STA 235 · Week 3 · model diagnostics

What error patterns do to prediction intervals

Every row uses the same fitted linear model. Read the prediction plot alongside the four familiar regression diagnostics: residuals versus fitted, normal Q-Q, a residual histogram, and scale-location.

Structural violation: a nonlinear relationship

The data follow a curve, but the fitted model is forced to use one straight line.

Forced linear prediction versus the true curved mean

forced linear fit95% prediction intervaltrue curved meanmissed future outcome
Residuals vs fitted
Normal Q-Q
Residual histogram
Scale-location

Regular errors: normal and equal spread

The usual prediction interval is calibrated for this data-generating story.

Prediction interval and future outcomes

fitted line95% prediction intervalcovered future outcomemissed future outcome
Residuals vs fitted
Normal Q-Q
Residual histogram
Scale-location

Violation: non-normal, right-skewed errors

Occasional large positive errors make the residual distribution asymmetric.

Prediction interval and future outcomes

fitted line95% prediction intervalcovered future outcomemissed future outcome
Residuals vs fitted
Normal Q-Q
Residual histogram
Scale-location

Violation: unequal variance

Errors are small at low x and grow sharply as x increases.

Prediction interval and future outcomes

fitted line95% prediction intervalcovered future outcomemissed future outcome
Residuals vs fitted
Normal Q-Q
Residual histogram
Scale-location