Section 33 Model Diagnostics


Estimates of Residuals

\[ \large \hat\epsilon_{i} = y_i - \hat y_{i} \]


33.1 Plots

  • Properties of residual as \(\large \epsilon_{ij} \sim NID(0, \sigma^2)\)

  • Various diagnostic tools are available in order to check these assumptions:

    - Histogram of residuals: should be approximately $\sim NID(0, \sigma^2)$

- QQ plot should show the distribution of residual as Normal

- Residuals versus fitted values

- Residuals versus explanatory variables

- Residuals versus any other variables not included in the model

33.2 Standard residual plots using lm object

plot(fm)

33.3 Customised residual plots

33.3.1 Calculate predicted values and residuals

pred <- predict(object = fm, data=BP)

res <- fm$residuals


33.3.2 Histogram of residuals

pred <- predict(object = fm, data=BP)

res <- fm$residuals

hist(x=res, freq = TRUE, 
     main = '', 
     xlab = 'Residuals', 
     ylab = 'Frequency', 
     axes = TRUE, 
     col = 'lightblue', 
     lty = 1, border = 'purple')

hist(x=res, breaks=10, freq = FALSE, 
     xlab = 'Residuals', 
     ylab = 'Density', 
     axes = TRUE, 
     col = 'lightblue', 
     lty = 1, border = 'purple')
lines(density(res), col = 'red', lwd = 2, lty = 1)

33.3.3 QQ plot of residuals

qqnorm(y=res, main='Normal QQ of residuals',
     xlab='Theoretical Quantiles',
     ylab='Residuals',
     col='blue', pch=20)
qqline(y=res, lty=1, lwd=2, col='red')


33.3.4 Scatter plot of residual against fitted value

plot(x=pred, y=res, pch=20, col='purple')
# lo <- loess(res ~ pred)
# lines(predict(lo), col='red', lty=1, lwd=2)
abline(a = 0, b = 0, lty=2, col='red')