tan or atan link function

[rmcelreath]

tan or atan link function

[rmcelreath]

I suggest using Stan directly for now. Sorry. But I do have a hack below that might be alright for you.

More detail:

I don't understand why you're getting a unknown distribution error. I'd have to see the code you are attempting. But this kinda works:

m <- map2stan(
    alist(
        b <- 0 + 1*tan(bu),
        bu ~ dunif("-pi()/2","pi()/2")
    ),
    data=list(y=1),
    sample=TRUE )

The major issue here is that variable assignments in map2stan are assumed to be indexed by the outcome variable cases. In the above example, there is just one case, so b gets a single value for each bu sample. But that won't be true in ordinary regression examples. Maybe just doing b[1] in the code would work though? Haven't tried.

All of this feels kludgy and probably dangerous. But it might work fine in your current context.

On the horizon, I have been planning to just allow users to insert raw Stan code into any of Stan's blocks. So a parameter transform could be done like:

TV : beta <- mu + tau*tan(beta_unif)

That would define beta in the transformed variables block in Stan, so it could be used in the model block.

As for adding link functions, in principle that is trivial. We just need more entries in this list:
https://github.com/rmcelreath/rethinking/blob/master/R/map2stan.r#L33

I will put this on my to-do list. I already have a plan to add cloglog there. So might as well as common trig functions.

[jgabry]

I don't remember if we put this in the Stan manual, but sometimes it can
also be helpful to formulate the Cauchy as a ratio of normals. Maybe that's
an easier parameterization to do in the current version of rethinking (not
sure)?

On Thursday, July 28, 2016, Richard McElreath [email protected]
wrote:

Reopened #45 #45.

—
You are receiving this because you are subscribed to this thread.
Reply to this email directly, view it on GitHub
#45 (comment), or mute
the thread
https://github.com/notifications/unsubscribe-auth/AHb4Qx66_75QH2j4lp3flVlC-C_5n0Ltks5qaGAWgaJpZM4JW3NY
.

[tom-christie]

Great, thank you to both of you. I'll try both suggestions. Appreciate the help!