glmmTMB ordinal support

[strengejacke]

Fixes #1220

[jmgirard]

Thanks for starting this! Two things I noticed testing it on the wine data:

1. In get_parameters(), out[-1, ] runs after the thresholds are prepended, so it drops 1|2 and keeps (Intercept).
2. The threshold rows of vcov(full = TRUE) are on glmmTMB's internal softmax scale, so the SEs don't match clm (0.99 vs 0.55 for 1|2); they need the delta method (see Standard errors of category threshold parameters for the new ordinal family glmmTMB/glmmTMB#1323).

I have a branch that handles both, plus get_statistic() (which currently errors on the row-count mismatch with summary()), get_predicted() via type = "probs" with clm-style intervals, get_variance() (latent residual variance, as for clmm), and tests against clm/clmm:

main...jmgirard:insight:glmmtmb-ordinal

Happy to open it as a PR, or you can cherry-pick from it into this one, whichever you prefer. One thing on the parameters side: .se_fixed_effects_glmmTMB() takes SEs from coef(summary()), which lacks the thresholds, so model_parameters() errors; falling back to insight::get_varcov() for the ordinal family would fix it.

[strengejacke]

Great, thanks a lot! Let me check how get_parameters.glmmTMB() handles ordinal families, and if this is in line with ordinal::clm() and similar (i.e. thresholds are also returned).

Also, I noticed that vcov() for glmmTMB-ordinal returns fixed effects first, and then threshold estimates. We have to make sure this is in line with get_parameters(), so standard errors will correctly match in parameters.

[jmgirard]

Good questions, luckily both should be covered in the branch already.

get_parameters() returns the thresholds first and then the slopes, with the same names and order as ordinal::clm()/clmm(); find_parameters(m_tmb) is identical() to find_parameters(m_clm) in test-glmmTMB-ordinal.R. For the covariance matrix, you're right that glmmTMB's vcov() puts the fixed effects first, and the threshold rows there are also on the internal softmax scale, so get_varcov() doesn't use it directly. It builds the delta-method matrix itself (R/utils_glmmtmb_ordinal.R) in the same order as get_parameters(), and the tests check the SEs against clm()/clmm(). On the parameters side, easystats/parameters#1253 takes SEs from get_varcov() matched by name, so model_parameters() matches clmm() column for column there as well.

On CI: the ordinal work is lint- and spelling-clean now; what's still red comes from main (the brms tests on Windows, styling on test-brms_dpars.R, and the older lints and spelling words). Happy to help clean those up in a separate PR if useful.

[strengejacke]

Short note: Please don't focus too much on CI-tests, they're not fully reliable, AFAIK, and also yield some false-positives. Main formatting is now done by air, so not all lintr-comments must be followed (although there are always some useful comments).

[strengejacke]

Let me start a copilot review, I'll also review this PR later in office.

[Copilot]

🟡 Changes recommended

Unresolved moderate issues affect prediction dispatch, covariance transformation, and prediction argument forwarding.

Once you've addressed the issues Copilot identified, you can request another Copilot review.

Pull request overview

Adds glmmTMB::ordinal() support across parameter extraction, covariance/statistic calculations, predictions, and variance estimates.

Changes:

• Adds threshold-aware parameters and delta-method covariance handling.
• Supports probability and classification predictions.
• Adds tests and documentation updates.

File summaries

File	Summary / review note
tests/testthat/test-glmmTMB-ordinal.R	Adds ordinal model coverage.
R/utils_glmmtmb_ordinal.R	Adds ordinal helpers; moderate issue with forwarding prediction arguments.
R/get_varcov.R	Adds ordinal covariance handling; moderate issue with supplied robust covariances.
R/get_statistic.R	Adds ordinal Wald statistics.
R/get_predicted_mixed.R	Adds ordinal prediction dispatch; moderate issue missing the expected alias.
R/get_parameters_mixed.R	Adds threshold parameter extraction.
R/find_parameters_mixed.R	Adds ordinal parameter discovery.
R/compute_variances.R	Adds ordinal variance handling.
NEWS.md	Documents ordinal support; nit regarding placement under the development section.
inst/WORDLIST	Adds softmax vocabulary.

Review details

Suppressed comments (2)

R/get_predicted_mixed.R:120

• expected is an established alias that .get_predicted_args() normalizes to expectation (R/get_predicted_args.R:108-110), but it is absent from this dispatch list. Consequently, get_predicted(x, predict = "expected") skips this ordinal handler and the generic glmmTMB path returns a transformed scalar link prediction rather than the per-category probability table. Include expected here.

    ordinal_types <- c(
      "expectation",
      "expected",
      "response",
      "prediction",

R/utils_glmmtmb_ordinal.R:138

• This ordinal path drops the ... arguments before calling predict(), even though get_predicted.glmmTMB() forwards those arguments for other families and the public API documents them as prediction-method arguments. Options such as na.action or other supported predict.glmmTMB() controls are therefore silently ignored for ordinal models; preserve the remaining dots when constructing this call while overriding the arguments managed by insight.

  rez <- stats::predict(
    x,
    newdata = my_args$data,
    type = "probs",
    re.form = my_args$re.form,
    allow.new.levels = my_args$allow_new_levels,
    se.fit = !classification

• Files reviewed: 10/10 changed files
• Comments generated: 2
• Review effort level: Lite

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.

[strengejacke]

Great, thanks a lot! Looks good to me!

[strengejacke]

Will run a final local check, and then merge