Re: [stan-users] mixed effect model: ADVI and Maximum Likelihood results are very different

[Bob Carpenter]

[moved to stan-dev]

> On Jun 23, 2016, at 3:04 PM, Andrew Gelman <gel...@stat.columbia.edu> wrote:
>
> Hey, this is an example of a mistake that would've been caught by my parser!
>
>> On Jun 23, 2016, at 5:39 PM, Andrew Gelman <gel...@stat.columbia.edu> wrote:
>>
>> Don't do this:
>> sigma2 ~ inv_gamma(0.001,0.001) ;
>>
>> Instead, just put a prior on sigma directly, for example a half-normal or half-Cauchy with peak at 0.


I take it the nagging means you want me to make this my top
priority? I can push everything else down my stack (the top
of which used to be adding random number generators to transformed
data).

Question 1: What should the messages say?

Question 2: Should this be a compiler option (e.g., "--andrew-mode")
or always on?

Question 3: If an option, what do we call it?

Question 4: If not always on, how do we control this?

Question 5: If not always on, do we allow our other message
types to be turned off the same way?

Time estimate:

always on, someone formulates exact messages for me: 1 day

not always on, need to design control, need to fish through
all the interfaces: who knows?

- Bob

P.S. A biggy that's a lot harder to code is catching undefined
variables, which also bites a lot of people.

[Ben Goodrich]

On Friday, June 24, 2016 at 11:42:58 AM UTC-4, Bob Carpenter wrote:

P.S.  A biggy that's a lot harder to code is catching undefined
variables, which also bites a lot of people.

I think one it is possible to declare and define on the same line, we should make it illegal to declare without defining on that same line (except in data and parameters). It should still be possible to redefine something later in a block, but that change would eliminate a lot of errors when people declare something in the model block and try to use it without defining it.

Ben

[Andrew Gelman]

No, this does not have to be top priority at all! We're still working on the list. I just sent you that note cos we'd been discussing the parser lately, so I thought it was a good reminder that, when we do get to this, it will be useful! I'll respond to these items on the wiki rather than by email.
A

> --
> You received this message because you are subscribed to the Google Groups "stan development mailing list" group.
> To unsubscribe from this group and stop receiving emails from it, send an email to stan-dev+u...@googlegroups.com.
> For more options, visit https://groups.google.com/d/optout.