prior_summary() |> dplyr::select() alters the content in the prior column

[ASKurz]

It looks like there's a bug when you try to subset the columns from prior_summary() with dplyr::select(). Some of the information in the prior column changes. To give a sense, here's a simplified version of fit2 from the inhaler documentation.

# Load
library(tidyverse)
library(brms)

# Fit
fit2.1 <- brm(rating ~ treat + (1 | subject),
              data = inhaler, 
              family = cumulative(),
              prior = set_prior("normal(0, 5)"),
              cores = 4, seed = 1)

# Before `select()`
prior_summary(fit2.1)

                prior     class      coef   group resp dpar nlpar lb ub       source
          normal(0,5)         b                                                 user
          normal(0,5)         b     treat                               (vectorized)
 student_t(3, 0, 2.5) Intercept                                              default
 student_t(3, 0, 2.5) Intercept         1                               (vectorized)
 student_t(3, 0, 2.5) Intercept         2                               (vectorized)
 student_t(3, 0, 2.5) Intercept         3                               (vectorized)
 student_t(3, 0, 2.5)        sd                                    0         default
 student_t(3, 0, 2.5)        sd           subject                  0    (vectorized)
 student_t(3, 0, 2.5)        sd Intercept subject                  0    (vectorized)

To my eye, this all looks correct. Now see what happens when we use select().

prior_summary(fit2.1) |> 
  select(prior, class, coef, group) 

                prior     class      coef   group    source
          normal(0,5)         b                   (unknown)
               (flat)         b     treat         (unknown)
 student_t(3, 0, 2.5) Intercept                   (unknown)
               (flat) Intercept         1         (unknown)
               (flat) Intercept         2         (unknown)
               (flat) Intercept         3         (unknown)
 student_t(3, 0, 2.5)        sd                   (unknown)
               (flat)        sd           subject (unknown)
               (flat)        sd Intercept subject (unknown)

Several rows have now changed to (flat) in the prior column. The issue appears the same if you instead use the get_prior(fit2.1) or fit2.1$prior methods. It also persists if you use the base bracket notation in place of select() (prior_summary(fit2.1)[, 1:4]).

Though my primary interest is with the prior column, I find the behavior of the source column surprising, too. When I tried to drop that column with select(), it still appeared in the print output, but now with all (unknown).

[paul-buerkner]

not the content changes. just what is printed. when changing the class to a regular data.frame the special printing is no longer applied. that is what you are seeing A. Solomon Kurz ***@***.***> schrieb am Mi., 2. Apr. 2025, 17:33:

…

It looks like their's a bug when you try to subset the columns from prior_summary() with dplyr::select(). Some of the information in the prior column changes. To give a sense, here's a simplified version of fit2 from the inhaler documentation. # Load library(tidyverse) library(brms) # Fit fit2.1 <- brm(rating ~ treat + (1 | subject), data = inhaler, family = cumulative(), prior = set_prior("normal(0, 5)"), cores = 4, seed = 1) # Before `select()` prior_summary(fit2.1) prior class coef group resp dpar nlpar lb ub source normal(0,5) b user normal(0,5) b treat (vectorized) student_t(3, 0, 2.5) Intercept default student_t(3, 0, 2.5) Intercept 1 (vectorized) student_t(3, 0, 2.5) Intercept 2 (vectorized) student_t(3, 0, 2.5) Intercept 3 (vectorized) student_t(3, 0, 2.5) sd 0 default student_t(3, 0, 2.5) sd subject 0 (vectorized) student_t(3, 0, 2.5) sd Intercept subject 0 (vectorized) To my eye, this all looks correct. Now see what happens when we use select(). prior_summary(fit2.1) |> select(prior, class, coef, group) prior class coef group source normal(0,5) b (unknown) (flat) b treat (unknown) student_t(3, 0, 2.5) Intercept (unknown) (flat) Intercept 1 (unknown) (flat) Intercept 2 (unknown) (flat) Intercept 3 (unknown) student_t(3, 0, 2.5) sd (unknown) (flat) sd subject (unknown) (flat) sd Intercept subject (unknown) Several rows have now changed to (flat) in the prior column. The issue appears the same if you instead use the get_prior(fit2.1) or fit2.1$prior methods, instead. — Reply to this email directly, view it on GitHub <#1761>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/ADCW2AG562LHTYYRDLBANCD2XPYKDAVCNFSM6AAAAAB2JX24HKVHI2DSMVQWIX3LMV43ASLTON2WKOZSHE3DMNJYGE4TSOI> . You are receiving this because you are subscribed to this thread.Message ID: ***@***.***> [image: ASKurz]*ASKurz* created an issue (paul-buerkner/brms#1761) <#1761> It looks like their's a bug when you try to subset the columns from prior_summary() with dplyr::select(). Some of the information in the prior column changes. To give a sense, here's a simplified version of fit2 from the inhaler documentation. # Load library(tidyverse) library(brms) # Fit fit2.1 <- brm(rating ~ treat + (1 | subject), data = inhaler, family = cumulative(), prior = set_prior("normal(0, 5)"), cores = 4, seed = 1) # Before `select()` prior_summary(fit2.1) prior class coef group resp dpar nlpar lb ub source normal(0,5) b user normal(0,5) b treat (vectorized) student_t(3, 0, 2.5) Intercept default student_t(3, 0, 2.5) Intercept 1 (vectorized) student_t(3, 0, 2.5) Intercept 2 (vectorized) student_t(3, 0, 2.5) Intercept 3 (vectorized) student_t(3, 0, 2.5) sd 0 default student_t(3, 0, 2.5) sd subject 0 (vectorized) student_t(3, 0, 2.5) sd Intercept subject 0 (vectorized) To my eye, this all looks correct. Now see what happens when we use select(). prior_summary(fit2.1) |> select(prior, class, coef, group) prior class coef group source normal(0,5) b (unknown) (flat) b treat (unknown) student_t(3, 0, 2.5) Intercept (unknown) (flat) Intercept 1 (unknown) (flat) Intercept 2 (unknown) (flat) Intercept 3 (unknown) student_t(3, 0, 2.5) sd (unknown) (flat) sd subject (unknown) (flat) sd Intercept subject (unknown) Several rows have now changed to (flat) in the prior column. The issue appears the same if you instead use the get_prior(fit2.1) or fit2.1$prior methods, instead. — Reply to this email directly, view it on GitHub <#1761>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/ADCW2AG562LHTYYRDLBANCD2XPYKDAVCNFSM6AAAAAB2JX24HKVHI2DSMVQWIX3LMV43ASLTON2WKOZSHE3DMNJYGE4TSOI> . You are receiving this because you are subscribed to this thread.Message ID: ***@***.***>

[lunafazio]

when changing the class to a regular data.frame the special printing is no longer applied. that is what you are seeing

That's not exactly true. dplyr does keep the brmsprior class, which is why (flat) is printed instead of the actual content of those entries which is just an empty string.

There are a few other weird things that can happen when manipulating the brmsprior frame in ways that alter its structure. I suppose this could be avoided with some sort of validation that warns the user when they break structure or proactively unclasses the object as draws_df do when they lose their required metadata columns.

[paul-buerkner]

Yeah that is a good point. Unclassing brmsprior objects upon certain columns being changes/dropped could be sensible.

[ASKurz]

@paul-buerkner, I'm not sure if this is implied by your last comment, but from my standpoint it would be really great if code like this

prior_summary(fit2.1) |> 
  data.frame()

                 prior     class      coef   group resp dpar nlpar lb ub  source
1          normal(0,5)         b                                            user
2                              b     treat                               default
3 student_t(3, 0, 2.5) Intercept                                         default
4                      Intercept         1                               default
5                      Intercept         2                               default
6                      Intercept         3                               default
7 student_t(3, 0, 2.5)        sd                                    0    default
8                             sd           subject                       default
9                             sd Intercept subject                       default

carried the filled-in cells in the prior column down to the blank cells below, similar to what happens in the print() method.

prior_summary(fit2.1) |> 
  print()

                prior     class      coef   group resp dpar nlpar lb ub       source
          normal(0,5)         b                                                 user
          normal(0,5)         b     treat                               (vectorized)
 student_t(3, 0, 2.5) Intercept                                              default
 student_t(3, 0, 2.5) Intercept         1                               (vectorized)
 student_t(3, 0, 2.5) Intercept         2                               (vectorized)
 student_t(3, 0, 2.5) Intercept         3                               (vectorized)
 student_t(3, 0, 2.5)        sd                                    0         default
 student_t(3, 0, 2.5)        sd           subject                  0    (vectorized)
 student_t(3, 0, 2.5)        sd Intercept subject                  0    (vectorized)

Same request about what happens with the lb (and presumably the ub) column. It'd be great if the values carried down in the data frame, as they do with the print() method.

[paul-buerkner]

I see. Could make sense. I will have to think about what the best behavior is in that regard.