Assess the impact of filtering the current data on estimates

[ben18785]

Test different filtering methods from basic thinning to more sophisticated filters.

[MichaelClerx]

Saw one example today where very noisy data created bias in the log-likelihood estimate! Got less when using more samples (understandably) so that'll be one thing to look out for

[MichaelClerx]

These guys don't like it
http://onlinelibrary.wiley.com/doi/10.1111/j.2041-210X.2011.00131.x/abstract
blog post about this paper:
http://doingbayesiandataanalysis.blogspot.co.uk/2011/11/thinning-to-reduce-autocorrelation.html

[mirams]

Thinning the data and thinning the MCMC chains are quite different things ;-)

[ben18785]

Yep, thinning chains is used to reduce memory footprint.

…

On 24 Nov 2017, at 14:14, Gary Mirams ***@***.***> wrote: Thinning the data and thinning the MCMC chains are quite different things ;-) — You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub, or mute the thread.

[mirams]

or to hide badly mixing ones?!?

[MichaelClerx]

Woops!

[MichaelClerx]

We do thinning in the current MCMC examples, should we remove that?
It's also a default option for the adaptive covariance mcmc routine

[ben18785]

I wouldn't make it a default option, but think we want to retain the functionality for memory footprint issues.

…

On 24 Nov 2017, at 14:21, Michael Clerx ***@***.***> wrote: We do thinning in the current MCMC examples, should we remove that? It's also a default option for the adaptive covariance mcmc routine — You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub, or mute the thread.

[MichaelClerx]

Hey @ben18785 , seems to me we're doing this one (kind of) in e.g. Richard's problem. Should we maybe close this as out of scope? Unless we can turn it into a solid feature request for PINTS