data subsampling var context pointers

[Alp Kucukelbir]

hi all (but mostly bob, because he volunteered),

could you point me towards the right direction for the data subsampling business?

thanks in advance.

cheers
alp

[Bob Carpenter]

The easiest thing to do is to hack the var_context data structure that
Stan uses for data I/O.

I have no idea how you're going to do this in any generality, though.
I still don't understand exactly what you're trying to do when you
say there is a number N of data points and a subsample size M <= N.
The thing to do is think in terms of the behavior you want and try
to write out a simple functional spec (emphasis on both simple and on
functional --- this is a "what should it do?" not a "how will it do
it?" --- the latter's a technical spec).

Suppose I have two predictors

real x[N];
real z[N];

and an outcome

real y[N];

How am I going to know which data structures are relevant? We just
have numbers at the point we get to var_context rather than names, so
some additional info is almost certainly going to be needed.

Then what about group-level information, like say we have a group

int jj[N];

where that picks out the hierarchical group tow hich data item N belongs.
We'll need to pull that along. Then some groups might now have zero data,
so we have to be careful with things like priors to make sure they're
proper and the posterior doesn't blow up.

- Bob

> --
> 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.