Faster convergence

Interfaces brms
1

I am using BRMS to run a three level Bayesian model below. When I run the model using default priors, it gives results in about 5 hours.
I have now changed to the priors as indicated below and it in more than 2 days we are at 33% of 3000 iterations. Is there anything I could do to quicken the running of the code?
prior ← c(set_prior(“cauchy(0,10)”, class = “sds”, coef = “s(Year)”),
set_prior(“cauchy(0,10)”, class = “sds”, coef = “s(timeInt)”),
set_prior(“cauchy(0,10)”, class = “sd”, group = “Cluster:House”),
set_prior(“cauchy(0,10)”, class = “sd”, group = “Cluster”),
set_prior(“normal(0,10)”, class = “b”, coef = “GenderFemale”),
set_prior(“normal(0,10)”, class = “b”, coef = “TypeRural”),
set_prior(“normal(0,10)”, class = “b”, coef = “RegionEasternCape”),
set_prior(“normal(0,10)”, class = “b”, coef = “RegionFreeState”),
set_prior(“normal(0,10)”, class = “b”, coef = “RegionGauteng”),
set_prior(“normal(0,10)”, class = “b”, coef = “RegionKwazuluMNatal”),
set_prior(“normal(0,10)”, class = “b”, coef = “RegionLimpopo”),
set_prior(“normal(0,10)”, class = “b”, coef = “RegionMpumalanga”),
set_prior(“normal(0,10)”, class = “b”, coef = “RegionNorthernCape”),
set_prior(“normal(0,10)”, class = “b”, coef = “RegionNorthWest”),
set_prior(“normal(0,10)”, class = “b”, coef = “EduHigher”),
set_prior(“normal(0,10)”, class = “b”, coef = “EduPrimary”),
set_prior(“normal(0,10)”, class = “b”, coef = “EduSecondary”),
set_prior(“normal(0,10)”, class = “b”, coef = “EthnicityColoured”),
set_prior(“normal(0,10)”, class = “b”, coef = “EthnicityIndianDAsian”),
set_prior(“normal(0,10)”, class = “b”, coef = “EthnicityOther”),
set_prior(“normal(0,10)”, class = “b”, coef = “EthnicityWhite”))

Bayesm ← brm(y ~ s(timeInt)+ s(Year)+Gender+Type+Region+Edu+Ethnicity+(1|Cluster/House),
data = dtLong,
family = bernoulli(link = “logit”),
warmup = 1000,
iter = 3000,
chains = 1, thin = 10,control = list(adapt_delta = 0.85),
inits = “0”,
cores = 4,
seed = 123,prior=prior)

2

try:

  1. Using CmdStan as a backend: brm(..., backend = "cmdstanr")
  2. QR Decomposition: bf(formula, family = XX, decomp = "QR")
  3. Un-normalizing PDFs: brm(..., normalize = FALSE)

All of those are well documented at the brms reference page.

EDIT: Oh, and if it helps, please post the speed-up here for us…

3

Thank you.

Will vectorizing the priors help?

Regards

4

How do I get that?

5
  1. CmdStan you can find it here: R Interface to CmdStan • cmdstanr
  2. QR decomposition in the bf() function reference: Set up a model formula for use in brms — brmsformula • brms
  3. Normalization of PDF in the brm() function reference: Fit Bayesian Generalized (Non-)Linear Multivariate Multilevel Models — brm • brms

Regarding prior vectorization I cannot answer, because I honestly don’t know. But wouldn’t hurt trying:

prior = set_prior("normal(0, 10)", class = "b", coef = c("GenderFemale", "TypeRural", ...)
6

Do prior predictive checks.