Report brms refitting models

[jgeller112]

Maybe I am missing something, but is there a way to avoid always refitting brms models using the report function? I have pretty large models and using report will either error out or run forever.

[strengejacke]

@rempsyc @mattansb @DominiqueMakowski I'm not sure, but I guess this is due to computing Bayes factors when retrieving the performance-table? We should try to find the bottle neck here, maybe using the profile package.

@jgeller112 do you have a model you can share, so we can try to reproduce your issue? Else, maybe you find a model out of our pre-compiled models from here: https://easystats.github.io/circus/reference/index.html and just name one of those example models that behaves similarly.

[mattansb]

I think this is because report standardizes coefficients using the refit method (which it 100% should not do since it does not adjust the priors accordingly).

[jgeller112]

Here you go:

library(tidyverse)
library(report)

fit_c <- read_rds("https://osf.io/hmqpc/download")

report(fit_c)

[strengejacke]

If I understand the profiling right, one issue is logLik():

And another bottle neck is rstan::summary():

Given that already the simple call to summary() takes ~ 7 seconds on my machine, this looks like the issue is due to the very complex model. The profile-file is attached (I used profvis), maybe we find ways to remove some of those time-consuming things.

report_brms.zip

[DominiqueMakowski]

I think this is because report standardizes coefficients using the refit method (which it 100% should not do since it does not adjust the priors accordingly).

yes, the standardization method should be changeable to something else than "refit" (?)

[strengejacke]

@jgeller112 Can you then try to run report(fit_c, include_effectsize = FALSE)?

[jgeller112]

I get

Error in report_table.brmsfit(x, include_effectsize = FALSE, ...) : 
  formal argument "include_effectsize" matched by multiple actual arguments

[ZoeGold]

Has there been a fix for this issue in the meantime? I am running into the exact same issue with my brms models (and not just one model, but a wide range of them)

[utkuturk]

I am having the same problem with brms. Even though model fit with a file specified, it refits all of the models I am using, even the simplest intercept only models. Is there a way to prohibit this behavior? Or something that I could do while fitting the models, so it does not refits it?

[rempsyc]

Same as #568 and fixed in #574. Can you try again with report version 0.6.1.5? Using

easystats::install_latest(packages = "report")

[rempsyc]

Seems like there's only a modest improvement. But I suppose that is only useful if we knew how much time running the original model takes, because now we have no idea whether it takes as much time, more time, or less time?

library(bench)
library(tidyverse)
library(report)
library(rstan)
packageVersion("report")
#> [1] '0.6.1.4'

mark(
  fit_c <- read_rds("https://osf.io/hmqpc/download"),
  report_model(fit_c, effectsize_method = "refit"),
  report_model(fit_c),
  check = FALSE
)

#> # A tibble: 3 × 6
#>   expression                             min median `itr/sec` mem_alloc `gc/sec`
#>   <bch:expr>                          <bch:> <bch:>     <dbl> <bch:byt>    <dbl>
#> 1 "fit_c <- read_rds(\"https://osf.i… 25.18s 25.18s 0.0397     978.51MB  0      
#> 2 "report_model(fit_c, effectsize_me…  4.19h  4.19h 0.0000663     1.2TB  0.00597
#> 3 "report_model(fit_c)"                3.76h  3.76h 0.0000740    1.18TB  0.00481

Created on 2025-09-21 with reprex v2.1.1

Though I see this message when running report_model() on this object:

The desired updates require recompiling the model
Start sampling
#> Running MCMC with 4 sequential chains, with 4 thread(s) per chain...
#> Chain 1 Iteration: 1 / 5000 [ 0%] (Warmup)

Question is whether we can avoid recompiling the model or whether this is unavoidable because of the extra information we are asking.

[rempsyc]

Oops, I was still on the old version. On version 0.6.1.5, there is a major improvement: an almost 4x reduction from ~4h to ~50 min:

library(bench)
library(tidyverse)
library(easystats)
library(rstan)
packageVersion("report")
#> [1] '0.6.1.5'

mark(
  fit_c <- read_rds("https://osf.io/hmqpc/download"),
  report_model(fit_c, effectsize_method = "refit"),
  report_model(fit_c),
  max_iterations = 1,
  check = FALSE
)
#> The desired updates require recompiling the model
#> Start sampling
#> Running MCMC with 4 sequential chains, with 4 thread(s) per chain...
#> 
#> Chain 1 Iteration:    1 / 5000 [  0%]  (Warmup) 
[...]
#> Chain 4 Iteration: 5000 / 5000 [100%]  (Sampling) 
#> Chain 4 finished in 2612.8 seconds.
#> The desired updates require recompiling the model
#> Start sampling
#> Running MCMC with 4 sequential chains, with 4 thread(s) per chain...
#> 
#> Chain 1 Iteration:    1 / 5000 [  0%]  (Warmup) 
[...]
#> Chain 4 Iteration: 5000 / 5000 [100%]  (Sampling) 
#> Chain 4 finished in 2671.7 seconds.
#> 
#> # A tibble: 3 × 6
#>   expression                             min median `itr/sec` mem_alloc `gc/sec`
#>   <bch:expr>                          <bch:> <bch:>     <dbl> <bch:byt>    <dbl>
#> 1 "fit_c <- read_rds(\"https://osf.i… 22.35s 22.35s 0.0447     914.81MB  0      
#> 2 "report_model(fit_c, effectsize_me…  4.18h  4.18h 0.0000664     1.2TB  0.00571
#> 3 "report_model(fit_c)"               49.73m 49.73m 0.000335     1.25TB  0.0255
Created on 2025-09-22 with [reprex v2.1.1](https://reprex.tidyverse.org/)

And also when running report alone, the model isn't refitted so there is no recompiling the model and new chains, etc.

[rempsyc]

Closing this issue as I believe it is solved now (no more refitting of the model by default), but let me know if it's not solved for you and we can reopen it