Time fixed effects (time dummies) and time serie variables

[YH jordaan]

Hello everyone,

I'm working on a research where I want to investigate the relationship between the (market) liquidity of banks with respect to their credit ratings (dummy variable ''good bank'').

My variable of interest is the dummy variable ''good bank'' (independent variable), but of course I have also multiple control variables. These control variables include bank specific controls, like size of the bank or profit. These variables are in a panel data form. However, I also include macroeconomic control variables like the (change) in inflation, a crisis dummy variable or the money market rate.

My question is the following: When I include time fixed effects in my estimation, beside the entity (in my case the ''banks'') fixed effects, the results indicate that I have to exclude the time serie variables (like inflation, money market rate) etc, because they are highly insignificant. I know that one can't solely base the decision to exclude a variable because it is insignificant, but I just mentioned myself that when one uses time fixed effects, which controls for specific time effects, it is maybe unnecessary (and therefore wrong) to also control for time effects variables (inflation etc). Is this idea correct?

Thanks in advance for your help

Yannick

PS: small side questions what I think has to do with the ''label'' option. My dates in the regression output are visible in a number format (like 167) while the variable is in yearQuarter format in my dataset. Does this has to do with labeling?

PS2: I attached the regression result where my main and side questions are based on. Is there an efficient way to post it in the regular post? I tried copy table as HTML, but this did not work.

[YH jordaan]

I see that my attached photo is only visible when you double click on it and zoom in a litte bit. If one needs this photo to answer my question and finds it frustrating to see it only in this way, please let me know how I can add it in a good way, so that I can you provide you with the necessary information.

[Clyde Schechter]

I can't read your attachment at all. (That is not unusual--screen shot attachments are often unreadable on this Forum.) The way to post Stata results here is to copy (not copy table, not copy as picture, just plain old copy) from the Results window (or from your log file) and then paste into a code block. If you do not know how to set up a code block, see #12 in the FAQ for explicit instructions.

While it wouldn't hurt to see your results before giving advice, in general terms, lack of statistical significance is NOTa reason to exclude a variable from a model. It can still exert a confounding effect even if it is not statistically significant.

To answer the question in your PS, I don't know what "label" option you are referring to: none of the commands you have discussed using has such an option. Suffice it to say that if you have a variable that you expect to appear as a quarterly date and it is showing as an integer like 167, the command -format variable %tq- will cause it to display as a quarterly date.

[Carlo Lazzaro]

YH:
I do share Clyde's previous advice.
I tried to take a look at your (very hardly readable) screenshot (for the future, please use code delimiters to post what you typed and what Stata gave you back. See FAQ on how to do it. Thanks.), and what I got is that most of your coefficients are not significant (not ony the ones you complain about); besides, the R2 is quite low.
Are you sure that your model is well specified?
I would take a look at the literature and see what others did in the past when dealing with the same research topic.

[YH jordaan]

I took an (another) look at the FAQ. Hope it will be displayed fine this time:

. areg marketliq goodbank Dsize  Dracr Dprofit Dllreserve Dloangrowth Dcrisisdummy   changeinflation  DiffLibor changeFedF
> und i.datum, absorb (gvkey)
note: 197.datum omitted because of collinearity
note: 215.datum omitted because of collinearity
note: 217.datum omitted because of collinearity

Linear regression, absorbing indicators           Number of obs   =       3100
                                                  F(  59,   2941) =       6.11
                                                  Prob > F        =     0.0000
                                                  R-squared       =     0.1162
                                                  Adj R-squared   =     0.0687
                                                  Root MSE        =     0.0166

---------------------------------------------------------------------------------
      marketliq |      Coef.   Std. Err.      t    P>|t|     [95% Conf. Interval]
----------------+----------------------------------------------------------------
       goodbank |  -.0027453   .0010636    -2.58   0.010    -.0048307   -.0006598
          Dsize |   .0538683   .0050781    10.61   0.000     .0439113    .0638254
          Dracr |   .0019582   .0002886     6.78   0.000     .0013922    .0025241
        Dprofit |  -.1102577   .0918334    -1.20   0.230     -.290322    .0698065
     Dllreserve |   .6840738   .1133466     6.04   0.000      .461827    .9063205
    Dloangrowth |  -.0003817   .0004657    -0.82   0.412    -.0012947    .0005314
   Dcrisisdummy |  -.0397147   2298.354    -0.00   1.000    -4506.586    4506.507
changeinflation |   .0811879   4852.458     0.00   1.000    -9514.478     9514.64
      DiffLibor |   .0191369   4450.648     0.00   1.000    -8726.682     8726.72
  changeFedFund |  -.0126189   6079.186    -0.00   1.000     -11919.9    11919.88
                |
          datum |
           166  |   .0105497   5029.107     0.00   1.000    -9860.916    9860.938
           167  |  -.0328287   4528.734    -0.00   1.000    -8879.842    8879.776
           168  |   .0072159   1012.395     0.00   1.000    -1985.068    1985.083
           169  |   .0299684   1663.821     0.00   1.000    -3262.341    3262.401
           170  |   .0114673   381.6218     0.00   1.000    -748.2614    748.2844
           171  |   .0156432   3032.267     0.00   1.000    -5945.566    5945.597
           172  |    .085064   3879.029     0.00   1.000    -7605.803    7605.973
           173  |   .0176199   2298.533     0.00   1.000     -4506.88    4506.915
           174  |   .0414836   1879.921     0.00   1.000    -3686.052    3686.135
           175  |   .0428268    2702.99     0.00   1.000    -5299.901    5299.987
           176  |  -.0147404   1002.053    -0.00   1.000     -1964.81    1964.781
           177  |   .0211855   249.0215     0.00   1.000     -488.253    488.2954
           178  |  -.0019061   922.9896    -0.00   1.000    -1809.773    1809.769
           179  |  -.0089829   622.1029    -0.00   1.000     -1219.81    1219.792
           180  |  -.0091135   1257.488    -0.00   1.000    -2465.655    2465.637
           181  |   .0319232   1293.596     0.00   1.000    -2536.413    2536.477
           182  |  -.0003694   822.8716    -0.00   1.000    -1613.463    1613.462
           183  |  -.0087508   926.4192    -0.00   1.000    -1816.505    1816.487
           184  |  -.0441192     3383.4    -0.00   1.000    -6634.116    6634.028
           185  |  -.0314509   2446.151    -0.00   1.000    -4796.373     4796.31
           186  |  -.0191877   281.6666    -0.00   1.000     -552.303    552.2646
           187  |   .0122858   527.9066     0.00   1.000    -1035.092    1035.116
           188  |   -.002992   147.0993    -0.00   1.000     -288.431     288.425
           189  |   .0109907   423.9553     0.00   1.000    -831.2682    831.2902
           190  |  -.0235796   2694.525    -0.00   1.000    -5283.369    5283.322
           191  |  -.0064075   2191.248    -0.00   1.000    -4296.542    4296.529
           192  |   .0381699   4481.434     0.00   1.000    -8787.027    8787.103
           193  |  -.0137332   2808.528    -0.00   1.000    -5506.894    5506.866
           194  |   .0162607   7924.802     0.00   1.000    -15538.71    15538.74
           195  |   .0497187   5749.609     0.00   1.000    -11273.62    11273.72
           196  |   .0173758   8753.303     0.00   1.000     -17163.2    17163.24
           197  |          0  (omitted)
           198  |   -.004657   745.4412    -0.00   1.000    -1461.644    1461.635
           199  |   .0852336   4745.058     0.00   1.000    -9303.887    9304.058
           200  |   .0784374   3973.893     0.00   1.000    -7791.815    7791.972
           201  |   .0462066   3324.563     0.00   1.000    -6518.661    6518.753
           202  |   .0483099   3021.786     0.00   1.000    -5924.982    5925.079
           203  |   .0217023   995.0387     0.00   1.000    -1951.021    1951.065
           204  |   .0123024   897.8384     0.00   1.000    -1760.443    1760.468
           205  |  -.0231794   1681.802    -0.00   1.000    -3297.652    3297.606
           206  |  -.0062636   525.6116    -0.00   1.000     -1030.61    1030.598
           207  |  -.0197457   990.3939    -0.00   1.000    -1941.955    1941.916
           208  |  -.0045479   262.7079    -0.00   1.000    -515.1145    515.1054
           209  |   .0121346   231.5889     0.00   1.000    -454.0807     454.105
           210  |   .0068878   543.0426     0.00   1.000    -1064.775    1064.789
           211  |   .0026581   510.5523     0.00   1.000    -1001.074    1001.079
           212  |   .0352491   2247.569     0.00   1.000    -4406.932    4407.002
           213  |   .0108938   887.1538     0.00   1.000    -1739.494    1739.516
           214  |   .0204061   648.3934     0.00   1.000     -1271.33    1271.371
           215  |          0  (omitted)
           216  |    .004117   49.92206     0.00   1.000    -97.88161    97.88984
           217  |          0  (omitted)
                |
          _cons |  -.0908358   5463.355    -0.00   1.000    -10712.48     10712.3
----------------+----------------------------------------------------------------
          gvkey |       F(99, 2941) =      0.322   1.000         (100 categories)

@ Carlo:
I did an extensive literature research. The ''problem'' here is that normally, the value of market liquidity at time T is being used in the related literature. However, I would like to investigate the difference in the use of market liquidity. So my variable marketliq is specified as marketliquidity(t) - marketliquidity (t-1). Unfortunately, there is not much literature about this approach. When i use the ''standard'' market liqudiity at time T as dependent variable, my model performs much better (adj R^2 of around 80%).

Side question to you: The adjuster R^2 does not have to be high in order to have ''good'' model right?

[YH jordaan]

Clyde Schechter: Regarding the label option I mentioned: the variable ''datum'' is in %tq format and displays year and quarter (e.g 2001Q1). You helped me out with this in a previous forum post of me. However, when I use this in my regression by incorperating i.datum, the dates appear as integers in my regression results.

[Carlo Lazzaro]

YH:
thanks for fixing the formatting issues and providing further details.
You differenced the depvar but not the indepvars: perhaps the problem rests on this methodological choice (which is not backed up by the existing literature, if I understand your statemente correctly).
Perhaps a difference in difference approach might sound more reasonable, but I'm not an expert in your research field.
The outcome of your model seems to point out that -datum- an -gvkey- are far from being significant; hence, even though ruling out a predictors on its face-value significance only is rarely the way to go, in this instance it is probably worth thinking it over and switch towards something more parsimonious..
As far as your side question is concerned, you're right. There's no need for R2 and adjusted R2 (which, unlike R2, is useful for models comparisons) to be "that high" (whatever if means): a usual reference standard is their value in similar regression models focused on the same research topic.

[Clyde Schechter]

So, yes, to get the regression output to show the quarterly dates as quarterly dates instead of plain numbers, you need to construct a value label and apply it to the variable datum. Here's some code that can do that:

levels of datum, local(data)
capture label drop datum
foreach d of local data {
     label define datum `d' `"`:display %tq `d''"',  modify
}
label values datum datum

Now if you do a regression with i.datum Stata will show them as quarterly dates.

[YH jordaan]

Thank you both for your input.

The results presented above changed a lot when I deleted Dcrisisdummy (so that is the quarterly difference in a crisis period dummy) to just the ''normal'' crisis dummy variable.

However, last question regarding this topic to be 100% sure: One can see that particular date fixed effects appear to be significant, while others are not. Does the same analogy here apply that one should not exclude variables on the basis of significance? Or is this an other case when including time fixed effects, because one suspect a time fixed effects on their dependentt variable at a particular time (and then it does not make sense to include also other time fixed effects)?

[Carlo Lazzaro]

YH:
if I get your question right, variation between some years may turn out significant, whereas the opposite may creep up for other years. I would consider that as possible regression outcome would not worry about it.

[YH jordaan]

Alright, but should I also exclude those insignificant time effects from the regression equation? I.e. not presenting them in the results of my thesis? Or does the same general rule here apply: don't delete variables just because they turn out to be insignificant?

[Carlo Lazzaro]

YH:
don't delete variables just because they turn out to be insignificant!