[BUG] Incorrect Posterior Probability with dnUniformInteger and mvRandomIntegerWalk

[minghao-du]

Describe the bug
When using dnUniformInteger and mvRandomIntegerWalk, it is not possible to obtain the correct posterior probability.

To Reproduce
I generated 1000 samples ($y_i, i=1,...,1000$) from a Bernoulli distribution with $p=0.1$ and conducted inference on $p$.
$$y_i \sim \text{Bernoulli}(p)$$

$$p_\text{int} \sim \text{UniformInteger}(0,10)$$

$$p \coloneqq \frac{p_\text{int}}{10}$$

However, the inferred posterior distribution of $p_\text{int}$ is almost the same as the prior, uniformly distributed between $0$ and $10$, rather than being mostly $1$.

When inferring directly using a uniform distribution, there is no problem with the inference.
$$y_i \sim \text{Bernoulli}(p)$$

$$p_\text{int} \sim \text{Uniform}(0,10)$$

$$p \coloneqq \frac{p_\text{int}}{10}$$

The scripts used and the output are in this file.
UniInt_bug.zip

Expected behavior
When inferring using a UniformInteger distribution, the inferred $p_\text{int}$ is $1$.

Computer info
I am using a Mac, and the version of RevBayes I am using is development (rapture-2388-ge4baa8).

[bredelings]

Note that the Posterior and Likelihood are constant. So they aren't getting updated when p_int changes. This is the underlying source of the problem.

I tweaked the script to replace the collection of Bernoulli observations with a single Binomial observation, and the problem doesn't go away. So this could possibly be a problem with mvRandomIntegerWalk( )... but I'm not sure how the move can prevent the likelihood and prior from changing.

[bjoelle]

This seems tied to the model() call

• when using model(uni_int) as in the example script, p_estimated and the Bernouilli distribution do not get integrated into the model (so uni_int updates but then nothing else does)
• when using model(p_estimated) or model(y), the graph does get included correctly, but uni_int is set as constant and loses its prior distribution (so uni_int doesn't update and neither does anything else)
• when using model(data), it gives the error "Cannot find node with name 'uni_int' in the model but received a move working on it." so I'm assuming it just lost uni_int entirely

Overall, there seems to be something weird going on with understanding the graph, tied to (I assume) deterministic nodes and/or integer parameters

[bredelings]

Interestingly, it looks like this problem seems not to happen if the integer distribution is dnPoisson( ) instead of dnUniformInteger( ).

[bredelings]

Here's what we see if dnPoisson is used (and the sample size is changed to 10:

[bredelings]

Here is what we see if dnUniformInteger is used: