Beta family: Error due to 1s being generated

[IreneSophia]

I want to use this package with the standard Beta function. However, whenever I'm trying to simulate the data, I always get this error:

Error in data_response.brmsframe(bframe, data, check_response = check_response, : Family 'beta' requires response smaller than 1.

Any ideas what's happening here? Using SBC version 0.5.0.9000 and brms version 2.23.0.

Here's a minimal example:

library(SBC)
library(brms)

df = data.frame(Outcome = rep(0.5, times = 100))

gen = SBC_generator_brms(Outcome ~ 1, data = df, family = Beta,
                         thin = 50, warmup = 10000, refresh = 2000,
                         generate_lp = TRUE)
dat = generate_datasets(gen, 100)

[lunafazio]

Hi, the issue here is that you left the prior argument unspecified, which means the brms defaults are used, namely:

                prior     class coef group resp dpar nlpar lb ub tag  source
 student_t(3, 0, 2.5) Intercept                                      default
    gamma(0.01, 0.01)       phi                             0        default

As these priors are very disperse, the data generation step samples parameter values which are extreme enough to produce beta samples that get rounded to exact 0s and 1s. Because you've set generate_lp = TRUE, the brms::log_lik function is called and that's where the brms data validation step kicks in to reject the simulated data.

[IreneSophia]

Hi, thanks for the response.

I tried a lot of different priors before posting, but none worked. To keep the minimal example as short as possible, I did not include them.

Would you have any suggestions for Beta priors that do not lead to any 0s or 1s when generating the data?

I'm a bit confused by your last sentence: Does this imply that no data validation step is used, if I set generate_lp = FALSE? Because I tried that and still encountered the same issues.

[lunafazio]

I should have included an example, sorry!

First, it's worth noting that the default parametrization of the Beta in brms is in terms of the logit-scale mean and precision. Here's an example of a prior that samples tightly around a mean of 0.5 with a fixed precision of 10:

library(SBC)
library(brms)

df = data.frame(Outcome = rep(0.5, times = 100))

p = get_prior(Outcome ~ 1, data = df, family = Beta)
p[1, "prior"] = "normal(0, 0.01)"
p[2, "prior"] = "constant(10)"

gen = SBC_generator_brms(
  Outcome ~ 1, data = df, family = Beta, prior = p,
  thin = 50, warmup = 10000, refresh = 2000,
  generate_lp = TRUE
)

dat = generate_datasets(gen, 100)

When I'm not sure about what a sensible prior is, my general approach is precisely what you see above: I fix all parameters but one (fixing all requires specifying the "Fixed param" algorithm instead of using the regular sampler) until I find a reasonable starting point and then relax the priors around that initial point until I'm satisfied with the result.

As for the generate_lp argument, changing it to FALSE and keeping your example unchanged otherwise will specifically prevent the error from happening. You still get 0s and 1s in your simulated datasets, but I thought I'd mention this because it allows you to complete the generate_datasets step and examine the actual datasets being generated.