That is the query posed in paper by Baker, Larcker and Wang (2022). I summarize their key arguments under.

The validity of…[the DiD]…strategy rests on the central assumption that the noticed pattern in management models’ outcomes mimic the pattern in remedy models’ outcomes had they not acquired remedy. As the authors write:

First, DiD estimates are unbiased in settings with a single remedy interval, even when there are dynamic remedy results. Second, DiD estimates are additionally unbiased in settings with staggered timing of remedy task and homogeneous remedy impact throughout corporations and over time. Finally, when analysis settings mix staggered timing of remedy results and remedy impact heterogeneity, staggered DiD estimates are seemingly biased.

Oftentimes, DiD is applied utilizing an extraordinary least squares (OLS) regression based mostly mannequin as follows:

When there are greater than two teams and greater than and a pair of time intervals, regression-based DiD fashions sometimes depend on two-way fastened impact (TWFE) of the shape:

Where the primary two coefficients are unit and time interval
fastened results. Note that earlier analysis from Goodman-Bacon
(2021) exhibits that static types of the TWFE DiD is definitely a “weighted
common of all potential two-group/two-period DiD estimators within the knowledge.”

When remedy results can change over time (“dynamic
remedy results”), staggered DiD remedy impact estimates can really
receive the other signal of the true ATT, even when the researcher had been in a position to
randomize remedy task (thus the place the parallel-trends assumption
holds).

The motive for it is because Goodman-Bacon
(2021) exhibits that the static TWFE DiD is definitely consists of three elements:

• Variance-weighted common remedy impact on
  the handled (VWATT)
• Variance-weighted common counterfactual developments
  (VWCT)
• Weighted sum of the change within the common
  remedy on the handled inside a treatment-timing group’s post-period and
  round a later-treated unit’s remedy window (ΔATT)

The first time period is the time period of curiosity.  If the parallel developments happens, then VWCT =0.  The final time period arises as a result of, beneath static
TWFE DiD, already-treated teams as successfully used as comparability teams for later-treated
teams.  If DiD is estimated in a
two-period mannequin, nevertheless, this time period disappears and there’s no bias. Alternatively,
if remedy results are static (i.e., not altering over time after the
intervention), then ΔATT = 0 and TWFE DiD is legitimate. 

The challenges, nevertheless, happens when remedy results are
dynamic.  In this case ΔATT
≠
0 and the TWFE DiD is biased.

So what may be carried out? The authors provide 3 options:

• Callaway and Santa’Anna (2021). Here, the authors permit one to estimate remedy impact for a specific group (remedy at time g) utilizing observations at time τ and g-1 from a clear set of controls.  These are principally not-yet handled, last-treated, or never-treated teams. 
• Sun and Abraham (2021).  The same methodology is used as in CS, however always-treated models are dropped, and the one models that can be utilized as efficient controls are these which are never-treated or last-treated. Further, this strategy is absolutely parametric.
• Stacked regression estimators. Cengiz (2019) implements this strategy.  The objective is to “create event-specific “clean 2 × 2” datasets, together with the end result variable and controls for the handled cohort and all different observations which are “clean” controls throughout the remedy window (e.g., not-yet-, last-, or never-treated models). For every clear 2 × 2 dataset, the researcher generates a dataset-specific figuring out variable. These event-specific knowledge units are then stacked collectively, and a TWFE DiD regression is estimated on the stacked dataset, with dataset-specific unit- and time-fixed results… In essence, the stacked regression estimates the DiD from every of the clear 2 × 2 datasets, then applies variance weighting to mix the remedy results throughout cohorts effectively.”

While there was numerous math on this put up, if researchers apply these various DiD estimators, the authors correctly suggest that “researchers should justify their choice of ‘clean’ comparison groups—not-yet treated, last treated, or never treated—and articulate why the parallel-trends assumption is likely to apply”.

You can learn the complete article right here.