Aitken Estimator

The generalized least squares estimator used when regression errors are correlated or have non-constant variance.

The Aitken estimator is another name for the generalized least squares estimator. It is used in linear regression when the error terms are not spherical, meaning they are heteroskedastic, correlated, or both.

$$$$

The Basic Problem

Start with the linear model:

[ y = X\beta + \varepsilon, \qquad E[\varepsilon|X] = 0, \qquad Var(\varepsilon|X) = \Omega ]

If \Omega is proportional to the identity matrix, ordinary least squares is efficient. If not, OLS may remain unbiased under standard conditions, but it no longer uses the data in the most efficient way.

The Aitken Or GLS Formula

When \Omega is known, the Aitken estimator is:

[ \hat{\beta}_{GLS} = (X’\Omega^{-1}X)^{-1}X’\Omega^{-1}y ]

This reweights the regression to account for unequal variance and correlation across observations.

Why It Matters

The point of GLS is that some observations contain more usable information than others once error structure is taken into account. A well-specified GLS estimator can reduce estimator variance relative to OLS.

If \Omega is diagonal, GLS becomes weighted least squares. If \Omega must be estimated rather than known, the method becomes feasible GLS.

Knowledge Check

### What is the Aitken estimator another name for? - [x] Generalized least squares - [ ] Ordinary least squares - [ ] Maximum likelihood only - [ ] The sample mean > **Explanation:** In econometrics, the Aitken estimator is the GLS estimator for linear regression with non-spherical errors. ### Why might OLS be inefficient in this setting? - [ ] Because OLS cannot be used with explanatory variables - [x] Because the error variance-covariance matrix is not proportional to the identity matrix - [ ] Because GLS always gives the same answer as OLS - [ ] Because unbiasedness is impossible in regression > **Explanation:** When errors are correlated or heteroskedastic, OLS does not weight information optimally. ### When does GLS collapse to weighted least squares? - [ ] When the regression has no intercept - [ ] When all regressors are binary - [x] When the error covariance matrix is diagonal but variances differ across observations - [ ] When sample size is infinite > **Explanation:** A diagonal covariance matrix means there is heteroskedasticity without correlation, which is the weighted-least-squares case.