Support ordinal (argument mode for emmeans backend

[strengejacke]

@jmgirard is there something I am missing, or is this the expected output? For polr::MASS, the result is the same. Not sure I'm missing a combination of mode and type arguments, or how are ordinal models handled?

data(housing, package = "MASS")
m <- glmmTMB::glmmTMB(Sat ~ Infl + Type + Cont, data = housing, family = glmmTMB::ordinal())

emmeans::emmeans(m, "Type", mode = "prob")
#>  Type       prob SE  df asymp.LCL asymp.UCL
#>  Tower     0.333  0 Inf     0.333     0.333
#>  Apartment 0.333  0 Inf     0.333     0.333
#>  Atrium    0.333  0 Inf     0.333     0.333
#>  Terrace   0.333  0 Inf     0.333     0.333
#> 
#> Results are averaged over the levels of: Sat, Infl, Cont 
#> Confidence level used: 0.95

^{Created on 2026-09-30 with reprex v2.1.1}

[jmgirard]

The 0.333 values are expected emmeans behavior. With mode = "prob", emmeans adds the response (Sat) to the reference grid as a pseudo-factor, and emmeans(m, "Type", mode = "prob") averages over it. The three category probabilities sum to 1, so each average is 1/3. emmeans(m, ~ Type | Sat, mode = "prob") gives the per-category probabilities, and these match estimate_means(m, "Type") with the marginaleffects backend (difference below 1e-15 for polr and clm). The cum.prob, exc.prob, and linear.predictor modes add a cut pseudo-factor in the same way, so get_emmeans() needs to add that pseudo-factor to specs. Separately, on main the emmeans backend returns latent-scale values labeled "Probability" for polr, because emmeans ignores type = "response" for these models. I will open a PR into this branch with a fix.