Error: One-dimensional unit vector is discrete about rstanarm

[jgabry]

@bgoodri Can you take a look at this? It's also breaking some code from Regression and Other stories (see below for more details).

@invertdna Thanks for reporting this. I think this is a bug introduced with the latest update when we switched to using Stan 2.26. This happens when there's only 1 predictor (it should work fine if you add another variable on the RHS of the formula) when using stan_lm or stan_polr.

@bgoodri any idea why this fails with 2.26? This causes some of the examples from Regression and Other Stories to error that used to run fine. For example, this stan_polr code from ROS now errors with 2.26 because it only has 1 predictor:

https://github.com/avehtari/ROS-Examples/blob/a049a1044aef17bb90664d3b4735fc7ad6ff5575/Storable/storable.Rmd#L54-L55

The unit vector array in polr.stan is declared with

array[K > 1] unit_vector[K] u;

so if there's only 1 predictor (K=1) it would be array[0] unit_vector[0] u. But I don't see why that should be a problem since it's only used in transformed parameters if K > 1 and it used to be fine in earlier versions. Also the error message says that the target distribution must be continuous, which doesn't seem relevant here:

Exception: Found dimension size one in unit vector declaration. 
One-dimensional unit vector is discrete but the target distribution 
must be continuous. variable=u; dimension size expression=K 

[jgabry]

The solution may be

array[K > 1] unit_vector[K > 1 ? K : 2] u;

which makes it size 2 when K=1 (it won't even get used so we can just do this to avoid error when K=1)