Why the Event-Study Plot Needs Diagnosis
A conventional event-study graph displays estimated coefficients at periods before and after treatment begins. Its shape can look like a direct record of treatment dynamics: near-zero leads followed by rising positive effects, for example. But under staggered adoption, each coefficient comes from a regression that combines observations from different cohorts and dates. Some comparisons may use already-treated units, and the cohorts observed at one event time may differ from those observed at another.
The central diagnostic question is therefore not just whether a coefficient is positive or statistically distinguishable from zero. It is which observations and comparisons produce it? A conventional two-way fixed-effects (TWFE) coefficient is generally a regression contrast, not automatically the average treatment effect for units at that event time. Treatment-effect heterogeneity and changing event-time support can make its pattern misleading.
The Conventional Event-Study Regression
Let \(G_i\) denote unit \(i\)'s adoption period, with \(G_i=\infty\) for a never-treated unit. Event time is \(k=t-G_i\) for a treated unit in calendar period \(t\). A conventional event study includes indicators for selected event times and omits one period as the reference. A common specification omits \(k=-1\):
Here \(\alpha_i\) and \(\lambda_t\) are unit and calendar-time fixed effects. The indicator for event time \(k\) is one only for treated units observed at that relative time. In a finite sample, the sum contains only the event-time indicators included in the specification; the omitted \(k=-1\) category is absorbed into the fixed effects and other regressors. Thus each \(\beta_k\) is a regression coefficient relative to the omitted period, not a raw mean outcome for cohort members at \(k\).
This matters because TWFE uses both within-unit changes and cross-unit differences to separate the event-time indicators from the fixed effects. A treated unit can help identify one coefficient while also serving as a comparison for another. Depending on adoption timing and included event-time indicators, some comparisons involve units that have already been treated. Earlier material on TWFE decompositions, negative weights, already-treated controls, and lead-lag bias explains why such comparisons can be problematic.
What the Regression Coefficient Combines
The Frisch–Waugh–Lovell result makes the role of the other regressors explicit. To describe the coefficient on the indicator \(X_k=1\{G_i<\infty,\ t-G_i=k\}\), first regress \(X_k\) on the unit and time fixed effects and all other included event-time indicators. Let \(\widetilde X_k\) be the residual from that regression. Then the coefficient from ordinary least squares is:
This formula is a short derivation of why the coefficient is not generally a simple treated-versus-untreated mean difference. The residualized indicator can be positive for some observations and negative for others. Its variation reflects which units and dates distinguish event time \(k\) from the fixed effects and the other event-time indicators. If outcomes contain heterogeneous treatment effects, the regression can consequently combine effects from different cohorts and event times, with weights that do not have to match a desired average treatment effect.
For causal interpretation, the relevant assumptions still matter. In particular, without anticipation means treatment does not affect outcomes before the chosen onset, and parallel trends requires the untreated potential outcomes of the relevant cohorts to have comparable changes over time. These assumptions do not, by themselves, turn every conventional TWFE event-study coefficient into a cohort-specific causal effect. When effects vary across cohorts or event times, the regression's implicit comparisons can mix those effects even if untreated trends satisfy the required condition.
Worked Example: An Already-Treated Comparison Can Distort a Later Effect
Consider a hypothetical balanced panel with an early cohort adopting in period 1, a late cohort adopting in period 3, and a never-treated cohort. Suppose the untreated outcome rises by 1 unit per period for every cohort. The early cohort's treatment effect is 5 units in period 1 and 8 units in periods 2 and 3. The late cohort's effect at adoption in period 3 is 4 units. These numbers are illustrative, not observed study results.
From period 1 to period 2, the early cohort's outcome change is the untreated change plus its change in treatment effect:
Over the same periods, the late cohort is not yet treated, so its outcome change is 1 unit. The early-versus-late difference in changes is \(4-1=3\) units. This contrast reflects the early cohort's treatment-effect increase from 5 to 8 units, not a new treatment effect of 3 units at a single event time.
Now compare the late cohort's change from period 2 to period 3 with the early cohort's change. The late cohort changes by \(1+4=5\) units. The early cohort, already treated in both periods, changes by \(1+(8-8)=1\) unit. The late-minus-early difference in changes is:
In this particular numerical example, the contrast happens to equal the late cohort's 4-unit onset effect because the early cohort's effect does not change between periods 2 and 3. If the early cohort's effect changed over that interval, its change would enter the comparison. These simple contrasts illustrate comparisons that can contribute to a TWFE regression; they are not a claim that any one contrast equals the full regression coefficient.
Why Leads Can Look Like Pretrends
A lead coefficient is often treated as a diagnostic for differences in untreated trends before adoption. That interpretation requires care. A lead indicator marks an observation before that unit's own adoption, but the coefficient is estimated in a regression where other observations may already be treated. If their treatment effects change over time, comparisons involving those observations can create nonzero lead coefficients even when the late cohort has no treatment effect before adoption and untreated outcomes follow parallel trends.
A nonzero lead is therefore not automatically proof of a violation of parallel trends. Conversely, small or imprecise leads do not establish that the assumptions needed for a causal interpretation hold. The coefficients can reflect limited support, the choice of included leads and lags, treatment-effect heterogeneity, or actual differences in untreated trends. A pretrend diagnostic must be interpreted in light of the comparisons that identify its coefficients.
Worked Example: A Nonzero Lead From an Already-Treated Cohort
Use the hypothetical cohorts and untreated trend from the previous example. At period 1, the early cohort's treatment effect is 5 units; at period 2, it is 8 units. The late cohort remains untreated in both periods. Suppose the late cohort's untreated outcomes increase by 1 unit from period 1 to period 2, so it has no pre-treatment effect.
The early cohort's observed outcome change is \(1+(8-5)=4\) units. The late cohort's change is 1 unit. If a lead comparison for the late cohort uses the early cohort as a comparison, the late-minus-early contrast is:
The negative contrast is generated by the early cohort's 3-unit increase in its treatment effect, not by a change in the late cohort's untreated outcome trend. This calculation does not say that the conventional regression's lead coefficient must equal \(-3\); the full coefficient also depends on all included observations and regressors. It shows why a lead can be contaminated and why its comparison structure needs inspection.
Changing Support Can Change the Meaning of the Plot
Event-time support is the set of cohorts observed at each relative time. Early adopters can contribute many post-treatment periods; late adopters may contribute few. If the sample ends at a fixed calendar date, later event times are often supported only by early cohorts. As a result, even an idealized average at \(k=0\) may describe a different cohort mix from an average at \(k=2\). A rising or falling plot can then reflect both changing effects and changing composition.
The cohort-time treatment effect for cohort \(g\) in period \(t\) is commonly written \(ATT(g,t)\). At event time \(k\), the corresponding effects are \(ATT(g,g+k)\) for cohorts with data at that event time. A dynamic summary could average these cohort-time effects, but its interpretation depends on which cohorts are included and how they are weighted. A conventional TWFE coefficient need not use the same weights or even correspond to that summary.
Worked Example: Apparent Dynamics With Changing Cohort Composition
Suppose two hypothetical cohorts have the following treatment effects, measured in percentage points. Both are observed at event time 0, but only cohort A is observed at event time 2 because the panel ends before cohort B reaches that event time.
| Cohort | Effect at event time 0 | Effect at event time 2 |
|---|---|---|
| A | 2 percentage points | 10 percentage points |
| B | 8 percentage points | Not observed |
For illustration, give the two supported cohorts equal weight at event time 0. The average is:
At event time 2, only cohort A is supported, so the available-cohort average is 10 percentage points. The plotted change is \(10-5=5\) percentage points. Yet for cohort A alone, the change from event time 0 to event time 2 is \(10-2=8\) percentage points. The two changes differ because the event-time-0 average includes cohort B while the event-time-2 average does not. This arithmetic illustrates a support problem; it is not a formula for a conventional TWFE coefficient.
A Practical Diagnostic
Before giving a causal interpretation to a conventional event-study plot, inspect the regression specification and the data support together. The following sequence focuses on what the reported coefficients can and cannot reveal.
Identify the omitted category, any combined tail periods, and the calendar periods available for each lead and lag.
For every event time, list which adoption cohorts contribute observations. Note when late lags are supported by only a subset of cohorts.
Determine whether the design and specification allow already-treated units to help identify a lead or lag coefficient. Do not assume the omitted period makes all other comparisons untreated.
Ask whether effects plausibly vary across cohorts or over time since adoption. If they do, a conventional coefficient may combine different effects rather than estimate one transparent event-time average.
Report the coefficient as a TWFE event-study contrast unless its causal target and identifying comparisons are justified. Treat leads as diagnostics, not as automatic proof for or against parallel trends.
Common Mistakes
- Calling \(\widehat\beta_k\) the average effect at event time \(k\) by default. The coefficient is a regression contrast. With heterogeneous effects, its implicit weights and contributing comparisons may not match that average.
- Assuming every lead compares untreated units. A unit can be untreated at its own lead event time while other units contributing to the coefficient are already treated.
- Reading a change in coefficients as a pure within-cohort dynamic. The cohorts supporting one lag may differ from those supporting another, so composition can change alongside effects.
- Treating insignificant leads as confirmation of parallel trends. Insignificance can reflect uncertainty or limited support, and the lead coefficients themselves may be contaminated.
Conventional TWFE event studies remain useful descriptions of a fitted regression and can be informative under restrictive conditions, such as sufficiently homogeneous treatment effects and appropriate untreated comparisons. The key is to state those conditions rather than infer them from the shape of a graph. When the conventional coefficients do not have a clear causal interpretation, report their limitations and use an estimator designed for the causal target of interest.
Check Your Understanding
Answer each question by referring to the regression comparisons, not just the appearance of a coefficient plot.
- In a conventional TWFE event study, what does the omitted event-time category do, and why does it not make every other coefficient a simple comparison with untreated units?
- How can an already-treated cohort contribute to a nonzero lead coefficient when the cohort marked by that lead has no treatment effect yet?
- Suppose only early adopters are observed at a long post-treatment lag. What complication does this create when comparing that lag with the adoption-period coefficient?
- What does the residualized-indicator formula reveal about the comparisons behind an event-study coefficient?
- Why do near-zero estimated leads not, by themselves, establish the assumptions required for a causal interpretation?