[FernandoRios]

Yes it would
you cannot add this type of fixed effect, since they would be colinear with the County fixed effect

[Jongeun Park]

Hi FernandoRios, can I include state-year fixed effects? The TWFE version I'd like to run would look like the code below.

use https://friosavila.github.io/playingwithstata/drdid/mpdta.dta, clear

gen state=0 in 1/500
replace state=1 in 501/1000
replace state=2 in 1001/1500
replace state=3 in 1501/2000
replace state=4 in 2001/2500

bysort countyreal: gen treatment=(year>=first_treat)
replace treatment=0 if first_treat==0

reghdfe lemp treatment, absorb(countyreal state#year)

You mentioned in the previous post that the covariates are interacted with year variable. Then, can I just add "state" as covariate and let it be interacted with year? If yes, should I add "i.state" or "state"?

csdid2 lemp state, ivar(countyreal) time(year) gvar(first_treat)
csdid2 lemp i.state, ivar(countyreal) time(year) gvar(first_treat)

I also noticed csdid and csdid2 have different results with state. Is there any reason?

. csdid2 lemp state, ivar(countyreal) time(year) gvar(first_treat)
Producing Long Gaps by default
Using method dripw
----+--- 1 ---+--- 2 ---+--- 3 ---+--- 4 ---+--- 5
............
. estat event
------------------------------------------------------------------------------
             | Coefficient  Std. err.      z    P>|z|     [95% conf. interval]
-------------+----------------------------------------------------------------
     Pre_avg |   .0169285   .0173361     0.98   0.329    -.0170496    .0509065
    Post_avg |  -.1280806   .0264225    -4.85   0.000    -.1798678   -.0762935
         tm4 |   .0030356   .0248078     0.12   0.903    -.0455869     .051658
         tm3 |    .024782   .0184956     1.34   0.180    -.0114686    .0610327
         tm2 |   .0229677   .0145274     1.58   0.114    -.0055055     .051441
         tp0 |  -.0237983   .0121392    -1.96   0.050    -.0475907   -5.90e-06
         tp1 |  -.0718211   .0200747    -3.58   0.000    -.1111669   -.0324753
         tp2 |  -.2301203    .046353    -4.96   0.000    -.3209706   -.1392701
         tp3 |  -.1865828   .0478486    -3.90   0.000    -.2803644   -.0928012
------------------------------------------------------------------------------

. csdid lemp state, ivar(countyreal) time(year) gvar(first_treat)
............

. estat event
ATT by Periods Before and After treatment
Event Study:Dynamic effects
------------------------------------------------------------------------------
             | Coefficient  Std. err.      z    P>|z|     [95% conf. interval]
-------------+----------------------------------------------------------------
     Pre_avg |   .0012933   .0077151     0.17   0.867    -.0138279    .0164146
    Post_avg |  -.3417521   .0760209    -4.50   0.000    -.4907504   -.1927538
         Tm3 |    .028662   .0146971     1.95   0.051    -.0001438    .0574679
         Tm2 |  -.0018143   .0133592    -0.14   0.892    -.0279978    .0243692
         Tm1 |  -.0229677   .0145274    -1.58   0.114     -.051441    .0055055
         Tp0 |  -.0350765   .0155179    -2.26   0.024     -.065491   -.0046619
         Tp1 |  -.1653035   .0504029    -3.28   0.001    -.2640915   -.0665156
         Tp2 |  -.6531053   .1461919    -4.47   0.000    -.9396362   -.3665744
         Tp3 |  -.5135231   .1376213    -3.73   0.000     -.783256   -.2437903
------------------------------------------------------------------------------

[FernandoRios]

a) No you cannot add State Fixed effects, because those are already collinear with County Fixed effects.
b) Because they are colinear, they are "affecting" how the Prppensity score is being estimated. (CSDID and CSDID2 try to do this differently)
c) you cannot use "state" alone, because that is a categorical variable not continuous.
F

[Jongeun Park]

a) No you cannot add State Fixed effects, because those are already collinear with County Fixed effects.
b) Because they are colinear, they are "affecting" how the Prppensity score is being estimated. (CSDID and CSDID2 try to do this differently)
c) you cannot use "state" alone, because that is a categorical variable not continuous.
F

Thanks for the reply. I meant state-year fixed effects, not just state fixed effects. I'd like to add countyreal and i.state#i.year fixed effects instead of countyreal and year fixed effects. Is this still not working?

[FernandoRios]

Same reson
you cant add that interaction

[Jongeun Park]

Same reson
you cant add that interaction

I get it. Thank you very much for your help.

[OLivia Jie Liu]

Hi @FernandoRios, I have two questions about csdid and csdid2. I appreciate your help.

1. My tvar and gvar are like this:

tab year implementation_time

           |                        implementation_time
      year |      2008       2009       2010       2011       2012       2013 |     Total
-----------+------------------------------------------------------------------+----------
      2011 |       123        202        314        227        252        138 |    10,942 
      2013 |       152        224        272        202        262        160 |    11,622 
      2015 |       140        255        300        243        249        179 |    12,068 
      2018 |       149        242        314        263        268        181 |    12,407 
      2020 |       151        253        297        253        265        171 |    12,098 
-----------+------------------------------------------------------------------+----------
     Total |       715      1,176      1,497      1,188      1,296        829 |    59,137 


           |                        implementation_time
      year |      2014       2015       2016       2017       2018       2019 |     Total
-----------+------------------------------------------------------------------+----------
      2011 |       267      1,837        796      5,425        407        522 |    10,942 
      2013 |       342      1,954        874      5,727        424        521 |    11,622 
      2015 |       342      2,020        895      5,850        462        574 |    12,068 
      2018 |       358      1,946        935      6,110        481        609 |    12,407 
      2020 |       335      2,017        887      5,973        489        589 |    12,098 
-----------+------------------------------------------------------------------+----------
     Total |     1,644      9,774      4,387     29,085      2,263      2,815 |    59,137 


           | implementa
           | tion_time
      year |      2020 |     Total
-----------+-----------+----------
      2011 |       432 |    10,942 
      2013 |       508 |    11,622 
      2015 |       559 |    12,068 
      2018 |       551 |    12,407 
      2020 |       418 |    12,098 
-----------+-----------+----------
     Total |     2,468 |    59,137

csdid doesn't work for this dataset because years are not equally spaced. So I was thinking about doing some change to the year and implementation time to make year evenly spaced. This is an example of what I'm doing:

replace year=2001 if year==2011
replace year=2003 if year==2013
replace year=2005 if year==2015
replace year=2007 if year==2018
replace year=2009 if year==2020

tab year implementation_time
replace implementation_time=2001 if implementation_time<=2011
replace implementation_time=2002 if implementation_time==2012
replace implementation_time=2003 if implementation_time==2013
replace implementation_time=2004 if implementation_time==2014
replace implementation_time=2005 if implementation_time==2015

replace implementation_time=2006 if implementation_time==2016
replace implementation_time=2006 if implementation_time==2017
replace implementation_time=2007 if implementation_time==2018
replace implementation_time=2008 if implementation_time==2019
replace implementation_time=2009 if implementation_time==2020

I was wondering if this method is right, or is there any other method to deal with this problem?

2. Actually I found csdid2 works for the original dataset (with unequally spaced years) and produces results. However, when I change the year and implementation time (using the code in the first question), I found that csdid and csdid2 produce different results. That's kinda confusing because I think csdid and csdid2 are just different in speed. So I'm also confusing if the result I obtained using csdid2 and the original dataset (with unequally spaced years) is right and can be interpreted.

Below is the result produced by csdid:

. csdid retire $x, ivar(ID) time(year) gvar(implementation_time) method(dripw)  notyet  cluster(CITY) agg(simple) 
Units always treated found. These will be ignored
Panel is not balanced
Will use observations with Pair balanced (observed at t0 and t1)
.xxx...x...x...x...x.xxx.xxxxxxx
Difference-in-difference with Multiple Time Periods

                                                        Number of obs = 13,858
Outcome model  : least squares
Treatment model: inverse probability
                                 (Std. err. adjusted for 100 clusters in CITY)
------------------------------------------------------------------------------
             | Coefficient  Std. err.      z    P>|z|     [95% conf. interval]
-------------+----------------------------------------------------------------
         ATT |      0.034      0.030     1.14   0.254       -0.025       0.093
------------------------------------------------------------------------------
Control: Not yet Treated

See Callaway and Sant'Anna (2021) for details

. estat group 
ATT by group
------------------------------------------------------------------------------
             | Coefficient  Std. err.      z    P>|z|     [95% conf. interval]
-------------+----------------------------------------------------------------
    GAverage |      0.052      0.032     1.63   0.103       -0.011       0.115
       G2002 |      0.098      0.010     9.89   0.000        0.079       0.118
       G2003 |     -0.066      0.038    -1.73   0.083       -0.142       0.009
       G2004 |      0.110      0.057     1.94   0.053       -0.001       0.220
       G2005 |     -0.032      0.048    -0.66   0.506       -0.127       0.063
       G2006 |      0.086      0.049     1.75   0.080       -0.010       0.183
------------------------------------------------------------------------------

. estat calendar
ATT by Calendar Period
------------------------------------------------------------------------------
             | Coefficient  Std. err.      z    P>|z|     [95% conf. interval]
-------------+----------------------------------------------------------------
    CAverage |      0.028      0.025     1.14   0.256       -0.020       0.076
       T2003 |      0.095      0.020     4.64   0.000        0.055       0.135
       T2005 |     -0.079      0.051    -1.54   0.123       -0.179       0.021
       T2007 |      0.068      0.034     2.02   0.044        0.002       0.133
------------------------------------------------------------------------------

This is the result produced by csdid2:

. csdid2 retire $x, ivar(ID) tvar(year) gvar(implementation_time) method(dripw)  notyet  cluster(CITY)  agg(simple)
Producing Long Gaps by default
Using method dripw
Panel is not balanced
Will use observations with Pair balanced (observed at t0 and t1)
----+--- 1 ---+--- 2 ---+--- 3 ---+--- 4 ---+--- 5 
x.xxxxxxxx.xx.x.xxxxx.x.xx.xxxxxxx.xxxx
Difference-in-difference with Multiple Time Periods
Outcome model  : least squares
Treatment model: inverse probability
------------------------------------------------------------------------------
             |               Robust
             | Coefficient  std. err.      z    P>|z|     [95% conf. interval]
-------------+----------------------------------------------------------------
   SimpleATT |      0.096      0.040     2.42   0.016        0.018       0.174
------------------------------------------------------------------------------

. estat group 
------------------------------------------------------------------------------
             | Coefficient  Std. err.      z    P>|z|     [95% conf. interval]
-------------+----------------------------------------------------------------
    GAverage |      0.094      0.042     2.22   0.026        0.011       0.177
       g2002 |      0.097      0.010     9.78   0.000        0.078       0.117
       g2004 |      0.128      0.061     2.08   0.037        0.007       0.248
       g2006 |      0.092      0.049     1.87   0.062       -0.005       0.188
------------------------------------------------------------------------------

. estat calendar
------------------------------------------------------------------------------
             | Coefficient  Std. err.      z    P>|z|     [95% conf. interval]
-------------+----------------------------------------------------------------
    TAverage |      0.079      0.026     3.00   0.003        0.027       0.131
       t2003 |      0.097      0.010     9.78   0.000        0.078       0.117
       t2005 |      0.040      0.065     0.60   0.546       -0.089       0.168
       t2007 |      0.100      0.045     2.23   0.026        0.012       0.188
------------------------------------------------------------------------------

[FernandoRios]

So I think there will be something else going on
first it’s better if you show the full results rather than the aggregates to see which effects are estimated
second it may be that your covariates are collinear which causes the difference between both commanda

[OLivia Jie Liu]

So I think there will be something else going on
first it’s better if you show the full results rather than the aggregates to see which effects are estimated
second it may be that your covariates are collinear which causes the difference between both commanda

Thanks for your reply! So is it reliable if I just use the results from csdid2 and the original dataset? This is the full results:

. csdid retire $x, ivar(ID) time(year) gvar(implementation_time) method(dripw)  notyet  cluster(CITY) 
Units always treated found. These will be ignored
Panel is not balanced
Will use observations with Pair balanced (observed at t0 and t1)
.xxx...x...x...x...x.xxx.xxxxxxx
Difference-in-difference with Multiple Time Periods

                                                        Number of obs = 13,858
Outcome model  : least squares
Treatment model: inverse probability
                                 (Std. err. adjusted for 100 clusters in CITY)
------------------------------------------------------------------------------
             | Coefficient  Std. err.      z    P>|z|     [95% conf. interval]
-------------+----------------------------------------------------------------
g2002        |
 t_2001_2003 |      0.098      0.010     9.89   0.000        0.079       0.118
 t_2001_2005 |      0.000  (omitted)
 t_2001_2007 |      0.000  (omitted)
 t_2001_2009 |      0.000  (omitted)
-------------+----------------------------------------------------------------
g2003        |
 t_2001_2003 |      0.089      0.050     1.78   0.075       -0.009       0.186
 t_2001_2005 |     -0.275      0.057    -4.80   0.000       -0.388      -0.163
 t_2001_2007 |     -0.276      0.148    -1.86   0.063       -0.567       0.015
 t_2001_2009 |      0.000  (omitted)
-------------+----------------------------------------------------------------
g2004        |
 t_2001_2003 |     -0.187      0.072    -2.60   0.009       -0.328      -0.046
 t_2003_2005 |      0.040      0.065     0.60   0.546       -0.089       0.168
 t_2003_2007 |      0.169      0.081     2.07   0.038        0.009       0.328
 t_2003_2009 |      0.000  (omitted)
-------------+----------------------------------------------------------------
g2005        |
 t_2001_2003 |      0.037      0.041     0.89   0.372       -0.044       0.117
 t_2003_2005 |     -0.085      0.056    -1.51   0.130       -0.195       0.025
 t_2003_2007 |      0.013      0.051     0.26   0.798       -0.086       0.112
 t_2003_2009 |      0.000  (omitted)
-------------+----------------------------------------------------------------
g2006        |
 t_2001_2003 |     -0.008      0.032    -0.25   0.805       -0.070       0.054
 t_2003_2005 |     -0.004      0.044    -0.10   0.919       -0.091       0.082
 t_2005_2007 |      0.086      0.049     1.75   0.080       -0.010       0.183
 t_2005_2009 |      0.000  (omitted)
-------------+----------------------------------------------------------------
g2007        |
 t_2001_2003 |      0.001      0.095     0.01   0.988       -0.186       0.188
 t_2003_2005 |      0.000  (omitted)
 t_2005_2007 |      0.000  (omitted)
 t_2005_2009 |      0.000  (omitted)
-------------+----------------------------------------------------------------
g2008        |
 t_2001_2003 |     -0.030      0.073    -0.42   0.676       -0.173       0.112
 t_2003_2005 |      0.000  (omitted)
 t_2005_2007 |      0.000  (omitted)
 t_2007_2009 |      0.000  (omitted)
-------------+----------------------------------------------------------------
g2009        |
 t_2001_2003 |      0.000  (omitted)
 t_2003_2005 |      0.000  (omitted)
 t_2005_2007 |      0.000  (omitted)
 t_2007_2009 |      0.000  (omitted)
------------------------------------------------------------------------------
Control: Not yet Treated

. csdid2 retire $x, ivar(ID) tvar(year) gvar(implementation_time) method(dripw)  notyet  cluster(CITY)  
Producing Long Gaps by default
Using method dripw
Panel is not balanced
Will use observations with Pair balanced (observed at t0 and t1)
----+--- 1 ---+--- 2 ---+--- 3 ---+--- 4 ---+--- 5 
x.xxxxxxxx.xx.x.xxxxx.x.xx.xxxxxxx.xxxx

. estat attgt
------------------------------------------------------------------------------
             | Coefficient  Std. err.      z    P>|z|     [95% conf. interval]
-------------+----------------------------------------------------------------
g2002        |
  t2001_2003 |      0.097      0.010     9.78   0.000        0.078       0.117
-------------+----------------------------------------------------------------
g2004        |
  t2001_2003 |      0.187      0.072     2.60   0.009        0.046       0.328
  t2003_2005 |      0.040      0.065     0.60   0.546       -0.089       0.168
  t2003_2007 |      0.200      0.083     2.43   0.015        0.039       0.362
-------------+----------------------------------------------------------------
g2006        |
  t2001_2005 |      0.034      0.044     0.76   0.448       -0.053       0.121
  t2003_2005 |      0.002      0.045     0.06   0.955       -0.085       0.090
  t2005_2007 |      0.092      0.049     1.87   0.062       -0.005       0.188
-------------+----------------------------------------------------------------
g2008        |
  t2003_2007 |     -0.220      0.099    -2.22   0.026       -0.414      -0.026
------------------------------------------------------------------------------

[FernandoRios]

Ok csdid2 is working better than csdid
first your data is problematic because it seems to have a gap of two years but you see a group treated in an even year. That is messing things up internally
what I would do is look why some groups are not estimated, use fewer covariates and check with the simplest method (reg)

[OLivia Jie Liu]

Ok csdid2 is working better than csdid
first your data is problematic because it seems to have a gap of two years but you see a group treated in an even year. That is messing things up internally
what I would do is look why some groups are not estimated, use fewer covariates and check with the simplest method (reg)

Thanks! That's helpful. Reg does produces more results.

Also, I deleted some covariates, and found more results.

1. This is the result before deleting some covariates:

. csdid2 retire $x, ivar(ID) tvar(year) gvar(implementation_time) method(dripw)  notyet  cluster(CITY) agg(attgt)
Producing Long Gaps by default
Using method dripw
Always Treated units have been excluded
Panel is not balanced
Will use observations with Pair balanced (observed at t0 and t1)
----+--- 1 ---+--- 2 ---+--- 3 ---+--- 4 ---+--- 5 
x.xxxxxxxxx.xx.xx.xxxxx.x.xxxxxxxxxxxxxxx.xxxxx
Difference-in-difference with Multiple Time Periods
Outcome model  : least squares
Treatment model: inverse probability
------------------------------------------------------------------------------
             |               Robust
             | Coefficient  std. err.      z    P>|z|     [95% conf. interval]
-------------+----------------------------------------------------------------
g2012        |
  t2011_2013 |      0.097      0.010     9.78   0.000        0.078       0.117
-------------+----------------------------------------------------------------
g2014        |
  t2011_2013 |      0.187      0.072     2.60   0.009        0.046       0.328
  t2013_2015 |      0.040      0.065     0.60   0.546       -0.089       0.168
  t2013_2018 |      0.200      0.083     2.43   0.015        0.039       0.362
-------------+----------------------------------------------------------------
g2016        |
  t2011_2015 |     -0.028      0.044    -0.64   0.524       -0.115       0.059
  t2013_2015 |      0.011      0.025     0.44   0.657       -0.038       0.061
-------------+----------------------------------------------------------------
g2019        |
  t2013_2018 |     -0.220      0.099    -2.22   0.026       -0.414      -0.026
------------------------------------------------------------------------------

This is the results after deleting some covariates:

. csdid2 retire $x, ivar(ID) tvar(year) gvar(implementation_time) method(dripw)  notyet  cluster(CITY) agg(attgt)
Producing Long Gaps by default
Using method dripw
Always Treated units have been excluded
Panel is not balanced
Will use observations with Pair balanced (observed at t0 and t1)
----+--- 1 ---+--- 2 ---+--- 3 ---+--- 4 ---+--- 5 
x.x.xx.xxxxx.xx.xx.xxxxx.x.xxx.xxxxxxxxx.x.x.xxx
Difference-in-difference with Multiple Time Periods
Outcome model  : least squares
Treatment model: inverse probability
------------------------------------------------------------------------------
             |               Robust
             | Coefficient  std. err.      z    P>|z|     [95% conf. interval]
-------------+----------------------------------------------------------------
g2012        |
  t2011_2013 |      0.094      0.011     8.33   0.000        0.072       0.116
  t2011_2015 |     -0.000      0.019    -0.01   0.990       -0.037       0.037
  t2011_2018 |      0.243      0.068     3.55   0.000        0.109       0.377
-------------+----------------------------------------------------------------
g2014        |
  t2011_2013 |      0.195      0.070     2.78   0.005        0.058       0.332
  t2013_2015 |      0.039      0.067     0.58   0.559       -0.093       0.171
  t2013_2018 |      0.148      0.070     2.11   0.035        0.010       0.285
-------------+----------------------------------------------------------------
g2016        |
  t2011_2015 |     -0.026      0.038    -0.68   0.498       -0.101       0.049
  t2013_2015 |      0.009      0.025     0.37   0.709       -0.039       0.058
  t2015_2018 |      0.203      0.066     3.07   0.002        0.073       0.332
-------------+----------------------------------------------------------------
g2019        |
  t2011_2018 |     -0.289      0.074    -3.89   0.000       -0.435      -0.144
  t2013_2018 |     -0.244      0.077    -3.15   0.002       -0.396      -0.093
  t2015_2018 |      0.253      0.106     2.39   0.017        0.046       0.461
------------------------------------------------------------------------------

.

Does this mean that those deleted covariates are collinear and affect the results?

2. My first_treat occurs in 2008-2020, with 1 year interval. But I found it only has results for g2012, 2014, 2016, 2018. Does tihis mean other years are not estimated (omitted)? If I want to include other gvar years, can I try to change year 2013 to 2012, 2015 to 2014, 2017 and 2018 to 2016, 2020 to 2019 so that it can estimate all the groups?

Below is my tvar and gvar:

. tab year implementation_time

           |                                              implementation_time
      year |      2008       2009       2010       2011       2012       2013       2014       2015       2016       2017 |     Total
-----------+--------------------------------------------------------------------------------------------------------------+----------
      2011 |       123        202        314        227        252        138        267      1,837        796      5,425 |    10,942 
      2013 |       152        224        272        202        262        160        342      1,954        874      5,727 |    11,622 
      2015 |       140        255        300        243        249        179        342      2,020        895      5,850 |    12,068 
      2018 |       149        242        314        263        268        181        358      1,946        935      6,110 |    12,407 
      2020 |       151        253        297        253        265        171        335      2,017        887      5,973 |    12,098 
-----------+--------------------------------------------------------------------------------------------------------------+----------
     Total |       715      1,176      1,497      1,188      1,296        829      1,644      9,774      4,387     29,085 |    59,137 


           |       implementation_time
      year |      2018       2019       2020 |     Total
-----------+---------------------------------+----------
      2011 |       407        522        432 |    10,942 
      2013 |       424        521        508 |    11,622 
      2015 |       462        574        559 |    12,068 
      2018 |       481        609        551 |    12,407 
      2020 |       489        589        418 |    12,098 
-----------+---------------------------------+----------
     Total |     2,263      2,815      2,468 |    59,137

[Richard Thomas Boylan]

The csdid2_plot is hard to read when printing it on a black and white page. Any suggestions on how to make it more readable?
Here is an example.
use https://friosavila.github.io/playing...rdid/mpdta.dta, clear
csdid2 lemp lpop , ivar(countyreal) time(year) gvar(first_treat) method(dripw)
estat event
csdid2_plot, scheme(s1mono)

Now, this graph is readable on the screen. However, when it is printed out, it is only barely readable.

[FernandoRios]

try playing with "scale" option
csdid2_plot, scheme(s1mono) scale(1.5)

[Richard Thomas Boylan]

This is better, thanks!

[OLivia Jie Liu]

Hi my data has gvar from 2008 to 2020, and tvar in 2011, 2013, 2015, 2018, 2020. After deleting data when tvar=2020 and changing gvar to 0 if it is larger than 2018, I found the agg(simple) results are exactly the same. After deleting 2020 data, the attgt also does not estimate cohort 2019. I'm confused why deleting the data in the whole year would not affect the result of agg(simple).This is the attgt result before deleting 2020 data:

. estat attgt
------------------------------------------------------------------------------
             | Coefficient  Std. err.      z    P>|z|     [95% conf. interval]
-------------+----------------------------------------------------------------
g2012        |
  t2011_2013 |      0.094      0.011     8.33   0.000        0.072       0.116
  t2011_2015 |     -0.000      0.019    -0.01   0.990       -0.037       0.037
  t2011_2018 |      0.243      0.068     3.55   0.000        0.109       0.377
-------------+----------------------------------------------------------------
g2014        |
  t2011_2013 |      0.195      0.070     2.78   0.005        0.058       0.332
  t2013_2015 |      0.039      0.067     0.58   0.559       -0.093       0.171
  t2013_2018 |      0.148      0.070     2.11   0.035        0.010       0.285
-------------+----------------------------------------------------------------
g2016        |
  t2011_2015 |     -0.026      0.038    -0.68   0.498       -0.101       0.049
  t2013_2015 |      0.009      0.025     0.37   0.709       -0.039       0.058
  t2015_2018 |      0.203      0.066     3.07   0.002        0.073       0.332
-------------+----------------------------------------------------------------
g2019        |
  t2011_2018 |     -0.289      0.074    -3.89   0.000       -0.435      -0.144
  t2013_2018 |     -0.244      0.077    -3.15   0.002       -0.396      -0.093
  t2015_2018 |      0.253      0.106     2.39   0.017        0.046       0.461
------------------------------------------------------------------------------

And this is the result after deleting:

. estat attgt
------------------------------------------------------------------------------
             | Coefficient  Std. err.      z    P>|z|     [95% conf. interval]
-------------+----------------------------------------------------------------
g2012        |
  t2011_2013 |      0.094      0.011     8.33   0.000        0.072       0.116
  t2011_2015 |     -0.000      0.019    -0.01   0.990       -0.037       0.037
  t2011_2018 |      0.243      0.068     3.55   0.000        0.109       0.377
-------------+----------------------------------------------------------------
g2014        |
  t2011_2013 |      0.195      0.070     2.78   0.005        0.058       0.332
  t2013_2015 |      0.039      0.067     0.58   0.559       -0.093       0.171
  t2013_2018 |      0.148      0.070     2.11   0.035        0.010       0.285
-------------+----------------------------------------------------------------
g2016        |
  t2011_2015 |     -0.026      0.038    -0.68   0.498       -0.101       0.049
  t2013_2015 |      0.009      0.025     0.37   0.709       -0.039       0.058
  t2015_2018 |      0.203      0.066     3.07   0.002        0.073       0.332
------------------------------------------------------------------------------