Index: The Book of Statistical ProofsStatistical Models ▷ Univariate normal data ▷ Multiple linear regression ▷ Maximum likelihood estimation

Theorem: Given a linear regression model with correlated observations

$\label{eq:MLR} y = X\beta + \varepsilon, \; \varepsilon \sim \mathcal{N}(0, \sigma^2 V) \; ,$

the maximum likelihood estimates of $\beta$ and $\sigma^2$ are given by

$\label{eq:MLE-MLE} \begin{split} \hat{\beta} &= (X^\mathrm{T} V^{-1} X)^{-1} X^\mathrm{T} V^{-1} y \\ \hat{\sigma}^2 &= \frac{1}{n} (y-X\hat{\beta})^\mathrm{T} (y-X\hat{\beta}) \; . \end{split}$

Proof: With the probability density function of the multivariate normal distribution, the linear regression equation \eqref{eq:MLR} implies the following likelihood function

$\label{eq:MLR-LF} \begin{split} p(y|\beta,\sigma^2) &= \mathcal{N}(y; X\beta, \sigma^2 V) \\ &= \sqrt{\frac{1}{(2\pi)^n |\sigma^2 V|}} \cdot \exp\left[ -\frac{1}{2} (y - X\beta)^\mathrm{T} (\sigma^2 V)^{-1} (y - X\beta) \right] \end{split}$

and, using $\lvert \sigma^2 V \rvert = (\sigma^2)^n \lvert V \rvert$, the log-likelihood function

$\label{eq:MLR-LL1} \begin{split} \mathrm{LL}(\beta,\sigma^2) = &\log p(y|\beta,\sigma^2) \\ = &- \frac{n}{2} \log(2\pi) - \frac{n}{2} \log (\sigma^2) - \frac{1}{2} \log |V| \\ &- \frac{1}{2 \sigma^2} (y - X\beta)^\mathrm{T} V^{-1} (y - X\beta) \; . \end{split}$

Substituting the precision matrix $P = V^{-1}$ into \eqref{eq:MLR-LL1} to ease notation, we have:

$\label{eq:MLR-LL2} \begin{split} \mathrm{LL}(\beta,\sigma^2) = &- \frac{n}{2} \log(2\pi) - \frac{n}{2} \log(\sigma^2) - \frac{1}{2} \log(|V|) \\ &- \frac{1}{2 \sigma^2} \left( y^\mathrm{T} P y - 2 \beta^\mathrm{T} X^\mathrm{T} P y + \beta^\mathrm{T} X^\mathrm{T} P X \beta \right) \; . \end{split}$

The derivative of the log-likelihood function \eqref{eq:MLR-LL2} with respect to $\beta$ is

$\label{eq:dLL-dbeta} \begin{split} \frac{\mathrm{d}\mathrm{LL}(\beta,\sigma^2)}{\mathrm{d}\beta} &= \frac{\mathrm{d}}{\mathrm{d}\beta} \left( - \frac{1}{2 \sigma^2} \left( y^\mathrm{T} P y - 2 \beta^\mathrm{T} X^\mathrm{T} P y + \beta^\mathrm{T} X^\mathrm{T} P X \beta \right) \right) \\ &= \frac{1}{2 \sigma^2} \, \frac{\mathrm{d}}{\mathrm{d}\beta} \left( 2 \beta^\mathrm{T} X^\mathrm{T} P y - \beta^\mathrm{T} X^\mathrm{T} P X \beta \right) \\ &= \frac{1}{2 \sigma^2} \left( 2 X^\mathrm{T} P y - 2 X^\mathrm{T} P X \beta \right) \\ &= \frac{1}{\sigma^2} \left( X^\mathrm{T} P y - X^\mathrm{T} P X \beta \right) \end{split}$

and setting this derivative to zero gives the MLE for $\beta$:

$\label{eq:beta-MLE} \begin{split} \frac{\mathrm{d}\mathrm{LL}(\hat{\beta},\sigma^2)}{\mathrm{d}\beta} &= 0 \\ 0 &= \frac{1}{\sigma^2} \left( X^\mathrm{T} P y - X^\mathrm{T} P X \hat{\beta} \right) \\ 0 &= X^\mathrm{T} P y - X^\mathrm{T} P X \hat{\beta} \\ X^\mathrm{T} P X \hat{\beta} &= X^\mathrm{T} P y \\ \hat{\beta} &= \left( X^\mathrm{T} P X \right)^{-1} X^\mathrm{T} P y \end{split}$

The derivative of the log-likelihood function \eqref{eq:MLR-LL1} at $\hat{\beta}$ with respect to $\sigma^2$ is

$\label{eq:dLL-ds2} \begin{split} \frac{\mathrm{d}\mathrm{LL}(\hat{\beta},\sigma^2)}{\mathrm{d}\sigma^2} &= \frac{\mathrm{d}}{\mathrm{d}\sigma^2} \left( - \frac{n}{2} \log (\sigma^2) - \frac{1}{2 \sigma^2} (y - X\hat{\beta})^\mathrm{T} V^{-1} (y - X\hat{\beta}) \right) \\ &= - \frac{n}{2} \, \frac{1}{\sigma^2} + \frac{1}{2 (\sigma^2)^2} (y - X\hat{\beta})^\mathrm{T} V^{-1} (y - X\hat{\beta}) \\ &= - \frac{n}{2 \sigma^2} + \frac{1}{2 (\sigma^2)^2} (y - X\hat{\beta})^\mathrm{T} V^{-1} (y - X\hat{\beta}) \end{split}$

and setting this derivative to zero gives the MLE for $\sigma^2$:

$\label{eq:s2-MLE} \begin{split} \frac{\mathrm{d}\mathrm{LL}(\hat{\beta},\hat{\sigma}^2)}{\mathrm{d}\sigma^2} &= 0 \\ 0 &= - \frac{n}{2 \hat{\sigma}^2} + \frac{1}{2 (\hat{\sigma}^2)^2} (y - X\hat{\beta})^\mathrm{T} V^{-1} (y - X\hat{\beta}) \\ \frac{n}{2 \hat{\sigma}^2} &= \frac{1}{2 (\hat{\sigma}^2)^2} (y - X\hat{\beta})^\mathrm{T} V^{-1} (y - X\hat{\beta}) \\ \frac{2 (\hat{\sigma}^2)^2}{n} \cdot \frac{n}{2 \hat{\sigma}^2} &= \frac{2 (\hat{\sigma}^2)^2}{n} \cdot \frac{1}{2 (\hat{\sigma}^2)^2} (y - X\hat{\beta})^\mathrm{T} V^{-1} (y - X\hat{\beta}) \\ \hat{\sigma}^2 &= \frac{1}{n} (y - X\hat{\beta})^\mathrm{T} V^{-1} (y - X\hat{\beta}) \end{split}$

Together, \eqref{eq:beta-MLE} and \eqref{eq:s2-MLE} constitute the MLE for multiple linear regression.

Sources:

Metadata: ID: P78 | shortcut: mlr-mle | author: JoramSoch | date: 2020-03-11, 12:27.