Closed bob-carpenter closed 8 years ago
Julian King points out:
p(y | lambda) = Poisson(y | lambda) / (1 - PoissonCDF(0 | lambda))
so log p = log_Poisson(y | lambda) - log(1 - PoissonCDF(0 | lambda))
= log_Poisson(y | lambda) - log(1 - exp(-lambda)
= log_Poisson(y | lambda) - log1m_exp(-lambda)
From Joachim Vandekerckhove on stan-users:
Sorry I'm a little late to this party, I had a long summer of travel. The reason we originally implemented the lower-bound hits as negative reaction times was to have a cheap way of encoding what is essentially a bivariate distribution (strictly positive RT and binary response). Because both the RT and the outcome are random variables, the current implementation (with only one bound) isn't entirely satisfactory if at any point we want to generate random numbers from the distribution.
I always liked the negative-RT solution even though it's a little counter-intuitive. I'd probably prefer it in Stan since it's how it works in JAGS... but the more conventional solution would be to make it actually bivariate, and let it operate on a [RT response] pair, where the first is a strictly positive real and the second is a boolean.
Howard Zail points out on stan-users:
Most of the hyperlinks in the Stan Manual seem to be broken. In particular, I am looking for the http://mc-stan.org/examples.html page. This link is also broken on the mc-stan website.
Organize examples from manual on stan-dev/example-models better
This is really going to have to be done on that repo, I think, because they're no longer part of the manual directory itself.
José Rojas Echenique reports:
Section 1.6 (p25) and 1.8 (p26) are both titled Variational Inference.
Also,
( moved here from #1629 )
Ashley Ford reported in another issue: https://github.com/stan-dev/stan/issues/1637
My understanding now is that the actual integration time is
stepsize * round(int_time / stepsize)
usually slightly less than int_time
There is a typo in the manual 2.8.0, section '48.6. Multivariate Student-t Distribution', in the PDF formula It is: \Gamma \chi ((... it should be: \Gamma ((...
From Andre Pfeuffer on stan-users:
model { eta ~ normal(0, sigma); }
. Also see: https://github.com/stan-dev/stan/issues/485
From Ryan Batt on issue #1691:
Should y[2,4] = 0
instead be y[2,4]=1
?
Page 136 of Manual v2.80, the very end of the caption for Figure 12.1. It's about the data base representation of a sparse matrix.
From Ryan Batt on issue #1692:
At the top of page 138 of manual 2.80, The 6th element of z
should be 12.9 (not 129):
Also, I just realized this is Figure 12.2.
csr_to_dense_matrix
argument types in function spec; argument w
should be matrix
Page 224 of the manual, section 21 (Reproducibility). The word "on" is repeated twice in the following:
It doesn’t matter if you use a stable release version of Stan or the version with a particular Git hash tag. The same goes for all of the interfaces, compilers, and so on on.
This is the first paragraph after the list.
Page 84:
The data is declared in the same way as the other time-series regressions. Here the are parameters for the mean output mu and error scale sigma, as well as regression coefficients phi for the autoregression and theta for the moving average component of the model.
It's unclear what the purpose of this paragraph is, it seems to point out the obvious. The second sentence is particularly unclear and at a minimum suffers from a typo ("the are").
p. 37, definition of cholesky_factor_corr is unclear. "length of each row is 1." Length so commonly refers to the number of elements that maybe it would be good, for clarify, to say instead that each row is a unit vector.
Section 5.8 "Hierarchical Logistic Regression", last line on page 56:
... an approach would no pooling assigns each level l its own coefficient ...
"would" should be changed to "with", I believe.
In the ARMA(1,1) models (p. 84): MA and ARMA models are not identifiable if the roots of the characteristic polynomial for the MA part lie inside the unit circle, so it's necessary to add the constraint
real<lower = -1, upper = 1> theta;
When I run the model as it appears in the manual, without the constraint, using synthetic data from arma.sim, the simulation can sometimes find modes for (theta,phi) outside the [-1,1] interval, which creates a multiple mode problem in the posterior and also causes the NUTS treedepth to get very large (often > 10). Adding the constraint both improves the accuracy of the posterior and dramatically reduces the treedepth, which speeds up the simulation considerably (typically by much more than an order of magnitude).
Further, unless one thinks that the process is really non-stationary, it's worth adding an additional constraint
read<lower = -1, upper = 1>phi;
to ensure causality (stationarity).
David Manheim on stan-users suggests:
Very minor:
The parameter for segment
on a row_vector
is v
in the specification, and rv
in the description.
@brendan-R: thanks, I'm moving this to 2.9.0++ since 2.9.0 already got tagged.
This is where updates for the manual for the next release go if they are not related to a pull request (new features, bug fixes, etc.)