One-sided hypothesis test with Cauchy as prior

Hi all,

I’d like to run a one-sided hypothesis test but specified a Cauchy distribution for the overall effect (see below). Specifically, I’d like to test whether TypeOutcome.2 > 0. Since brms::hypothesis() may pose problems with such a distribution, I wanted to ask if there’s an alternative to do this.

Many thanks,

prior2 <- c(
  brms::prior(cauchy(0, (1/sqrt(2))), coef = Intercept), 
  brms::prior(normal(0, 1), class = b), 
  brms::prior(cauchy(0, 0.1), class = sd)  

bmod_H1 <- brm(
  CohensD | se(SECohensD) ~ 0 + Intercept + TypeOutcome2.c + (1|ArticleID) + (1|StudyID),
  data = Data_aggregated_H1,
  prior = prior2,
  sample_prior = TRUE,
  save_all_pars = FALSE,
  chains = 4,
  warmup = 5000,
  iter = 20000,
  cores = parallel::detectCores(),
  control = list(adapt_delta = .99)
  • Operating System: Windows >= 8 x64 (build 9200)
  • brms Version: 2.12.0

You can use the bayes_factor method for this purpose.

Thanks for your quick reply! Then, I would compare the above model to one without “TypeOutcome2.c”? But isn’t that a two-sided hypothesis test?

Which two hypothesis do you want to compare exactly?

I’d like to test whether TypeOutcome2.c > 0 against TypeOutcome2.c <= 0.
As far as I understood, bayes_factor tests TypeOutcome2.c = 0 against TypeOutcome2.c != 0 if I specify the above model with TypeOutcome2.c against a model without it, but that’s not what I’d like to do.

bayes_factor compares whatever two models you through at it. So you can have one model with prior(normal(0, 1), class = b, lb = 0) and another model with prior(normal(0, 1), class = b, ub = 0).

When you do it that way, please make sure that you use 1 + TypeOutcome2.c instead of 0 + Intercept + TypeOutcome2.c to avoid the Intercept getting a sign restriction as well through the above prior.

Great, thank you very much!
I have one final question about the truncation. Do I understand it correctly that the truncation upper = 0 (see screenshot) now only applies to b TypeOutcome2.c? Or do I need to apply the non-linear workaround and specify different nlpar?

prior2 <- c(
brms::prior(cauchy(0, (1/sqrt(2))), coef = Intercept),
brms::prior(normal(0, 1), class = b, lb = 0),
brms::prior(cauchy(0, 0.1), class = sd)

You need to do the non linear workaround or follow the advice from my former post.

Ah yes, now I fully get your previous post. Thanks!