if condition in xtreg

[Evgeniya Mitrokhina]

Hello, everyone!

I do not understund why the output shows me minimum number of observation that equels 5, however, I establish condition consider only those where it is more that 10. COuld enyone help me? I am using xtreg in Stata 13.1.

Here I create variable that counts observations per group (which is country)

egen gdp_count = count(l_gdp_percap), by(ISO3N)

This is the model I'd like to estimete

xtreg l_gdp_percap ib1.reg_magnr  L(1/3).l_gdp_percap  ib0.persol2  i.decade if gdp_count>10, fe cl(ISO3N)

Fixed-effects (within) regression               Number of obs      =      3409
Group variable: ISO3N                           Number of groups   =       117

R-sq:  within  = 0.9654                         Obs per group: min =         2
       between = 0.9997                                        avg =      29.1
       overall = 0.9968                                        max =        50

                                                F(14,116)          =   3000.34
corr(u_i, Xb)  = 0.8411                         Prob > F           =    0.0000

                                 (Std. Err. adjusted for 117 clusters in ISO3N)
-------------------------------------------------------------------------------
              |               Robust
 l_gdp_percap |      Coef.   Std. Err.      t    P>|t|     [95% Conf. Interval]
--------------+----------------------------------------------------------------
    reg_magnr |
    Military  |  -.0243957   .0104707    -2.33   0.022    -.0451341   -.0036572
      Morchy  |   .0055622   .0160169     0.35   0.729    -.0261613    .0372857
  Multiparty  |  -.0165441    .009361    -1.77   0.080    -.0350848    .0019965
Single Party  |  -.0172222   .0124459    -1.38   0.169    -.0418729    .0074285
              |
 l_gdp_percap |
          L1. |   1.209961   .0800573    15.11   0.000     1.051398    1.368525
          L2. |  -.1723034   .0931483    -1.85   0.067    -.3567953    .0121885
          L3. |  -.0700679   .0287383    -2.44   0.016    -.1269876   -.0131481
              |
      persol2 |
           1  |   .0149075   .0070386     2.12   0.036     .0009666    .0288484
           2  |   .0141714   .0089652     1.58   0.117    -.0035853    .0319281
              |
       decade |
        1970  |   .0029361   .0036517     0.80   0.423    -.0042966    .0101689
        1980  |  -.0088728   .0044202    -2.01   0.047    -.0176276   -.0001179
        1990  |  -.0090631   .0050484    -1.80   0.075    -.0190621    .0009359
        2000  |   .0154033   .0073271     2.10   0.038      .000891    .0299157
        2010  |   .0184441   .0089666     2.06   0.042     .0006846    .0362035
              |
        _cons |   .2525653   .0527426     4.79   0.000     .1481018    .3570287
--------------+----------------------------------------------------------------
      sigma_u |  .04344562
      sigma_e |  .06886767
          rho |  .28468218   (fraction of variance due to u_i)
-------------------------------------------------------------------------------

. 

[Carlo Lazzaro]

Evgenia:
lagged variables may be an explanation of your min=2 observations.
As an aside, if you include lagged dependent variable among predictors (as it seems from you post), your estimates will be biased. I would take a look at -xtabond-, instead.

[Evgeniya Mitrokhina]

Thank you, Carlo!

I also thought about the lagged variables, but yearsterday everything worked fine!
And the ferquences of the the dependent variable that countes observations per group (country) seem to be fine.

----------------------
 table count

  count |      Freq.
----------+-----------
        0 |        570
       12 |         54
       13 |         54
       14 |         54
       15 |         54
       17 |         54
       18 |         54
       19 |        432
       20 |        108
       21 |         54
       22 |         54
       23 |         54
       24 |        594
       25 |         54
       26 |        108
       27 |        162
       28 |        108
       29 |        108
       30 |        108
       32 |         54
       33 |        108
       34 |        478
       37 |         54
       38 |         54
       39 |        207
       44 |        270
       46 |        162
       47 |         54
       48 |        162
       49 |        270
       53 |         54
       54 |      4,482
---------------------- 

Thanks for advice with -xtabond-, this is the nex step, firstly I decided to check FE.
I'm also scared that the imposibility to establish condition for number of observations will cause bad estimates using GMM (because of too many instruments, I have large T sample).

Actually, even estimatin the model without lagged dependent variable I still cannot establish the condition.

 xtreg l_gdp_percap ib1.reg_magnr  ib0.persol2  i.decade if gdp_count>10, fe cl(ISO3N)

Fixed-effects (within) regression               Number of obs      =      3659
Group variable: ISO3N                           Number of groups   =       117

R-sq:  within  = 0.2277                         Obs per group: min =         5
       between = 0.0061                                        avg =      31.3
       overall = 0.0366                                        max =        53

                                                F(11,116)          =     10.41
corr(u_i, Xb)  = 0.0088                         Prob > F           =    0.0000

                                 (Std. Err. adjusted for 117 clusters in ISO3N)
-------------------------------------------------------------------------------
              |               Robust
 l_gdp_percap |      Coef.   Std. Err.      t    P>|t|     [95% Conf. Interval]
--------------+----------------------------------------------------------------
    reg_magnr |
    Military  |  -.0218377   .2179559    -0.10   0.920    -.4535268    .4098514
      Morchy  |   .1438201   .3388362     0.42   0.672    -.5272876    .8149279
  Multiparty  |  -.0514717   .2193221    -0.23   0.815    -.4858668    .3829234
Single Party  |   .0916895   .2368069     0.39   0.699    -.3773364    .5607154
              |
      persol2 |
           1  |   .1894304   .0975784     1.94   0.055    -.0038359    .3826967
           2  |   .1638425   .1010949     1.62   0.108    -.0363886    .3640737
              |
       decade |
        1970  |    .224083   .0387758     5.78   0.000     .1472826    .3008834
        1980  |   .2830303   .0677675     4.18   0.000     .1488082    .4172524
        1990  |   .3352445   .1037122     3.23   0.002     .1298295    .5406596
        2000  |   .6098887   .1218926     5.00   0.000      .368465    .8513124
        2010  |   .8137106   .1345645     6.05   0.000     .5471885    1.080233
              |
        _cons |   6.856443    .244648    28.03   0.000     6.371887    7.340999
--------------+----------------------------------------------------------------
      sigma_u |  1.2208027
      sigma_e |  .33484565
          rho |  .93003249   (fraction of variance due to u_i)
------------------------------------------------------------------------------- 

[Carlo Lazzaro]

Evgenia:
have you checked that you were working on the same yesterday's dataset?

[Evgeniya Mitrokhina]

Yep, I did, that was the same one, but I created the variable differently

egen count = total(e(sample)), by( ISO3N )

which was wrong.

And the today's version

egen gdp_count = count(l_gdp_percap), by(ISO3N)

I want to create a variable that counts observations per country (ISO3N) for GDP (l_gdp_percap) and then use only these ones where there are more than 10 observations of the dependent variable per country.

[Carlo Lazzaro]

Evgenia:
do you mean that you could have experienced the same problem yesterday had you create the count variable in the same way?

[Evgeniya Mitrokhina]

I ment that today I realised that count variable was constructed incorrectly, that is why today I created anotehr one, which is gdp_count and it doesn't work in regression (I theink teh second one is correct meanin that O would like to count observations per country (ISO3N) for GDP (l_gdp_percap), am I right?)

[Carlo Lazzaro]

Evgeniya:
yes, the second one is correct for your research goal.

[Evgeniya Mitrokhina]

But still it doesn't work with xtreg

[Andrew Musau]

Evgeniya, in addition to Carlo's helpful advice in #2, consider that an observation includes all entries of variables in your model and not just of the dependent variable. Stata like most other software programs employs listwise deletion i.e., an entire observation is excluded from analysis if any single value is missing. I illustrate with a simple example:

. clear

. webuse grunfeld

. keep if time<5 & company<4
(188 observations deleted)

. replace mvalue=. if inrange(_n, 6, 7)
(2 real changes made, 2 to missing)

. egen count= count(invest), by(company)

. l, sepby(company)

     +----------------------------------------------------------+
     | company   year   invest   mvalue   kstock   time   count |
     |----------------------------------------------------------|
  1. |       1   1935    317.6   3078.5      2.8      1       4 |
  2. |       1   1936    391.8   4661.7     52.6      2       4 |
  3. |       1   1937    410.6   5387.1    156.9      3       4 |
  4. |       1   1938    257.7   2792.2    209.2      4       4 |
     |----------------------------------------------------------|
  5. |       2   1935    209.9   1362.4     53.8      1       4 |
  6. |       2   1936    355.3        .     50.5      2       4 |
  7. |       2   1937    469.9        .    118.1      3       4 |
  8. |       2   1938    262.3   1801.9    260.2      4       4 |
     |----------------------------------------------------------|
  9. |       3   1935     33.1   1170.6     97.8      1       4 |
 10. |       3   1936       45   2015.8    104.4      2       4 |
 11. |       3   1937     77.2   2803.3      118      3       4 |
 12. |       3   1938     44.6   2039.7    156.2      4       4 |
     +----------------------------------------------------------+



\\\ NOTICE THAT THE MINIMUM NOBS= 2 BELOW BECAUSE COMPANY 2 HAS 2 MISSING 
VALUES FOR MVALUE IN SPITE OF NO MISSING VALUE FOR THE OUTCOME INVEST 


. xtreg invest mvalue kstock, fe cl(company)

Fixed-effects (within) regression               Number of obs     =         10
Group variable: company                         Number of groups  =          3

R-sq:                                           Obs per group:
     within  = 0.8618                                         min =          2
     between = 0.4286                                         avg =        3.3
     overall = 0.5674                                         max =          4

                                                F(2,2)            =      16.88
corr(u_i, Xb)  = 0.3970                         Prob > F          =     0.0559

                                (Std. Err. adjusted for 3 clusters in company)
------------------------------------------------------------------------------
             |               Robust
      invest |      Coef.   Std. Err.      t    P>|t|     [95% Conf. Interval]
-------------+----------------------------------------------------------------
      mvalue |   .0498509   .0089992     5.54   0.031     .0111306    .0885711
      kstock |  -.0633788   .1704306    -0.37   0.746    -.7966824    .6699247
       _cons |   77.49927   25.62287     3.02   0.094    -32.74703    187.7456
-------------+----------------------------------------------------------------
     sigma_u |   117.1291
     sigma_e |   21.88049
         rho |  .96628001   (fraction of variance due to u_i)
------------------------------------------------------------------------------

.