Posterior Predictions with a Wiener model - very slow?

Interfaces brms
,
1

Hello,

Does anybody have any advice on generating predictions from a BRMS model with a Wiener distribution? When I try, the process seems very slow. It could be that my own intuition is very wrong, and everything is fine, but I thought it would be good to double check I’m not doing something dumb!

At the moment I am using the tidybayes package.

The dataset I use to fit the model has 22k rows (data from a psychology experiment)
I am aggregating over repetitions before running the predictions, which reduces the amount of data to 2.3k rows.

If I run the below (with n = 1), the code takes around 6 seconds to run, and the predicted values look quite nice when compared to the empirical median reaction time.

start_time <- Sys.time()

d %>% group_by(id, bk, co, st) %>%
  summarise(median_rt = median(rt),
            mean_resp = mean(resp),
            .groups = "drop") %>%
  add_predicted_draws(m,
                      re_formula = NULL,
                      n = 1,
                      negative_rt = TRUE,
                      prediction = "prt") %>%
  select(-.chain, -.iteration, -.draw) -> p

print(Sys.time() - start_time)

Although, if I then rm(p) and rerun the code, it takes much longer. On the last try, I hit the stop button after waiting for 2 minutes! Restarting R and running again leads to 11 seconds to run the code.

Ideally, I think I want to run the above with higher n.

My PC has 16GB RAM and is running Windows.

If it helps, the model formula is:

formula <- bf(rt | dec(resp) ~ 0 + bk:co:st + (0 + bk:co:st|p|id), 
              bs ~ 0 + bk:co + (0 + bk:co|p|id),
              bias ~ 0 + bk:co + (0 + bk:co|p|id),
              ndt ~ 1 + (1|p|id))

Thank you for your help

1 Like
2

This implies that you would expect to wait about 6 hours to get predictions for all 4000 posterior samples (which is the default). So waiting for 2 minutes doesn’t seem unexpected. Unfortunately, the prediction code in brms can be quite slow.

Best of luck with your model!

3

I just ran into this practicing with Paul’s fit_drift1 from his Bayesian IRT paper (https://arxiv.org/pdf/1905.09501), and I can confirm executing pp_check(fit_drift1, ndraws = 30) was shockingly slow. So slow that I manually quit after several minutes. This was with brms 2.22.0.

4

This is because the packages that brms uses for predictions of wiener DDMs are slow for random number generation. Sometimes the Rwiener backend is slow while the rtdists backend is faster. you can change the backend via option(“wiener_backend”, “rtdists”)

1 Like
5

Fast wiener_rng Alternative to wiener_rng in Stan - #11 by spinkney

1 Like
6

Good to know; thanks!