Add G-computation correction for emmeans, contrasts argument by apanagos-lilly · Pull Request #567 · openpharma/mmrm

apanagos-lilly · 2026-05-21T13:11:04Z

Adds emmeans_gcomp_vars to mmrm_control(). When set, emmeans() returns the average treatment effect with standard errors that account for covariate variability across subjects. Default is NULL, when NULL, mmrm runs identically as before. Related to #560, though the counterfactual path does not support variance correction and has 50x runtime overhead only addressable through emmeans changes.

mmrm() and fit_mmrm() now accept a contrasts argument for specifying contrast matrices or functions, matching lm(). When a contrast matrix includes levels not present in the fitting data, those levels are preserved and marked as aliased, enabling prediction on new data

…oint estimates.

…o emmmeans_gcomp_subject_data

danielinteractive

Nice start, thanks @apanagos-lilly ! Please see comments below.

For next time, let's please try to have separate PRs for the separate enhancements as much as possible (here it would be 3 as far as I can see). That usually speeds up reviews, resolution and merges.

One additional high level comment:
I wonder if there is a nice way we can print to the user when they use emmeans on an mmrm with g-computation arguments that average treatment effect estimates are calculated. To basically highlight this as message, or hook into the show/print of emmeans results or similar, to make it very visible for the user.

danielinteractive · 2026-05-22T01:05:24Z

    person("Daniel D.", "Sjoberg", , "sjobergd@gene.com", role = "aut",
           comment = c(ORCID = "0000-0003-0862-2018")),
    person("Nikolas Ivan", "Krieger", , "nikolas.krieger@experis.com", role = "aut",
           comment = c(ORCID = "0000-0002-4581-3545")),


please add yourself as author or contributor here

danielinteractive · 2026-05-22T01:07:27Z

+      visit_var = visit_var,
+      subject_var = subject_var,
+      subject_groups = subject_groups,
+      cov_type = "us"


should we still check that additional arguments are ignored?

Reverted with emp_start removal

danielinteractive · 2026-05-22T01:08:01Z

+    subject_groups = fit$tmb_data$subject_groups,
+    cov_type = fit$formula_parts$cov_type
+  )
+  expect_length(result, 2L)


can we please keep comparing to the numbers as before? I want to see that this stays stable

Reverted with emp_start removal

danielinteractive · 2026-05-22T01:09:20Z

+  )
+})
+
+# emp_start per covariance type ----


this should be moved upwards such that all emp_start tests are together in one section of this file

Reverted with emp_start removal

danielinteractive · 2026-05-22T01:43:48Z

+  # Start with the model-based covariance.
+  V <- component(object, "beta_vcov")
+
+  if (!is.null(object$emmeans_gcomp_vars)) {


can we move this into helper function and apply here? otherwise this thing gets too long

Yes, placed in h_emm_basis_gcomp()

danielinteractive · 2026-05-22T01:44:07Z

+    aliased_pos <- which(is.na(beta_hat))
+
+    # Apply G-computation correction if specified.
+    visit_var <- object$formula_parts$visit_var


will come back for review once algorithm vignette is there

vignette has been added, thanks

danielinteractive · 2026-05-22T01:45:49Z


+  contrasts_preserve_vars <- NULL
+  if (!is.null(contrasts)) {
+    formula_factors <- intersect(


ok we need some explanations here. either in comments or in the function documentation. please also refer to how lm() is doing this, is it the same or different and why

Added comments above here which explains how lm() is doing this and how that's insufficient and how contrasts fixes it

danielinteractive · 2026-05-22T01:47:33Z

+#' `emp_start` supports all non-spatial covariance types.
+#' It uses linear regression to first obtain the coefficients
+#' and uses the residuals to obtain the empirical
+#' variance-covariance matrix. This is then decomposed into


can you please include more algorithm details here? as this is the only place where this is documented

Reverted with emp_start removal

danielinteractive · 2026-05-22T01:49:55Z

+# mmrm contrasts argument ----
+
+test_that("mmrm accepts contrasts argument with contrast function", {
+  fit <- mmrm(


can we also check that the result is as expected? I mean the coefficients for ARMCD

check added using the coefficient for ARMCD

Co-authored-by: Daniel Sabanes Bove <danielinteractive@users.noreply.github.com>

(also update some stray Rd's)

…message specifying the g-compute and emmeans occurrence

… to `suppressMessage`

apanagos-lilly · 2026-05-29T19:08:23Z

@danielinteractive, thanks for all your great feedback. First, yes I agree, keeping these separate would have been better. We reverted out the pr which added the emp_start additions so now it is only g-computation and the contrast. The vignette for g-compute has been added and a message() too when it is called.

danielinteractive · 2026-06-01T03:54:43Z

Thanks @apanagos-lilly can you please check the open comments above, if they are no longer applicable please resolve them. Please let me know once all comments resolved and ready for my second review

…niel

apanagos-lilly · 2026-06-02T19:39:13Z

One additional high level comment: I wonder if there is a nice way we can print to the user when they use emmeans on an mmrm with g-computation arguments that average treatment effect estimates are calculated. To basically highlight this as message, or hook into the show/print of emmeans results or similar, to make it very visible for the user.

Added a message() in emm_basis.mmrm()

apanagos-lilly · 2026-06-02T19:40:08Z

@danielinteractive, I believe all comments have been tended to. Please let me know if I missed any. Thanks!

jmiahjones-lilly and others added 8 commits April 17, 2026 15:41

Expanded emp_start to other covariance structures

03e1ab5

Add g-computation, supporting g-computation corrections for SEs and p…

fbd2bcc

…oint estimates.

fix test bug and add test cov

823734a

Added and tested 'contrast' argument

c10904f

Merge branch '562_emp_start'

df99ae7

[skip actions] Bump version to 0.3.17.9004

f5dba54

move g-computation to mmrm_control, change gcomp_subject_data input t…

6b4dd91

…o emmmeans_gcomp_subject_data

[skip actions] Bump version to 0.3.17.9005

9904116

apanagos-lilly requested a review from danielinteractive as a code owner May 21, 2026 13:11

danielinteractive and others added 2 commits May 22, 2026 09:00

trigger checks

6176538

[skip actions] Bump version to 0.3.17.9006

ace6d7c

danielinteractive reviewed May 22, 2026

View reviewed changes

jmiahjones-lilly and others added 18 commits May 26, 2026 14:50

Added vignette for model robust estimator

037ab7f

[skip actions] Bump version to 0.3.17.9007

1d314aa

Update NEWS.md

3769ebc

Co-authored-by: Daniel Sabanes Bove <danielinteractive@users.noreply.github.com>

[skip actions] Bump version to 0.3.17.9008

23c5495

Update R/fit.R

995e629

Co-authored-by: Daniel Sabanes Bove <danielinteractive@users.noreply.github.com>

[skip actions] Bump version to 0.3.17.9009

8e22220

Update documentation for contrasts and revert emp_start changes

4b049be

[skip actions] Bump version to 0.3.17.9010

d60f209

fix bug in vingette and tmb.R

58e0887

[skip actions] Bump version to 0.3.17.9011

48ddf7f

incorporate pr 567 feedback into gcomp

50e6835

remove validate, add comments explaining tmb_data drop

0bb6115

(also update some stray Rd's)

Move in-line to helper (h_emm_basis_gcomp()) and add function call …

9008f70

…message specifying the g-compute and emmeans occurrence

[skip actions] Bump version to 0.3.17.9012

104ca95

add hard coded ref numbers from ind. validation, flip expect_silent…

bbbdc82

… to `suppressMessage`

[skip actions] Bump version to 0.3.17.9013

4bc125b

update authors

9daa9cd

[skip actions] Bump version to 0.3.17.9014

3f72cbf

apanagos-lilly and others added 2 commits May 29, 2026 15:48

fix test bug

cbb9124

[skip actions] Bump version to 0.3.17.9015

1f34958

apanagos-lilly and others added 2 commits June 2, 2026 15:31

additional contrasts comments and test update per pr feedback from Da…

198ed70

…niel

[skip actions] Bump version to 0.3.17.9016

1f28fb1

apanagos-lilly changed the title ~~Add G-computation correction for emmeans, contrasts argument, and expanded emp_start~~ Add G-computation correction for emmeans, contrasts argument Jun 2, 2026

Uh oh!

Conversation

apanagos-lilly commented May 21, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

danielinteractive left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

apanagos-lilly commented May 29, 2026

Uh oh!

danielinteractive commented Jun 1, 2026

Uh oh!

apanagos-lilly commented Jun 2, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

apanagos-lilly commented Jun 2, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

apanagos-lilly commented May 21, 2026 •

edited

Loading

apanagos-lilly commented Jun 2, 2026 •

edited

Loading