Initial values rejected for ordinal model with default priors

I am trying to fit an ordinal model using brms with default priors. To my surprise I received messages that initial values have been rejected because the log probability evaluates to log(0). After many initialization attempts for each chain the sampling does eventually start. However, I am concerned about the rejections. Am I doing something wrong? This is my code:

T1$Rating <-factor(T1$Rating, ordered=TRUE)

family_1B ← brmsfamily(family = “cumulative”, link = “probit”)
formula_1B ← bf(Rating | thres(gr = SubjID) ~ 1 + TrialType + (1 | TrialType:StimType) + (0 + TrialType | SubjID))
fit1B ← brm(data = T1, family = family_1B, formula = formula_1B)

There are 24,192 observations in T1. Each row contains the following:

  • Rating consists of numbers 1, 2, 3, and 4
  • SubjID consists of string identifiers for each of the 14 subjects
  • TrialType consists of string identifiers for each of the 3 different trial types
  • StimType consists of string identifiers for each of the 3*18=54 different stimuli (18 per TrialType)

I added " | thres(gr = SubjID) " in an attempt to deal with subjects’ different response styles

Why are the initial values rejected? Is this normal? Should I be worried?

*brms version 2.19.0

In my experience, this happens on occasion with cumulative models. If your chains eventually worked, I wouldn’t worry about it. However, if you ever run across a model where the problem persists, you might have to set init = 0, or manually set the initial values for your parameters.

1 Like