[Clyde Schechter]

I woud only add that, to the best of my knowledge/experience is quite unfrequent that real-world panel datasets can be handled with -regress- fruitfully.

That has been my experience, too.

[Daniel Kiory]

@ Clyde and @ Carlo, Oke thank you. I thought dat with regress and i.year is basicly is a panel estimation method, but with only time fixed effects. Does this mean that a panel estimation method with only time fixed effects is not possible. So that I have to use xtreg y1 x1 xn i.year,fe which include time-and cross sectional effects.

The problem with this last option is that the estimation output contains dots instead of values in some cases.

Random effects are not an option since a Hausman test already indicated that with a p-value 0f 0.000. So I wil use fixed effects.

@ Clyde, I thought it did not make sense because there was only 1 dummy. I thought that this maybe was not enough to say I included time fixed effects

[Clyde Schechter]

Does this mean that a panel estimation method with only time fixed effects is not possible.

It's possible. But it's usually not valid. As you have, it seems, already run the -xtreg, fe- model, take a look at that F-test that all u_i are zero in the bottom line. It will probably reject the all u_i = 0 null hypothesis--it's a very rare panel data set where that null hypothesis is not rejected. Only if that null hypothesis is not rejected would it be valid to use a model without panel level effects.

The problem with this last option is that the estimation output contains dots instead of values in some cases.

Why don't you show that output, and perhaps someone can help you figure out what it means and whether it's a problem or not?

I thought it did not make sense because there was only 1 dummy. I thought that this maybe was not enough to say I included time fixed effects

But you have included a complete set of adjustments for time-fixed effects, because x6-x8 are proxies for that. They are there, they're just under a different name. You can't distinguish the effects of x6-x8 per se from the effects of time, nor can you separately estimate them, but they are all adequately adjusted for.

[Daniel Kiory]

If I do this: xtreg y1 x1 x2 x3 x4 x5 x6 x7 x8 i.year, fe vce(robust)

The output contains no dots, it is perfectly fine:

. xtreg y1 x1 x2 x3 x4 x5 x6 x7 x8 i.year, fe vce(robust)
note: 2015.year omitted because of collinearity
note: 2016.year omitted because of collinearity

Fixed-effects (within) regression               Number of obs     =     67,452
Group variable: bankid                          Number of groups  =      4,818

R-sq:                                           Obs per group:
     within  = 0.0685                                         min =         14
     between = 0.0022                                         avg =       14.0
     overall = 0.0093                                         max =         14

                                                F(9,4817)         =          .
corr(u_i, Xb)  = -0.9027                        Prob > F          =          .

                             (Std. Err. adjusted for 4,818 clusters in bankid)
------------------------------------------------------------------------------
             |               Robust
          y1 |      Coef.   Std. Err.      t    P>|t|     [95% Conf. Interval]
-------------+----------------------------------------------------------------
          x1 |    .010848   .0010579    10.25   0.000      .008774    .0129221
          x2 |   -.212691   .1155646    -1.84   0.066    -.4392504    .0138684
          x3 |   -.004891   .0023192    -2.11   0.035    -.0094377   -.0003443
          x4 |   .0147497   .0048281     3.05   0.002     .0052845    .0242149
          x5 |   -.001771   .0001367   -12.95   0.000    -.0020391    -.001503
          x6 |   .0095694   6.841967     0.00   0.999    -13.40381    13.42295
          x7 |   .0044677    19.0627     0.00   1.000    -37.36714    37.37607
          x8 |  -.0001731   51.41708    -0.00   1.000    -100.8011    100.8008
             |
        year |
       2004  |  -.0152893   86.10412    -0.00   1.000    -168.8187    168.7881
       2005  |  -.0171848   141.0533    -0.00   1.000     -276.546    276.5117
       2006  |  -.0112041   70.46392    -0.00   1.000    -138.1527    138.1303
       2007  |  -.0189596   115.2941    -0.00   1.000    -226.0481    226.0102
       2008  |  -.0180756   183.2691    -0.00   1.000    -359.3092    359.2731
       2009  |   .0614708   116.7237     0.00   1.000    -228.7703    228.8933
       2010  |   .0398587   6.002331     0.01   0.995    -11.72745    11.80717
       2011  |   .0212817    9.73061     0.00   0.998    -19.05516    19.09772
       2012  |   .0045678   1.674019     0.00   0.998    -3.277273    3.286409
       2013  |   .0090123   16.77495     0.00   1.000    -32.87756    32.89558
       2014  |  -.0064524   1.705452    -0.00   0.997    -3.349917    3.337013
       2015  |          0  (omitted)
       2016  |          0  (omitted)
             |
       _cons |  -.2184157   56.09421    -0.00   0.997    -110.1887    109.7518
-------------+----------------------------------------------------------------
     sigma_u |  .09193293
     sigma_e |   .1240455
         rho |  .35453172   (fraction of variance due to u_i)
------------------------------------------------------------------------------

But if i use another dependent variable (I use two different dependent variables, not at the same time of course, to compare results). I get this

. xtreg y2 x1 x2 x3 x4 x5 x6 x7 x8 i.year, fe vce(robust)
note: 2015.year omitted because of collinearity
note: 2016.year omitted because of collinearity

Fixed-effects (within) regression               Number of obs     =     67,452
Group variable: bankid                          Number of groups  =      4,818

R-sq:                                           Obs per group:
     within  = 0.1093                                         min =         14
     between = 0.0063                                         avg =       14.0
     overall = 0.0323                                         max =         14

                                                F(5,4817)         =          .
corr(u_i, Xb)  = -0.8078                        Prob > F          =          .

                             (Std. Err. adjusted for 4,818 clusters in bankid)
------------------------------------------------------------------------------
             |               Robust
          y2 |      Coef.   Std. Err.      t    P>|t|     [95% Conf. Interval]
-------------+----------------------------------------------------------------
          x1 |   .0010844   .0003195     3.39   0.001     .0004581    .0017106
          x2 |  -.1043181   .0481647    -2.17   0.030    -.1987429   -.0098934
          x3 |   .0008831   .0004661     1.89   0.058    -.0000307     .001797
          x4 |  -.0157483   .0015555   -10.12   0.000    -.0187977   -.0126988
          x5 |   .0019073   .0000524    36.43   0.000     .0018047      .00201
          x6 |  -.0068211          .        .       .            .           .
          x7 |  -.0090473          .        .       .            .           .
          x8 |   .0025418          .        .       .            .           .
             |
        year |
       2004  |   .0224684          .        .       .            .           .
       2005  |   .0085545          .        .       .            .           .
       2006  |   .0065398          .        .       .            .           .
       2007  |   .0145562          .        .       .            .           .
       2008  |   .0111794          .        .       .            .           .
       2009  |  -.0824662          .        .       .            .           .
       2010  |  -.0173424          .        .       .            .           .
       2011  |  -.0040334          .        .       .            .           .
       2012  |   -.004375          .        .       .            .           .
       2013  |   .0071509          .        .       .            .           .
       2014  |   .0096708          .        .       .            .           .
       2015  |          0  (omitted)
       2016  |          0  (omitted)
             |
       _cons |   .0969779          .        .       .            .           .
-------------+----------------------------------------------------------------
     sigma_u |  .03135754
     sigma_e |  .05403998
         rho |  .25189358   (fraction of variance due to u_i)
------------------------------------------------------------------------------

Any idea why those dots appear? These are summary statistics of y1 and y2, maybe this helps. For the record: I use vce(robust) standards errors because there might be heteroskedasticy and/or serial correlation in the residuals

. xtsum y1 y2

Variable         |      Mean   Std. Dev.       Min        Max |    Observations
-----------------+--------------------------------------------+----------------
y1       overall |  .0026815   .1264788  -3.032915   3.982771 |     N =   67452
         between |             .0257269  -.2354076   .1743918 |     n =    4818
         within  |             .1238351  -3.026223   3.984894 |     T =      14
                 |                                            |
y2       overall |  .0017804   .0561652  -.9032218   .9659844 |     N =   67452
         between |             .0105348  -.0600262    .061931 |     n =    4818
         within  |             .0551686  -.8629883   .9309759 |     T =      14

[Clyde Schechter]

Well, one circumstance where I would expect to see this is if y2 takes on the same value for all bankids in every given year. Does that happen in your data?

[Daniel Kiory]

No this is not the case. y2 for a particular bankid is different over years. Also, the value of y2 of the bankid's within a year are not the same.

It could be that ocassionally values are the same, but this is a coincidence so I do not believe this is the reason,

[Daniel Kiory]

I think it has something to do with the vce(robust) standard errors.

If I do: xtreg y2 x1 x2 x3 x4 x5 x6 x7 x8 i.year, fe
I get the following results:

. xtreg y2 x1 x2 x3 x4 x5 x6 x7 x8 i.year, fe
note: 2015.year omitted because of collinearity
note: 2016.year omitted because of collinearity

Fixed-effects (within) regression               Number of obs     =     67,452
Group variable: bankid                          Number of groups  =      4,818

R-sq:                                           Obs per group:
     within  = 0.1093                                         min =         14
     between = 0.0063                                         avg =       14.0
     overall = 0.0323                                         max =         14

                                                F(19,62615)       =     404.30
corr(u_i, Xb)  = -0.8078                        Prob > F          =     0.0000

------------------------------------------------------------------------------
          y2 |      Coef.   Std. Err.      t    P>|t|     [95% Conf. Interval]
-------------+----------------------------------------------------------------
          x1 |   .0010844   .0000977    11.10   0.000      .000893    .0012757
          x2 |  -.1043181   .0127098    -8.21   0.000    -.1292293   -.0794069
          x3 |   .0008831   .0002856     3.09   0.002     .0003234    .0014429
          x4 |  -.0157483   .0009214   -17.09   0.000    -.0175541   -.0139424
          x5 |   .0019073   .0000303    63.05   0.000      .001848    .0019666
          x6 |  -.0068211   29.54744    -0.00   1.000    -57.91985    57.90621
          x7 |  -.0090473   81.13548    -0.00   1.000    -159.0347    159.0166
          x8 |   .0025418   224.0549     0.00   1.000    -439.1454    439.1505
             |
        year |
       2004  |   .0224684   371.5275     0.00   1.000    -728.1722    728.2172
       2005  |   .0085545   601.1465     0.00   1.000     -1178.24    1178.257
       2006  |   .0065398   308.1431     0.00   1.000    -603.9545    603.9676
       2007  |   .0145562    489.674     0.00   1.000    -959.7474    959.7766
       2008  |   .0111794   809.9781     0.00   1.000    -1587.547     1587.57
       2009  |  -.0824662   501.3769    -0.00   1.000    -982.7821    982.6171
       2010  |  -.0173424   26.29954    -0.00   0.999    -51.56448     51.5298
       2011  |  -.0040334    42.4661    -0.00   1.000    -83.23766    83.22959
       2012  |   -.004375    7.28164    -0.00   1.000     -14.2764    14.26765
       2013  |   .0071509   70.62048     0.00   1.000    -138.4091    138.4234
       2014  |   .0096708   7.492325     0.00   0.999     -14.6753    14.69464
       2015  |          0  (omitted)
       2016  |          0  (omitted)
             |
       _cons |   .0969779   245.5125     0.00   1.000     -481.108     481.302
-------------+----------------------------------------------------------------
     sigma_u |  .03135754
     sigma_e |  .05404041
         rho |  .25189057   (fraction of variance due to u_i)
------------------------------------------------------------------------------
F test that all u_i=0: F(4817, 62615) = 1.15                 Prob > F = 0.0000

However, I do not know why this yields different results. I really wanted to use the vce(robust) st errors but maybe I should forget about this..

[Daniel Kiory]

Dear Clyde and Carlo,

I would really thank you for your reactions, answers and insights you gave me on this forum. It really helped me.

After al, the i.year dummies appear to be insignificant (after testing with testparm i.year), which surprises me enourmously, and is contradicting economic theory, but it is shown by the data.

Also a Chow test which tested if coefficents are different in different periods (I compared two different periods) did indicate a significant differnce. This was, besides economic theory, one of the reasons to introduce time fixed effects. As final question: does this Chow test and the insignifance of the i.year dummies contractict each othter?

I will use xtreg, fe as of now.

Thank you

[Carlo Lazzaro]

Daniel:
see -help j_robustsingular- about the missing F-test with the -robust()- option (the same would happen with -cluster()-, under -xtreg-).
About the reported discrepancy between your data and the economic theory, two remarks:
- statistical significance of a given predictor should be read keeping in mind the regression model performed (is it exacly in line with the theory of your research field? Sometimes real-world data base let your expectations down and there's nothing you can do to work that downside around. Outside the statistcal textbooks example, everything loses its normality!) and the fact that predictors act like a team: they adjust each other. hence, with a different spcification -i.years- may turn out to be statistically significant (which is not a good or bad thing per se, though);
- you should also consider that -xtreg, fe. focuses on the within variation. Your resulst seem to tell you that, when adjusted for the remaining predictors, time passing by does not seem to produce the evidence of a statistical significant effect on the dependent variable.

[Daniel Kiory]

Carlo,

I do not really understand what you mean at the first point about the difference beween my data and the economic theory. Do you mean that it may be that my statistical model is not perfectly in line with the theory about this subject?

Of course, sometimes the results let expectations down, but not finding a significance is as wel a result I guess. I think that because of the insignificance of the time dummies I should exclude them. But that is mostly because I am used to make decisions based on data, if I would keep them, I need a really good explanation I guess.

[Carlo Lazzaro]

Daniel:
q1) Not quite. I meant something I remind to myself first, that is being sure that the regression model (or whatever I'm doing) gives a fair and true view of the data generating process underlying the population from which the sample I'm working on was drawn;
q2) I do share your thought that being or not statistical significance is anyway a result (I'm not a "p-value less than" fan). Your approach to rule -i.year- out of the set of predictors is perfectly legal (others may have a different take on that, though).