Update `construct_default_priors_list` to accomodate bmmformulas more… by GidonFrischkorn · Pull Request #306 · venpopov/bmm

GidonFrischkorn · 2026-01-26T13:41:04Z

Fix default prior specification for multiple predictors and interaction-only formulas

Issues

Bug 1: Multiple predictors without intercept
Formulas like kappa ~ 0 + set_size + session only set priors on the first predictor (set_size), leaving subsequent predictors with flat priors.

Bug 2: Interaction-only with intercept
Formulas like kappa ~ 1 + set_size:session resulted in flat class-level priors for interaction terms instead of inheriting from the effects prior.

Solution

Restructured logic to set class-level effects prior only when intercept is present OR for multi-predictor formulas without intercept
Added condition to check interactions_count > 0 alongside fixed_effects_count > 0 for setting class-level effects priors
For multi-predictor formulas without intercept (e.g., 0 + A + B), first predictor now gets main prior, remaining predictors get effects prior (matching brms' statistical interpretation)

Tests

Added comprehensive coverage validating:

Multi-predictor handling with/without intercept (0 + A + B, 1 + A + B)
Interaction-only formulas (1 + A:B, 0 + A:B)
Proper main vs effects prior distinction across patterns
Invariance to factor contrast coding (contr.treatment, contr.sum, contr.helmert, contr.equalprior)

[x] Confirm that all tests passed
[x] Confirm that devtools::check() produces no errors

Release notes

Bug fixes

Fixed default prior specification for formulas with multiple predictors without intercept (e.g., 0 + A + B) - now correctly assigns main prior to first predictor and effects prior to remaining predictors
Fixed default prior specification for interaction-only formulas with intercept (e.g., 1 + A:B) - now correctly sets class-level effects prior instead of flat priors

… boradly

- Remove tests that rely on bayestestR

venpopov · 2026-03-02T22:01:32Z

Are you sure there is a problem to be solved? When I tested the supposed problem cases I get correct priors on the develop branch:

library(bmm)
library(dplyr)
#> 
#> Attaching package: 'dplyr'
#> The following objects are masked from 'package:stats':
#> 
#>     filter, lag
#> The following objects are masked from 'package:base':
#> 
#>     intersect, setdiff, setequal, union
data <- oberauer_lin_2017
data$group <- factor(rep(c('A', 'B'), times = nrow(data)/2))
data$cond <- factor(rep(c('x', 'y'), each = nrow(data)/2))

model <- mixture2p('dev_rad')

# Check what bmm sets
print(filter(default_prior(bmf(kappa ~ 0 + group:cond), data, model), nlpar == 'kappa'))
#> No formula for parameter thetat provided. Only a fixed intercept will be estimated.
#>         prior class         coef group resp dpar nlpar   lb   ub tag
#>  normal(2, 1)     b groupA:condx                 kappa <NA> <NA>    
#>  normal(2, 1)     b groupA:condy                 kappa <NA> <NA>    
#>  normal(2, 1)     b groupB:condx                 kappa <NA> <NA>    
#>  normal(2, 1)     b groupB:condy                 kappa <NA> <NA>    
#>  normal(2, 1)     b                              kappa <NA> <NA>    
#>        source
#>  (vectorized)
#>  (vectorized)
#>  (vectorized)
#>  (vectorized)
#>          user
print(filter(default_prior(bmf(kappa ~ 0 + group + cond), data, model), nlpar == 'kappa'))
#> No formula for parameter thetat provided. Only a fixed intercept will be estimated.
#>         prior class   coef group resp dpar nlpar   lb   ub tag       source
#>  normal(0, 1)     b  condy                 kappa <NA> <NA>     (vectorized)
#>  normal(0, 1)     b                        kappa <NA> <NA>             user
#>  normal(2, 1)     b groupA                 kappa <NA> <NA>             user
#>  normal(2, 1)     b groupB                 kappa <NA> <NA>             user
print(filter(default_prior(bmf(kappa ~ 0 + group + group:cond), data, model), nlpar == 'kappa'))
#> No formula for parameter thetat provided. Only a fixed intercept will be estimated.
#>         prior class         coef group resp dpar nlpar   lb   ub tag
#>  normal(0, 1)     b groupA:condy                 kappa <NA> <NA>    
#>  normal(0, 1)     b groupB:condy                 kappa <NA> <NA>    
#>  normal(0, 1)     b                              kappa <NA> <NA>    
#>  normal(2, 1)     b       groupA                 kappa <NA> <NA>    
#>  normal(2, 1)     b       groupB                 kappa <NA> <NA>    
#>        source
#>  (vectorized)
#>  (vectorized)
#>          user
#>          user
#>          user
# Check what bmm sets
print(filter(default_prior(bmf(kappa ~ 0 + set_size:session), data, model), nlpar == 'kappa'))
#> No formula for parameter thetat provided. Only a fixed intercept will be estimated.
#>         prior class               coef group resp dpar nlpar   lb   ub tag
#>  normal(2, 1)     b set_size1:session1                 kappa <NA> <NA>    
#>  normal(2, 1)     b set_size1:session2                 kappa <NA> <NA>    
#>  normal(2, 1)     b set_size2:session1                 kappa <NA> <NA>    
#>  normal(2, 1)     b set_size2:session2                 kappa <NA> <NA>    
#>  normal(2, 1)     b set_size3:session1                 kappa <NA> <NA>    
#>  normal(2, 1)     b set_size3:session2                 kappa <NA> <NA>    
#>  normal(2, 1)     b set_size4:session1                 kappa <NA> <NA>    
#>  normal(2, 1)     b set_size4:session2                 kappa <NA> <NA>    
#>  normal(2, 1)     b set_size5:session1                 kappa <NA> <NA>    
#>  normal(2, 1)     b set_size5:session2                 kappa <NA> <NA>    
#>  normal(2, 1)     b set_size6:session1                 kappa <NA> <NA>    
#>  normal(2, 1)     b set_size6:session2                 kappa <NA> <NA>    
#>  normal(2, 1)     b set_size7:session1                 kappa <NA> <NA>    
#>  normal(2, 1)     b set_size7:session2                 kappa <NA> <NA>    
#>  normal(2, 1)     b set_size8:session1                 kappa <NA> <NA>    
#>  normal(2, 1)     b set_size8:session2                 kappa <NA> <NA>    
#>  normal(2, 1)     b                                    kappa <NA> <NA>    
#>        source
#>  (vectorized)
#>  (vectorized)
#>  (vectorized)
#>  (vectorized)
#>  (vectorized)
#>  (vectorized)
#>  (vectorized)
#>  (vectorized)
#>  (vectorized)
#>  (vectorized)
#>  (vectorized)
#>  (vectorized)
#>  (vectorized)
#>  (vectorized)
#>  (vectorized)
#>  (vectorized)
#>          user
print(filter(default_prior(bmf(kappa ~ 0 + set_size + session), data, model), nlpar == 'kappa'))
#> No formula for parameter thetat provided. Only a fixed intercept will be estimated.
#>         prior class      coef group resp dpar nlpar   lb   ub tag       source
#>  normal(0, 1)     b  session2                 kappa <NA> <NA>     (vectorized)
#>  normal(0, 1)     b                           kappa <NA> <NA>             user
#>  normal(2, 1)     b set_size1                 kappa <NA> <NA>             user
#>  normal(2, 1)     b set_size2                 kappa <NA> <NA>             user
#>  normal(2, 1)     b set_size3                 kappa <NA> <NA>             user
#>  normal(2, 1)     b set_size4                 kappa <NA> <NA>             user
#>  normal(2, 1)     b set_size5                 kappa <NA> <NA>             user
#>  normal(2, 1)     b set_size6                 kappa <NA> <NA>             user
#>  normal(2, 1)     b set_size7                 kappa <NA> <NA>             user
#>  normal(2, 1)     b set_size8                 kappa <NA> <NA>             user
print(filter(default_prior(bmf(kappa ~ 0 + set_size + set_size:session), data, model), nlpar == 'kappa'))
#> No formula for parameter thetat provided. Only a fixed intercept will be estimated.
#>         prior class               coef group resp dpar nlpar   lb   ub tag
#>  normal(0, 1)     b set_size1:session2                 kappa <NA> <NA>    
#>  normal(0, 1)     b set_size2:session2                 kappa <NA> <NA>    
#>  normal(0, 1)     b set_size3:session2                 kappa <NA> <NA>    
#>  normal(0, 1)     b set_size4:session2                 kappa <NA> <NA>    
#>  normal(0, 1)     b set_size5:session2                 kappa <NA> <NA>    
#>  normal(0, 1)     b set_size6:session2                 kappa <NA> <NA>    
#>  normal(0, 1)     b set_size7:session2                 kappa <NA> <NA>    
#>  normal(0, 1)     b set_size8:session2                 kappa <NA> <NA>    
#>  normal(0, 1)     b                                    kappa <NA> <NA>    
#>  normal(2, 1)     b          set_size1                 kappa <NA> <NA>    
#>  normal(2, 1)     b          set_size2                 kappa <NA> <NA>    
#>  normal(2, 1)     b          set_size3                 kappa <NA> <NA>    
#>  normal(2, 1)     b          set_size4                 kappa <NA> <NA>    
#>  normal(2, 1)     b          set_size5                 kappa <NA> <NA>    
#>  normal(2, 1)     b          set_size6                 kappa <NA> <NA>    
#>  normal(2, 1)     b          set_size7                 kappa <NA> <NA>    
#>  normal(2, 1)     b          set_size8                 kappa <NA> <NA>    
#>        source
#>  (vectorized)
#>  (vectorized)
#>  (vectorized)
#>  (vectorized)
#>  (vectorized)
#>  (vectorized)
#>  (vectorized)
#>  (vectorized)
#>          user
#>          user
#>          user
#>          user
#>          user
#>          user
#>          user
#>          user
#>          user

GidonFrischkorn · 2026-03-03T07:02:11Z

The bugs are real, but your examples don't expose them for two reasons:

You only test kappa, which already has an effects prior on develop. The bug affects parameters without effects priors (like thetat).

Compare your single-formula approach vs specifying both parameters:

library(bmm)
library(dplyr)
data <- oberauer_lin_2017
model <- mixture2p("dev_rad")

# Your approach — looks fine because kappa has effects prior defined:
pr <- default_prior(bmf(kappa ~ 0 + set_size + session), data, model)
filter(pr, nlpar == "kappa", coef == "session2")
#>  normal(0, 1)     b session2  kappa  (vectorized)

# But specify thetat with the same formula — thetat gets flat priors:
pr2 <- default_prior(
  bmf(kappa ~ 0 + set_size + session, thetat ~ 0 + set_size + session),
  data, model
)
filter(pr2, coef == "session2")
#>  (flat)     b session2  kappa default
#>  (flat)     b session2 thetat default

thetat has no effects prior defined on develop, so it gets completely flat priors for all regression coefficients. Same issue applies to mu1 (mixture models) and mu (SDM).

You tested kappa ~ 0 + set_size:session (no intercept), but the code bug is in kappa ~ 1 + set_size:session (with intercept).

# No intercept + interaction — works fine (your test):
pr <- default_prior(bmf(kappa ~ 0 + set_size:session), data, model)
filter(pr, nlpar == "kappa", coef == "")
#>  normal(2, 1)     b    kappa  user

# WITH intercept + interaction — all interactions get flat priors:
pr <- default_prior(bmf(kappa ~ 1 + set_size:session), data, model)
filter(pr, nlpar == "kappa", class == "b")
#>  (flat)     b                    kappa default
#>  (flat)     b set_size1:session1 kappa default
#>  ...all 16 interaction terms are (flat)...
#>  normal(2, 1)  b Intercept       kappa user

The cause: construct_default_priors_list checks fixed_effects_count > 0 to decide whether to set the class-level effects prior, but interactions have order > 1 so they're not counted. The fix changes this to fixed_effects_count > 0 || interactions_count > 0.

venpopov · 2026-03-03T11:02:54Z

I see. There are a couple of things going on here:

I tested with kappa because that's what the original issue describes! So there is no bug 1 in constructing the priors - rather just missing effect priors for some parameters like theta. When you include theta yes you don't get effect priors on it but that's not specific to "Multiple predictors without intercept" - just reflects that we hadn't put effect priors on theta.

The fact that it shows also flat for kappa in your example is an artefact of filtering the prior df and how brms deals with this - the prior on kappa effects is only defined in the generic "b" row, and the summary "pretends it is defined for all b rows, but if you filter it it shows the underlying empty field. Without filtering it looks like:

          prior     class      coef group resp   dpar  nlpar   lb   ub tag       source
 logistic(0, 1)    theta2                                    -Inf  Inf          default
   normal(0, 1)         b  session2                    kappa <NA> <NA>     (vectorized)
         (flat)         b                             thetat                    default
         (flat)         b  session2                   thetat               (vectorized)
   normal(0, 1)         b                              kappa <NA> <NA>             user
   normal(2, 1)         b set_size1                    kappa <NA> <NA>             user
   normal(2, 1)         b set_size2                    kappa <NA> <NA>             user
   normal(2, 1)         b set_size3                    kappa <NA> <NA>             user
   normal(2, 1)         b set_size4                    kappa <NA> <NA>             user
   normal(2, 1)         b set_size5                    kappa <NA> <NA>             user
   normal(2, 1)         b set_size6                    kappa <NA> <NA>             user
   normal(2, 1)         b set_size7                    kappa <NA> <NA>             user
   normal(2, 1)         b set_size8                    kappa <NA> <NA>             user

So yes, missing effects on theta, but not on kappa. Thetat still misses priors even in the simplest formula that has effects:

pr2 <- default_prior(
  bmf(kappa ~ 0 + set_size + session, thetat ~ session),
  data, model
)
pr2
---
          prior     class      coef group resp   dpar  nlpar   lb   ub tag       source
 logistic(0, 1)    theta2                                    -Inf  Inf          default
   normal(0, 1)         b  session2                    kappa <NA> <NA>     (vectorized)
         (flat)         b                             thetat                    default
         (flat)         b  session2                   thetat               (vectorized)
   normal(0, 1)         b                              kappa <NA> <NA>             user
   normal(2, 1)         b set_size1                    kappa <NA> <NA>             user
   normal(2, 1)         b set_size2                    kappa <NA> <NA>             user
   normal(2, 1)         b set_size3                    kappa <NA> <NA>             user
   normal(2, 1)         b set_size4                    kappa <NA> <NA>             user
   normal(2, 1)         b set_size5                    kappa <NA> <NA>             user
   normal(2, 1)         b set_size6                    kappa <NA> <NA>             user
   normal(2, 1)         b set_size7                    kappa <NA> <NA>             user
   normal(2, 1)         b set_size8                    kappa <NA> <NA>             user
 logistic(0, 1)         b Intercept                   thetat <NA> <NA>             user
    constant(0) Intercept                         mu1        <NA> <NA>             user
    constant(0) Intercept                         mu2        <NA> <NA>             user
 constant(-100) Intercept                      kappa2        <NA> <NA>             user

I'm not saying that we shouldn't add effect priors to the parameters that lack them, but I was confused by the description of the bug because I don't see a bug.

I tested kappa ~ 0 + set_size:session (no intercept) because kappa ~ 1 + set_size:session is not a proper model specification. It includes both an intercept and a separate parameter for every condition combination, which is not identifiable and shouldn't be run at all! Here's the model matrix for such a formula:

dat <- expand.grid(f1 = letters[1:3], f2 = LETTERS[1:2])
dat$y <- rnorm(nrow(dat), mean = as.numeric(dat$f1) * as.numeric(dat$f2))
model.matrix(y ~ 1 + f1:f2, data = dat)
#>   (Intercept) f1a:f2A f1b:f2A f1c:f2A f1a:f2B f1b:f2B f1c:f2B
#> 1           1       1       0       0       0       0       0
#> 2           1       0       1       0       0       0       0
#> 3           1       0       0       1       0       0       0
#> 4           1       0       0       0       1       0       0
#> 5           1       0       0       0       0       1       0
#> 6           1       0       0       0       0       0       1
#> attr(,"assign")
#> [1] 0 1 1 1 1 1 1
#> attr(,"contrasts")
#> attr(,"contrasts")$f1
#> [1] "contr.treatment"
#> 
#> attr(,"contrasts")$f2
#> [1] "contr.treatment"

When you try to estimate such a model, you get a singularity and NA values, and at least with lm it arbitrarily chooses the last level to be dropped and implicitly acts as the intercept:

dat <- expand.grid(id = 1:5, f1 = letters[1:3], f2 = LETTERS[1:2])
dat$y <- rnorm(nrow(dat), mean = as.numeric(dat$f1) * as.numeric(dat$f2))

lm(y ~ 1 + f1:f2, data = dat) |> summary()
#> 
#> Call:
#> lm(formula = y ~ 1 + f1:f2, data = dat)
#> 
#> Residuals:
#>     Min      1Q  Median      3Q     Max 
#> -1.9306 -0.6243 -0.0380  0.5917  3.7008 
#> 
#> Coefficients: (1 not defined because of singularities)
#>             Estimate Std. Error t value Pr(>|t|)    
#> (Intercept)   6.4577     0.5747  11.237 4.81e-11 ***
#> f1a:f2A      -5.8709     0.8127  -7.224 1.83e-07 ***
#> f1b:f2A      -4.0398     0.8127  -4.971 4.48e-05 ***
#> f1c:f2A      -3.5928     0.8127  -4.421 0.000181 ***
#> f1a:f2B      -4.8769     0.8127  -6.001 3.40e-06 ***
#> f1b:f2B      -2.5487     0.8127  -3.136 0.004482 ** 
#> f1c:f2B           NA         NA      NA       NA    
#> ---
#> Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
#> 
#> Residual standard error: 1.285 on 24 degrees of freedom
#> Multiple R-squared:  0.7257, Adjusted R-squared:  0.6686 
#> F-statistic:  12.7 on 5 and 24 DF,  p-value: 4.245e-06

This is degenerate and lm tries to fix it. I don't even know what happens if you try to run it with brms.

venpopov · 2026-03-03T12:01:45Z

I'm sorry that I phrased my original messages so confrontationally. I couldn't sleep and it came off sharper than I intended - I was just trying to understand what the issue is so that I know what the solution is aiming to do to review it properly. I am super grateful for all the work you've done on these PRs - they both extend the modeling options substantially and make the user experience much better and easier.

GidonFrischkorn · 2026-03-03T14:19:38Z

No worries. I totally understand that you want to understand the issue.

I have to think about this a bit and not handle this in between tasks. So, let me get back to you, once I have found time to sort out what you have raised and what my initial concerns were.

I see your point that 1 + A:B is not appropriate and is an inadequate model specification and thus we should probably not include it as a test case.

Will get back to you ASAP.

Update construct_default_priors_list to accomodate bmmformulas more…

d1b21b9

… boradly

GidonFrischkorn linked an issue Jan 26, 2026 that may be closed by this pull request

[bug] Default Priors Not Applied to Interaction Terms in Formula Pattern 0 + A + A:B #305

Open

GidonFrischkorn requested a review from venpopov January 26, 2026 13:41

GidonFrischkorn added 2 commits January 26, 2026 15:08

Address warning during tests.

62957e5

Update test-default-priors.R

5520982

- Remove tests that rely on bayestestR

GidonFrischkorn added this to the 1.3.0 milestone Jan 26, 2026

GidonFrischkorn self-assigned this Feb 1, 2026

GidonFrischkorn added the bug Something isn't working label Feb 1, 2026

GidonFrischkorn and others added 4 commits February 11, 2026 11:53

Merge branch 'develop' into 305-bug-default-priors

440eb6d

Merge branch 'develop' into 305-bug-default-priors

55f7234

Update helpers-prior.R

1f9b5ed

Merge branch 'develop' into 305-bug-default-priors

43f522a

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Update `construct_default_priors_list` to accomodate bmmformulas more…#306

Update `construct_default_priors_list` to accomodate bmmformulas more…#306
GidonFrischkorn wants to merge 7 commits intodevelopfrom
305-bug-default-priors

GidonFrischkorn commented Jan 26, 2026

Uh oh!

venpopov commented Mar 2, 2026

Uh oh!

GidonFrischkorn commented Mar 3, 2026

Uh oh!

venpopov commented Mar 3, 2026

Uh oh!

venpopov commented Mar 3, 2026 •

edited

Loading

Uh oh!

GidonFrischkorn commented Mar 3, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

Conversation

GidonFrischkorn commented Jan 26, 2026

Fix default prior specification for multiple predictors and interaction-only formulas

Issues

Solution

Tests

Release notes

Uh oh!

venpopov commented Mar 2, 2026

Uh oh!

GidonFrischkorn commented Mar 3, 2026

Uh oh!

venpopov commented Mar 3, 2026

Uh oh!

venpopov commented Mar 3, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

GidonFrischkorn commented Mar 3, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

venpopov commented Mar 3, 2026 •

edited

Loading