2

In a generalized Difference in Difference for staggered year event, from this discussion, we need to include the event year to the sample.

For example, if Korea passed the law in 1997, so post*treat should receive the value of 0 for 1995 and 1996 while it receives the value of 1 for 1997,1998,1999,2000,2001,2002.

I am wondering if I can exclude the year 1997 in this case out of my sample, is there any justification for this exclusion. I mean, post*treat equals to 0 for 1995 and 1996 while it receives value of 1 for 1998,1999,2000,2001,2002.

If this exclusion is wrong, so I am wondering why it is wrong.

Phil Nguyen
  • 539

1 Answers1

3

Unless the model is fully saturated with treatment indicators to show the full dynamic response to treatment, then I see no justifiable reason to exclude a period, let alone excluding the immediate adoption year. And even when we do include a full series of lead/lag indicators, we typically exclude a pre-treatment year. You could drop the year before treatment, or a more distant period. However, excluding the initial adoption year is very rare in practice.

I suppose if you suspect effects kick in with a one year lag, then excluding the year of adoption seems warranted, but again it's rare. Maybe it took a long time for treated firms in Korea to adapt to the new law/policy. Even then, I wouldn't recommend excluding a period unless the model was saturated. If only a finite number of pre-treatment "leads" were considered (i.e., 2 leads), then all the years before 1995 may serve as a reference.

Thomas Bilach
  • 5,999
  • 2
  • 11
  • 33
  • Thanks Thomas. However, regarding "And even when we do include a full series of lead/lag indicators, we typically exclude a pre-treatment year", is there any reference or justification for this statement then? – Phil Nguyen May 30 '21 at 23:31
  • Apart from that, you said "If only a finite number of pre-treatment "leads" were considered (i.e., 2 leads), then all the years before 1995 may serve as a reference". Can I ask what does this sentence mean, especially the word reference in this case. – Phil Nguyen May 30 '21 at 23:36
  • 2
    The justification is that we need a reference period. In most impact evaluations the authors typically drop the period immediately before treatment. You could use a more distant period but this is not as common, at least in the literature I've encountered on this topic. – Thomas Bilach Jun 01 '21 at 21:49
  • 2
    And by "reference" I mean the pre-treatment epoch. In a typical event study framework evaluators may assess effects in all periods before and after treatment. In essence, we're assessing the evolution of effects in reference to the omitted period. – Thomas Bilach Jun 01 '21 at 21:52
  • A very clear and thorough explanation, just the last question. So "we're assessing the evolution of effects in reference to the omitted period", so the reference here mean pre-treatment period and omitted period is post-treatment period? – Phil Nguyen Jun 02 '21 at 09:38
  • 2
    The period we omit is our reference period. In applied work, evaluators typically drop a pre-treatment period. The period you omit is up to you, but the period immediately before treatment starts is a popular choice. Does that make sense? – Thomas Bilach Jun 02 '21 at 21:53
  • 1
    It makes totally sense now, thank you so much – Phil Nguyen Jun 02 '21 at 22:17
  • @ThomasBilach I am interested in this discussion. I am wondering if you mean your reference period is the pre-treatment period and omitted period is the time you end your pre-treatment period? For example of OP, Korea passed the laws in 1997, so the reference period is 1995->1996, and 1996 is omitted period? – Phil Nguyen Jun 02 '21 at 23:06
  • I need more information to answer this question. Are we only considering a finite number of leads and lags? It depends on the context. – Thomas Bilach Jun 15 '21 at 16:58