Event study using "never treated" units - DID multiple periods

[Marcela Vieira]

Hello everyone,

I am applying DID with multiple periods using csdid and "never treated" units as controls. I have two questions:

1. How do the csdid estimates work when there are a few "never treated" units (e.g. 6 units) compared to "treated units" (e.g. 50)? Would the credibility of my estimates be compromised?
2. I am testing for parallel trend four periods before the first treatment took place using "estat event, window (-4 -1)"
My results are:

. estat event, window (-4 -1)
ATT by Periods Before and After treatment
Event Study:Dynamic effects
------------------------------------------------------------------------------
| Coefficient Std. err. z P>|z| [95% conf. interval]
-------------+----------------------------------------------------------------
Tm4 | .0349455 .0268908 1.30 0.194 -.0177596 .0876505
Tm3 | .0111339 .02827 0.39 0.694 -.0442742 .0665421
Tm2 | -.0125245 .0296415 -0.42 0.673 -.0706208 .0455719
Tm1 | .029379 .0433705 0.68 0.498 -.0556256 .1143836
------------------------------------------------------------------------------


Does this mean that they follow a parallel trend? P-value is not significant for all four pre-treatment periods...

Apologies for the silly questions, I am just getting familiar with the model.

Many thanks,

[FernandoRios]

Hi Marcela
My feeling is that you do have too few observations to talk about statistical inference, specially if you use asymptotic standard errors.
Perhaps, for a case like this, the Uniform band standard errors will be better.
For that, you will have to either use the RIF files, and csdid_stat event, wboot
or request the "agg(events) wboot" when calling into csdid.
HTH
Fernando

[Marcela Vieira]

Hi Fernando,

Thank you for your help!
I have done what you suggested using the following code:

csdid log_outcome if 1967<year<1977 & region=="SW" | 1967<year<1977 & region=="NA" |1967<year<1977 & region=="S" | 1967<year<1977 & region==" CA", ivar(codes) time (year) gvar(first_treat) method(drimp) [notyet] wboot agg(event)

Difference-in-difference with Multiple Time Periods

Number of obs = 1,078
Outcome model : weighted least squares
Treatment model: inverse probability tilting
----------------------------------------------------------------------
| Coefficient Std. err. t [95% conf. interval]
-------------+--------------------------------------------------------
T-13 | -.1396999 .0504126 -2.77 -.2813729 .0019731
T-12 | -.101237 .0203058 -4.99 -.1583018 -.0441722
T-11 | -.1000672 .0332995 -3.01 -.1936478 -.0064865
T-10 | .0331675 .0662329 0.50 -.1529647 .2192997
T-9 | .0598475 .0316777 1.89 -.0291753 .1488703
T-8 | .0080236 .0216322 0.37 -.0527687 .0688159
T-7 | .0131258 .0183201 0.72 -.0383586 .0646103
T-6 | .0066517 .0169129 0.39 -.0408782 .0541817
T-5 | .0472592 .0187259 2.52 -.0053657 .0998841
T-4 | .0013172 .021684 0.06 -.0596206 .062255
T-3 | .0067839 .0141842 0.48 -.0330774 .0466452
T-2 | -.0073786 .022493 -0.33 -.07059 .0558327
T-1 | .0270854 .0232495 1.16 -.0382519 .0924227
T+0 | .0139724 .0195173 0.72 -.0408765 .0688212
T+1 | .0885572 .0397497 2.23 -.0231501 .2002644
T+2 | .2047421 .0599511 3.42 .0362634 .3732207
T+3 | .2407023 .0914567 2.63 -.0163156 .4977202
T+4 | .2489177 .087475 2.85 .0030895 .4947459
T+5 | .1734025 .0773296 2.24 -.0439144 .3907193
T+6 | .2009591 .0913902 2.20 -.0558718 .45779
T+7 | .2845614 .0598582 4.75 .1163437 .452779
T+8 | .2346028 .0832782 2.82 .0005686 .468637
----------------------------------------------------------------------
Control: Not yet Treated

My questions are:

1. My panel data is from 1968 to 1976, so I thought I would get T-4 to T4. Why do I get T-13 to T+8?
2. Could I say that they follow a pre parallel trend at any period?

I am sorry if I completely misinterpreted the results...

Many thanks,

[FernandoRios]

Hi Marcela
So for question 1: WHat is happening is that your treatment may have occurred at different points in time.
Something that you can do, for example is calculate this "relative times" using the following:
gen reltime=first_treat-year if first_treat!=0
if you then look into reltime, and see for which units will be entering the T-13 or T+8.
For Question 2: Except for the t-10 or lower, i does seem to me that you have no pretrends. Im guessing you can also do "estat pretrend" or reestimate the model excluding using something like:
if inrange(reltime,-10,8) | reltime==0 In addition to the other sample restrictions. and then rerun the estat pretrend.

Hope this helps
Fernando