In izahn/smargins: Simulated Average Marginal Effects

knitr::opts_knit$set(
        stop_on_error = 2L
)

Specify x variable names or values?

library(smargins)

Given the following model:

m <- lm(mpg ~ hp + wt + gear, data = mtcars)

suppose that we wish to examine the effect of hp by calculating expected values at several levels of hp while averaging over the observed values of wt and gear.

Which do you like better?

smargins_v1(m, hp = quantile(hp))

or

smargins_v2(m, "hp")

?

Now suppose that we want to choose the values of hp ourselves:

smargins_v1(m, hp = seq(50, 350, by = 50))

or

smargins_v2(m, "hp", xvalues = list(hp = seq(50, 350, by = 50)))

Now which do you prefer?

smargins_v1 is more verbose for the simple case, but less so when the user does wish to specify values. Which one should we pick? Or is there some combination or alternative that would be even better?

Interface for comparisons

Currently there is a scompare function that calculates pairwise differences for each combination of values on a particular predictor variable. For example, we can compare the differences between expected values for each level of gear as follows:

summary(scompare(smargins(m, gear = sort(unique(gear))), var="gear"))

It would be nice to offer some flexibility here so that comparisons other than pairwise can be made. There are a few different interfaces we might choose from.

Some options include:

• matrix based, like car::lht, and https://github.com/goshevs/effcomp
• character representation, like car::lht
• function based, like emmeans::contrast
• function based, but function does computation directly

A few examples follow below.

Matrix interface for comparisons

summary(scompare_v1(smargins(m,
                             gear = sort(unique(gear))),
                    var = "gear",
                    contrast = c(-1, 0, 1)))

summary(scompare_v1(smargins(m,
                             gear = sort(unique(gear))),
                    var = "gear",
                    contrast = c(-0.5, -0.5, 1.0)))

Simple function based interface to comparisons

summary(scompare_v2(smargins(m,
                             gear = sort(unique(gear))),
                    var = "gear",
                    contrast.fun = function(a, b, c) a - b + 0*c))

Naming is a problem...

summary(scompare_v2b(smargins(m,
                             gear = sort(unique(gear))),
                    var = "gear",
                    contrast.funs = list("3 vs 4" = function(a, b, c) a - b + 0*c)))

summary(scompare_v2b(smargins(m,
                             gear = sort(unique(gear))),
                    var = "gear",
                    contrast.funs = list("3 and 4 vs 5" = function(a, b, c) .5*a + .5*b + - c)))

After looking at these examples I'm not sure we have a clear winner yet. Perhaps there is another way that I haven't thought of yet? If you think of one please let me know by opening an issue at https://github.com/izahn/smargins/issues

izahn/smargins documentation built on Sept. 11, 2019, 2:08 p.m.