How to predict y from x with a regression line

Substitute the x-value, then keep the units

A linear regression model uses an explanatory variable $x$ to predict a response variable $y$, and the AP course writes it as $\hat{y} = a + bx$, where $\hat{y}$ (y-hat) is the predicted response, $a$ is the y-intercept, and $b$ is the slope. Making a prediction is one substitution:

1. Identify which variable is $x$. It is the explanatory variable, the one you are predicting from.
2. Put its value in place of $x$ in the equation.
3. Multiply by the slope first, then add or subtract the intercept.
4. Attach the units of $y$ and call the number predicted.

A cross-country coach fit the line $\hat{y} = 28.5 - 0.35x$ to data from 24 runners, where $x$ is average weekly training miles and $y$ is 5K race time in minutes. For a runner averaging 20 miles per week:

Write the result as a sentence: the model predicts a 5K time of 21.5 minutes for a runner averaging 20 miles per week. Dropping the units turns a complete answer into a bare number, and dropping the word "predicted" claims more than the model supports.

Read the equation before you substitute

Most prediction errors happen before any arithmetic. Three things are worth confirming every time.

Which variable is which. The line predicts $y$ from $x$ and only in that direction. If the equation is $\hat{y} = 28.5 - 0.35x$ with $x$ in weekly miles, then substituting a race time in place of $x$ produces a meaningless number. Check the definitions given in the problem, not the letters alone.

The sign of the slope. Here $b = -0.35$, so the term is subtracted. The slope reads as a rate: for each additional weekly training mile, the model predicts the 5K time decreases by 0.35 minutes. Interpreting that coefficient in context is its own skill, covered in how to interpret the slope of a regression line.

The precision of the coefficients. If a calculator reports $b = -0.3512$, substitute -0.3512 and round only the final prediction. Rounding the slope to -0.35 first and then multiplying by a large $x$ can shift the answer noticeably.

If you also need to find $a$ and $b$ in the first place, least-squares regression line walks through both from summary statistics.

Compute a residual to see how far the prediction missed

A prediction is only useful if you can measure how well it did. That measurement is the residual, the difference between what actually happened and what the model predicted:

One runner in the coach's data averaged 30 miles per week and ran a 5K in 19.2 minutes. The prediction at $x = 30$ is $\hat{y} = 28.5 - 0.35(30) = 28.5 - 10.5 = 18.0$ minutes, so the residual is $19.2 - 18.0 = 1.2$ minutes.

The sign carries the meaning, and the AP course states the convention directly. A positive residual means the model underpredicted, so the observed value came in above the prediction. A negative residual means the model overpredicted. This runner's positive residual of 1.2 minutes says she ran 1.2 minutes slower than the model predicted for her training volume.

Residuals are also how you check whether a line belonged there at all. Plot every residual against $x$ and look for a pattern; apparent randomness supports the linear model, while curvature says a line was the wrong choice. See how to interpret residual plots for what each shape means, and why square the residuals for how they define the line itself.

Extrapolation: the x-values you cannot trust

The line was fit using a particular stretch of x-values, and it earns credibility only there. The coach's 24 runners averaged between 12 and 38 miles per week.

Interpolation means predicting at an x-value inside that stretch. Both 20 and 30 miles per week qualify, so those predictions rest on data.

Extrapolation means predicting beyond it, and the AP course warns that a predicted value gets less reliable the further you push. Nothing in the data says the straight-line pattern continues outside 12 to 38 miles, and often it plainly does not.

Push it far enough and the model says something impossible. At 100 miles per week:

The model predicts a 5K time of -6.5 minutes, a runner finishing six and a half minutes before the starting gun. The arithmetic is fine and the answer is nonsense, which is the whole warning: a regression equation will hand you a number for any $x$ you feed it, and it never objects. Compare the two cases in interpolation vs extrapolation.

There is no fixed cutoff for how far is too far. Stay inside the observed range when you can, and label any prediction outside it as an extrapolation whose reliability you cannot vouch for.

A prediction is a model output, not a guarantee

Even a well-behaved prediction from the middle of the data is an estimate of the average response at that x-value, not a promise about any single individual. Two runners both averaging 30 miles per week will not run the same 5K time, yet the line gives them the same $\hat{y} = 18.0$ minutes.

The scatter around the line is the part the model does not explain, and residuals are exactly its size. The coefficient of determination $r^2$ reports the share of the variation in $y$ that the linear relationship with $x$ does explain, so the remainder is variation the prediction cannot account for. Even $r^2 = 0.90$ leaves 10 percent of the variation unexplained, and because that 10 percent is a share of the variance rather than of the spread, the typical miss is larger than it sounds: the leftover standard deviation is $\sqrt{0.10} \approx 0.32$ of the original spread in $y$, so about a third of it survives the model. See r vs r-squared for what that number does and does not cover.

A prediction also says nothing about cause. The line describes how $y$ tends to move with $x$ in the observed data. Concluding that adding 10 weekly miles would cause a given runner's time to drop 3.5 minutes requires an experiment, not a regression fit to observational data. Correlation vs causation covers the gap.

So report a prediction with three hedges built in: the word predicted, the units, and the x-range it came from.

Common mistakes when predicting

These slips cost points on otherwise correct work.

• **Writing $y$ where $\hat{y}$ belongs.** The hat is what marks the number as a prediction. Without it you are claiming an observed value.
• Losing the units. "21.5" is not an answer; "21.5 minutes" is.
• Substituting into the wrong slot. Plugging a y-value into the equation and solving for $x$ does not give the least-squares prediction of $x$, because the line was built to minimize errors in the $y$ direction only.
• Rounding the coefficients before substituting. Round the final prediction, not the inputs.
• Extrapolating without saying so. If $x$ falls outside the data range, say the prediction is an extrapolation and is unreliable.
• **Treating a high $r^2$ as a license to extrapolate.** A strong fit inside the data says nothing about behavior outside it.
• Computing the residual backward. It is observed minus predicted. Reversing the order flips every sign and reverses the underpredict and overpredict reading.

Check your work and practice

Enter a data set in the regression calculator to get $a$, $b$, $r$, and $r^2$, then make a prediction by hand and confirm it against the equation the tool reports. Doing both keeps the arithmetic honest.

To see how fragile a prediction can be, move one point around in the influential point interactive and watch the predicted values shift with the line. Then work the regression practice set, which mixes predictions, residuals, and interpretation in AP-style multi-part questions.