Hierarchical example prior distribution
In the model comparison section of the hierarchical analysis example, I suggest a HalfGaussian
prior on the sigma0
value. As sigma0
is a scale parameter a more sensible prior would be a LogUniform
, so I should suggest that.
A couple of related things:
- if running with the numerical likelihood, when using many pulsars then number of grid points at which the individual likelihoods are evaluated becomes important (particularly for when there are many pulsars with posterior support that ranges over a few orders of magnitude). So, I should add information about increasing the number of interpolation bins in these cases,
or perhaps add the ability to have the bins on a logarithmic scale(it seems that past me was ahead of the game and this is already on option using thegridtype
argument). -
When storing the interpolated likelihoods asActually, it is probably worth changing things so that the individual interpolated likelihoods are not all calculated on the same grid, but on individual grids that are specific to the scale of the actual posterior bulk. For inference with many pulsars, having a single grid (as described in the first bullet) can result in extremely spiking likelihoods, which having individual grids may alleviate.splrep
outputs, the knots are the same for each case, but I store all of them. The amount of memory could be reduced by only storing the knots once.
Edited by Matthew Pitkin