Switch to avg slope reporting in modelsummary vignette

[grantmcdermott]

Minor issue, but I think the group contrast examples starting here are potentially confusing, since the tables are simply reporting the number of observations in each group rather than the slopes coefficients, which are almost certainly going to be of more interest for tabling output.

Maybe switch to avg_comparisons?

library(marginaleffects)
library(modelsummary)

dat <- mtcars
dat$gear <- as.factor(dat$gear)
mod <- glm(vs ~ gear + mpg, data = dat, family = binomial)

cmp <- avg_comparisons(mod)
modelsummary(cmp, shape = term + contrast ~ model) 

gear	mean(4) - mean(3)	0.037
		(0.137)
	mean(5) - mean(3)	-0.340
		(0.099)
mpg	mean(+1)	0.061
		(0.013)
Num.Obs.		32
AIC		26.2
BIC		32.1
Log.Lik.		-9.101
F		2.389
RMSE		0.31

^{Created on 2024-09-05 with reprex v2.1.1}

[vincentarelbundock]

I would call these group contrasts "counterfactual comparisons." For example, mean(4)-mean(3) is computed by:

1. Copying the entire 32 observations, and setting gear to 4
2. Copying the entire 32 observations, and setting gear to 3
3. Computing predictions in both sets
4. Taking the average difference between those predictions

So each contrast actually uses all 32 observations, which makes me feel like the "Num.Obs." row makes sense.

If you want to compute differences between subgroup average predictions, I'd do something like:

avg_predictions(mod, by = "gear", hypothesis = ~ sequential)

[grantmcdermott]

(Ugh, I just realised that I raised this in the wrong repo; should be in marginaleffects. Rushing too much this morning. Sorry!)

Cool, cool. Happy to close, but I will just say that reporting grouped avg_slopes/comparisons is one of the canonical use-cases of marginaleffects+modelsummaary IMO, since we are often interested in the modulating effect across groups. I think the docs do an excellent job demonstrating this elsewhere; it would be nice to see an example of it in the dedicated "Presentation tables" vignette. (I was motivated to raise this issue after helping a colleague that was stuck trying to do this.)