Residual Plots - College Statistics

A residual plot for a linear regression model contains one point with an extremely large negative residual. All other residuals are clustered in a narrow band around zero. From the residual plot alone, what can be concluded about this single point?
Answer: The model's prediction for this point was substantially higher than the actual observed value. A residual is $y_{actual} - y_{predicted}$ . An extremely large negative residual means this difference is a large negative number, which implies that $y_{predicted}$ was much larger than $y_{actual}$ . Therefore, the model substantially overestimated the response. The residual plot itself does not provide enough information to determine leverage, influence, or whether the point is an error.
An analyst examines two residual plots from two different linear regression models. Plot A shows a clear sine-wave pattern. Plot B shows a random cloud of points that gets wider as the fitted values increase. Which statement best distinguishes the problems in Model A and Model B?
Answer: Model A violates the linearity assumption, while Model B violates the constant variance assumption. Any systematic, non-random pattern in the residuals (like a curve or sine wave) suggests the relationship between the variables is not linear. Therefore, Model A violates the linearity assumption. A pattern where the vertical spread of the residuals changes as the fitted values change (like a fan or cone) suggests that the variance of the errors is not constant. Therefore, Model B violates the constant variance (homoscedasticity) assumption.
The residual plot for a linear regression of y on x shows that the residuals are consistently positive for small x-values, negative for medium x-values, and positive again for large x-values. Which of the following is the most likely appearance of the original scatterplot of y versus x with the least-squares regression line superimposed?
Answer: The points form a U-shaped curve, and the regression line lies below the points at the ends and above them in the middle. A positive residual means observed y > predicted y (the point is above the line). A negative residual means observed y < predicted y (the point is below the line). The description matches points that are above the line for small and large x, and below the line for medium x. This pattern occurs when a straight line is fit to a U-shaped (concave up) scatter of points.
A linear regression of weight on height for a sample of adults produces a residual plot with a pronounced fan shape, opening to the right. A biologist suggests transforming the response variable, weight, to improve the model. Which transformation is most likely to be helpful in this situation, and why?
Answer: Taking the natural logarithm of the weight, because this will compress larger values more than smaller values, potentially stabilizing the variance. A fan-shaped residual plot signifies non-constant variance (heteroscedasticity), where the error variance grows with the predicted value. To correct this, a transformation that 'pulls in' the larger values more than the smaller values is often used. The logarithmic transformation is a common and effective choice for this purpose, as it can stabilize the variance. Squaring would make the problem worse. Standardization does not change the fundamental pattern of the residuals. A reciprocal is a strong transformation usually used for different kinds of patterns.
The least-squares regression line to predict sales ( $\hat{y}$ ) from advertising spending ( $x$ ) is given by $\hat{y} = 500 + 3.5x$ . The corresponding residual plot is shown. For the data point corresponding to an advertising spend of $x=200$ , what was the actual sales figure ( $y$ )? Refer to the residual plot shown.
Answer: 1150 This is a multi-step problem. First, calculate the predicted sales ( $\hat{y}$ ) for $x=200$ : $\hat{y} = 500 + 3.5(200) = 500 + 700 = 1200$ . Second, from the plot, find the residual for the point at $x=200$ , which is -50. Third, use the formula $\text{residual} = y - \hat{y}$ . So, $-50 = y - 1200$ . Solving for $y$ gives $y = 1200 - 50 = 1150$ .
For any simple linear regression model fit using the method of least squares, which of the following statements about the corresponding residual plot is always true?
Answer: The sum, and therefore the mean, of the residuals will be equal to zero. A mathematical property of the least-squares regression method is that it minimizes the sum of squared errors. A consequence of this optimization is that the sum of the simple residuals is always zero (or extremely close to it, allowing for rounding error). The other conditions (normality, constant variance, equal numbers of positive/negative residuals) are desirable properties of a well-fitting model but are not guaranteed to be true for every least-squares fit.
A researcher fits a simple linear regression model to predict crop yield (in kg/hectare) from the amount of a new fertilizer applied (in kg/hectare). A plot of the residuals versus the fitted values shows a distinct U-shaped pattern. Which of the following is the most appropriate conclusion?
Answer: The model systematically overpredicts crop yield for moderate fertilizer amounts and underpredicts it for very low or very high amounts. A U-shaped residual plot indicates that the linear model is not appropriate and a non-linear relationship exists. The 'U' shape means residuals are positive (underprediction) at the extremes (low and high x-values) and negative (overprediction) in the middle range. Thus, the model overpredicts for moderate amounts and underpredicts for low and high amounts.
An economist models the price of a house based on its square footage using a simple linear regression. The resulting residual plot shows the absolute values of the residuals tending to increase as the fitted house price increases. This pattern is often described as a 'fan' or 'cone' shape. What is the primary implication of this observation for the model?
Answer: The model's predictions are less reliable for more expensive houses than for less expensive ones. A fan-shaped residual plot indicates heteroscedasticity, meaning the variance of the residuals is not constant. In this case, the variance of the errors increases as the predicted value increases. This means the model's predictions have more uncertainty (are less reliable) for higher-priced, larger houses.
After fitting a linear model, $\hat{y} = b_0 + b_1x$ , a researcher plots the residuals against a second potential explanatory variable, $z$ , which was not included in the original model. This plot reveals a strong negative linear relationship. What is the most likely implication?
Answer: The variable $z$ contains additional information that could improve the model's predictions if it were included. When residuals from a model show a pattern when plotted against a new variable, it means that the new variable can explain some of the leftover variance that the original model could not. A strong pattern suggests that this new variable, $z$ , is a useful predictor and should be considered for inclusion in a multiple regression model to improve its accuracy.
A simple linear regression model was fit to a dataset. The residual plot is shown. What does the point with a fitted value of approximately 60 indicate? Refer to the residual plot shown.
Answer: For this observation, the model's prediction was a significant underestimate of the actual value. The point in question has a large positive residual (approximately +8). A residual is calculated as (actual value - predicted value). A positive residual means the actual value was greater than the predicted value, so the model underestimated the response for this observation.
A linear regression is performed to predict blood pressure from age. An examination of the residual plot reveals that for individuals between the ages of 20 and 30, nearly all the residuals are positive. What is the most direct interpretation of this finding?
Answer: The linear model tends to underestimate the blood pressure of people in their 20s. When analyzing residual plots in linear regression, you're looking for patterns that reveal how well your model fits the data. Residuals are the differences between actual observed values and predicted values from your regression line. If nearly all residuals are positive for people aged 20-30, this means the actual blood pressure values are consistently higher than what the linear model predicts for this age group. Since residual = actual value - predicted value, positive residuals indicate the model's predictions are too low. Therefore, answer D is correct: the linear model tends to underestimate the blood pressure of people in their 20s. Let's examine why the other options miss the mark. Option A focuses on variability (spread) rather than systematic bias - having all positive residuals indicates consistent underestimation, not increased variability. Option B gets the direction backward: if residuals are positive, the model is predicting values that are too low, not too high (overestimating would create negative residuals). Option C discusses the strength of the relationship, but systematic bias in one age group doesn't necessarily indicate a weaker correlation - it suggests the linear model isn't capturing the true pattern well for that subset. Remember this key insight: when you see a cluster of residuals with the same sign in a specific range, think "systematic bias." Positive residuals mean underestimation, negative residuals mean overestimation. This pattern often suggests the relationship might not be perfectly linear across all values of the predictor variable.
A linear model is created to predict students' final exam scores from the number of hours they studied. The model's residual plot exhibits a strong 'fan-out' pattern (heteroscedasticity), where the residuals are more spread out for students who studied longer. If this model is used to create a 95% prediction interval for the score of a student who studied for many hours, what can be said about this interval?
Answer: The interval will likely be too narrow, understating the true uncertainty in the prediction. Standard prediction intervals are calculated assuming constant variance (homoscedasticity). They use an estimate of error variance that is an average across all data points. When heteroscedasticity is present (variance increases with x), this average variance will be too small for high-x values. Consequently, the calculated prediction interval for a student who studied many hours will be narrower than it should be, giving a false sense of precision.
An analyst evaluates a residual plot that shows a random cloud of points centered at $y=0$ with no change in vertical spread. However, the density of the points is much greater for fitted values between 10 and 20 than for fitted values between 40 and 50. What is a reasonable inference from this observation?
Answer: The original dataset likely had a higher concentration of data points in the lower range of the explanatory variable. When analyzing residual plots in regression, you're checking whether the model meets key assumptions. A residual plot shows residuals (actual minus predicted values) on the y-axis and fitted values on the x-axis. The ideal plot shows a random scatter around zero with constant spread. The key insight here is distinguishing between the pattern of residuals and their density. This plot shows good pattern characteristics—random scatter around zero with constant vertical spread—but varying point density across different fitted value ranges. This density variation reflects the original data distribution, not model problems. Option D is correct because when you have more original data points in lower ranges of your explanatory variable, those points get transformed into fitted values that cluster in corresponding ranges. The regression line maps your original x-values to fitted values, so dense regions in your original data create dense regions in the residual plot. Option A is wrong because a curved relationship would show systematic patterns in the residuals (like a U-shape), not just varying density. Option B incorrectly identifies a homoscedasticity violation—the vertical spread remains constant, so variance is consistent. Option C misses that the residuals are centered at zero; varying density doesn't affect the mean. Strategy tip: On residual plot questions, separate pattern problems (which indicate model violations) from density differences (which reflect your original data distribution). Pattern issues like curves, funneling, or systematic deviations from zero indicate model problems, while uneven point density simply mirrors where your original data was concentrated.
A chemist develops a linear model to predict the reaction rate based on temperature and finds a strong coefficient of determination ( $R^2 = 0.92$ ). However, a plot of the residuals versus temperature shows a clear parabolic (curved) pattern. Which conclusion is most justified?
Answer: Despite the high $R^2$ , the linear model is inappropriate because the relationship is fundamentally non-linear. A residual plot is used to check the assumptions of the linear model. A clear pattern, such as a parabola, indicates that the linearity assumption is violated. Even if the linear model can explain a large portion of the variance (high R-squared), it is not the correct model form. It will produce systematic, predictable errors. The pattern points to a non-linear relationship, not heteroscedasticity (which would be a change in vertical spread).
A student fits a simple linear model to a set of data and finds that the residual plot shows no discernible pattern; the points appear as a random horizontal band. However, the student is concerned because the correlation coefficient is only $r = 0.30$ . What should the student conclude?
Answer: The linear model is appropriate, but the predictive relationship between the variables is weak. When evaluating linear regression models, you need to understand the distinction between model appropriateness and model strength. The residual plot tells you whether a linear model is appropriate, while the correlation coefficient tells you how strong the relationship is. A random, horizontal band pattern in the residual plot is exactly what you want to see—it indicates that the linear model assumptions are met and that linear regression is the appropriate modeling approach. The residuals show no systematic patterns, curved relationships, or increasing/decreasing variance, confirming that the linear model correctly captures the form of the relationship between variables. The correlation coefficient $r = 0.30$ indicates a weak positive linear relationship. While this means the model won't make highly accurate predictions, it doesn't invalidate the appropriateness of using linear regression. Many real-world relationships are legitimately weak but still linear in nature. Answer A is incorrect because residual plots should be random regardless of correlation strength—even weak linear relationships produce random residual patterns when properly modeled. Answer B misunderstands the role of residual plots; they assess model appropriateness, not the correlation coefficient. Answer C suggests an unnecessary step since the residual plot already confirms that linear modeling is appropriate, and switching to non-linear models won't necessarily improve correlation. Remember this key distinction: residual plots diagnose whether your chosen model type fits the data structure, while correlation coefficients measure the strength of that relationship. A good model can have weak predictive power—that's often just the nature of the data, not a modeling error.
A linear model is used to predict daily ice cream sales from the daily high temperature. The data were collected over 90 consecutive summer days. A plot of the residuals versus the observation day (1 to 90) shows that long stretches of positive residuals are followed by long stretches of negative residuals. What regression assumption is most clearly violated?
Answer: Independence of the residuals. Plotting residuals against time or observation order is a way to check for independence. A pattern where residuals are not randomly scattered but instead show trends or cycles indicates that the error in one observation is related to the error in another. This pattern, where positive values tend to follow positive values and negative follow negative, is called positive autocorrelation and is a violation of the independence assumption.
A marketing analyst suspects that the effect of advertising spending on sales is non-linear, with diminishing returns at higher spending levels. After fitting a simple linear model, which diagnostic plot would be the most direct and effective tool for confirming this suspicion?
Answer: A plot of residuals versus fitted values to detect systematic patterns in the model's errors. When you suspect a non-linear relationship after fitting a linear model, you need to examine whether your model is systematically missing the true pattern in the data. The most effective way to detect this is through residual analysis. A residuals versus fitted values plot (option B) is your best diagnostic tool here. If the relationship is truly non-linear with diminishing returns, your linear model will consistently overpredict at low and high spending levels while underpredicting in the middle range. This creates a curved pattern in the residual plot—you'll see negative residuals at the extremes and positive residuals in the middle, forming a distinctive U-shape or inverted U-shape. This systematic pattern directly reveals that your linear model is missing the non-linear structure. Option A is incorrect because a histogram of sales data only shows the distribution of your response variable, not the relationship between variables. Option C, while useful for initial exploration, isn't the most effective after you've already fitted a model—the residual plot provides clearer evidence of model inadequacy by removing the linear trend. Option D checks the normality assumption, which is important for inference but won't help you detect non-linearity in the underlying relationship. Study tip: Remember that residual plots are your primary tool for model validation. When residuals show patterns (curves, trends, or systematic deviations), your model is missing something important about the data structure. Random scatter indicates a well-fitting model, while patterns signal the need for model modifications.
A statistician fits a linear regression model and obtains an R-squared value of 0.22. The corresponding residual plot shows a seemingly random scatter of points with a uniform vertical spread centered on zero. Which statement is the most valid interpretation of these results?
Answer: The linear model is appropriate for the data, but the explanatory variable does not account for a large portion of the variability in the response. A random residual plot suggests that the assumptions of linearity and constant variance are met, meaning the linear model form is appropriate. However, R-squared measures the proportion of variance explained. A low R-squared (0.22) means the explanatory variable has weak predictive power. These two facts are not contradictory; a model can be correctly specified yet have low predictive power.
Consider the residual plot shown, which resulted from fitting a simple linear regression model. Which statement provides the most complete and accurate diagnosis of the model's deficiencies? Refer to the residual plot shown.
Answer: The model is inadequate due to both a non-linear relationship and non-constant variance of the errors. This residual plot shows two distinct problems. First, there is a clear curvilinear (U-shaped) pattern, which indicates that the linearity assumption is violated. Second, the vertical spread of the points is much smaller for fitted values in the middle of the range and larger at the extremes, indicating that the constant variance assumption (homoscedasticity) is also violated. A complete diagnosis must identify both issues.
On a standard residual plot where residuals are plotted against the fitted values ( $\hat{y}$ ), the horizontal line drawn at a residual value of 0 represents which of the following?
Answer: The cases where the model's prediction was perfectly accurate. A residual is defined as the observed value (y) minus the predicted value ( $\hat{y}$ ). A residual of 0 means that $y - \hat{y} = 0$ , which implies $y = \hat{y}$ . Therefore, any point lying on the horizontal line at 0 represents an observation for which the model's prediction was exactly equal to the actual outcome.

Start this quiz