alpaca

Info

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.

News

alpaca v0.3.4 (Release Date: 2022-08-10)

Changes:

• Added vcov.APEs() generic to extract the covariance matrix after getAPEs().
• Improved the finite sample performance of bias corrections for the average partial effects in case of perfectly classified observations.
• Bias corrections for the average partial effects, i.e. getAPEs() after biasCorr(), do not require an offset algorithm anymore.
• The default option 'n.pop' in getAPEs() has been changed. Now the estimated covariance consists of the delta method part only, i.e. correction factor = 0.
• Improved the numerical stability of the bias corrections.
• biasCorr() now also supports one-way fixed effects models.
• Added bias corrections for 'cloglog' and 'cauchit'.
• 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.
• Improved the numerical stability of the step-halving in feglm().
• Changed the projection in the MAP algorithm.
• The default option 'center.tol' in feglmControl() has been lowered to better handle fitting problems that are not well-behaved.
• Added optional 'weights' argument to feglm() and feglm.nb().
• Updated documentation.

alpaca v0.3.3 (Release Date: 2020-10-30)

Changes:

• Stopping condition of 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.
• Step-halving in feglmFit() and feglmOffset() is now similar to glm.fit2().
• Fixed an error in the covariance (influence function) of getAPEs().
• Updated some references in the documentation and vignette.
• Fixed some typos in the documentation and vignette.

alpaca v0.3.2 (Release Date: 2020-01-12)

Changes:

• Added option 'panel.structure' to 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.
• Added option 'sampling.fe' to 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().
• Added an example of the three-way bias correction suggested by Hinz, Stammann, and Wanner (2019) to the vignette.
• The control parameter 'trace' now also returns the current parameter values as well as the residual deviance.
• Fixed an error in getAPEs() related to the estimation of the covariance.
• Fixed a bug in the internal function that is used to estimate spectral densities.

alpaca v0.3.1 (Release Date: 2019-05-24)

Changes:

• All routines now use 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.
• Fixed a bug in feglm() that prevented to use I() for the dependent variable.
• Fixed an error in getAPEs() related to the covariance.
• The last line of print.summary.feglm() now ends with a line break (#6 @zauster).
• The internal function feglmFit() now correctly sets 'conv' if the algorithm does not converge (#5 @zauster).

alpaca v0.3 (Release Date: 2019-05-14)

Changes:

• Added feglm.nb() for negative binomial models.
• Added post-estimation routine biasCorr() for analytical bias-corrections (currently restricted to logit and probit models with two-way error component).
• Added post-estimation routine 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.
• Changed stopping criteria to the one used by glm().
• Vignettes revised.

alpaca v0.2 (Release Date: 2018-07-23)

ATTENTION: Syntax changed slightly. Have a look at the vignettes or help files.

Changes:

• various improvements (glm architecture, clustered standard errors, speed improvements).
• Syntax now more similiar to glm().

alpaca v0.1.3 (Release Date: 2018-03-08)

Changes:

• added option "cloglog" to argument family.
• added checks and routines to ensure that the model response is correctly encoded.

alpaca v0.1.2 (Release Date: 2018-03-04)

Changes:

• factor() should now work as intended.

alpaca v0.1.1 (Release Date: 2018-01-21)

Changes:

• added option "probit" to argument family.
• some performance tweaks.
• extract names of the fixed effects and getFEs() returns a named vector.
• adjusted computation of starting values.
• computation of the update step (irls) made numerically more stable.

Bugfix:

• construction of the regressor matrix such that factor variables are correctly dummy encoded.
• dropping perfectly classified observations for binomial models should now work as intended (thanks to jmboehm@github).