Brms: cumulative model\ rejecting initial values issue

Modeling
1

Before moving into full model, I tried to run following simple model:
m1<-brm(Response ~ Avglength,
data = sdf,
family = cumulative(link=“probit”, threshold=“flexible”),
chains=2,
iter=100,
cores=2,
seed = 1,
init_r=0)
Every time i run the model, i got the same issue which was: rejecting initials.Log probability evaluates to log(0), i.e. negative infinity.Stan can’t start sampling from this initial value. I tried with and without init_r=0, but got the same issue.
My response has 11 levels and independent variable was length of fish (numeric value).

2

Does using init_r=0 and init=0 achieve the same thing? Setting init_r is about the range of randomly generated inits…

3

Yes. Gives the same result.

4

Both give following error:
Chain 1: Rejecting initial value:
Chain 1: Log probability evaluates to log(0), i.e. negative infinity.
Chain 1: Stan can’t start sampling from this initial value.
[1] “Error in sampler$call_sampler(args_list[[i]]) : Initialization failed.”
[1] “error occurred during calling the sampler; sampling not done”

5

Maybe @Solomon has some suggestions, as I think he might make frequent use of these types of models in brms

1 Like
6

My first step would be to move away from the default priors.

7

Hi Solomon:

This was my model:
m2<-brm(Response ~1+Avglength,
data = sdf1,
family = cumulative(link=“probit”, threshold=“flexible”),
prior = c(prior(normal(0,2), class=Intercept),
prior(normal(0,10), class=b)),
chains=2,
iter=1000,
init_r=2)

where Response has satisfaction level from 0 to 10 and Avglength is the size of fish individual angler caught. Error are below:

Chain 1: Stan can’t start sampling from this initial value.
Chain 1:
Chain 1: Initialization between (-2, 2) failed after 100 attempts.
Chain 1: Try specifying initial values, reducing ranges of constrained values, or reparameterizing the model.
[1] “Error in sampler$call_sampler(args_list[[i]]) : Initialization failed.”
[1] “error occurred during calling the sampler; sampling not done”

8

You might spend some time thinking more about setting individual priors for the individual thresholds. Also, normal(0, 2) is really broad on the probit scale.

9

As you suggested in your blog, I calculated the right-handed thresholds and use that information to assign the individual prior, see below. Still I got the same result. Are there an other things to consider it?

fit1 ← brm(
data = cdata3,
family = cumulative(probit),
Response ~ 1,
prior = c(prior(normal(-1.34, 1), class = Intercept, coef = 1),
prior(normal(-0.91, 1), class = Intercept, coef = 2),
prior(normal(-0.60, 1), class = Intercept, coef = 3),
prior(normal(-0.35, 1), class = Intercept, coef = 4),
prior(normal(-0.11, 1), class = Intercept, coef = 5),
prior(normal(0.11, 1), class = Intercept, coef = 6),
prior(normal(0.35, 1), class = Intercept, coef = 7),
prior(normal(0.60, 1), class = Intercept, coef = 8),
prior(normal(0.91, 1), class = Intercept, coef = 9),
prior(normal(1.34, 1), class = Intercept, coef = 10)),
cores = 2,
chains=1,
iter=100)

Error:Chain 1: Rejecting initial value:
Chain 1: Log probability evaluates to log(0), i.e. negative infinity.
Chain 1: Stan can’t start sampling from this initial value.

what could be issues here?

10

Hmmm. Maybe try setting init_r = 0.2.

11

Changing the priors doesn’t effect the inits, so fiddling with the priors, while definitely a good idea, won’t tend to resolve this issue.