ordered

ordered is a parsnip extension to enable additional classification models for ordinal outcomes (e.g., “low”, “medium”, “high”). While there are several model/engine combinations in the parsnip package that can be used, this package adds:

• cumulative link (cumulative logit) ordinal regression via MASS::polr()
• cumulative link ordinal regression via ordinal::clm() (Christensen, 2023)
• generalized linear and generalized additive ordinal regression models of cumulative link, adjacent categories, continuation ratio, and stopping ratio families via VGAM::vglm() and VGAM::vgam() (Yee, 2015)
• regularized elastic net ordinal regression models of cumulative link, adjacent categories, continuation ratio, and stopping ratio families via ordinalNet::ordinalNet() (Wurm, Hanlon, and Rathouz, 2021)
• regularized cumulative probability ordinal regression models via rms::lrm() and rms::orm() (Harrell, 2015)
• regularized elastic net continuation ratio ordinal regression via glmnetcr::glmnetcr() (Archer and Williams, 2012)
• ordinal classification trees via rpartScore::rpartScore() (Galimberti, Soffritti, and Di Maso, 2012)
• latent variable ordinal forests via ordinalForest::ordfor() (Hornung, 2020)
• conditional probability ordered forests via orf::orf() (Lechner and Okasa, 2025)

Some features in other tidymodels packages are useful for ordinal outcomes:

• The partykit engines for parsnip::decision_tree() and parsnip::rand_forest(), provided by bonsai, use the ordinal nature of ordered factors to train the model, yielding different fits than for unordered factors.
• The yardstick package has yardstick::kap() for weighted and unweighted Kappa statistics (the former being of more interest). Also, yardstick::classification_cost() can utilize more complex cost structures and uses the class probabilities for estimation.

Installation

You can install the latest released version of ordered from CRAN via

install.packages("ordered")

and the development version of ordered from GitHub via

# install.packages("pak")
# pak::pak("corybrunson/ordered")
install.packages("ordered", repos = c('https://corybrunson.r-universe.dev', 'https://cloud.r-project.org'))

or

# install.packages("remotes")
# remotes::install_github("corybrunson/ordered")
install.packages("ordered", repos = c('https://corybrunson.r-universe.dev', 'https://cloud.r-project.org'))

Available models, engines, and prediction types

ordered provides new engines for several models, including all engines for ordinal_reg():

model	engine	class	prob	linear_pred
decision_tree	rpartScore	✔	✖	✖
gen_additive_mod	vgam	✔	✔	✔
ordinal_reg	polr	✔	✔	✖
ordinal_reg	clm	✔	✔	✔
ordinal_reg	lrm	✔	✔	✔
ordinal_reg	orm	✔	✔	✔
ordinal_reg	vglm	✔	✔	✔
ordinal_reg	ordinalNet	✔	✔	✔
ordinal_reg	glmnetcr	✔	✔	✖
rand_forest	ordinalForest	✔	✔	✖
rand_forest	orf	✔	✔	✖

Tunable model arguments

ordinal_reg recognizes several tunable arguments shared by at least two engines in ordered:

engine	ordinal_link	odds_link	threshold_structure	parallel_reg	penalty	mixture
polr	✔	✖	✖	✖	✖	✖
clm	✔	✖	✔	✔	✖	✖
lrm	✖	✖	✖	✖	✔	✖
orm	✔	✖	✖	✖	✔	✖
vglm	✔	✔	✔	✔	✖	✖
ordinalNet	✔	✔	✖	✔	✔	✔
glmnetcr	✖	✖	✖	✖	✔	✔

Example

Here is a simple example using computational chemistry data to predict the permeability of a molecule:

library(dplyr)
library(ordered)

data(caco, package = "QSARdata")

caco_dat <-
  inner_join(caco_Outcome, caco_Dragon, by = "Molecule") %>%
  as_tibble() %>%
  select(
    class = Class,
    mol_weight = QikProp_mol_MW,
    volume = QikProp_volume,
    ClogP
  )
caco_train <- caco_dat[-(1:10), ]
caco_test  <- caco_dat[ (1:10), ]

ord_rf_spec <- 
  # you should really use many more trees and score sets
  rand_forest(mtry = 2, trees = 100) %>%
  set_mode("classification") %>%
  set_engine("ordinalForest", nsets = 100)

set.seed(382)
ord_rf_fit <- ord_rf_spec %>% fit(class ~ ., data = caco_train)
augment(ord_rf_fit, new_data = caco_test)
#> # A tibble: 10 × 8
#>    .pred_class .pred_L .pred_M .pred_H class mol_weight volume  ClogP
#>    <ord>         <dbl>   <dbl>   <dbl> <ord>      <dbl>  <dbl>  <dbl>
#>  1 L            0.374    0.368  0.258  M           123.   445.  0.799
#>  2 M            0.289    0.512  0.199  L           290.   856.  0.534
#>  3 M            0.170    0.809  0.0212 M           519.  1576.  1.02 
#>  4 M            0.209    0.740  0.0503 M           533.  1606.  1.58 
#>  5 M            0.126    0.762  0.113  M           505.  1517.  1.71 
#>  6 M            0.0657   0.917  0.0176 M           519.  1547.  2.27 
#>  7 M            0.213    0.747  0.0401 M           517.  1600.  1.78 
#>  8 M            0.111    0.859  0.0306 M           531.  1631.  2.34 
#>  9 M            0.0403   0.932  0.0277 M           517.  1572.  2.81 
#> 10 L            0.569    0.428  0.003  L           588.  1799. -1.85

Code of Conduct

Please note that the ordered project is released with a Contributor Code of Conduct. By contributing to this project, you agree to abide by its terms.

To voice support for or volunteer to contribute additional engines, please comment on this issue.

Acknowledgments

Jason Cory Brunson was supported in part by NIH/NIDCD R01DC022733.