Using xtreg: Number of observations too small?

[Samuel Pfaff]

Hello everyone,
I have a total of 202 observations for the US on a monthly basis from 2004-2017. I want to regress the monthly search volume (DSVI_country) on one independent variable (total_c), including year_month fixed effects.

* Example generated by -dataex-. For more info, type help dataex
clear
input str13 country float(country_id year_month DSVI_country total_c)
"united states" 1 529  12.667176 140
"united states" 1 530   9.074037 162
"united states" 1 531 -18.523067 205
"united states" 1 532  13.853467 565
"united states" 1 533  -6.838091 549
end
format %tm year_month

I declared my data as panel data:

. xtset country_id year_month

Panel variable: country_id (strongly balanced)
 Time variable: year_month, 2004m2 to 2020m11
         Delta: 1 month

. xtdes

country_id:  1, ., ..., 1                                    n =          1
year_month:  2004m2, 2004m3, ..., 2020m11                    T =        202
           Delta(year_month) = 1 month
           Span(year_month)  = 202 periods
           (country_id*year_month uniquely identifies each observation)

Distribution of T_i:   min      5%     25%       50%       75%     95%     max
                       202     202     202       202       202     202     202

     Freq.  Percent    Cum. |  Pattern*
 ---------------------------+--------------------------------------------------------------------------------------------------------
        1    100.00  100.00 |  22222222222222222222222222222222222222222222222222222222222222222222222222222222222222222222222222222
 ---------------------------+--------------------------------------------------------------------------------------------------------
        1    100.00         |  XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
 ------------------------------------------------------------------------------------------------------------------------------------
 *Each column represents 2 periods.

Originally, I wanted to run the regression using xtreg. However, stata returns an error:

. xtreg DSVI_country total_c i.year_month, fe
no observations
r(2000);

I confirmed that I have no missing values and all values are non-string. Could it be that I have too few observations? Or is xtreg not working since I have a T>N panel dataset?

I tried using xtregar as an alternative, however, it does not seem to work either:

. xtregar DSVI_country total_c i.year_month, fe
time variable may not be included in varlist
r(198);

Running the code above without "i.year_month" worked, but I am not sure whether in this case I am still including my year-month fixed effects?

I would be very glad if someone could provide me with a hint on how I could run this regression including year-month fixed effects.

Thank you so much in advance for your time!

Samuel

[Felix Bittmann]

I am a bit confused, why would you run xtreg with exactly one group, one country? Maybe a time series analysis makes more sense for this purpose?

[Samuel Pfaff]

Dear Felix,
thanks for your reponse and sorry for causing confusion!
I thought about it again and I think I realized my mistake. Since I only have one group (one country), using xtreg (or xtregar) doesn't make any sense, right?
Since I only have one country, this would probably also mean that including year-month fixed effects doesn`t make any sense either, am I right?

Thanks for your clarifications!

Samuel

[Felix Bittmann]

Well I think it depends, if you assume that some months or seasons have distinct effects, you can still estimate them. Did you simply try a linear OLS model (regress)?

[Samuel Pfaff]

Dear Felix,
that makes sense, thank you!

Yes, I tried a simple regress command and tried to include the year-month fixed effects.
However, it seems that I have an issue with collinearity..

 regress DSVI_country total_c i.year_month
note: 730.year_month omitted because of collinearity.

      Source |       SS           df       MS      Number of obs   =       202
-------------+----------------------------------   F(201, 0)       =         .
       Model |  67354.1955       201    335.0955   Prob > F        =         .
    Residual |           0         0           .   R-squared       =    1.0000
-------------+----------------------------------   Adj R-squared   =         .
       Total |  67354.1955       201    335.0955   Root MSE        =         0

------------------------------------------------------------------------------
DSVI_country | Coefficient  Std. err.      t    P>|t|     [95% conf. interval]
-------------+----------------------------------------------------------------
     total_c |   1.583448          .        .       .            .           .
             |
  year_month |
        530  |    -38.429          .        .       .            .           .

Do you have any further idea on what else I could try?

Thanks in advance!

Samuel

[Felix Bittmann]

This happens if you months are enumerated from, say 1 to 202 so each month is individual. Apparently, this wont work. But if you are interested in seasonal effects, you can code all Januarys as 1, Februarys as 2 and so on. Then you can estimate the effect.