AP Statistics: Residuals and Residual Plots

In statistics, fitting a linear model is only half the battle. The real test is determining whether that line is a good description of the relationship in your data. Residuals—the differences between observed and predicted values—are your primary diagnostic tool. By calculating and visualizing these prediction errors, you move beyond simply drawing a best-fit line to rigorously assessing its adequacy, accuracy, and the very assumptions that make linear regression valid. Mastering residuals and residual plots is what separates a superficial analysis from a truly critical statistical evaluation.

The Residual: Defining the Prediction Error

A residual is the vertical distance between an observed data point and the point predicted by the regression line. It is calculated using a simple but powerful formula:

$$\text{residual}=\text{observed }y-\text{predicted }\hat{y}$$

Or, more compactly: $e=y-\hat{y}$.

The sign of the residual is crucial. A positive residual means the observed value lies above the regression line (the model underestimated), while a negative residual means the observed value lies below the line (the model overestimated). A residual of zero is a perfect prediction. For example, imagine you have a regression line predicting a student's final exam score ($\hat{y}$) based on their midterm score. If a student scored 88 on the final (observed $y$) but the model predicted 85 ($\hat{y}$), their residual is $88-85=+3$. The model underestimated their performance by 3 points.

To find a predicted value $\hat{y}$, you substitute the corresponding $x$-value into the least-squares regression equation: $\hat{y}=a+bx$, where $a$ is the $y$-intercept and $b$ is the slope. Calculating residuals for all points is the first step in diagnosing your model.

Constructing and Interpreting Residual Plots

While the regression equation and correlation coefficient $r$ give you a summary, a residual plot provides a detailed visual diagnostic. This scatterplot graphs the explanatory variable ($x$) on the horizontal axis and the residuals ($e$) on the vertical axis. Sometimes, $\hat{y}$ is plotted on the horizontal axis instead; the interpretation is the same.

The fundamental question a residual plot answers is: Does the linear model capture all the systematic pattern in the data? To assess this, you look for the absence of pattern.

• What a "Good" Plot Shows: A residual plot that confirms the appropriateness of a linear model displays no discernible pattern. The residuals should be randomly scattered in a horizontal band centered around zero, as shown in the figure below. This randomness suggests that the line has accounted for all the non-random, linear trend, leaving only unpredictable variation.

• Patterns That Signal Problems: Distinct patterns in the residual plot are red flags indicating the linear model is incomplete or inappropriate.
• A Curved Pattern (e.g., a U-shape or arch): This is the most common signal that the relationship is nonlinear. A quadratic, exponential, or other nonlinear model may be needed.
• A Funnel Shape (increasing or decreasing spread as $x$ increases): This indicates non-constant variance or heteroscedasticity. It violates the linear regression condition of equal variability around the line for all values of $x$.
• Systematic Clustering: Groups of consistently positive or negative residuals for certain ranges of $x$ also suggest a missed pattern.

The key takeaway is this: patterns in the residual plot reveal problems with the model (linearity, constant variance), while random scatter supports its use.

The Residual Standard Deviation: Quantifying Prediction Error

The residual plot gives a visual sense of error spread, but the residual standard deviation ($s$, or sometimes $s_e$) provides a precise numerical measure of typical prediction error. It is calculated as:

$$s=\sqrt{\frac{\sum ({\text{residuals}}^2)}{n-2}}=\sqrt{\frac{\sum (y-\hat{y}{)}^2}{n-2}}$$

You can think of it as the standard deviation of the prediction errors. The denominator is $n-2$ because two parameters (the slope $b$ and intercept $a$) were estimated from the data. A smaller $s$ indicates that the data points are, on average, closer to the regression line, meaning predictions will generally be more accurate. A larger $s$ indicates more scatter and less precise predictions. In our test score example, if $s=4.5$ points, you could expect most predictions to be within about $\pm 9$ points (roughly $\pm 2s$) of the actual scores. This metric is essential for putting confidence intervals around predictions.

Identifying Influential Points

Not all data points affect the regression line equally. An influential point is an observation that, if removed, would significantly change the slope, intercept, or correlation of the regression line. To be influential, a point typically has to be an outlier in the x-direction (a high leverage point). A point with high leverage that also doesn't follow the trend of the rest of the data (a large residual) is especially powerful.

Consider a dataset on car weight and highway mileage. Most points show a clear negative trend. Now, imagine adding a data point for an extremely heavy vehicle (high leverage in *x$) that, surprisingly, gets good gas mileage (so it lies far from the trend line, creating a large residual). This single point would "pull" the regression line toward it, flattening the slope. Its removal would result in a steeper, likely more accurate, line. You can identify these points visually on a scatterplot and confirm their influence by comparing regression results with and without the point. You should never automatically remove influential points, but you must investigate them for measurement error and report their impact on your analysis.

Common Pitfalls

1. Mistaking "No Pattern" for "No Spread": A good residual plot shows no systematic pattern, but it will still show vertical scatter. The goal is randomness, not a perfectly flat line of points. The amount of scatter is quantified by $s$, and its existence doesn't invalidate the model; it just tells you the inherent variability in your predictions.
2. Using the Wrong Axis or Misinterpreting Curvature: Always double-check the axes of your residual plot. The vertical axis must be the residuals, not the observed $y$-values. Plotting $y$ vs. $x$ will always show the original trend, defeating the diagnostic purpose. Furthermore, a curved pattern doesn't mean the model is "bad"—it means a linear model is bad. It's a directive to try a different type of model.
3. Overlooking the Impact of Influential Points: Failing to check for influential points can lead to a model that describes your overall data poorly. A single point can distort the story. Always create a scatterplot of the original data to visually screen for points with high leverage and large residuals.
4. Confusing $s$ with the Standard Deviation of $y$: The residual standard deviation $s$ measures error after accounting for the linear relationship with $x$. The standard deviation of the $y$-values alone ($s_y$) measures total variation in $y$. A key goal of regression is that $s$ should be substantially smaller than $s_y$, indicating that knowing $x$ helps predict $y$.

Summary

• A residual ($e=y-\hat{y}$) is the difference between an observed data point and the value predicted by the regression model. It is the cornerstone of model diagnostics.
• A residual plot (residuals vs. $x$ or $\hat{y}$) is the primary visual tool for checking the linearity and constant variance assumptions of regression. Random scatter supports the linear model; clear patterns (curves, funnels) indicate it is inadequate.
• The residual standard deviation ($s$) numerically summarizes the average distance of data points from the regression line. It is your measure of typical prediction error and is used to construct prediction intervals.
• Influential points, often outliers in the x-direction with large residuals, can disproportionately alter the regression results. Their presence must be identified and their impact analyzed.