Clustering Standard Error

[Prashant Bhandari]

I am working on finding the relationship between FDI inflow and military expenditure. I have panel data of 62 developing countries from 1990 to 2018. These countries are divided into five geographical regions. I was wondering if I could cluster my errors to the geographical regions instead of clustering by country. Wikipedia article mentions that 30-50 clusters are preferred but I could not see a particular reason behind this.

[Fei Wang]

Prashant, you may cluster at the region level, but it's not necessary and may be too conservative. Clustering at country level would be standard for your case.

[Maxence Morlet]

In addition to Fey's informative response, here is a veryuseful article you may want to read:

https://www.nber.org/papers/w24003

[Carlo Lazzaro]

Prahsant:
in addition to previous helpful guidance, if you have in mind to go -fe-, you may want to consider the community-contributed module -reghdfe- (just type -search reghdfe- from within Stata to spot and install it) that allows multi-way clustering (vs. -xtreg,fe-), as detailed in the folllowing toy-example:

. use "https://www.stata-press.com/data/r16/nlswork.dta"
(National Longitudinal Survey.  Young Women 14-26 years of age in 1968)

. xtreg ln_wage age, fe vce(cluster race)

Fixed-effects (within) regression               Number of obs      =     28510
Group variable: idcode                          Number of groups   =      4710

R-sq:  within  = 0.1026                         Obs per group: min =         1
       between = 0.0877                                        avg =       6.1
       overall = 0.0774                                        max =        15

                                                F(1,2)             =    761.45
corr(u_i, Xb)  = 0.0314                         Prob > F           =    0.0013

                                   (Std. Err. adjusted for 3 clusters in race)
------------------------------------------------------------------------------
             |               Robust
     ln_wage |      Coef.   Std. Err.      t    P>|t|     [95% Conf. Interval]
-------------+----------------------------------------------------------------
         age |   .0181349   .0006572    27.59   0.001     .0153072    .0209626
       _cons |   1.148214   .0190883    60.15   0.000     1.066083    1.230344
-------------+----------------------------------------------------------------
     sigma_u |  .40635023
     sigma_e |  .30349389
         rho |  .64192015   (fraction of variance due to u_i)
------------------------------------------------------------------------------

. reghdfe ln_wage age, abs(idcode) vce(cluster race birth_yr)
(dropped 551 singleton observations)
(converged in 1 iterations)

HDFE Linear regression                            Number of obs   =     27,959
Absorbing 1 HDFE group                            F(   1,      2) =     325.59
Statistics robust to heteroskedasticity           Prob > F        =     0.0031
                                                  R-squared       =     0.6540
                                                  Adj R-squared   =     0.5936
Number of clusters (race)    =          3         Within R-sq.    =     0.1026
Number of clusters (birth_yr) =         14        Root MSE        =     0.3035

                          (Std. Err. adjusted for 3 clusters in race birth_yr)
------------------------------------------------------------------------------
             |               Robust
     ln_wage |      Coef.   Std. Err.      t    P>|t|     [95% Conf. Interval]
-------------+----------------------------------------------------------------
         age |   .0181349    .001005    18.04   0.003     .0138106    .0224592
------------------------------------------------------------------------------

Absorbed degrees of freedom:
---------------------------------------------------------------+
 Absorbed FE |  Num. Coefs.  =   Categories  -   Redundant     |
-------------+-------------------------------------------------|
      idcode |            0            4159           4159 *   |
---------------------------------------------------------------+
* = fixed effect nested within cluster; treated as redundant for DoF computation

[Prashant Bhandari]

Thank you so much for all the helpful responses!