Cluster standard errors with fixed effects

[Jan Geurst]

Hi,

I am using a fixed effects model, but I am wondering why my results are significant when I use vce(robust) and become insignificant when I use vce(cluster company). The coefficients remain the same only the degrees of freedom are different. Can someone explain to me the difference between vce(robust) and vce(cluster company), and why my coefficients become insignificant? Further can someone explain to me what the degrees of freedom mean in this analysis?

Below are my two outputs:

. reghdfe bda announcement_eligible announcement l_lnassets l_roa, absorb(t incorp indus c) vce(robust)
(converged in 3 iterations)
note: announcement omitted because of collinearity

HDFE Linear regression                            Number of obs   =      3,944
Absorbing 4 HDFE groups                           F(   3,   3426) =       3.17
Statistics robust to heteroskedasticity           Prob > F        =     0.0235
                                                  R-squared       =     0.8970
                                                  Adj R-squared   =     0.8815
                                                  Within R-sq.    =     0.0393
                                                  Root MSE        =     0.1509

---------------------------------------------------------------------------------------
                      |               Robust
                  bda |      Coef.   Std. Err.      t    P>|t|     [95% Conf. Interval]
----------------------+----------------------------------------------------------------
announcement_eligible |   .0123368   .0070637     1.75   0.081    -.0015127    .0261863
         announcement |          0  (omitted)
           l_lnassets |  -.0151639   .0218286    -0.69   0.487    -.0579624    .0276345
                l_roa |  -.8755667   .3432176    -2.55   0.011    -1.548499   -.2026348
---------------------------------------------------------------------------------------

Absorbed degrees of freedom:
------------------------------------------------------------------------+
          Absorbed FE |  Num. Coefs.  =   Categories  -   Redundant     |
----------------------+-------------------------------------------------|
                    t |            8               8              0     |
               incorp |           15              16              1     |
                indus |           55              56              1 ?   |
                    c |          437             493             56 ?   |
------------------------------------------------------------------------+
? = number of redundant parameters may be higher

 reghdfe bda announcement_eligible announcement l_lnassets l_roa, absorb(t incorp indus c) vce(cluster c)
(converged in 3 iterations)
note: announcement omitted because of collinearity

HDFE Linear regression                            Number of obs   =      3,944
Absorbing 4 HDFE groups                           F(   3,    492) =       1.62
Statistics robust to heteroskedasticity           Prob > F        =     0.1827
                                                  R-squared       =     0.8970
                                                  Adj R-squared   =     0.8795
                                                  Within R-sq.    =     0.0393
Number of clusters (c)       =        493         Root MSE        =     0.1521

                                             (Std. Err. adjusted for 493 clusters in c)
---------------------------------------------------------------------------------------
                      |               Robust
                  bda |      Coef.   Std. Err.      t    P>|t|     [95% Conf. Interval]
----------------------+----------------------------------------------------------------
announcement_eligible |   .0123368   .0149413     0.83   0.409    -.0170198    .0416935
         announcement |          0  (omitted)
           l_lnassets |  -.0151639   .0309545    -0.49   0.624    -.0759833    .0456554
                l_roa |  -.8755667   .5327938    -1.64   0.101    -1.922399    .1712652
---------------------------------------------------------------------------------------

Absorbed degrees of freedom:
------------------------------------------------------------------------+
          Absorbed FE |  Num. Coefs.  =   Categories  -   Redundant     |
----------------------+-------------------------------------------------|
                    t |            8               8              0     |
               incorp |           15              16              1     |
                indus |           55              56              1 ?   |
                    c |            0             493            493 *   |
------------------------------------------------------------------------+
? = number of redundant parameters may be higher
* = fixed effect nested within cluster; treated as redundant for DoF computation

[Marcos Almeida]

Perhaps this text from Sergio Correia may be helpful.


In the site you also get this information:

Warning: in a fe panel regression, using robust will lead to inconsistent standard errors if for every fixed effect, the other dimension is fixed. For instance, in an standard panel with individual and time fixed effects, we require both the number of individuals and time periods to grow asymptotically. If that is not the case, an alternative may be to use clustered errors, which as discussed below will still have their own asymptotic requirements. For a discussion, see Stock and Watson, "Heteroskedasticity-robust standard errors for fixed-effects panel-data regression," Econometrica 76 (2008): 155-174

cluster clustervars estimates consistent standard errors even when the observations are correlated within groups.

[Clyde Schechter]

In fact, since version 13, Stata has disallowed the use of -vce(robust)- in conjunction with -xtreg, fe-. When -vce(robust)- is specified with -xtreg, fe-, Stata calculates the cluster robust standard errors, clustering on the panel variable that was declared in the -xtset- command.

[Jeff Tree]

In fact, since version 13, Stata has disallowed the use of -vce(robust)- in conjunction with -xtreg, fe-. When -vce(robust)- is specified with -xtreg, fe-, Stata calculates the cluster robust standard errors, clustering on the panel variable that was declared in the -xtset- command.

Hi Clyde,

If I may be so bold and revisit this post just to make sure I'm absolutely clear about the fe/clustering option...

If we have the following,

xtset country year
xtnbreg y x1 x2 x3 x4 x5, fe vce(cluster country)

Then, in addition to fixed effects, Stata will automatically cluster errors on country, correct?

(The reason for my inquiry is because I'm dusting off some old work I completed during grad school many years ago and have encountered an error (vcetype 'cluster' not allowed) which I'm assuming is due to the fact I'm using a much newer version of Stata than I did back in 2010.)

[Carlo Lazzaro]

Jeff:
the issue here is that -xtnbreg- does not allow -vce(cluster panelid) standard errors; you may want to explore some different -correlation- options and/or -bootstrap- standard errors.