I have tried to look for an answer online but so far I have not succeeded. Is there a Bayesian equivalence of the Breusch-Godfrey test to no autocorrelation or of the Ljung-Box test? I have found some references on the Durbin-Watson test but as is well known this test is not of much use in applied work.

In a classical framework one would estimate the model of interest and then run an auxiliary regression to test for autocorrelation up to lag $p$. How is this done in a Bayesian setting?

I could think of obtaining the distribution of residuals and then do an auxiliary regression of these. I could then test the significance of the coefficients on lags using a Bayesian F-test or using the marginal likelihood to compare models. However, running the auxiliary regression would require a prior. Something which I do not have. Could the prior mean of the residuals be set to zero as that is what I expect them to be?

Further, would it be enough to just select a model based on the marginal likelihood and just disregard a test for autocorrelation?


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.