[Feature request] add the ability to easily display hypothesis tests and subgroup means

[rferrali]

Thanks for writing this great package -- and the also great marginaleffects package. I often display OLS estimates of causal effects, in subgroups. So, my models often look like y ~ treatment + controls, and then y ~ treatment * subgroup_factor + controls.

I typically want to report (1) treatment + treatment:subgroup_factor_1, treatment + treatment:subgroup_factor_2, ... at the bottom of the table, as well as (2) the mean outcome of the control group, by subgroup. So mean(y[subgroup_factor == 1]), ...

For (1), marginaleffects::hypotheses is the go to to conduct the hypothesis test. I then format this as the main estimates; i.e., estimate^stars, and (SD) right below. However, displaying using the add_rows argument is quite tedious.

For (2), I usually either use add_lines or glance_custom. Typically, add_lines rather than glance_custom as the subgroups vary from table to table. So, setting the setting globally using glance_custom is a bit dangerous.

I wonder how we could do better? I don't think that this use case is general enough to be an additional argument of modelsummary. Perhaps using another function that could be piped after modelsummary?

Happy to write this as a separate package, but I would appreciate hearing thoughts before getting started.

[vincentarelbundock]

That's an interesting use-case.

My guess is that the easiest (or more "direct") implementation would be to write a function that accepts the same list of models, and then automatically builds the required input for the add_rows argument.

So you would just do this preprocessing, and then call modelsummary() as usual.

[rferrali]

That's what I'm currently doing. The problem with this approach is that it doesn't handle formatting gracefully, especially when displaying standard errors in parenthesis underneath (I typically use tex output, so that function basically wraps all numbers within \num{}, and adds parentheses; but then, it doesn't display well when using HTML output, e.g., during development). Of course, one could add parameters to the function to tweak the output, but in the end, all the parameters set in modelsummary (e.g., stars, number of significant digits, ...) would need to be re-entered. I guess the main challenge is to manage to inherit the modelsummary parameters without having the user re-type them.

On Wed, Nov 27, 2024 at 4:53 PM Vincent Arel-Bundock < @.***> wrote:

That's an interesting use-case.

My guess is that the easiest (or more "direct") implementation would be to write a function that accepts the same list of models, and then automatically builds the required input for the add_rows argument.

So you would just do this preprocessing, and then call modelsummary() as usual.

— Reply to this email directly, view it on GitHub https://github.com/vincentarelbundock/modelsummary/issues/838#issuecomment-2504228314, or unsubscribe https://github.com/notifications/unsubscribe-auth/AGL2RAIW33O7OLUKEJILNCT2CXTJNAVCNFSM6AAAAABSS7I2PWVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMZDKMBUGIZDQMZRGQ . You are receiving this because you authored the thread.Message ID: @.***>

-- Romain Ferrali Postdoctoral associate NYU Abu Dhabi UAE O: +971 50 675 4347

http://rferrali.net

[vincentarelbundock]

Hard to say without a Minimal Working Example (emphasis on the minimal since the description is very complex).