Support new models

[vincentarelbundock]

IF THE MODEL YOU WOULD LIKE TO SUPPORT IS NOT LISTED BELOW, PLEASE OPEN A NEW ISSUE.

It is often very easy to add support for new models. If you would like to help us do it (thanks!!!), please read this vignette on marginaleffects extensions: https://vincentarelbundock.github.io/marginaleffects/articles/extensions.html

marginaleffects only supports modeling packages released on CRAN. If you would like to add support for a non-CRAN package, your code can be hosted in the user-submitted (b
ut unsupported) code library in the vignette above.

If you cannot contribute code to support a new model, and if that model is not listed below, please open a new issue with a working example using a small publicly available dataset.

Seems possible

• ordinal::clmm2: Support ordinal::clmm and ordinal::clmm2 #1003
• GLMMadaptive: Support model: GLMMadaptive #1411
• twopartm: Support for twopartm models #1494

Challenges

Existing issues (with some notes)

• bife: bife Support Not Working Properly #949
• earth (MARS): Support earth mars #1186
• flexsurv: Support: flexsurv #320
• gamm4: Support gamm4 #947
• logitr: See code in sandbox/ and Support: logitr #516
• mrmm: Support mmrm #963 [DRAFT] adding support for the {mmrm} package #1000
• MuMIn: Feature Request: Support for model class "averaging" from MuMIn::model.avg() in R #1154
• panelr: Support: panelr #210
• stats::nls: Support: nls #326
• systemfit: Support: systemfit #836

predict problem

• alpaca: Support alpaca package #786
• deming: Support: deming #1601
• aov: Support: aov and aovlist #1606
• bayesX
• censReg
• coxme: Support: coxme #645
• coxphw: Support timereg or coxphw packages #467
• coxrobust
• cplm
• endorse: Support: endorse #1144
• gee
• glmm
• glmnl
• ivprobit
• MASS::ridgelm
• mgcv::gamm: Support: mgcv::gamm #187
• ordinal::clm2: no prediction for each level
• ordinal::clmm: no predict method at all
• pglm
• plm::pgmm
• robust::glmRob: breaks with newdata
• timereg: Support timereg or coxphw packages #467
• geessbin: Support for geessbin #1529

vcov problem

• complmrob
• bfsl no vcov method at all
• lqmm: breaks with newdata Support: lqmm #1072

Categorical / Multinomial DV

• VGAM::vglm: not just multinomial
• VGAM::vgam

Not currently planned

Spatial regression

Spatial regression models typically rely on very different data structures, and are currently considered out of scope for marginaleffects

• Support for SSN2 models #1052
• spaMM::HLfit:
• Support for SSN2 models #1052 (comment)
• Support for Spatial Regression Models #1067
• sdmTMB: Support: sdmTMB #1136

Multi-row data structure

Some multinomial models require dataset to be presented with one observation-choice combination per row. The whole marginaleffects infrastrcture is designed to create contrasts 1 row at a time, so we cannot support these.

• mlogit
• MNP Support: mnp package #1382

No formula interface

• glinternet: Support: glinternet #934

Very different object structure

• metafor: Very different input data. Likely very difficult to support. Would need code submission from an expert contributor. Support: metafor (rma, rma.mv) #604
• lavaan: Vincent has no experience with this package. We could consider this, but it would require serious code contributions. Support: lavaan #499
• R-INLA: Very different and unlikely. Support: R-INLA #342
• Apollo for choice models. Seems extremely idiosyncratic, and does not appear to supply any of the standard methods such as vcov(), coef(), or predict(). Supporting this is likely to be a major challenge. Vincent is unlikely to do this himself without significant help (or more likely a quasi-complete PR) from a contributor. https://cran.r-project.org/web/packages/apollo/index.html
• cmprsk::crr: Seems very different. Vincent will not implement this unless someone is willing to help.

Already supported via tidymodels or mlr3

• glmnet: no formula interface and already supported through tidymodels and mlr3. Support: glmnet #923
• SuperLearner: Support: SuperLearner #834

Not on CRAN

• rblimp Support: rblimp #1621
• gamlss2: Support model: gamlss2 #1412: checked on 2025-06-17
• bigmlogit: checked on 2022-06-02
• mixor: checked on 2022-03-10
• mnlogit: checked on 2023-01-11
• tobit1: checked on 2023-01-11
• ZBZIMM: Support: ZBZIMM #541

Supported by other packages

• Source: emmeans
  https://cran.r-project.org/web/packages/emmeans/vignettes/models.html
• Source: easystats
  • https://easystats.github.io/insight/
• Source: ggeffects
  • bamlss, cgam, cgamm, gamm, gamm4, MixMod, nlmer, survreg, sjstats::svyglm.nb, vgam.

[strengejacke]

quantreg::rq

You may take a look at insight::get_varcov.rq(). Maybe you could even rely on insight::get_varcov(), since it automatically does some "fixing" (like fixing negativ matrices or rank deficiency).

[vincentarelbundock]

quantreg::rq

You may take a look at insight::get_varcov.rq(). Maybe you could even rely on insight::get_varcov(), since it automatically does some "fixing" (like fixing negativ matrices or rank deficiency).

That's fantastic! I didn't know about this and will definitely take a look, and almost certainly adopt it.

[vincentarelbundock]

quantreg::rq

You may take a look at insight::get_varcov.rq(). Maybe you could even rely on insight::get_varcov(), since it automatically does some "fixing" (like fixing negativ matrices or rank deficiency).

Worked like charm: add907e . Thanks @strengejacke !

[dustingilbreath]

If requests are open, I'd love for svyolr and svymultinom to be supported!

[vincentarelbundock]

If requests are open, I'd love for svyolr and svymultinom to be supported!

Ooops, sorry I missed this post @dustingilbreath !

I am interested in adding support for this. Unfortunately, the svyolr object does not have a predict() method, so it cannot be supported at the moment. I have contacted the package maintainer a while ago, and he seemed positive about the chances that this feature could eventually be published. See here: #157

I am not familiar with svymultinom model; it does not seem to be a part of the survey package. Where is it from? Do you know if they have a predict() method with a newdata argument?

[dustingilbreath]

No worries and thanks for getting back to me!

First, that's great on svyolr! It would make my life much much easier :)

Second, svymultinom isn't integrated into the survey package, but is rather in the svyrepmisc package:

https://rdrr.io/github/carlganz/svrepmisc/man/svymultinom.html

On whether it has a predict method with a newdata argument, it passes arguments to multinom, so maybe? I'm a terrible programmer if being honest.

[vincentarelbundock]

@dustingilbreath is svymultinom supplied by a package on CRAN or just Github?

FWIW, I only plan to support packages which have officially been released on CRAN. A lot of the Github packages are experimental, and it can be challenging to support packages with unstable user interface. Also, being released on CRAN means there is at least a minimal level of testing, and suggests that there may be enough of a user-base to be worth the work.

[dustingilbreath]

Unfortunately just Github. It may be my ignorance, but I haven't seen a multinomial model that's adjusted for complex survey design anywhere else, which is where my question stems from at the end of the day.

In any case, many thanks for getting back to me, the great package, and the hopefully coming svyolr addition!

[TysonStanley]

Love the new stuff that keeps coming out here. I'm curious if there are plans to support mice (multiple imputation) models in some way. I wouldn't mind thinking through this but wondered if that was already in the works.

[vincentarelbundock]

Love the new stuff that keeps coming out here. I'm curious if there are plans to support mice (multiple imputation) models in some way. I wouldn't mind thinking through this but wondered if that was already in the works.

No current plans, but I'm not sure how this would work. Shouldn't we estimate marginal effects in each of the imputed datasets, and then combine the results using Rubin's rules? I anticipate challenges in trying to make marginaleffects "support" mice objects. For instance, do they even have the methods we need, such as predict()?