Issue with mi() and residual correlations

Interfaces brms
1

Please also provide the following information in addition to your question:

  • Operating System: Mac OS High Sierra v10.13.6
  • brms Version: 2.8.0

Hello!

I am somewhat new at all this, so please bear with me, but I recently encountered a problem that I had not seen before.

PROBLEM:

My imputation model continually produces this error message:

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.
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.

However, just approximately 1 month ago, this model worked perfectly without this error message.

MODEL

I am doing research on hormones, and because of the problems with collection, there are a few NAs in my data. I decided to impute these values with a similar model to below:

Missing_1 <- bf(log_hormone_1 | mi() ~ weight + age + (1|q|Group) + Species * Sex)

Missing_2 <- bf(log_hormone_2 | mi() ~ weight + age + (1|q|Group) + Species * Sex)

Missing_3 <- bf(log_hormone_3 | mi() ~ weight + age + (1|q|Group) + Species * Sex)

Missing_4 <- bf(log_hormone_4 | mi() ~ weight + age + (1|q|Group) + Species * Sex)

Model_1 <- brm(Missing_1 + Missing_2 + Missing_3 + Missing_4 + set_rescor(TRUE),
                    data = MyData,
                    prior = c(
                      prior(normal(0,2), class="Intercept"),
                      prior(normal(0,1), class="b"),
                      prior(cauchy(0,1), class="sd", resp="Missing_1"),
                      prior(cauchy(0,1), class="sd", resp="Missing_2"),
                      prior(cauchy(0,1), class="sd", resp="Missing_3"),
                      prior(cauchy(0,1), class="sd", resp="Missing_4"),
                      prior(lkj(2), class="cor")),
                      chains=3, cores=3, warmup=2000, iter=6000, 
                      control = list(max_treedepth = 15, adapt_delta=0.99))

Although this model worked just fine previously, now I get that error message mentioned above.

Oddly enough, this model CAN work now if I remove the residual correlation terms.

I tried alternative formulations such as:

Model_1 <- brm(mvbf(Missing_1, Missing_2, Missing_3, Missing_4, rescor = TRUE)

But this was also unsuccessful.

SUMMARY

I looked around and noticed that the mi() term has been updated within the last month, so I imagine that my problems have begun occurring because of this update.

I am just wondering:

  1. Why does the model not work when the residual correlations are included?

  2. How should I rephrase my model to make it more compatible to the newer mi() code?

Thank you very much for your time and efforts with brms!

2

Can you try it again with the GitHub version of brms? I believe I have fixed something like that before a few weeks ago.

3

Thanks for getting back to me!

I just tried the Github version (2.8.9) and still encountered the same problem when setting “set_rescor” to “TRUE”.

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.

So no luck so far!

4

Does the following work for you?

dat <-
  data.frame(a = rnorm(100),
         b = c(NA, rnorm(99)))

bfa <- bf(a ~ 1)
bfb <- bf(b | mi() ~ 1)

fit4 <-
  brm(data = dat,
      mvbf(bfa, bfb, rescor = TRUE))
5

No, it does not seem to work. Here is the end of the error message:

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.
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."
6

But the exact same model works on my machine with the current github version. Perhaps you still have the old brms version loaded in your R session? A restart of R could help and if it still fails afterwards, please also post your session info.

7

Ah! That worked! Also, my main model appears to be working as well!

Thanks again!