1

Suppose we perform linear regression in some data and the model is correct (Y is a linear combination of X plus a normal iid error term). We know by assumption that the $e_i\sim\mathcal N(0,\sigma^2)$. But what about the residuals $\hat{e}_i=Y_i-X_i\hat b$? what is their distribution and why?

1 Answers1

1

The 'hat' matrix $H=X(X^TX)^{-1}X^T$ is so named because $\hat Y=HY$. This means $Y-\hat Y = (I-H)Y$, where $I$ is an identity matrix.

If you multiple an $N(0,V)$ vector by a matrix $A$ you get an $N(0,A^TVA)$ vector. Applying this with $A=I-H$ and $V=\sigma^2I$, we have $$\hat e\sim N\left(0, \sigma^2(I-H)^T(I-H)\right)$$ So, the residuals are multivariate Normally distributed, but they aren't independent and they don't all have the same variance

Thomas Lumley
  • 38,062