logit HDFE and panel structure

[Ariel Soto-Caro]

Hello everyone.
I have a question that I hope you can help me with.
I have a logistic regression with 2 FE. The problem is, the fixed effects are for year and county (US), but the dataset is at the individual-level. So, I got an error (because I have several individuals per year and county):

xtset YEAR COUNTY
. repeated time values within panel

I did this trying to run a -xtlogit- model...

And also, one of FE (county) it's high-dimensional, so when I try the standard -logit- regression takes 2 hours to run.

I have something like this (for the standard logit):

logit depvar var1-var10 i.var11 i.var12 i.YEAR i.COUNTY [pw=w], cluster(COUNTY)

In conclusion, I don't know how to face this problem, of how to run a logit regression with FE and one HDFE.
I really appreciate any help on this.

[Joro Kolev]

Xtset your data on the high dimensional fixed effect as follows:

xtset COUNTY

and then control for the low dimensional fixed effects with factor variable notation i.YEAR

E.g., without the i.YEAR

. sysuse auto
(1978 Automobile Data)

. xtset rep
       panel variable:  rep78 (unbalanced)

. xtlogit fore price mpg head, nolog

Random-effects logistic regression              Number of obs     =         69
Group variable: rep78                           Number of groups  =          5

Random effects u_i ~ Gaussian                   Obs per group:
                                                              min =          2
                                                              avg =       13.8
                                                              max =         30

Integration method: mvaghermite                 Integration pts.  =         12

                                                Wald chi2(3)      =       7.64
Log likelihood  = -29.467606                    Prob > chi2       =     0.0540

------------------------------------------------------------------------------
     foreign |      Coef.   Std. Err.      z    P>|z|     [95% Conf. Interval]
-------------+----------------------------------------------------------------
       price |   .0001784   .0001377     1.30   0.195    -.0000915    .0004484
         mpg |   .1506742   .0804532     1.87   0.061    -.0070111    .3083595
    headroom |  -.7949371   .5415162    -1.47   0.142    -1.856289     .266415
       _cons |  -3.405605   2.967324    -1.15   0.251    -9.221452    2.410243
-------------+----------------------------------------------------------------
    /lnsig2u |    .783666   1.228934                     -1.625001    3.192333
-------------+----------------------------------------------------------------
     sigma_u |   1.479691   .9092213                      .4437471    4.934081
         rho |   .3995881   .2948428                      .0564737     .880953
------------------------------------------------------------------------------
LR test of rho=0: chibar2(01) = 5.44                   Prob >= chibar2 = 0.010

.

[Ariel Soto-Caro]

Thank you so much, Joro Kolev . I have a further question though. Did you intentionally omit the option -, fe- in your example? Because, as I stated in my question, both YEAR and COUNTY are FE. Thanks!

[Joro Kolev]

No, I did not intentionally omit the fixed effects. I just forgot that the default in -xtlogit- is the random effects, and I did not look at the output carefully to see the difference.

So of course you need to do -xtlogit, fe- in your case.

And for the year I did not put it because in the auto data there is no year.

In your regression you would do something like

xtset county

xtlogit y x i.year, fe

Thank you so much, Joro Kolev . I have a further question though. Did you intentionally omit the option -, fe- in your example? Because, as I stated in my question, both YEAR and COUNTY are FE. Thanks!

[Ariel Soto-Caro]

Thank you!
Sadly I'm getting an error:

. 10,150 (group size) take 9,114 (# positives) combinations results in numeric overflow;
computations cannot proceed

But I will figure it out. At least, now I know how to run my model.
Thanks!!

No, I did not intentionally omit the fixed effects. I just forgot that the default in -xtlogit- is the random effects, and I did not look at the output carefully to see the difference.

So of course you need to do -xtlogit, fe- in your case.

And for the year I did not put it because in the auto data there is no year.

In your regression you would do something like

xtset county

xtlogit y x i.year, fe