4
$\begingroup$

Jeffrey Wooldridge, a famous econometrician, posed the following question to Bayesians on twitter:

I think frequentists and Bayesians are not yet on the same page, and it has little to do with philosophy. It seems some Bayesians think a proper response to clustering standard errors is to specify an HLM. But in the linear case, HLM leads to GLS, not OLS.

Moreover, a Bayesian would take the HLM structure seriously in all respects: variance and correlation structure and distribution. I'm happy to use an HLM to improve efficiency over pooled estimation, but I would cluster my standard errors, anyway. A Bayesian would not.

There still seems to be a general confusion that fully specifying everything and using a GLS or joint MLE is a costless alternative to pooled methods that use few assumptions. And the Bayesian approach is particular unfair to pooled methods.

One only needs to think of something like a simple time series regression with serial correlation. I think there are four common things one might do.

  1. OLS with usual (nonrobust) SEs
  2. OLS with Newey-West SEs
  3. Prais-Winsten with usual SEs
  4. P-W with N-W SEs

In my view, choice (3) is almost as bad as (1). Choices (2) and (4) make sense, with (4) requiring strict exogeneity. But at least we're then comparing apples with apples.

Again, what is the Bayesian version of (4) after priors and distributional assumptions are imposed?

What is the answer to Wooldridge's question?

Glossary of acronyms:

HLM = Hierarchical Linear Model, GLS = Generalized Least Squares, OLS = Ordinary Least Squares, SE = Standard Error, MLE = Maximum Likelihood Estimator.

Improve this question
$\endgroup$
5
  • 1
    $\begingroup$ This question is the cat's pajamas. $\endgroup$
    – dimitriy
    Commented Mar 23, 2021 at 2:36
  • $\begingroup$ The Tweets start at twitter.com/jmwooldridge/status/1368557505018757131 I've corrected what I take as trivial typos in the original. $\endgroup$
    – Nick Cox
    Commented Mar 23, 2021 at 3:05
  • $\begingroup$ I have always understood GLS (more particularly IGLS) as an estimator for HMLs/MLMs (i.e. models). For example, the "standard" estimation approach in MLwiN is IGLS (or RIGLS which is a similar strategy to RMLE, as I understand it), and includes variance and covariance estimates. But one can also use MCMC estimation in MLwiN (including with a link to JAGs to get fully Bayesean with one's modeling of prior information). Am I misunderstanding something (i.e. GLS = estimator, and HLM = model specification/structure)? $\endgroup$
    – Alexis
    Commented Mar 23, 2021 at 4:20
  • $\begingroup$ Is this question, more generally, "How can a Bayesian 'fix up' her standard errors when she doesn't believe her model assumptions?" $\endgroup$
    – JTH
    Commented Mar 24, 2021 at 1:01
  • $\begingroup$ Is there a particular reason for why he suggests model 4 (Prais-Winsten and Newey-West) as sensible for simple time series data? It seems unnecessary to double-correct for serial correlation. $\endgroup$
    – Durden
    Commented Jun 21, 2023 at 21:21

0

Reset to default

Your Answer

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