I have data of some physiological measure, represented as a vector of 185 measurements taken every 2 seconds. I can model this response in two different ways, and I wish to compare the fit of my two competing models so that I can extract a Bayes factor for the ratio of the likelihoods of obtaining this data given the first or the second model. My approach so far was to fit a BLUE linear regression model to the data (with only one regressor, namely the expected response under the first or the second model), and to use the formula for the log-likelihood of a regression model with the MLE coefficients ($-N/2\times(log(2\times\pi)+log(\sigma^{2})+1)$). I then subtracted the two terms to get the log of my Bayes factor.

But this assumes my samples are independent, when in reality they are serially autocorrelated. I guess that shouldn't bias my BF, but it should nonetheless polarize it. How can I overcome this other than thinning down my vector? I guess I should somehow correct my $N$?

  • $\begingroup$ Please tell me if I am understanding you correctly. You are attempting to calculate a Bayesian estimator from a Pearson-Neyman parameter estimate and a Fisherian likelihood estimate, is that correct? Also, regardless of dependence, your formula is incorrect. Are you really thinking about the Bayesian Information Criterion and not the Bayes Factor? Finally, if your concern is model selection and autocorrelation in the sample space, then why are you not just constructing this as a Bayesian problem in the first place? $\endgroup$ – Dave Harris May 29 '17 at 6:16
  • $\begingroup$ Hi, thanks! Since both models have the same number of parameters, comparing the MLEs of the two models is identical to computing the Bayes factor using the formula ($exp(BIC_{0}-BIC_{1}/2)). I chose to use the MLE estimate instead of specifying a full Bayesian model just to save myself the trouble of finding appropriate priors for my parameters and estimating the evidence for each model. Also - can you point me to the error in my formula (this is what I obtained when I applied the likelihood formula to the BLUE)? Thanks again! $\endgroup$ – TanZor May 29 '17 at 6:33

That's right: linear regression is not appropriate for these data, and 'thinning' them out may not solve the problem (although you can check this using autocorrelation testing). Depending on the details of your question of interest, you can either derive the rate of change (taking the difference between each measurement and its immediate predecessor, which in some cases could also be autocorrelated) or use spatial detrending of various levels of complexity e.g. Ermagun et al. 2017 Parts of your design are unclear as presented (e.g. how many independent parameters are you considering, what are the two 'different ways'), so I cannot add anything to the second part of the question, but perhaps this prior post and this would be helpful.


Your Answer

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

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