web.archive.org

linear model: Definition from Answers.com

In statistics the linear model is given by

$Y = X \beta + \varepsilon$

where Y is an n×1 column vector of random variables, X is an n×p matrix of "known" (i.e. observable and non-random) quantities, whose rows correspond to statistical units, β is a p×1 vector of (unobservable) parameters, and ε is an n×1 vector of "errors", which are uncorrelated random variables each with expected value 0 and variance σ².

Much of the theory of linear models is associated with inferring the values of the parameters β and σ². Typically this is done using the method of maximum likelihood, which in the case of normal errors is equivalent (by the Gauss-Markov theorem) to the method of least squares.

Assumptions

Multivariate normal errors

Often one takes the components of the vector of errors to be independent and normally distributed, giving Y a multivariate normal distribution with mean Xβ and co-variance matrix σ² I, where I is the identity matrix. Having observed the values of X and Y, the statistician must estimate β and σ².

Rank of X

We usually assume that S is of full rank p, which allows us to invert the p × p matrix $X^{\top} X$ . The essence of this assumption is that the parameters are not linearly dependent upon one another, which would make little sense in a linear model. This also ensures the model is identifiable.

Methods of inference

Maximum likelihood

β

The log-likelihood function (for ε_i independent and normally distributed) is

$l(\beta, \sigma^2; Y) = -\frac{n}{2} \log (2 \pi \sigma^2) - \frac{1}{2\sigma^2} \sum_{i=1}^n \left(Y_i - x_i^{\top} \beta \right)^2$

where $x_i^{\top}$ is the ith row of X. Differentiating with respect to β_j, we get

$\frac{\partial l}{\partial \beta_j} = \frac{1}{\sigma^2} \sum_{i=1}^n x_{ij} \left( Y_i - x_i^{\top} \beta \right)$

so setting this set of p equations to zero and solving for β gives

$X^{\top} X \hat{\beta} = X^{\top} Y.$

Now, using the assumption that X has rank p, we can invert the matrix on the left hand side to give the maximum likelihood estimate for β:

$\hat{\beta} = (X^{\top} X)^{-1} X^{\top} Y$ .

We can check that this is a maximum by looking at the Hessian matrix of the log-likelihood function.

σ²

By setting the right hand side of

$\frac{\partial l}{\partial \sigma^2} = -\frac{n}{2\sigma^2} + \frac{1}{2 \sigma^4} \sum_{i=1}^n \left(Y_i - x_i^{\top} \beta \right)^2$

to zero and solving for σ² we find that

$\hat{\sigma}^2 = \frac{1}{n} \sum_{i=1}^n \left(Y_i - x_i^{\top} \hat{\beta} \right)^2 = \frac{1}{n} \| Y - X \hat{\beta} \|^2.$

Accuracy of maximum likelihood estimation

Since we have that Y follows a multivariate normal distribution with mean Xβ and co-variance matrix σ² I, we can deduce the distribution of the MLE of β:

$\hat{\beta} = (X^{\top} X)^{-1} X^{\top} Y \sim N_p (\beta, (X^{\top}X)^{-1} \sigma^2 ).$

So this estimate is unbiased for β, and we can show that this variance achieves the Cramér-Rao bound.

A more complicated argument^[1] shows that

$\hat{\sigma}^2 \sim \frac{\sigma^2}{n} \chi^2_{n-p};$

since a chi-squared distribution with n − p degrees of freedom has mean n − p, this is only asymptotically unbiased.

Generalizations

Generalized least squares

If, rather than taking the variance of ε to be σ²I, where I is the n×n identity matrix, one assumes the variance is σ²M, where M is a known matrix other than the identity matrix, then one estimates β by the method of "generalized least squares", in which, instead of minimizing the sum of squares of the residuals, one minimizes a different quadratic form in the residuals — the quadratic form being the one given by the matrix M⁻¹:

${\min_{\beta}}\left(y-X\beta\right)'M^{-1}\left(y-X\beta\right)$

This has the effect of "de-correlating" normal errors, and leads to the estimator

$\widehat{\beta}=\left(X'M^{-1}X\right)^{-1}X'M^{-1}y$

which is the best linear unbiased estimator for β. If all of the off-diagonal entries in the matrix M are 0, then one normally estimates β by the method of weighted least squares, with weights proportional to the reciprocals of the diagonal entries. The GLS estimator is also known as the Aitken estimator, after Alexander Aitken, the Professor in the University of Otago Statistics Department who pioneered it.^[2]

Generalized linear models

Generalized linear models, for which rather than

E(Y) = Xβ,

one has

g(E(Y)) = Xβ,

where g is the "link function". The variance is also not restricted to being normal.

An example is the Poisson regression model, which states that

Y_i has a Poisson distribution with expected value e^γ+δx_i.

The link function is the natural logarithm function. Having observed x_i and Y_i for i = 1, ..., n, one can estimate γ and δ by the method of maximum likelihood.

General linear model

The general linear model (or multivariate regression model) is a linear model with multiple measurements per object. Each object may be represented in a vector.

References

^ A.C. Davidson Statistical Models. Cambridge University Press (2003).
^ http://web.uvic.ca/econ/aitken.html