Repository navigation
updates to ARM model for normal_id_glm - variable name "cov" should be "x_as_mat" or something like that #216
Description
Activity
apologies for blaming the wrong person - @WardBrian - why the name "cov"?
possible names: design_matrix, predictors, xs? xs seems best.
why this:
data { int<lower=1> N; vector[N] x; vector[N] y; } transformed data { matrix[N, 1] xs = [x']'; }instead of:
matrix[N, 1] xs = to_matrix(x);is
to_matrixless efficient? is it so inefficient that we don't want it in the example models?I didn’t rename them, my name might show up in the blame because I ran the canonicalizer on the models a few months ago
As for why the coding was done that way, I don’t believe that will be more efficient than to_matrix. The implementation of to_matrix for vector types is just a forwarding constructor
@WardBrian apologies - I was right the first time - #200
I've spent the past 30 minutes trying to find announcements or guidance on when
normal_id_glmwas introduced. it looks to me like a bunch of devs added stuff and assumed that what they added was just obvious. for our users, myself included, it's not. this is an ongoing problem - we get fabulous contributions to the code from sophisticated developers but the users need it explained to them like they're 7 - not that there's anything wrong with being 7.As for why the coding was done that way, I don’t believe that will be more efficient than to_matrix. The implementation of to_matrix for vector types is just a forwarding constructor
I think I see why this was done - in a few models, you need to past together several inputs, and you end up writing things like this:
data { int<lower=1> N; int<lower=1> J; // number of counties array[N] int<lower=1, upper=J> county; vector[N] u; vector[N] x; vector[N] y; } transformed data { matrix[N, 2] xs = [x', u']'; }to_matrixrequires something like this:matrix[N, 2] xs = to_matrix({x', u'})';clear? does it blend? will test.
the names of the chapter 12 ARM models make no sense, and the description in the README is wrong.
these were developed 9 years ago by a Columbia undergrad and no one had the time to check them."no-pooling" and "complete-pooling" should be non-hierarchical models.
any model with "chr" - which perhaps stands for "centered hierarchical regression" - just a guess -
is a "partial pooling" model.I doubt any of this is worth fixing, OTOH, keeping this around is an equally bad option. we should archive this altogether.
the ARM models which have been updated for normal_id_glm have introduced a transformed data variable named "cov" - not sure what this name is supposed to be, but it's definitely not a covariance matrix.
perhaps @andrjohns can explain?would it be OK to change to "x_mat" or something like that?