Prediction Versus Observed Values

Compare the Prediction With the Observation

A regression line gives a predicted response \(\hat{y}\) for a specified predictor value \(x\). For a data point that was actually observed, we can compare that prediction with the case’s observed response \(y\). The difference tells us how far the observed value is from the line and which side of the line it falls on.

In “Strong Correlation Does Not Mean a Good Model,” you met the term residual. Here, we focus on using the residual to describe the gap between one prediction and one observed response. Always match the observed \(y\) with the line’s prediction at that same case’s \(x\). Comparing values from different cases would not describe that case’s prediction gap.

The order of subtraction matters. A residual is not \(\hat{y}-y\): it is \(y-\hat{y}\). The sign tells you whether the observation is above or below the regression line. The size tells you how far the observation is from its predicted value, measured vertically on a scatterplot.

• Positive residual: \(y\) is greater than \(\hat{y}\). The point is above the line, and the prediction is lower than the observed response.
• Negative residual: \(y\) is less than \(\hat{y}\). The point is below the line, and the prediction is higher than the observed response.
• Zero residual: \(y\) equals \(\hat{y}\). The point lies on the line, so the predicted and observed responses match.

A residual is expressed in the units of \(y\), not \(x\). For example, if the response is travel time in minutes, a residual of \(4\) is \(4\) minutes. It is not a difference of four kilometers, nor is it automatically a percentage. If a response is recorded as a percentage, state whether the difference is in percentage points.

A Reliable Way to Find and Explain the Gap

For each observed case, use this sequence. It makes the comparison clear and helps prevent sign errors.

A prediction by itself is not a residual. For a future case, the actual response is not yet known, so there is no observed \(y\) to subtract. Once that response is observed, the difference can be calculated. A residual therefore describes the gap for an observed case; it does not promise how far a future response will be from its prediction.

Worked Examples

What the Residual Does—and Does Not—Say

A residual summarizes one observed case’s vertical difference from the regression line. Its sign and magnitude are useful, but they answer a limited question: how did this observed response compare with the value predicted by the line at this case’s predictor value?

• The sign gives direction. A positive residual means the response was greater than predicted; a negative residual means it was less than predicted.
• The magnitude gives the gap’s size. A residual of \(5\) and a residual of \(-5\) have the same distance from zero, but they point in opposite directions.
• The units come from the response. Subtract two \(y\)-values, so the residual uses the units in which \(y\) is measured.
• The residual is case-specific. It compares one observed data point with its own prediction, not the prediction with an unrelated response.
• A small residual does not prove the model is appropriate. It only says that this case’s observed response is close to the line’s prediction. As discussed in “Strong Correlation Does Not Mean a Good Model,” judge a linear model using the overall data pattern as well, not one point alone.

On a scatterplot, the predicted value \(\hat{y}\) lies on the regression line at the case’s \(x\)-value. The observed value \(y\) is the point’s vertical position. The residual is the signed vertical difference between those two responses. A point above the line has a positive residual; a point below it has a negative residual. This visual connection can help you check the sign of your calculation.

Residuals also do not establish that the predictor caused the response to differ. A residual describes a comparison with a fitted line, not the reason for an individual outcome. Keep the interpretation focused on the observed response and the model’s prediction, consistent with “Why Correlation Does Not Imply Causation.”

Common Mistakes and AP Exam Tips

• Reversing the subtraction. Writing \(\hat{y}-y\) reverses the sign. A full-credit calculation starts with \(y-\hat{y}\), observed minus predicted.
• Reporting only the sign. “The residual is positive” is not a complete contextual interpretation. State that the observed response was greater than the predicted response, and include the amount and units.
• Calling a negative residual a negative response. The response itself might be positive. A negative residual means only that the observed response is less than the prediction.
• Using predictor units for the gap. The residual uses \(y\)-units. If \(x\) is measured in hours and \(y\) in dollars, the residual is in dollars.
• Comparing the wrong case and prediction. Calculate \(\hat{y}\) at the observed point’s own \(x\)-value before subtracting.
• Claiming the prediction is what actually happened. Use wording such as “the line predicted” and “the observed response was.” Do not describe \(\hat{y}\) as the case’s guaranteed outcome.
• Confusing a residual with an absolute distance. The signed residual preserves whether the point is above or below the line. If you report only the distance, you lose that direction.

For a clear AP-style response, show the predicted value, calculate observed minus predicted, and interpret the sign and amount using the response’s units. Name the case or context so the reader knows what the gap describes.