lfe2fixest

Installation

Install the package directly from GitHub:

# install.packages("remotes")
remotes::install_github("grantmcdermott/lfe2fixest")

Motivation

The goal of lfe2fixest is to take R scripts that rely on the felm() function from the lfe package and convert them to their feols() equivalents from the fixest package.

Why would you want to do this?

Both lfe::felm() and fixest::feols() provide “fixed-effects” estimation routines for high-dimensional data. Both methods are also highly optimised. However, feols() is newer, tends to be significantly faster, and allows for a lot more functionality (e.g. a predict method). At the same time, the primary author of lfe has stopped developing the package. It has since been adopted, but is now essentially in pure maintenance mode.

The syntax between felm() and feols() is similar, albeit with one not necessarily providing a drop-in replacement for the other. For example, the following two lines of code are functionally equivalent versions of the same underlying model.¹

• felm(y ~ x1 + x2 | fe1 + fe2 | (x3 | x4 ~ z1 + z2) | cl1, data = dat)
• feols(y ~ x1 + x2 | fe1 + fe2 | c(x3, x4) ~ z1 + z2, cluster = ~cl1, data = dat)

The lfe2fixest package automates the translation between these kinds of models. It does its best to ignoring any non-relevant arguments and adjust for differing syntax options between felm() and feols(). The end goal is a converted R script that will produce exactly the same output, some minor caveats notwithstanding.

Example

A detailed example is provided in the introductory vignette. See vignette("lfe2fixest").

Caveats

While I’m confident that lfe2fixest will work out-of-the-box in most (9 out of 10?) situations, there are some minor caveats to bear in mind.

• feols() offers a variety of optimised features and syntax for things like varying slopes, multiple estimation, etc. that go beyond the standard R syntax (although the latter still works). lfe2fixest doesn’t try to exploit any of these specialised features. It’s more or less a literal translation of the felm() formula. The goal is to get you up and running with as little pain as possible, rather than eking every extra bit out of feols()’s already eye-watering performance.
• Similarly, because felm() and feols() do not share all of the same arguments, there are cases where the conversion can yield different results. An example would be in the case of multiway clustering where the felm() call does not specify the right cmethod option. (More here and here.)
• The conversion only handles (or attempts to handle) the actual model calls. No attempt is made to convert downstream objects or functions like regression table construction. Although, as demonstrated in the example, you should be okay if you use a modern table-generating package like modelsummary. But the same would not be true of stargazer, for example.
• I assume that users always provide a dataset in their model calls (i.e.  regressions with global variables are not supported).
• Similarly, the package only supports models that are explicitly written out in a script. If your models are constructed programatically (e.g. with Formula()) then the conversion probably won’t work.

¹ An IV regression with multiple endogenous variables and fixed effects, as well as clustered standard errors. But that’s besides the point.↩︎