Interpretation of time fixed effects ( time dummies )

[Lamie Akindele]

Hi Statalisters!!

Please I would like you to help with a certain problem I have : I noticed that two time variables 2007&2011 are significant in my regression (i'm guessing due to the financial and Eurozone crisis). But how do I interpret the coefficient and sign on each time variable for my paper and also do i mentoned the R-sq (im a bit confused as to its validity in this fixed effect model.

xtreg $ylist ubdur replacementrate uegen unionden  lmp_exp tax_wedge i.year, fe

Fixed-effects (within) regression Number of obs = 107
Group variable: id Number of groups = 10

R-sq: within = 0.8373 Obs per group: min = 7
between = 0.0185 avg = 10.7
overall = 0.0933 max = 12

F(17,80) = 24.21
corr(u_i, Xb) = -0.7157 Prob > F = 0.0000

---------------------------------------------------------------------------------
une_rt_a | Coef. Std. Err. t P>|t| [95% Conf. Interval]
----------------+----------------------------------------------------------------
ubdur | .0178783 .0093775 1.91 0.060 -.0007835 .0365401
replacementrate | -11.31995 7.618282 -1.49 0.141 -26.48081 3.84092
uegen | -1.547902 .4918411 -3.15 0.002 -2.526697 -.5691071
unionden | -.1612843 .1412985 -1.14 0.257 -.4424773 .1199086
lmp_exp | 4.840894 .3651561 13.26 0.000 4.11421 5.567578
tax_wedge | -.0481105 .1508529 -0.32 0.751 -.3483174 .2520964
|
year |
2001 | -.5060035 .5743391 -0.88 0.381 -1.648975 .6369677
2002 | -.2499434 .606667 -0.41 0.681 -1.457249 .9573623
2003 | -.2116938 .6104098 -0.35 0.730 -1.426448 1.00306
2004 | .11122 .5931715 0.19 0.852 -1.069229 1.291669
2005 | .3910291 .6053263 0.65 0.520 -.8136087 1.595667
2006 | .9255652 .6233933 1.48 0.142 -.3150269 2.166157
2007 | 1.151597 .626483 1.84 0.070 -.0951442 2.398338
2008 | .7085658 .6347024 1.12 0.268 -.5545322 1.971664
2009 | .2741273 .6757321 0.41 0.686 -1.070622 1.618877
2010 | 1.163628 .6834175 1.70 0.093 -.1964165 2.523672
2011 | 1.940984 .6812735 2.85 0.006 .5852064 3.296761
|
_cons | 23.06657 5.909102 3.90 0.000 11.30709 34.82606
----------------+----------------------------------------------------------------------------------------
sigma_u | 4.0626822
sigma_e | 1.0688825
rho | .93526085 (fraction of variance due to u_i)
F test that all u_i=0: F(9, 80) = 19.08 Prob > F = 0.0000









Thank you,

Lamie

[Carlo Lazzaro]

Lamie:
as far as I can see from your output (which s hardly readable; as per FAQ, please post it via code delimiters,.Thanks), the only year that reaches statistical significance is 2011.
What's your time-serie dimension in -xtset-?

[Lamie Akindele]

The time series dimension is :

panel variable: id (strongly balanced)
time variable: year, 2000 to 2011
delta: 1 year
Yes and sorry my bad i have tried to make it a bit neater, also I am unsure how to use the code delimiters

xtreg $ylist ubdur replacementrate uegen unionden  lmp_exp tax_wedge i.year, fe

Fixed-effects (within) regression               Number of obs      =       107
Group variable: id                              Number of groups   =        10

R-sq:  within  = 0.8373                         Obs per group: min =         7
       between = 0.0185                                        avg =      10.7
       overall = 0.0933                                        max =        12

                                                F(17,80)           =     24.21
corr(u_i, Xb)  = -0.7157                        Prob > F           =    0.0000

---------------------------------------------------------------------------------
       une_rt_a |      Coef.   Std. Err.      t    P>|t|     [95% Conf. Interval]
----------------+----------------------------------------------------------------
          ubdur |   .0178783   .0093775     1.91   0.060    -.0007835    .0365401
replacem |      -11.31995    7.618282    -1.49   0.141    -26.48081     3.84092
          uegen | -1.547902   .4918411    -3.15   0.002    -2.526697   -.5691071
       unionden |  -.1612843   .1412985   -1.14   0.257    -.4424773    .1199086
        lmp_exp |   4.840894   .3651561    13.26   0.000      4.11421    5.567578
      tax_wedge |  -.0481105   .1508529    -0.32   0.751    -.3483174    .2520964
                    |
           year  |    Coef.             Std. Err.      t          P>|t|     [95% Conf. Interval]
          2001  |  -.5060035   .5743391    -0.88      0.381    -1.648975    .6369677
          2002  |  -.2499434    .606667    -0.41       0.681    -1.457249     .9573623
          2003  |  -.2116938   .6104098    -0.35      0.730    -1.426448     1.00306
          2004  |     .11122   .5931715       0.19      0.852    -1.069229    1.291669
          2005  |   .3910291   .6053263     0.65     0.520    -.8136087    1.595667
          2006  |   .9255652   .6233933     1.48     0.142    -.3150269    2.166157
          2007  |   1.151597    .626483     1.84      0.070    -.0951442     2.398338
          2008  |   .7085658   .6347024     1.12      0.268    -.5545322    1.971664
          2009  |   .2741273   .6757321     0.41      0.686    -1.070622    1.618877
          2010  |   1.163628   .6834175     1.70      0.093    -.1964165    2.523672
          2011  |   1.940984   .6812735     2.85       0.006     .5852064    3.296761
                    |
          _cons |   23.06657   5.909102     3.90   0.000     11.30709    34.82606
----------------+----------------------------------------------------------------
        sigma_u |  4.0626822
        sigma_e |  1.0688825
            rho |  .93526085   (fraction of variance due to u_i)
---------------------------------------------------------------------------------
F test that all u_i=0:     F(9, 80) =    19.08               Prob > F = 0.0000

please is this much better, also please how do i interpret these time figures. ( are 2007, 2010 also significant ?)


Thank you,
Lamie

[Carlo Lazzaro]

Lamie.
thanks for giving it a try with code delimiters.
Again, as far as I can see, only 2011 seems significant. Anyway, I would get rid of of i.year, also because you have already included years in -xtset-.
Deleting i.year from the set of your predictors, will also shelter you from overfitting your regression.

[Phil Bromiley]

Carlo.

I thought that xtreg did not automatically includes time effects. My reading of the equations in the manual for xtreg is that it includes panel effects but not time [xtreg, Remarks and examples section, equation (1)]. If it includes time effects, then @Lamie should get nothing when he puts in i.year in just as Lamie would get nothing if he or she put in i.id (panel dummies).

What am I missing?

Phil

[Carlo Lazzaro]

Phil is correct (http://www.statalist.org/forums/foru...s-year-dummies) and I was too hurried in writing and careless in reading before posting.
But the issue seems to remain the same; I would trade the i.year for the risk of overfit.

[Lamie Akindele]

Hi Carlo and Phil thank you, so basically the advice is not to use the time dummies (although I have seen some papers use it in their models; country specific effects, time fixed effects, etc). Also please i wanted to ask a quick question: to estimate an OLS model with the Newey-West robust standard errors; what is this command please:

reg ylist xlist, robust

I know that this is just for the robust standard errors but please which is for Newey-West robust standard errors , which correct for both first-order autocorrelation and panel heteroscedasticity.

[Carlo Lazzaro]

Lamie:
you can also leave i.year in, provided they do not convey added value to your results.
For lag=0, the results produced by -newey- are the same reported by -regress- (as you can read in the -newey- entry in Stata .pdf manual, pages 384-385).
However, vce(robust) in -xtreg, fe- takes both heteroskedasticity and panel serial-correlation into account (as you can read in the -xtreg- entry in Stata .pdf manual, page 372).

[Anna Solovieva]

Hello all!
I am very new to stata. I am writing my master thesis and I have problems with the results. I am analyzing CEO overconfidence (dummy) and its effect on financial statements restatements (dummy). I run my regression and I get statistically significant results. However, I need to include year fixed effects. When I include them, my results turn to insignificant. How can i explain this? Should I use year fixed effects? I did Hausman test and I understood that I need to use random effects. Can you please help me what can I do in this case? Should I use year fixed effects or random effects? If I need to use random effects, should I include in this formula i.fyear? Can anyone provide me with the right formula or do I have any other problem? I really hope that someone can help me! Thank you.

Kind regards,

Anna Solovieva.

Attached you can find my results.

[Carlo Lazzaro]

Anna:
welcome to the list.
Just some general remarks:
- please read the FAQ about posting rules (among them posters are kindly requested not to post attachments like yours);
- please post what you typed and what Stata gave you back: it worths more than thousand lines devoted to explain what's the matter with your dataset/analysis;
- the fact that your results were statistical significant then and no more now, far from being a reason of concern, may depend on different causes. However, nobody knows your data but you (follow the previous remark, please);
- if you're dealing with panel data (as it would seem from your post), please take a look at the -xt- command suite in Stata .pdf manual.

[Anna Solovieva]

Thank you, Carlo, for your reply!
This was my first post, so I didn't know the rules. The results that I've got are presented below. First, the regression without year fixed effects and next, the regression with year fixed effects. My question is why these results turn to insignificant and how can I explain why they are insignificant?
Thank you.

logit RESTATE share_retainer bdsize Lev ceochair growth roa OUTSIDE fin BTM size, cl(gvkey)

Iteration 0: log pseudolikelihood = -1532.1454
Iteration 1: log pseudolikelihood = -1505.8667
Iteration 2: log pseudolikelihood = -1505.2272
Iteration 3: log pseudolikelihood = -1505.2268
Iteration 4: log pseudolikelihood = -1505.2268

Logistic regression Number of obs = 3,905
Wald chi2(10) = 50.99
Prob > chi2 = 0.0000
Log pseudolikelihood = -1505.2268 Pseudo R2 = 0.0176

(Std. Err. adjusted for 885 clusters in gvkey)
--------------------------------------------------------------------------------
| Robust
RESTATE | Coef. Std. Err. z P>|z| [95% Conf. Interval]
---------------+----------------------------------------------------------------
share_retainer | .253403 .1055007 2.40 0.016 .0466254 .4601806
bdsize | -.02591 .0309838 -0.84 0.403 -.0866372 .0348172
Lev | .9424567 .389678 2.42 0.016 .178702 1.706212
ceochair | .1031067 .1147953 0.90 0.369 -.1218879 .3281013
growth | .0179178 .1011673 0.18 0.859 -.1803665 .216202
roa | -1.999581 .5759427 -3.47 0.001 -3.128408 -.8707538
OUTSIDE | -1.508419 .3778756 -3.99 0.000 -2.249042 -.7677965
fin | -.5003123 .4563693 -1.10 0.273 -1.39478 .394155
BTM | .0206983 .2021937 0.10 0.918 -.3755941 .4169907
size | .0262115 .0510151 0.51 0.607 -.0737763 .1261993
_ cons | -1.009361 .4153236 -2.43 0.015 -1.82338 -.1953413
--------------------------------------------------------------------------------

Regression with year fixed effects.

. xtlogit RESTATE share_retainer bdsize Lev ceochair growth roa OUTSIDE fin BTM size i.fyear, fe
note: multiple positive outcomes within groups encountered.
note: 584 groups (1,894 obs) dropped because of all positive or
all negative outcomes.

Iteration 0: log likelihood = -580.817
Iteration 1: log likelihood = -556.64478
Iteration 2: log likelihood = -552.63768
Iteration 3: log likelihood = -551.86422
Iteration 4: log likelihood = -551.69452
Iteration 5: log likelihood = -551.6575
Iteration 6: log likelihood = -551.65216
Iteration 7: log likelihood = -551.6515
Iteration 8: log likelihood = -551.65137
Iteration 9: log likelihood = -551.65135

Conditional fixed-effects logistic regression Number of obs = 2,011
Group variable: gvkey Number of groups = 301

Obs per group:
min = 2
avg = 6.7
max = 12

LR chi2(21) = 248.90
Log likelihood = -551.65135 Prob > chi2 = 0.0000

--------------------------------------------------------------------------------
RESTATE | Coef. Std. Err. z P>|z| [95% Conf. Interval]
---------------+----------------------------------------------------------------
share_retainer | .2336912 .1822974 1.28 0.200 -.1236052 .5909875
bdsize | -.0253299 .0741963 -0.34 0.733 -.1707521 .1200922
Lev | .8522291 .9195946 0.93 0.354 -.9501432 2.654602
ceochair | .2490882 .1835222 1.36 0.175 -.1106088 .6087851
growth | .1267658 .1535595 0.83 0.409 -.1742053 .427737
roa | -.5597839 1.058199 -0.53 0.597 -2.633815 1.514248
OUTSIDE | -1.304018 .8189207 -1.59 0.111 -2.909073 .3010371
fin | .1619264 .6630058 0.24 0.807 -1.137541 1.461394
BTM | -.3965328 .385204 -1.03 0.303 -1.151519 .3584532
size | .1953976 .2816945 0.69 0.488 -.3567135 .7475086
|
fyear |
2001 | 14.7506 364.2983 0.04 0.968 -699.261 728.7622
2002 | 17.22636 364.2983 0.05 0.962 -696.7852 731.238
2003 | 15.23778 364.2983 0.04 0.967 -698.7738 729.2494
2004 | 15.43136 364.2983 0.04 0.966 -698.5803 729.443
2005 | 15.23058 364.2984 0.04 0.967 -698.7811 729.2422
2006 | 15.05912 364.2984 0.04 0.967 -698.9526 729.0708
2007 | 15.1631 364.2984 0.04 0.967 -698.8487 729.1749
2008 | 15.13758 364.2984 0.04 0.967 -698.8742 729.1494
2009 | 14.67802 364.2985 0.04 0.968 -699.3338 728.6899
2010 | 14.45841 364.2985 0.04 0.968 -699.5535 728.4703
2011 | 14.96391 364.2985 0.04 0.967 -699.048 728.9758
--------------------------------------------------------------------------------

.

[Carlo Lazzaro]

Anna:
thanks for providing further details (from the pedantic corner: for the future, please: start a new thread instead of queing up to a previous one; post what you typed and what Stata gave you back, as you laudably did, via CODE delimiters, please, see FAQ#12. Thanks).
Two remarks about you results:
- if you look at both # of observations, custers and groups you seem to have two different models (regression technicalities apart); hence, no surprise that you got different results;
- the choice between (pooled) -logit- and -xtlogit- may be guided by the statistical significance of the -rho- at the foot of the outcome table of .-xtlogit- (something that you did not get because it supports default standard errors only);
- looking at -xtlogit- output, including i.year does not seem to add nothing; your analysis can probably survive even without them;
- as you might already be aware of, -xtlogit, fe- estimates conditional fe (something rather different from, say, -xtreg, fe-).

[Anna Solovieva]

Hello, Carlo,
Thank you very much for your comments. I did Hausman test and I understood that I need to use random effects. However, I am confused whether I need to use in this regression of random effects year dummies? Should the regression be in this way or it is not practically useful? xtreg (variables) i.fyear, re?

Thank you again,

Anna.

[Carlo Lazzaro]

Anna:
your code for -xtreg, re- looks fine.
However, in your previous email you reported a binary dependent variable. It that still holds, there's no scope for -xtreg,re-, as -xtreg- is made for continous dependent variables.
if -hausman- outcome goes in -re- direction and your depvar is a binary one, your code should be:

xtlogit depvar indepvars i.year, re

[Anna Solovieva]

Carlo,
Thank you again for your important comments. I did the regression that you wrote me and again the coefficients are all insignificant. However, if I don't put i.fyear in the regression of xtlogit , re, they are significant. How can I explain this turn to insignificance? When should we not use year fixed effects?
You can see what I get. I tried to use this #buttom and I wrote the output of the regression in between but I am not sure if I did it in a right way.

xtlogit RESTATE share_retainer bdsize Lev ceochair growth roa OUTSIDE fin BTM size i.fyear, re

note: 2000.fyear != 0 predicts failure perfectly
      2000.fyear dropped and 139 obs not used

note: 2011.fyear omitted because of collinearity

Fitting comparison model:

Iteration 0:   log likelihood = -1511.8918  
Iteration 1:   log likelihood = -1419.8324  
Iteration 2:   log likelihood = -1376.0688  
Iteration 3:   log likelihood = -1372.4237  
Iteration 4:   log likelihood = -1372.4226  
Iteration 5:   log likelihood = -1372.4226  

Fitting full model:

tau =  0.0     log likelihood = -1372.4226
tau =  0.1     log likelihood = -1359.2847
tau =  0.2     log likelihood = -1348.8574
tau =  0.3     log likelihood = -1341.8843
tau =  0.4     log likelihood =  -1339.062
tau =  0.5     log likelihood = -1341.3903

Iteration 0:   log likelihood = -1339.0624  
Iteration 1:   log likelihood = -1326.3443  
Iteration 2:   log likelihood = -1326.2662  
Iteration 3:   log likelihood = -1326.2661  

Random-effects logistic regression              Number of obs     =      3,766
Group variable: gvkey                                Number of groups  =        879

Random effects u_i ~ Gaussian                   Obs per group:
                                                                        min =          1
                                                                       avg =        4.3
                                                                       max =         11

Integration method: mvaghermite                 Integration pts.  =         12

                                                               Wald chi2(20)     =     255.00
Log likelihood  = -1326.2661                    Prob > chi2       =     0.0000

--------------------------------------------------------------------------------
       RESTATE |      Coef.   Std. Err.      z    P>|z|     [95% Conf. Interval]
---------------+----------------------------------------------------------------
share_retainer |   .1256698   .1282727     0.98   0.327    -.1257401    .3770797
             bdsize |  -.0352695   .0386957    -0.91   0.362    -.1111116    .0405726
                  Lev |   .6192393   .4771389     1.30   0.194    -.3159358    1.554414
          ceochair |   .1206002   .1291943     0.93   0.351    -.1326159    .3738163
             growth |   .1475759   .1182969     1.25   0.212    -.0842818    .3794336
                   roa |  -.8499334    .755769    -1.12   0.261    -2.331213    .6313467
        OUTSIDE |  -.7775239   .5036085    -1.54   0.123    -1.764578    .2095306
                    fin |  -.2099515   .4637027    -0.45   0.651    -1.118792    .6988892
                BTM |  -.0113149   .2405921    -0.05   0.962    -.4828668     .460237
                s ize |   .0643763   .0602143     1.07   0.285    -.0536415    .1823941
                       |
              fyear |
             2000  |          0  (empty)
             2001  |  -.1636146   .3145838    -0.52   0.603    -.7801875    .4529583
             2002  |   2.507473    .245995    10.19   0.000     2.025332    2.989614
             2003  |   .4271213   .2634397     1.62   0.105     -.089211    .9434536
             2004  |   .5819854   .2455206     2.37   0.018     .1007738    1.063197
             2005  |   .4004141   .2445008     1.64   0.101    -.0787987    .8796269
             2006  |   .2267742   .2465361     0.92   0.358    -.2564277     .709976
             2007  |   .2225504   .2661468     0.84   0.403    -.2990877    .7441886
             2008  |   .0781869   .2508899     0.31   0.755    -.4135482     .569922
             2009  |   -.396996   .2664944    -1.49   0.136    -.9193153    .1253234
             2010  |  -.6070495   .2739569    -2.22   0.027    -1.143995   -.0701039
              2011 |          0  (omitted)
                       |
             _cons |  -2.439526   .5507618    -4.43   0.000    -3.518999   -1.360052
---------------+----------------------------------------------------------------
          /lnsig2u |   .1566606   .1777448                     -.1917128    .5050339
---------------+----------------------------------------------------------------
         sigma_u |    1.08148   .0961137                      .9085945    1.287261
                 rho |   .2622732   .0343911                      .2005981     .334965
--------------------------------------------------------------------------------
LR test of rho=0: chibar2(01) = 92.31                  Prob >= chibar2 = 0.000

Kind regards,

Anna.