1
$\begingroup$

My understanding that if errors are non-spherical, OLS is no longer the minimum variance linear unbiased estimator (assume the error terms are fully independent of all covariaties- so unbiasedness holds).

So we can use robust standard errors to still obtain consistent estimation of standard errors when they are not homoscedastic. Now, if we use robust standard errors and that's it- have we restored 'BLUE"? i.e. does this just give us consistent estimation of the standard errors, or does it also restore the 'Best', minimum variance property?

Improve this question
$\endgroup$
1
  • $\begingroup$ Hi: minimum variance could only occur if you actually knew what the true covariance structure was. Since, you don't, my understanding is that robust standard errors result in consistent estimation. minimum variance is generally applied in the OLS framework because you know the covariance structure which is diagonal. Hopefully others can say more. $\endgroup$
    – mlofton
    Sep 17, 2020 at 19:02

2 Answers 2

Reset to default
1
$\begingroup$

Now, if we use robust standard errors and that's it- have we restored 'BLUE"? i.e. does this just give us consistent estimation of the standard errors, or does it also restore the 'Best', minimum variance property?

It just gives you consistent estimation of the standard errors. It doesn't change the point estimates, so it can't possibly make them any more BLUE.

Improve this answer
$\endgroup$
0
$\begingroup$

First choice in practice, I would (by examining regression residues, for example) postulate a crude relationship to account for the change in volatility as a function of a model's variable(s). This is, in effect, an attempt to restore BLUE properties.

Common models in practice include, for example, weighted least-squares, which follows from noting that error magnitudes may historically, for select datasets, move roughly in accord with the size of a variable.

Interestingly, if you believe that underlying fundamentals (like in risk levels) have changed, you could quite originally postulate (based on sound arguments) that a regime change in volatility, starting at a particular time point, has occurred (so a set of more recent observations could be down-weighted).

Note: applying an appropriate data transformation (see discussion here), at times, may obviate the need for such a volatility model correction.

If the above provides marginal results, consider employing robust regression as an option.

Improve this answer
$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge that you have read and understand our privacy policy and code of conduct.

Not the answer you're looking for? Browse other questions tagged or ask your own question.