glmhdfe – Package for R

The R package glmhdfe allows for the estimation of generalized linear models with high dimensional fixed effects. The package makes use of a convenient property of some combinations of error term distributions and link functions, where the fixed effects have — conditional on all other estimated parameters — an explicit solution.

Consider the following equation that we want to estimate:

The first order conditions for fixed effects can be simplified to

For certain distribution and link combinations this yields explicit solutions for the estimated coefficient for the fixed effects , given estimates for beta and the other deltas. Specifically, this is the case for the Gaussian distribution with identity and log link, and for the Poisson, Gamma and Inverse Gaussian distributions with log link. This makes it possible to update the fixed effects separately from the estimation of the coefficients on variables of interest in every iteration of the IRLS procedure used to estimate beta, dramatically increasing the speed of the estimation procedure.

For more detail on the inner workings see the technical note. A Stata implementation is coming soon.

Implementation in R

The R package glmhdfe implements this "trick" and utilizes the powers of the data.table package for a fast implementation. For smaller datasets or other error term distributions we recommend the feglm command in Amrei Stammann's alpaca package that also allows high-dimensional fixed effects in GLM estimations.

Installation

Install from Github via the remotes package:

remotes::install_github("julianhinz/R_glmhdfe")

Examples

The glmhdfe function has a similar syntax as the felm function from the lfe package and the feglm function in the alpaca package:

glmhdfe(trade ~ fta | iso_o_year + iso_d_year + iso_o_iso_d | iso_o + iso_d + year,
        family = poisson(link = "log"),
        data = data)

The first part of the formula is specified as usual. The second part of the formula specifies the fixed effects dimensions, the third part, which is optional, the clustering of the standard errors.

Options

There are numerous options to tweak the estimation procedure:

• formula describes dependent variable, right-hand side variables of interest, sets of fixed effects and clustering of standard errors, e.g. as y ~ x | fe1 + fe2 | cluster1 + cluster2
• data specifies the data.table or data.frame with data used in the regression
• family specifies the estimator used, currently limited to gaussian(link = "identity"), gaussian(link = "log"), poisson(link = "log"), Gamma(link = "log")
• beta allows to include a vector of starting values, although, interestingly, this does not tend to speed up the estimation
• tolerance specifies the minimum change in the deviance at which the iteration breaks
• max_iterations specifies the maximum number of iterations
• accelerate specifies whether to use an acceleration algorithm, still quite buggy
• accelerate_iterations specifies the number of iterations before starting acceleration algorithm
• accelerate_aux_vector specifies whether to include the estimated fixed effects vectors in IRLS, which, interestingly, increases convergence speed
• compute_vcov asks whether to compute the variance-covariance matrix. It can also be computed ex-post when data from estimation is provided
• demean_variables if you don't want to compute the variance-covariance matrix right away, do you still want to demean variables to be used in estimation of variance-covariance matrix?
• demean_iterations specifies the number of iterations for the demeaning
• demean_tolerance specifies the minimum change in the diagonal of the Hessian at which the demeaning iteration breaks
• include_fe asks whether the estimated fixed effects should be returned
• include_data asks whether the data used in the estimation should be returned, which may be useful if the variance-covariance matrix will be computed ex-post
• include_data_vcov return data used in variance-covariance matrix estimation?
• skip_checks specifies, whether certain data integrity checks should be skipped before starting the procedure. Current option to skip are the detection of separation issues ("separation"), multicollinearity ("multicollinearity"), or missing data ("complete_cases")

• trace asks whether to show some information during the estimation
• verbose asks whether to show a bit more information during estimation for the impatient

Other functions

There is the usual battery of generic functions, like coef, summary, etc. Furthermore, if for some reason you want to (re-)estimate the variance-covariance matrix afterwards, or change the level of clustering, you can do so with the compute_vcov command:

compute_vcov(data, call, info)

You need to specify the data (best in the form of a glmhdfe_data object), call (for information on clustering and variable of interest), and info (for information on degrees of freedom, etc.).

Roadmap

• try eval on variables that are not part of data, e.g. for something like y ~ log(x)
• Inverse Gaussian with log link
• tests using testthat
• parallelization in Rcpp with omp
• detect collinearity with lmhdfe or other data.table routine, not felm because this slows it down
• run first IRLS with transformed lhs, otherwise beta guess only works for gaussian
• faster checks for separation issues, implement procedure recommended by Correia et al. (2019)

Bugs?

• This package is still in its early stages. Let us know if you find bugs or have suggestions for changes by opening an issue or e-mail to mail@julianhinz.com.