Extended TWFE

[erinlbritton]

Here's how to estimate treatment effects using extended TWFE (Wooldridge, 2021):

* gen time-varying treatment indicator
gen w = D // time-varying treatment indicator

tab w t // should be 35 units treated at t=7 and 35 treated at t=11

* gen ever-treated cohort indicators
egen d = max(w), by(i)
sort i t

gen dinf = (d==0) // d-infinity = never-treated cohort
gen d7 = (Ei==7)
gen d11 = (Ei==11)

* gen t dummies
foreach t in 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 {
	gen t`t' = (t==`t')
}

* estimate TE using Wooldridge ETWFE
* TEs vary by entry cohort and time period:
reg Y c.d7#c.t7 c.d7#c.t8 c.d7#c.t9 c.d7#c.t10 c.d7#c.t11 c.d7#c.t12 c.d7#c.t13 c.d7#c.t14 c.d7#c.t15 ///
	c.d11#c.t11 c.d11#c.t12 c.d11#c.t13 c.d11#c.t14 c.d11#c.t15 ///
	d7 d11 i.t, cluster(i)

* ATTs by cohort
lincom (c.d7#c.t7+c.d7#c.t8+c.d7#c.t9+c.d7#c.t10+c.d7#c.t11+c.d7#c.t12+c.d7#c.t13+c.d7#c.t14+c.d7#c.t15)/9 // ATT(g=7)
lincom (c.d11#c.t11+c.d11#c.t12+c.d11#c.t13+c.d11#c.t14+c.d11#c.t15)/5 // ATT(g=11)

* ATTs by event-time
lincom (c.d7#c.t7+c.d11#c.t11)/2 // 1 period since event
lincom (c.d7#c.t8+c.d11#c.t12)/2 // 2 periods since event
lincom (c.d7#c.t9+c.d11#c.t13)/2 // 3 periods since event
lincom (c.d7#c.t10+c.d11#c.t14)/2 // 4 periods since event
lincom (c.d7#c.t11+c.d11#c.t15)/2 // 5 periods since event
lincom (c.d7#c.t12) // 6 periods since event
lincom (c.d7#c.t13) // 7 periods since event
lincom (c.d7#c.t14) // 8 periods since event
lincom (c.d7#c.t15) // 9 periods since event

* compare to CS
csdid Y, ivar(i) time(t) gvar(gvar) notyet
estat group
estat event