Statistics › Section L · Linear regression and correlation › Topic L.5
Predicting with a regression equation
Fitting the line was preparation; using it is the point. But prediction comes in two kinds - the average outcome at some x, or one particular individual's - and they carry very different uncertainties. Naming which question you are answering is the skill.
The point prediction
Prediction with a fitted line is substitution: plug the \(x\) value into \(\hat{y} = b_0 + b_1x\) and report the result, with units, in context. The house-price line \(\hat{y} = 108 + 9.6x\) (\(y\) in thousands of dollars, \(x\) in hundreds of square feet) predicts \(108 + 9.6 \times 24 = 338.4\) thousand dollars for a 2,400-square-foot house. The point prediction is the centre of every prediction story - but the honest story attaches the uncertainty, and the uncertainty depends on which question was asked.
Two different questions
"What is the AVERAGE price of 2,400-square-foot houses?" and "What will THIS 2,400-square-foot house sell for?" use the same \(x\) and produce the same point prediction - but they are different questions with different uncertainties. The first asks about the mean response at that size: a quantity the line estimates fairly tightly. The second asks about one individual house, and individual houses scatter around the line by their own residuals - a spread that dwarfs the line's own wobble. Confusing the two produces confident nonsense: narrow intervals quoted for wildly uncertain outcomes.
Why the widths differ
Both uncertainties start with the fitted line's own wobble - the line came from a sample, and a fresh sample would tilt it slightly. For the mean response, that wobble is the whole story: the average price at a given size is pinned fairly tightly near the fitted value. For an individual prediction, the residual scatter adds on top: the residual standard deviation \(s_e\) measures how far single houses typically sit from the line, and that spread is usually several times the line's own uncertainty. Mean estimates sharpen with larger samples; individual predictions stay as wide as the scatter itself - no sample size makes a single house's price predictable.
Worked example: an individual house
The house line has residual standard deviation \(s_e = 36\) thousand dollars (a figure the residuals topic ahead returns to) - single houses typically sit about $36,000 from the line. Predict the price of one 2,000-square-foot house (\(x = 20\)): 1. Point: \(108 + 9.6 \times 20 = 300\) thousand dollars. 2. Individual spread: the house's actual price will sit within roughly two residual SDs of the prediction - about \(300 \pm 2 \times 36\), or 228 to 372 thousand dollars (software gives the exact prediction interval). 3. Check: that width is honest - individual houses vary a great deal around any size-based prediction, and $144,000 of plausible spread is what the data's scatter says.
Worked example: the average at that size
Now predict the AVERAGE price of all 2,000-square-foot houses (still \(x = 20\), from a sample of \(n = 50\) houses): 1. Point: the same 300 thousand dollars. 2. At the mean size, \(x = \bar{x} = 20\), the line's own wobble is much smaller - its standard error there is \(s_e/\sqrt{n} = 36/\sqrt{50} \approx 5.1\) thousand, so the average is pinned within about \(300 \pm 2 \times 5.1\), roughly 290 to 310 thousand dollars. 3. Contrast: the individual house's plausible range was 228 to 372; the average's is 290 to 310 - about seven times narrower (\(\sqrt{50} \approx 7.1\)), because averaging cancels the house-to-house scatter that dominates any single prediction.
Precision depends on where you ask
The line's wobble is not constant: the fitted line is anchored at the mean point \((\bar{x}, \bar{y})\) and tilts more freely at the ends, so mean-response intervals are narrowest at \(x = \bar{x}\) and widen as the prediction moves away from the centre of the data. At \(x = 20\) (the houses' mean size) the average's interval was roughly \(\pm 10\); at \(x = 24\) - near the data's edge - it is roughly 30% wider, and beyond the observed range the widening turns into extrapolation: the formula will still produce numbers, but the model's support vanishes. Predict where the data live, and trust the centre more than the edges.
Prediction hygiene
Four habits keep predictions honest. Interpolate: predict inside the observed range of \(x\), and treat edge predictions as provisional. Name the question: mean response or individual - before quoting any interval. Attach units and context: "about 338 thousand dollars, give or take" is a sentence; "338.4" is not. And report the uncertainty every time: a bare point prediction borrows the line's authority without its honesty.
Common mistakes to avoid
First, quoting a mean-response interval for an individual outcome - the classic overconfidence, roughly seven times too narrow in the house example. Second, extrapolating: the line speaks only where the data spoke. Third, trusting edge predictions as much as central ones - the interval's width grows away from \(\bar{x}\). Fourth, quoting points without uncertainty. Fifth, forgetting units. Sixth, expecting large samples to make individual predictions sharp: \(n\) tightens the line, never the scatter around it.
Key takeaways
A fitted line predicts by substitution: \(\hat{y} = b_0 + b_1x\), 338.4 thousand for the 2,400-square-foot house. Two questions share the point but not the uncertainty: the mean response at \(x\) is pinned by the line's own wobble (narrowest at \(\bar{x}\), widening outward - roughly \(\pm 10\) thousand at the houses' centre), while an individual outcome adds the residual scatter \(s_e\) on top (about \(\pm 72\) thousand for one house at \(s_e = 36\)) - some seven times wider, and no sample size shrinks it. Interpolate, name the question, attach units and uncertainty, and trust the data's centre over its edges.
Flashcards
Flashcard
Prediction by substitution
Test your understanding
Question 1 of 5
For the line ŷ = 108 + 9.6x, the point prediction at x = 24 is:
Question 2 of 5
Predicting ONE house's price rather than the average price at that size adds:
Question 3 of 5
With s_e = 36 thousand, the rough 95% range for ONE house predicted at 300 thousand is:
Question 4 of 5
Why is the interval for the AVERAGE price at a size so much narrower than for one house?
Question 5 of 5
Mean-response intervals from a regression line are narrowest:
