ash-orig

ash-orig copied to clipboard

based on a question asked by Adam Siepel: in cases where the theoretical null is not applicable, maybe it would be useful to feed the method null examples. Then we...

stephens999

github.com/stephens999/ashr now contains the R package and will be where we do future development

stephens999

we should do this so that the user can define g to be a mixture of uniforms, pass into ash

stephens999

We still got warning message on EM convergence from "ash". A reproducible example is [here](https://github.com/stephens999/ash/blob/master/issues/issue20/get.warning.ash.Robj). [This](https://github.com/stephens999/ash/blob/master/issues/issue20/get.warning.ash.R) is a code to run the example. Can you check if you get consistent...

heejungshim

check get_loglik output

2

it seems that get_loglik returns penalized likelihood when prior is specified. check if this is desired behaviour.

stephens999

the halfuniform method, with nullcomp, has two copies of the null component

2

this means the penalized loglik is different than it would be if it had one copy... probably should fix this

stephens999

it might be worth running "null" simulations to see what kind of penalty is necessary to get FDR control under the null.

stephens999

The nullcheck flag checks the likelihood achieved by EM against the likelihood for the "null" (all mass on the component with smallest variance). When we use a penalized likelihood (eg...

stephens999

would be good to add functionality to continue running from a previous specified solution (for when convergence is not reached)

stephens999