search

cboettig / nimbios-transients

Example data that may or may not contain transients
MIT License
2 stars 0 forks source link

theory: How divergent can the posterior distribution of a parameter from a given model be when estimated from two different realizations of the same model? #9

Open cboettig opened 5 years ago

cboettig commented 5 years ago

I did ask Chris about this, he suggested there should be some more rigorous statements about the divergence between posterior estimates for two trajectories from the same ensemble of replicates in the particle filtering literature (which is basically an ensemble of replicates result).

But basically the upshot I got was that we should at least expect to get broad / overlapping posteriors of parameter estimates from MCMC on two different replicate simulations of the same transient dynamics. If we're not seeing this, it's probably more of numerical issue in the MCMC not mixing well or what not.

Based on #4 I don't think we're seeing this problem (except in cases where we can also see poor convergence in the traces etc).

henryrscharf commented 5 years ago

Confirming that this weird behavior was almost certainly (i.e., probability 1) the result of a coding bug on my part. Whoops. 😞