Fixest vcov catches

[grantmcdermott]

I know you (we!) largely came out against hardcoding specific model classes in #280. But been I've running into this whilst teaching, so thought I might as well have a go at it. I won't be offended if you decide not to merge given the aforementioned stance, but I do think the current behaviour is a little sub-optimal given the popularity of fixest.

(Note: If you're game, I can probably do something similar for lfe::felm, which also requires a couple of control flow catches. estimatr is harder b/c --- unlike lfe::felm --- it doesn't include the iid vcov errors as part of the return object).

cc @elbersb

Examples:.

devtools::load_all('~/Documents/Projects/modelsummary/')
#> ℹ Loading modelsummary
library(fixest)

mod_lm =         lm(hp ~ mpg + drat, mtcars)
mod_feols =   feols(hp ~ mpg + drat, mtcars, vcov = ~vs)

mods = list('lm' = mod_lm, 'feols' = mod_feols)

msummary(mods, gof_omit = 'R2|IC|Log|F', output = 'markdown')

	lm	feols
(Intercept)	278.515	278.515
	(55.415)	(20.573)
mpg	-9.985	-9.985
	(1.792)	(4.110)
drat	19.126	19.126
	(20.198)	(32.644)
Num.Obs.	32	32
Std.Errors		Clustered (vs)

msummary(mods, vcov = 'iid', gof_omit = 'R2|IC|Log|F', output = 'markdown') 

	lm	feols
(Intercept)	278.515	278.515
	(55.415)	(55.415)
mpg	-9.985	-9.985
	(1.792)	(1.792)
drat	19.126	19.126
	(20.198)	(20.198)
Num.Obs.	32	32
Std.Errors	IID	IID

## Aside: this "doesn't" work...
msummary(mods, vcov = c('iid', NULL), gof_omit = 'R2|IC|Log|F', output = 'markdown')

	lm	feols
(Intercept)	278.515	278.515
	(55.415)	(55.415)
mpg	-9.985	-9.985
	(1.792)	(1.792)
drat	19.126	19.126
	(20.198)	(20.198)
Num.Obs.	32	32
Std.Errors	IID	IID

## ... but only b/c the vcov list drops the NULL element (as it should)
c('iid', NULL)
#> [1] "iid"

msummary(mods, vcov = 'robust', gof_omit = 'R2|IC|Log|F', output = 'markdown') 

	lm	feols
(Intercept)	278.515	278.515
	(46.992)	(46.992)
mpg	-9.985	-9.985
	(2.471)	(2.471)
drat	19.126	19.126
	(23.133)	(23.133)
Num.Obs.	32	32
Std.Errors	Robust	Robust

msummary(mods, vcov = 'HC1', gof_omit = 'R2|IC|Log|F', output = 'markdown')

	lm	feols
(Intercept)	278.515	278.515
	(42.037)	(42.037)
mpg	-9.985	-9.985
	(2.209)	(2.209)
drat	19.126	19.126
	(20.572)	(20.572)
Num.Obs.	32	32
Std.Errors	HC1	HC1

msummary(mods, vcov = ~ vs, gof_omit = 'R2|IC|Log|F', output = 'markdown')

	lm	feols
(Intercept)	278.515	278.515
	(20.573)	(20.573)
mpg	-9.985	-9.985
	(4.110)	(4.110)
drat	19.126	19.126
	(32.644)	(32.644)
Num.Obs.	32	32
Std.Errors	C: vs	C: vs

msummary(mods, vcov = ~ cyl, gof_omit = 'R2|IC|Log|F', output = 'markdown')

	lm	feols
(Intercept)	278.515	278.515
	(40.571)	(40.571)
mpg	-9.985	-9.985
	(3.571)	(3.571)
drat	19.126	19.126
	(34.092)	(34.092)
Num.Obs.	32	32
Std.Errors	C: cyl	C: cyl

^{Created on 2021-08-14 by the reprex package (v2.0.0)}

[vincentarelbundock]

My guess is the test failures are due to this line, because class(model) is a vector with two elements for these classes. So I think you need class(model)[1] %in% ... or maybe inherits(model, "fixest") || inherits(model, "fixest_multi")

[vincentarelbundock]

This looks fantastic, thanks! I'd be happy to merge when tests pass. I'm pretty sure I identified the culprit in my comment above, and it should be a very easy fix.

[grantmcdermott]

I updated with your hotfix and also added some better checking + flow, depending on the fixest version. Looks like everything is passing except for an annoying warning about an undeclared dependency (Matrix?) for lfe, which I added to the "hardcoded" test suite. Happy to remove for now if that's easiest.

[vincentarelbundock]

Fantastic!