Our package offers a fast algorithm to estimate generalized linear models with high-dimensional fixed effects. The linear predictor of such models takes the following form: \[ \boldsymbol{\eta} = \mathbf{Z} \boldsymbol{\gamma} = \mathbf{D} \boldsymbol{\alpha} + \mathbf{X} \boldsymbol{\beta} = \sum_{j=1}^{k} \mathbf{D}_j \boldsymbol{\alpha}_{j} + \mathbf{X} \boldsymbol{\beta} \, , \] where the matrix \(\mathbf{D}\) arises from dummy encoding of \(k\) high-dimensional categorical variables and \(\mathbf{X}\) contains the variables of interest. We refer to \(\boldsymbol{\beta}\) and \(\boldsymbol{\alpha}\) as structural parameters and fixed effects. The latter are essentially nuisance parameters that are used to control for unobserved heterogeneity.

Brute force estimation of this kind of models is often restricted to computational limitations (either due to memory or time limitations). We tackle this problem by providing a fast and memory efficient algorithm suggested by Stammann (2018) based on the combination of the Frisch-Waugh-Lovell theorem and the method of alternating projections. We restrict ourselves to non-linear models because Gaure (2013) already offers a great package for linear models. Further, in the case of binary choice models with only one high-dimensional fixed effects we recommend using the package `bife`

.

In the following we utilize an example from labor economics to demonstrate the capabilities of `feglm()`

. More precisely, we use a balanced micro panel data set from the *Panel Study of Income Dynamics* to analyze the intertemporal labor force participation of 1,461 married women observed for nine years. A similar empirical illustration is used in Fernández-Val (2009).

Before we start, we briefly inspect the data set to get an idea about its structure and potential covariates.

```
## ID LFP KID1 KID2 KID3 INCH AGE TIME
## 1: 1 1 1 1 1 58807.81 26 1
## 2: 1 1 1 0 2 41741.87 27 2
## 3: 1 1 0 1 2 51320.73 28 3
## 4: 1 1 0 1 2 48958.58 29 4
## 5: 1 1 0 1 2 53634.62 30 5
## 6: 1 1 0 0 3 50983.13 31 6
```

`ID`

and `TIME`

are individual and time-specific identifiers, `LFP`

is an indicator equal to one if a woman is in labor force, `KID1`

- `KID3`

are the number of children in a certain age group, `INCH`

is the annual income of the husband, and `AGE`

is the age of a woman.

First, we use a specification similar to Fernández-Val (2009) and estimate a static model of women’s labor supply where we control for individual and time-specific unobserved heterogeneity (so called individual and time fixed effects).

```
library(alpaca)
stat <- feglm(LFP ~ KID1 + KID2 + KID3 + log(INCH) | ID + TIME, psid, binomial("probit"))
summary(stat)
```

```
## binomial - probit link
##
## LFP ~ KID1 + KID2 + KID3 + log(INCH) | ID + TIME
##
## Estimates:
## Estimate Std. error z value Pr(> |z|)
## KID1 -0.676905 0.056301 -12.023 < 2e-16 ***
## KID2 -0.344383 0.049896 -6.902 5.13e-12 ***
## KID3 -0.007039 0.035344 -0.199 0.842
## log(INCH) -0.234136 0.054403 -4.304 1.68e-05 ***
## ---
## Signif. codes: 0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1
##
## residual deviance= 6069.65,
## null deviance= 8152.05,
## n= 5976, l= [664, 9]
##
## ( 7173 observation(s) deleted due to perfect classification )
##
## Number of Fisher Scoring Iterations: 6
```

As `glm()`

, the summary statistic of the model provides detailed information about the coefficients and some information about the model fit (`residual deviance`

and `null deviance`

). Furthermore, we report statistics that are specific to fixed effects models. More precisely, we learn that only 5,976 observations out of 13,149 contribute to the idenfication of the structural parameters. This is indicated by the message that 7,173 observations are deleted due to perfect classification. With respect to binary choice models those are observations that are related to women who never change their labor force participation status during the nine years observed. Thus those women were either always employed or unemployed.[Note that in this specification (with individual and time fixed effects) also observations related to a specific time period where all women are either in labor force or not can be dropped. However this is very unlikely in practice.] Overall the estimation results are based on 664 women observed for nine years.

Because coefficients itself are not very meaningful, econometricians are usually interested in so called partial effects (also known as marginal or ceteris paribus effects). A commonly used statistic is the average partial effect. `alpaca`

offers a post-estimation routine to estimate average partial effects and their corresponding standard errors.[The routine is currently restricted to probit and logit models but will be extended in the future.]

```
## Estimates:
## Estimate Std. error z value Pr(> |z|)
## KID1 -0.0880159 0.0373733 -2.355 0.0185 *
## KID2 -0.0447791 0.0197079 -2.272 0.0231 *
## KID3 -0.0009152 0.0050072 -0.183 0.8550
## log(INCH) -0.0304440 0.0161221 -1.888 0.0590 .
## ---
## Signif. codes: 0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1
```

A widespread reason that prevents the use of non-linear fixed effects models in practice is the so-called incidental parameter bias problem (*IPP*) first mentioned by Neyman and Scott (1948). Fortunately, for classical panel data sets, like in this example, there already exist several asymptotic bias-corrections tackling the *IPP* (see Fernández-Val and Weidner (2018) for an overview). Our package provides a post-estimation routine that applies the analytical bias-correction derived by Fernández-Val and Weidner (2016).[See footnote 2.] Technical details on how the bias-correction is accelerated using the method of alternating projections can be found in Czarnowske and Stammann (2019).[In our article, we also apply several asymptotic bias-corrections to a very similar empirical example where we use German micro data from the *German Socio-Economic Panel*.]

```
## binomial - probit link
##
## LFP ~ KID1 + KID2 + KID3 + log(INCH) | ID + TIME
##
## Estimates:
## Estimate Std. error z value Pr(> |z|)
## KID1 -0.596292 0.056301 -10.591 < 2e-16 ***
## KID2 -0.303358 0.049896 -6.080 1.2e-09 ***
## KID3 -0.006111 0.035344 -0.173 0.862733
## log(INCH) -0.207068 0.054403 -3.806 0.000141 ***
## ---
## Signif. codes: 0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1
##
## residual deviance= 6069.65,
## null deviance= 8152.05,
## n= 5976, l= [664, 9]
##
## ( 7173 observation(s) deleted due to perfect classification )
##
## Number of Fisher Scoring Iterations: 6
```

```
## Estimates:
## Estimate Std. error z value Pr(> |z|)
## KID1 -0.096502 0.033379 -2.891 0.00384 **
## KID2 -0.049094 0.017778 -2.761 0.00575 **
## KID3 -0.000989 0.005041 -0.196 0.84448
## log(INCH) -0.033511 0.014861 -2.255 0.02414 *
## ---
## Signif. codes: 0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1
```

Whereas analytical bias-corrections for static models get more and more attention in applied work, it is not well known that they can also be used for dynamic models with fixed effects.

Before we can adjust our static to a dynamic specification, we first have to generate a lagged dependent variable.

Contrary to the bias-correction for the static models, we need to additionally provide a bandwidth parameter (`L`

) that is required for the estimation of spectral densities (see Hahn and Kuersteiner (2011)). Fernández-Val and Weidner (2016) suggest to do a sensitivity analysis and try different values for `L`

but not larger than four.

```
dyn <- feglm(LFP ~ LLFP + KID1 + KID2 + KID3 + log(INCH) | ID + TIME, psid, binomial("probit"))
dyn.bc <- biasCorr(dyn, L = 1L)
summary(dyn.bc)
```

```
## binomial - probit link
##
## LFP ~ LLFP + KID1 + KID2 + KID3 + log(INCH) | ID + TIME
##
## Estimates:
## Estimate Std. error z value Pr(> |z|)
## LLFP 1.01608 0.04695 21.643 < 2e-16 ***
## KID1 -0.45389 0.06787 -6.687 2.27e-11 ***
## KID2 -0.15737 0.06031 -2.610 0.00907 **
## KID3 0.01562 0.04331 0.361 0.71839
## log(INCH) -0.18834 0.06165 -3.055 0.00225 **
## ---
## Signif. codes: 0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1
##
## residual deviance= 4777.58,
## null deviance= 6549.14,
## n= 4792, l= [599, 8]
##
## ( 1461 observation(s) deleted due to missingness )
## ( 6896 observation(s) deleted due to perfect classification )
##
## Number of Fisher Scoring Iterations: 6
```

```
## Estimates:
## Estimate Std. error z value Pr(> |z|)
## LLFP 0.186312 0.070340 2.649 0.00808 **
## KID1 -0.072324 0.027220 -2.657 0.00788 **
## KID2 -0.025075 0.011242 -2.230 0.02572 *
## KID3 0.002489 0.004967 0.501 0.61632
## log(INCH) -0.030011 0.013689 -2.192 0.02835 *
## ---
## Signif. codes: 0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1
```

`alpaca`

is also compatible with `linearHypothesis()`

of the `car`

package and offers the possibility to compute robust and multi-way clustered standard errors. For an example, see vignette “Replicating an Empirical Example of International Trade”.

Czarnowske, Daniel, and Amrei Stammann. 2019. “Binary Choice Models with High-Dimensional Individual and Time Fixed Effects.” *ArXiv E-Prints*.

Fernández-Val, Iván. 2009. “Fixed Effects Estimation of Structural Parameters and Marginal Effects in Panel Probit Models.” *Journal of Econometrics* 150 (1): 71–85.

Fernández-Val, Iván, and Martin Weidner. 2016. “Individual and Time Effects in Nonlinear Panel Models with Large N, T.” *Journal of Econometrics* 192 (1): 291–312.

———. 2018. “Fixed Effects Estimation of Large-T Panel Data Models.” *Annual Review of Economics* 10 (1): 109–38.

Gaure, Simen. 2013. “lfe: Linear Group Fixed Effects.” *The R Journal* 5 (2): 104–17.

Hahn, Jinyong, and Guido Kuersteiner. 2011. “Bias Reduction for Dynamic Nonlinear Panel Models with Fixed Effects.” *Econometric Theory* 27 (6): 1152–91.

Neyman, Jerzy, and Elizabeth L. Scott. 1948. “Consistent Estimates Based on Partially Consistent Observations.” *Econometrica* 16 (1): 1–32.

Stammann, Amrei. 2018. “Fast and Feasible Estimation of Generalized Linear Models with High-Dimensional K-Way Fixed Effects.” *ArXiv E-Prints*.