What does a residual plot tell you in regression analysis?

What does a residual plot tell you in regression analysis?

• A. The strength of correlation
• B. Whether the model's assumptions hold, particularly homoscedasticity and linearity.
• C. The overall model fit R-squared
• D. The presence of multicollinearity

Answer: (B)

A residual plot is a diagnostic tool for regression model adequacy. It focuses on the residuals—the differences between observed values and the model’s predictions—and what their pattern says about the assumptions behind the model. The key idea is that, if the model is appropriate, the residuals should look like random noise with no systematic pattern when plotted against the fitted values. This indicates two important things: the spread of residuals is roughly constant across all levels of the fitted values (homoscedasticity), and there’s no curved or other pattern suggesting that the relationship is not truly linear or that important terms are missing. If you see a clear pattern—such as a curve, increasing or decreasing spread with the fitted values, or clusters—that signals potential problems like nonlinearity or heteroscedasticity. In that case, the model may be misspecified and you might need to transform the response, add polynomial or interaction terms, or consider a different modeling approach. This diagnostic is not about the strength of association, nor the overall R-squared, nor about multicollinearity (which is typically assessed with VIFs or similar metrics).

A residual plot is a diagnostic tool for regression model adequacy. It focuses on the residuals—the differences between observed values and the model’s predictions—and what their pattern says about the assumptions behind the model. The key idea is that, if the model is appropriate, the residuals should look like random noise with no systematic pattern when plotted against the fitted values. This indicates two important things: the spread of residuals is roughly constant across all levels of the fitted values (homoscedasticity), and there’s no curved or other pattern suggesting that the relationship is not truly linear or that important terms are missing.

If you see a clear pattern—such as a curve, increasing or decreasing spread with the fitted values, or clusters—that signals potential problems like nonlinearity or heteroscedasticity. In that case, the model may be misspecified and you might need to transform the response, add polynomial or interaction terms, or consider a different modeling approach.

This diagnostic is not about the strength of association, nor the overall R-squared, nor about multicollinearity (which is typically assessed with VIFs or similar metrics).