Feature request: emmeans::emmeans() like contrasts, emmeans() output into {gtsummary}

[yuryzablotski]

Dear Daniel,

It would be amazing if it would be possible to include capabilities of “emmeans”, particularly, calculating “contrasts” into {gtsummary}?

What I meant by the contrasts is a simple difference between pairwise levels of categorical variables. Because if use tbl_regression(), we see the difference (in ORs) between the reference level and other levels, but not between levels II and III in the Grade predictor, or not the difference between Stage levels pairwisely (I never understood, why it’s missing in the base r function, like summary(model) ):

m <- glm(response ~ stage + age, trial, family = binomial)

tbl_regression(m, exponentiate = T, pvalue_fun = ~style_pvalue(.x, digits = 3))

emmeans::emmeans(m, pairwise ~ stage, type = "response", adjust = "none") %>% summary(infer = T)

$emmeans stage prob SE df asymp.LCL asymp.UCL null z.ratio p.value T1 0.361 0.0686 Inf 0.240 0.503 0.5 -1.923 0.0545 T2 0.247 0.0604 Inf 0.148 0.383 0.5 -3.429 0.0006 T3 0.347 0.0768 Inf 0.215 0.508 0.5 -1.864 0.0623 T4 0.309 0.0716 Inf 0.188 0.464 0.5 -2.395 0.0166

Confidence level used: 0.95 Intervals are back-transformed from the logit scale Tests are performed on the logit scale

$contrasts contrast odds.ratio SE df asymp.LCL asymp.UCL null z.ratio p.value T1 / T2 1.718 0.756 Inf 0.725 4.07 1 1.231 0.2185 T1 / T3 1.061 0.478 Inf 0.439 2.57 1 0.132 0.8948 T1 / T4 1.260 0.565 Inf 0.523 3.03 1 0.516 0.6062 T2 / T3 0.618 0.288 Inf 0.247 1.54 1 -1.033 0.3017 T2 / T4 0.733 0.343 Inf 0.293 1.83 1 -0.664 0.5066 T3 / T4 1.187 0.567 Inf 0.466 3.03 1 0.359 0.7192

Confidence level used: 0.95 Intervals are back-transformed from the log odds ratio scale Tests are performed on the log odds ratio scale

Besides, the first part of emmeans (see below) is also useful, and it would be really grate to be able to beautifully gt-summaries the results of both, $emmeans and $contrasts

The final tbl_regression(m) should look like this:

[ddsjoberg]

Can you please update your post with an example of what you want the final table to look like?

[larmarange]

I need to explore it. An approach could be a custom tidier handled adequately by broom.helpers

Note: currently traveling for work

[yuryzablotski]

Well, I did not find a better and simpler way to get to contrasts as with emmeans. And since I realized I use emmens every time I model, I wander, why it (post-hoc pairwise comparisons) isn't possible with classic model-results functions, like summary()? Anyway, it would be soooo amazing, if we would be able to use emmeans-arguments, like "type = "response" ", or "adjust = "none" " in {gtsummary}

[larmarange]

I have drafted quickly a possible solution through a new function tidy_add_pairwise_contrasts() in broom.helpers.

tidy_plus_plus() gains accordingly additional arguments

library(gtsummary)
library(broom.helpers)
#> 
#> Attachement du package : 'broom.helpers'
#> Les objets suivants sont masqués depuis 'package:gtsummary':
#> 
#>     all_continuous, all_contrasts

mod <- glm(response ~ age + trt + stage + grade, data = trial, family = binomial)

mod %>%
  tidy_plus_plus(
    add_pairwise_contrasts = TRUE, 
    exponentiate = TRUE
  ) %>%
  knitr::kable()

term	variable	var_label	var_class	var_type	var_nlevels	contrasts	contrasts_type	reference_row	label	n_obs	n_event	estimate	std.error	statistic	p.value	conf.low	conf.high
age	age	Age	numeric	continuous	NA	NA	NA	NA	Age	183	58	1.0195893	0.0115076	1.6858397	0.0918267	0.9971774	1.043416
trtDrug A	trt	Chemotherapy Treatment	character	dichotomous	2	contr.treatment	treatment	TRUE	Drug A	89	27	1.0000000	NA	NA	NA	NA	NA
trtDrug B	trt	Chemotherapy Treatment	character	dichotomous	2	contr.treatment	treatment	FALSE	Drug B	94	31	1.1500613	0.3232223	0.4325668	0.6653295	0.6106565	2.176274
Drug B / Drug A	trt	Chemotherapy Treatment	character	dichotomous	2	pairwise	pairwise	NA	Drug B / Drug A	NA	NA	1.1500613	0.3717255	0.4325668	0.6653295	0.6103707	2.166947
stageT1	stage	T Stage	factor	categorical	4	contr.treatment	treatment	TRUE	T1	50	18	1.0000000	NA	NA	NA	NA	NA
stageT2	stage	T Stage	factor	categorical	4	contr.treatment	treatment	FALSE	T2	51	13	0.5628967	0.4437259	-1.2950770	0.1952937	0.2318960	1.333426
stageT3	stage	T Stage	factor	categorical	4	contr.treatment	treatment	FALSE	T3	39	14	0.9049542	0.4573824	-0.2183534	0.8271538	0.3651204	2.213548
stageT4	stage	T Stage	factor	categorical	4	contr.treatment	treatment	FALSE	T4	43	13	0.7633870	0.4537622	-0.5950036	0.5518411	0.3093673	1.850377
T2 / T1	stage	T Stage	factor	categorical	4	pairwise	pairwise	NA	T2 / T1	NA	NA	0.5628967	0.2497718	-1.2950770	0.5660480	0.1800348	1.759952
T3 / T1	stage	T Stage	factor	categorical	4	pairwise	pairwise	NA	T3 / T1	NA	NA	0.9049542	0.4139101	-0.2183534	0.9963269	0.2794586	2.930459
T3 / T2	stage	T Stage	factor	categorical	4	pairwise	pairwise	NA	T3 / T2	NA	NA	1.6076737	0.7542474	1.0120081	0.7424042	0.4816755	5.365884
T4 / T1	stage	T Stage	factor	categorical	4	pairwise	pairwise	NA	T4 / T1	NA	NA	0.7633870	0.3463962	-0.5950036	0.9336213	0.2379440	2.449146
T4 / T2	stage	T Stage	factor	categorical	4	pairwise	pairwise	NA	T4 / T2	NA	NA	1.3561761	0.6346940	0.6509987	0.9152375	0.4075285	4.513093
T4 / T3	stage	T Stage	factor	categorical	4	pairwise	pairwise	NA	T4 / T3	NA	NA	0.8435643	0.4045959	-0.3546908	0.9847078	0.2460316	2.892314
gradeI	grade	Grade	factor	categorical	3	contr.treatment	treatment	TRUE	I	65	21	1.0000000	NA	NA	NA	NA	NA
gradeII	grade	Grade	factor	categorical	3	contr.treatment	treatment	FALSE	II	58	17	0.8395776	0.4027265	-0.4341815	0.6641566	0.3779389	1.845925
gradeIII	grade	Grade	factor	categorical	3	contr.treatment	treatment	FALSE	III	60	20	1.0384054	0.3892503	0.0968175	0.9228713	0.4824457	2.232749
II / I	grade	Grade	factor	categorical	3	pairwise	pairwise	NA	II / I	NA	NA	0.8395776	0.3381201	-0.4341815	0.9013229	0.3266955	2.157637
III / I	grade	Grade	factor	categorical	3	pairwise	pairwise	NA	III / I	NA	NA	1.0384054	0.4041996	0.0968175	0.9948455	0.4170288	2.585638
III / II	grade	Grade	factor	categorical	3	pairwise	pairwise	NA	III / II	NA	NA	1.2368188	0.5043448	0.5212242	0.8609693	0.4756077	3.216350

^{Created on 2022-11-12 with reprex v2.0.2}

[larmarange]

You can play with it here: https://github.com/larmarange/broom.helpers/pull/192

My feeling is that it could be displayed as usual by tbl_regression(). However, it would require additional arguments to be added to tbl_regression().

A more generic approach could be to have a way to pass additional arguments to tidy_plus_plus(). It could be simply by using ... approach. Otherwise, by providing a list with something like tidy_plus_plus_args

[yuryzablotski]

I can't get it to work. Unfortunately. It does not produce contrasts. Do I need a developer version of gtsummary or broom.helpers? Because with usual versions it does not produce contrasts? Something is missing. Do you have any suggeestion what it might be?

[larmarange]

So far, it is only implemented in a PR. So to install that version of broom.helpers, you need to use remotes::install_github("larmarange/broom.helpers#192")

Then, you will be able to test the function broom.helpers::tidy_plus_plus() with is used internally by tbl_regression(). So you could see how pairwise contrasts could be computed.

However at that stage, it would not be possible to test with tbl_regression(), as a change in gtsummary would also be required, to pass new arguments to tidy_plus_plus()

[larmarange]

As a proof of concept, I have prepared a PR in gtsummary allowing to pass additional arguments to tidy_plus_plus() cf. https://github.com/ddsjoberg/gtsummary/pull/1387

If you install also this PR with remotes::install_github("ddsjoberg/gtsummary#1387"), both changes are working together.

library(gtsummary)
#> #BlackLivesMatter
mod <- glm(response ~ age + trt + stage + grade, data = trial, family = binomial)
mod |> 
  tbl_regression(exponentiate = TRUE, add_pairwise_contrasts = TRUE) |> 
  as_kable()

Characteristic	OR	95% CI	p-value
Age	1.02	1.00, 1.04	0.092
Chemotherapy Treatment
Drug B / Drug A	1.15	0.61, 2.17	0.7
T Stage
T2 / T1	0.56	0.18, 1.76	0.6
T3 / T1	0.90	0.28, 2.93	>0.9
T3 / T2	1.61	0.48, 5.37	0.7
T4 / T1	0.76	0.24, 2.45	>0.9
T4 / T2	1.36	0.41, 4.51	>0.9
T4 / T3	0.84	0.25, 2.89	>0.9
Grade
II / I	0.84	0.33, 2.16	>0.9
III / I	1.04	0.42, 2.59	>0.9
III / II	1.24	0.48, 3.22	0.9

^{Created on 2022-11-19 with reprex v2.0.2}

[yuryzablotski]

Now it has worked! That's amazing! 🤩 Contrasts are working with gaussian family of glm also! I love all of gtsummary functions, but the possibility to get contrasts puts it on another level! Love it! Will use it daily! Thanks Joseph and Daniel!

Just for your information:

• it did not work with tbl_uvregression() and
• worked but with a warning (Results may be misleading) with interactions

Wishful thinking of my is - it would be really useful, if we could use the functionality of {emmeans} in {gtsummary} somehow, particularly, to determine which contrasts we'll use after an interaction:

emmeans::emmeans(mod, pairwise ~ stage | grade, type = "response", adjust = "none") %>% summary(infer = T)

I'll use it only without interactions so far, that is already cool enough!

[larmarange]

For interactions, the warning is coming from emmeans.

tbl_uvregression() is not covered but it could be complex to update it. I let @ddsjoberg comment on that. I do not know if it should be relevant/possible to pass additional args.

For your last point, you can pass additional args to emmeans:

library(gtsummary)
#> #BlackLivesMatter
library(broom.helpers)
#> 
#> Attachement du package : 'broom.helpers'
#> Les objets suivants sont masqués depuis 'package:gtsummary':
#> 
#>     all_continuous, all_contrasts

mod <- glm(
  response ~ stage + grade + trt,
  gtsummary::trial,
  family = binomial,
)

mod |> 
  tbl_regression(
    exponentiate = TRUE,
    add_pairwise_contrasts = TRUE,
    pairwise_variables = c("stage", "grade"),
    emmeans_args = list(by = "trt")
  ) |> 
  as_kable()

Characteristic	OR	95% CI	p-value
T Stage
T2 / T1 | Drug A	0.61	0.20, 1.86	0.7
T3 / T1 | Drug A	1.10	0.35, 3.45	>0.9
T3 / T2 | Drug A	1.82	0.56, 5.94	0.6
T4 / T1 | Drug A	0.81	0.27, 2.42	>0.9
T4 / T2 | Drug A	1.33	0.42, 4.18	>0.9
T4 / T3 | Drug A	0.73	0.23, 2.34	0.9
T2 / T1 | Drug B	0.61	0.20, 1.86	0.7
T3 / T1 | Drug B	1.10	0.35, 3.45	>0.9
T3 / T2 | Drug B	1.82	0.56, 5.94	0.6
T4 / T1 | Drug B	0.81	0.27, 2.42	>0.9
T4 / T2 | Drug B	1.33	0.42, 4.18	>0.9
T4 / T3 | Drug B	0.73	0.23, 2.34	0.9
Grade
II / I | Drug A	0.96	0.39, 2.37	>0.9
III / I | Drug A	1.13	0.47, 2.76	>0.9
III / II | Drug A	1.19	0.48, 2.95	0.9
II / I | Drug B	0.96	0.39, 2.37	>0.9
III / I | Drug B	1.13	0.47, 2.76	>0.9
III / II | Drug B	1.19	0.48, 2.95	0.9
Chemotherapy Treatment
Drug A	—	—
Drug B	1.24	0.67, 2.30	0.5

^{Created on 2022-11-25 with reprex v2.0.2}

However, I don't think we could cover all possibilities of emmeans.

[yuryzablotski]

Sure, I understand! And I am already very grateful for what you did so far!

The example above though me an error-message:

Error in dplyr::bind_rows(): ! Can't combine ..1$estimate and ..2$estimate <factor>. Run rlang::last_error() to see where the error occurred.

[larmarange]

You should update broom.helpers with the last version of the PR.

Joseph Larmarange

Le ven. 25 nov. 2022 à 16:11, yuryzablotski @.***> a écrit :

Sure, I understand! And I am already very grateful for what you did so far!

The example above though me an error-message:

Error in dplyr::bind_rows(): ! Can't combine ..1$estimate and ..2$estimate . Run rlang::last_error() to see where the error occurred.

— Reply to this email directly, view it on GitHub https://github.com/ddsjoberg/gtsummary/issues/1383#issuecomment-1327614723, or unsubscribe https://github.com/notifications/unsubscribe-auth/AAHL5I6L5LP27RNKSTTHVZDWKDJKXANCNFSM6AAAAAAR3RGMUU . You are receiving this because you commented.Message ID: @.***>