Setting distributional parameters to constant bug

[athowes]

We have two bugs I've seen when trying to set distirbutional parameters to be constant:

set.seed(1)

obs_time <- 25
sample_size <- 500

meanlog <- 1.8
sdlog <- 0.5

# The mean of the lognormal distribution is: exp(meanlog + 0.5 * sdlog^2)

sim_obs <- simulate_gillespie() |>
  simulate_secondary(
    dist = rlnorm,
    meanlog = meanlog,
    sdlog = sdlog
  ) |>
  observe_process() |>
  filter_obs_by_obs_time(obs_time = obs_time)

sim_obs <- sim_obs[sample(seq_len(.N), sample_size, replace = FALSE)]
prep_obs <- as_latent_individual(sim_obs)

Bug 1

epidist(data = prep_obs, formula = brms::bf(mu ~ 1, sigma = 1), seed = 1)

results in:

Error: The following priors do not correspond to any model parameter: 
Intercept_sigma ~ normal(0, 0.5)
Function 'default_prior' might be helpful to you.

Reason: in epidist_prior we hard code that the model has to have certain priors.

Bug 2

Even if we didn't have this prior problem, then we have the following bug:

epidist(data = prep_obs, formula = brms::bf(mu ~ 1, shape = 1), family = stats::Gamma(link = "log"))

Compiling Stan program...
Semantic error in '/var/folders/4r/hkp4v9fn3wx044_hk8qnsjxw0000gn/T/RtmpSfhylK/model-3c965f0b3c81.stan', line 61, column 14 to column 72:
   -------------------------------------------------
    59:      mu += Intercept;
    60:      mu = exp(mu);
    61:      target += latent_gamma_lpdf(Y | mu, shape, pwindow, swindow, vreal1);
                       ^
    62:    }
    63:    // priors including constants
   -------------------------------------------------

Ill-typed arguments supplied to function 'latent_gamma_lpdf':
(vector, vector, real, vector, vector, array[] real)
Available signatures:
(vector, vector, vector, vector, vector, array[] real) => real
  The third argument must be vector but got real

Reason: our custom likelihood expects a vector for all distributional parameters. If instead its a real then it throws an error.

[seabbs]

what happens if you do this instead by setting the prior to be constant? I am not convinced this is the most important functionality to support so would push it unless you have a strong view that its functionality we need?

[athowes]

Currently we do not support users passing in priors. Perhaps #71 (allowing users to pass in priors) is therefore more important to do. And then we wouldn't need to support this in the formula.

[seabbs]

I think #71 would support a broader range of use cases? Obviously it would be nice if this worked though

[athowes]

Yes #71 would support a broader range of use cases and I agree is more of a priority.

(Sorry, I wrote 71 but really it's #196 now)

[athowes]

I think this will be closed via #255 but I need to check, then I'll close this. Maybe add a test.

[athowes]

Nope, get error:

Error in `f()`:
ℹ No available prior to replace in old_prior found for:
<lower=0> sigma ~ lognormal(-0.7, 0.4)
Run `rlang::last_trace()` to see where the error occurred.

To fix!