Request: Estimate Probit models and ability to change default standard errors.

[dlindzee]

The fixest package is great and has improved my workflow and saved me a lot of computation time. But I was wondering if there is any way to estimate a probit binary response model? At the moment I am using the R package bife. But it would be nice to be able to do this using fixest, as I can do with logit models.

Furthermore, I was wondering if it was possible to change the default standard errors when estimating a model. For example

set.seed(12)
Bond = runif(100, 0, 0.55)
Day = runif(100, 0, 1.45)
Intervention = round(Day)
Type = round(Bond)
Agency = round(Bond + Day)

Data <- data.table(Agency, Bond, Day, Intervention, Type)

feols(Agency ~ Bond + Day | Intervention + Type , Data)

Gives the following output

OLS estimation, Dep. Var.: Agency
Observations: 100 
Fixed-effects: Intervention: 2,  Type: 2
Standard-errors: Clustered (Intervention) 
     Estimate Std. Error t value Pr(>|t|)    
Bond 1.339600   0.324546  4.1276 0.000079 ***
Day  0.956392   0.269115  3.5538 0.000593 ***
---
Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
Log-likelihood: -13.21   Adj. R2: 0.65434 
                       R2-Within: 0.52428 

Where the standard errors are clustered at the intervention level . However, with this is not what I necessarily want. In large datasets it also would seem to have a large effect on the computation time relative to the usual standard errors. Could you provide a way of overriding this behavior so I can calculate only the usual standard errors?

[dlindzee]

Sorry I realized that the one can estimate a probit model. But the other comment stands.

[lrberge]

Hello and thanks for using! :-)

Yes it's possible to estimate probit with feglm and a probit family. Note that there is no bias correction though. I don't publicize it too much because of the incidental parameter problem that applies there. Good that you found out!

On your other comment, it's actually something that I planned to do but never have done because I didn't know if people would use it in the end. Thanks to you, I now have implemented it.

The code below should work with the dev version of the package.

data(base_did)
est_no_FE  = feols(y ~ x1, base_did)
est_one_FE = feols(y ~ x1 | id, base_did)
est_two_FE = feols(y ~ x1 | id + period, base_did)

etable(est_no_FE, est_one_FE, est_two_FE)
#>                               est_no_FE          est_one_FE          est_two_FE
#> (Intercept)           1.989*** (0.1492)
#> x1                  0.9831*** (0.05012) 0.9615*** (0.04808) 0.9779*** (0.04579)
#> Fixed-Effects:      ------------------- ------------------- -------------------
#> id                                   No                 Yes                 Yes
#> period                               No                  No                 Yes
#> ___________________ ___________________ ___________________ ___________________
#> Observations                      1,080               1,080               1,080
#> S.E. type                      Standard           1-way: id           1-way: id
#> R2                              0.26304             0.38715             0.48696
#> Within R2                            --             0.26507             0.30776

# Changing the default standard-errors
setFixest_se(no_FE = "white", one_FE = "standard", two_FE = "twoway")
etable(est_no_FE, est_one_FE, est_two_FE)
#>                               est_no_FE          est_one_FE          est_two_FE
#> (Intercept)            1.989*** (0.149)
#> x1                  0.9831*** (0.05102) 0.9615*** (0.05138) 0.9779*** (0.03453)
#> Fixed-Effects:      ------------------- ------------------- -------------------
#> id                                   No                 Yes                 Yes
#> period                               No                  No                 Yes
#> ___________________ ___________________ ___________________ ___________________
#> Observations                      1,080               1,080               1,080
#> S.E. type                         White            Standard  2-way: id & period
#> R2                              0.26304             0.38715             0.48696
#> Within R2                            --             0.26507             0.30776

# Reseting the defaults
setFixest_se()

Note that you need to specify the default for three different cases: without fixed-effects, with one, and with two or more fixed-effects.

Thanks again!