vce(cluster) does not work in fixed effects when there are only two observations per group

[Xinchen Dai]

I am working with a balanced individual-year panel dataset. All the individuals are observed for two and only two years. I am trying to do

 xtreg y x, fe vce(cluster id)

But I get only the coefficients, the standard errors are not produced.

Things work fine when I use random effects or OLS to estimate the model, so I guess that to cluster the standard error at the group level at least three observations are needed for each group when using within estimator. But why? Or maybe my guess is wrong. Thanks ahead for your explanation.

[Andrew Musau]

so I guess that to cluster the standard error at the group level at least three observations are needed for each group when using within estimator.

No. There must be something peculiar with your data. Here is a counterexample.

webuse grunfeld, clear
keep if time<3
xtreg invest mvalue, fe cluster(company)

Res.:

. xtreg invest mvalue, fe cluster(company)

Fixed-effects (within) regression               Number of obs     =         20
Group variable: company                         Number of groups  =         10

R-sq:                                           Obs per group:
     within  = 0.4230                                         min =          2
     between = 0.7746                                         avg =        2.0
     overall = 0.7503                                         max =          2

                                                F(1,9)            =       8.54
corr(u_i, Xb)  = 0.5577                         Prob > F          =     0.0170

                               (Std. Err. adjusted for 10 clusters in company)
------------------------------------------------------------------------------
             |               Robust
      invest |      Coef.   Std. Err.      t    P>|t|     [95% Conf. Interval]
-------------+----------------------------------------------------------------
      mvalue |   .0567196   .0194114     2.92   0.017     .0128079    .1006313
       _cons |   36.49626   17.34454     2.10   0.065    -2.739813    75.73234
-------------+----------------------------------------------------------------
     sigma_u |  72.169136
     sigma_e |  30.053113
         rho |  .85221652   (fraction of variance due to u_i)
------------------------------------------------------------------------------

[Xinchen Dai]

This is a subsample from the sample I am working with. I don't see anything weird with it but clustered standard error is still not produced.

clear
input double pid float yearcode double lifesatisfaction float(ln_y lag_ln_y lag2_ln_y lead_ln_y)
110006101 3 3 10.661173  10.84843  9.910132  10.90412
110006101 4 4  10.90412 10.661173  10.84843  11.13896
110009102 3 4 10.355473    10.242  9.972481  10.23279
110009102 4 4  10.23279 10.355473    10.242  11.25785
110009106 3 5 10.355473    10.242  9.972481  10.23279
110009106 4 1  10.23279 10.355473    10.242  11.25785
110009107 3 3 10.355473    10.242  9.972481  10.23279
110009107 4 3  10.23279 10.355473    10.242  11.25785
110026101 3 5 11.700234 11.027735  10.91509 11.289782
110026101 4 5 11.289782 11.700234 11.027735   12.1021
110026102 3 4 11.700234 11.027735  10.91509 11.289782
110026102 4 4 11.289782 11.700234 11.027735   12.1021
110026104 3 4 11.700234 11.027735  10.91509 11.289782
110026104 4 3 11.289782 11.700234 11.027735   12.1021
110030101 3 5 11.184422 11.184422 10.596635 11.472103
110030101 4 5 11.472103 11.184422 11.184422 11.532728
110031101 3 4  8.188689  9.680344  9.680344 10.806355
110031101 4 4 10.806355  8.188689  9.680344 10.933107
110035101 3 4 11.066638 10.576432 10.453534  8.804875
110035101 4 3  8.804875 11.066638 10.576432 11.289782
end

what I run is

xtset pid yearcode

xtreg lifesatisfaction ln_y lag_ln_y lag2_ln_y lead_ln_y , fe vce(cluster pid)

One piece of information that is important: In this original sample, each individuals are followed for 5 years. My covariates include current income, and its one-period lead, and one- and two-period lags. Therefore, only observations of year 3 and year 4 have all these variables nonmissing. And the clustering fails when I add lag2_ln_y into the regression, which is the two-period lagged ln_y, no matter whether I run the regression directly with the original sample or first keep observations only in year 3 and year 4.

[Andrew Musau]

Stata is unable to calculate the cluster robust standard errors for some reason. With your sample data in #2, I will note that there are no such problems, but I cannot guess what happens with your full sample.

EDIT: I will also recommend using factor variables and not creating the lags and leads yourself.

xtreg lifesatisfaction ln_y l(1/2).(ln_y) f.ln_y , fe vce(cluster pid)

[Carlo Lazzaro]

Xinchen:
I cannot replicate your problem:

. xtset pid yearcode
       panel variable:  pid (strongly balanced)
        time variable:  yearcode, 3 to 4
                delta:  1 unit

. xtreg lifesatisfaction ln_y lag_ln_y lag2_ln_y lead_ln_y , fe vce(cluster pid)

Fixed-effects (within) regression               Number of obs     =         20
Group variable: pid                             Number of groups  =         10

R-sq:                                           Obs per group:
     within  = 0.2728                                         min =          2
     between = 0.2256                                         avg =        2.0
     overall = 0.0071                                         max =          2

                                                F(4,9)            =       1.84
corr(u_i, Xb)  = -0.7092                        Prob > F          =     0.2058

                                   (Std. Err. adjusted for 10 clusters in pid)
------------------------------------------------------------------------------
             |               Robust
lifesatisf~n |      Coef.   Std. Err.      t    P>|t|     [95% Conf. Interval]
-------------+----------------------------------------------------------------
        ln_y |  -.3212684   .4603154    -0.70   0.503    -1.362574    .7200373
    lag_ln_y |  -.3609143     .65565    -0.55   0.595    -1.844098    1.122269
   lag2_ln_y |   .6864056   .5923998     1.16   0.276    -.6536959    2.026507
   lead_ln_y |  -.8215082   .4954296    -1.66   0.132    -1.942248    .2992314
       _cons |   13.04777   16.60944     0.79   0.452     -24.5254    50.62094
-------------+----------------------------------------------------------------
     sigma_u |  1.1342179
     sigma_e |  1.0730268
         rho |  .52770159   (fraction of variance due to u_i)
------------------------------------------------------------------------------

.

PS: I agree in full with Andrew's helpful reply.

[Xinchen Dai]

Hello Carlo,

I was using Stata 14.0 SE, then I noticed that you are using Stata 16.0 SE. I don't have Stata 16, so I switch to Stata 15.1. Everything works fine now. This is weird, and makes me a bit embarrassed. Thanks both you and Andrew for taking the time to help me.

[Carlo Lazzaro]

Xinchen:
I cannot replicate your problem with Stata 14.2, too:

. input double pid float yearcode double lifesatisfaction float(ln_y lag_ln_y lag2_ln_y lead_ln_y)

            pid   yearcode  lifesati~n       ln_y   lag_ln_y  lag2_ln_y  lead_ln_y
  1.
. 110006101 3 3 10.661173  10.84843  9.910132  10.90412
  2.
. 110006101 4 4  10.90412 10.661173  10.84843  11.13896
  3.
. 110009102 3 4 10.355473    10.242  9.972481  10.23279
  4.
. 110009102 4 4  10.23279 10.355473    10.242  11.25785
  5.
. 110009106 3 5 10.355473    10.242  9.972481  10.23279
  6.
. 110009106 4 1  10.23279 10.355473    10.242  11.25785
  7.
. 110009107 3 3 10.355473    10.242  9.972481  10.23279
  8.
. 110009107 4 3  10.23279 10.355473    10.242  11.25785
  9.
. 110026101 3 5 11.700234 11.027735  10.91509 11.289782
 10.
. 110026101 4 5 11.289782 11.700234 11.027735   12.1021
 11.
. 110026102 3 4 11.700234 11.027735  10.91509 11.289782
 12.
. 110026102 4 4 11.289782 11.700234 11.027735   12.1021
 13.
. 110026104 3 4 11.700234 11.027735  10.91509 11.289782
 14.
. 110026104 4 3 11.289782 11.700234 11.027735   12.1021
 15.
. 110030101 3 5 11.184422 11.184422 10.596635 11.472103
 16.
. 110030101 4 5 11.472103 11.184422 11.184422 11.532728
 17.
. 110031101 3 4  8.188689  9.680344  9.680344 10.806355
 18.
. 110031101 4 4 10.806355  8.188689  9.680344 10.933107
 19.
. 110035101 3 4 11.066638 10.576432 10.453534  8.804875
 20.
. 110035101 4 3  8.804875 11.066638 10.576432 11.289782
 21.
. end

.
. xtset pid yearcode
       panel variable:  pid (strongly balanced)
        time variable:  yearcode, 3 to 4
                delta:  1 unit

.
. xtreg lifesatisfaction ln_y lag_ln_y lag2_ln_y lead_ln_y , fe vce(cluster pid)

Fixed-effects (within) regression               Number of obs     =         20
Group variable: pid                             Number of groups  =         10

R-sq:                                           Obs per group:
     within  = 0.2728                                         min =          2
     between = 0.2256                                         avg =        2.0
     overall = 0.0071                                         max =          2

                                                F(4,9)            =       1.84
corr(u_i, Xb)  = -0.7092                        Prob > F          =     0.2058

                                   (Std. Err. adjusted for 10 clusters in pid)
------------------------------------------------------------------------------
             |               Robust
lifesatisf~n |      Coef.   Std. Err.      t    P>|t|     [95% Conf. Interval]
-------------+----------------------------------------------------------------
        ln_y |  -.3212684   .4603154    -0.70   0.503    -1.362574    .7200373
    lag_ln_y |  -.3609143     .65565    -0.55   0.595    -1.844098    1.122269
   lag2_ln_y |   .6864056   .5923998     1.16   0.276    -.6536959    2.026507
   lead_ln_y |  -.8215082   .4954296    -1.66   0.132    -1.942248    .2992314
       _cons |   13.04777   16.60944     0.79   0.452     -24.5254    50.62094
-------------+----------------------------------------------------------------
     sigma_u |  1.1342179
     sigma_e |  1.0730268
         rho |  .52770159   (fraction of variance due to u_i)
------------------------------------------------------------------------------

.

As an aside, please note that the FAQ kindly request posters to specify beforehand which Stata release they're working with. Thanks.