I have a general query regarding informativeness of priors, since my laptops gone down and not able to run this on Stan to check (but from previous runs I think this was the case). If the priors used to specify a bayesian model are highly informative, e.g. really small standard deviations say lognormal(0.2,0.000000001) for all the parameters, will autocorrelation be non-existent because of the bias inflicted where the data virtually provides no information towards the estimation of the posterior distribution, so essentially just sampling from this prior distribution which will converge with very little autocorrelation? If this is true, why exactly does this affect autocorrelation? Pardon my lack of understanding, still getting to grips with bayesian statistics.

  • 1
    $\begingroup$ I believe this should answer your question: stats.stackexchange.com/a/201059/35989 does it? $\endgroup$ – Tim Aug 13 '18 at 12:52
  • $\begingroup$ yes thanks for the link, it was helpful but would like an answer that addresses autocorrelation and number of effective parameters in relation to a highly informative prior specification such as the one stated above. Pretty sure i saw n_eff be exactly the number of iterations when i ran a model in Stan using such priors, but when relaxed the standard deviation, lots of autocorrelation was present. Unless this was random.. $\endgroup$ – s.g Aug 13 '18 at 13:50
  • $\begingroup$ But this would depend on properties of your data... $\endgroup$ – Tim Aug 13 '18 at 13:54
  • $\begingroup$ hmm, I was just thinking that it may bias my results as it gives very little weight to the data if my lognormal priors had standard deviation 0.000000001... $\endgroup$ – s.g Aug 14 '18 at 21:45

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.