Bayesian_Cognitive_Modeling: extracted Stan code from R files; updated some deprecated Stan functions #233
Conversation
| @@ -0,0 +1,25 @@ | |||
| // #### Notes to Stan model ####################################################### | |||
There was a problem hiding this comment.
Cut all of these disclaimers. This is a very simple censored model! You don't have to go deep I the User's Guide---you can just refer people to the chapter on censoring.
Did that prior come from the book? We usually only recommend hard interval constraints (here 0.25--1.0) when there are physical constraints on the model.
| @@ -40,7 +16,7 @@ parameters <- c("theta") | |||
|
|
|||
| # The following command calls Stan with specific options. | |||
| # For a detailed description type "?rstan". | |||
| samples <- stan(model_code=model, | |||
| samples <- stan(file="ChaSaSoon.stan", | |||
There was a problem hiding this comment.
If you really want to go to town, converting everything to cmdstanr would be a welcome improvement. RStan just can't keep up with CRAN's demands, so it's usually 1--2 years behind the Stan releases. For instance, it's now at 2.32, which is about 1.5 years old and we're about to release 2.37.
| @@ -0,0 +1,42 @@ | |||
| // #### Notes to Stan model ####################################################### | |||
| // ## The model is not very effective. | |||
| // ## 1) Don't change seed or lower iterations. This model converges slowly so if | |||
There was a problem hiding this comment.
This is just bad advice. I know it's not yours, but please just cut all this stuff out.
If the model only converges with a given seed, it's essentially broken and the model should be fixed. Fixing seeds does not guarantee the same behavior across platforms (C++ compiler version, C++ compiler options, different OSes, different hardware, etc.---they all affect floating point behavior).
| } | ||
|
|
||
| transformed parameters { | ||
| real<lower=0> sigma; |
There was a problem hiding this comment.
This should be
real<lower=0> sigma = inv_sqrt(lambda);But overall, for new models, I'd just model sigma directly. It only looks so roundabout because it was forced to use conjugate priors in BUGS for efficiency.
| // Common Precision | ||
| lambda ~ gamma(0.001, 0.001); | ||
|
|
||
| // Which Side is Time of Change Point? |
There was a problem hiding this comment.
Don't document things that are obvious. I'd say all of this doc is useless.
Also, those priors on mu and lambda are a really bad idea---these attempts at being "uninformative" were standard in BUGS, but we've learned a lot in 30 years.
So I wonder if some of these translate models are doing more harm than good. We definitely don't recommend that people follow this strategy for priors.
| @@ -0,0 +1,42 @@ | |||
| // #### Notes to Stan model ####################################################### | |||
| // ## Stan has hypergeometric distribution implemented, so in one way the code | |||
There was a problem hiding this comment.
Again, just drop all this.
| // ## Survey example). | ||
| // ################################################################################ | ||
|
|
||
| // Planes |
There was a problem hiding this comment.
Drop this comment---it's outside the data block and it's not at all clear what it means!
| } | ||
|
|
||
| transformed data { | ||
| int<lower=x> tmin; |
| else | ||
| lp_parts[t] = log(1.0 / tmax) + hypergeometric_lpmf(k | n, x, t - x); | ||
| } | ||
|
|
There was a problem hiding this comment.
for (t in 1:tmax) {
lp_parts[t] = -log(tmax) + (t < tmin) ? negative_infinity() : hypergeom...;log(1 / x) = -log(x) and it avoids a division.
| } | ||
|
|
||
| generated quantities { | ||
| int<lower=tmin, upper=tmax> t; |
There was a problem hiding this comment.
int<lower=1, upper=tmax> t = categorical_logit_rng(lp_parts);
This is my first pull request ever, so apologies if I did something wrong.
In many of the example models for
Bayesian_Cognitive_Modelingthe Stan code was defined as a string in the R file. In the README, @martin-smira was requesting that someone extract this code and create separate Stan files for them. So that's what I did.I also changed some deprecated Stan functions, such as:
<-is now=,increment_log_prob(u)is nowtarget += u.