Lp__ and target +=

Modeling
1

Is that corret the value of “lp__” should always equal to the right hand side of “target +=”?

The problem is that I tried to reproduce the value of “lp__” in R but failed. I used “target+=” in Stan programming and I think there should not be a problem of Jacobian because I have already done the transformation in my coding. So I am wondering which part is wrong.

Thanks a lot for helping!

Btw, I didn’t paste my code here because it is too long…

2

lp__ will be equal to the sum of all target += calls if you don’t have any parameter constraints. If you have any constraints (e.g., <lower=0>), then Stan will automatically add the necessary Jacobian corrections

2 Likes
3

Thanks a lot for your help!

I used <lower = 0> in the “transformed parameters” block but not “parameters” block. Does that also lead to adding Jacobian corrections?

4

No, constraints in transformed parameters do not lead to corrections. However, in order to consistently satisfy that constraint without any constrained parameters in the parameters block, you must have applied a nonlinear transform. Assuming that your transform is nonlinear, then if your transformed parameter is anywhere used on the left-hand side of a sampling statement (or as the first argument in a lpdf/lpmf on the right-hand side of a target += statement), then you might need to apply your a Jacobian adjustment in order to get the inference that you intend.

Note that sampling statements with a ~ are really target += statements in disguise, so these will also influence lp__. Note also that some parameter types, including covariance matrices, are inherently constrained types that induce a Jacobian adjustment even without an explicitly declared parameter constraint. See the reference manual here for more on the constraint transforms and their Jacobians:

2 Likes
5

To add to @jsocolar’s excellent answer, we could give you a more concrete explanation if you posted your model and the R code being used to recreate lp

1 Like
6

Hi all. Thanks so much for helping! Just figured out why there is a big difference. I incorrectly deleted one line of code in my stan file…