical F quantile and both band slices were independently calculated under the explicitly stated simple-regression model.

Previous: Workflow engine Next: Workspace representation