An R-package for fitting glm’s with high-dimensional k-way fixed effects.
Provides a routine to partial out factors with many levels during the optimization of the log-likelihood function of the corresponding generalized linear model (glm). The package is based on the algorithm described in Stammann (2018) and is restricted to glm’s that are based on maximum likelihood estimation and non-linear. It also offers an efficient algorithm to recover estimates of the fixed effects in a post-estimation routine and includes robust and multi-way clustered standard errors. Further the package provides analytical bias corrections for binary choice models (logit and probit) derived by Fernandez-Val and Weidner (2016) and Hinz, Stammann, and Wanner (2020).
If you have any suggestions for improvements or questions, feel free to contact me.
The package is also available on CRAN.
Changes:
qr() by lm.wfit to fix memory
issueChanges:
vcov.APEs() generic to extract the covariance
matrix after getAPEs().getAPEs() after biasCorr(), do not require an
offset algorithm anymore.getAPEs() has been
changed. Now the estimated covariance consists of the delta method part
only, i.e. correction factor = 0.biasCorr() now also supports one-way fixed effects
models.feglm() and feglm.nb() do not return a
matrix of scores anymore. Instead they, optionally, return the centered
regressor matrix. The corresponding option in
feglmControl() is ‘keep.mx’. Default is TRUE.feglm().feglmControl() has
been lowered to better handle fitting problems that are not
well-behaved.feglm() and
feglm.nb().Changes:
feglm.nb() has been adjusted to
better match that of glm.nb().feglm.nb() now additionally returns ‘iter.outer’ and
‘conv.iter’ based on iterations of the outer loop. Previously ‘iter’ and
‘conv’ were overwritten.feglmFit() and
feglmOffset() is now similar to
glm.fit2().getAPEs().Changes:
biasCorr() and
getAPEs(). This option allows to choose between the two-way
bias correction suggested by Fernandez-Val and Weidner (2016) and the
bias corrections for network data suggested by Hinz, Stammann, and
Wanner (2020). Currently both corrections are restricted to probit and
logit models.getAPEs() to impose
simplifying assumptions when estimating the covariance matrix.feglm() now permits to expand functions with
poly() and bs() (#9 @tcovert).feglm() now uses an acceleration scheme suggested by
Correia, Guimaraes, and Zylkin (2019) that uses smarter starting values
for centerVariables().getAPEs() related to the estimation
of the covariance.Changes:
setDT() instead of
as.data.table() to avoid unnecessary copies (suggested in
#6 @zauster).feglm.nb() now returns ‘iter’ and ‘conv’ based on
iterations of the outer loop.feglm() that prevented to use
I() for the dependent variable.getAPEs() related to the
covariance.print.summary.feglm() now ends with a
line break (#6 @zauster).feglmFit() now correctly sets
‘conv’ if the algorithm does not converge (#5 @zauster).Changes:
feglm.nb() for negative binomial models.biasCorr() for analytical
bias-corrections (currently restricted to logit and probit models with
two-way error component).getAPEs() to estimate
average partial effects and the corresponding standard errors (currently
restricted to logit and probit models with two-way error
component).getFEs() now returns a list of named vectors. Each
vector refers to one fixed effects category.glm().ATTENTION: Syntax changed slightly. Have a look at the vignettes or help files.
Changes:
glm().Changes:
family.Changes:
factor() should now work as intended.Changes:
"probit" to argument
family.getFEs() returns
a named vector.Bugfix: