Interpreting Event-Study Regression Coefficients

Why Event-Study Coefficients Need Careful Interpretation

An event-study regression describes how outcomes evolve around a treatment date. Its coefficients are often plotted against event time, so a positive coefficient after adoption can look like a treatment effect at a glance. But the coefficient is defined relative to an omitted period, and its causal interpretation depends on how the treated group would have changed without treatment. Reading the event-time index alone is not enough.

This tutorial considers a common adoption date: every unit in the treated group begins treatment in the same calendar period, while comparison units remain untreated throughout the analysis. This setting makes the connection between the regression coefficients and difference-in-differences contrasts especially transparent.

Definition: Let \(D_i=1\) identify units assigned to treatment and \(D_i=0\) identify comparison units. Treated units adopt in period \(T_0\), and event time is \(k=t-T_0\). An event-study coefficient for event time \(k\) compares the treated group’s outcome at \(k\), relative to a chosen reference period, with the corresponding change in the comparison group.

The Regression and Its Reference Period

For panel data, a standard specification with reference event time \(k=-1\) is

Y_{it}=\alpha_i+\lambda_t+\sum_{k\ne -1}\beta_k \left(D_i\mathbf{1}\{t-T_0=k\}\right)+\varepsilon_{it}.

Here \(Y_{it}\) is the outcome for unit \(i\) in calendar period \(t\); \(\alpha_i\) is a unit fixed effect; and \(\lambda_t\) is a calendar-time fixed effect. The indicator \(\mathbf{1}\{t-T_0=k\}\) equals one when period \(t\) is \(k\) periods from adoption. The interaction with \(D_i\) means that the event-time indicators apply only to treated units. The term for \(k=-1\) is omitted, so the last period before adoption is the reference period.

Unit fixed effects absorb stable differences between units, and calendar-time fixed effects absorb shocks common to both groups in a given period. The omitted event-time term is necessary because including all event-time interactions alongside these fixed effects would make the regression coefficients not uniquely determined. The normalization sets the reference-period interaction coefficient to zero; it does not assert that the treatment effect in that period is zero.

Interpretation: \(\beta_k\) is a treated-versus-control difference in outcome changes from the omitted period to event time \(k\). It is not, by definition alone, the treated group’s outcome level, nor is it automatically a causal effect.

In a balanced panel with the stated groups and a complete set of event-time indicators for the periods being analyzed, the coefficient can be written as the following difference-in-differences contrast:

\beta_k= \left(\bar{Y}_{1,k}-\bar{Y}_{1,-1}\right) - \left(\bar{Y}_{0,k}-\bar{Y}_{0,-1}\right),

where \(\bar{Y}_{1,k}\) and \(\bar{Y}_{0,k}\) are the treated- and comparison-group mean outcomes at event time \(k\). In unweighted balanced data, this expression gives the coefficient directly. With weights, unbalanced panels, or a different estimation sample, the precise regression weighting can affect which group averages the coefficient represents; the comparison-from-reference interpretation remains central.

When a Coefficient Has a Causal Interpretation

To state the causal target, let \(Y_{it}(0)\) denote unit \(i\)’s outcome in period \(t\) if it remains untreated, and let \(Y_{it}(1)\) denote its outcome under the treatment regime that begins at \(T_0\). For treated units at event time \(k\), the average treatment effect on the treated is

\tau_k=E\!\left[Y_{i,T_0+k}(1)-Y_{i,T_0+k}(0)\mid D_i=1\right].

A post-adoption coefficient can identify \(\tau_k\) when the difference in untreated outcome changes between the treated and comparison groups would have been zero, and when treatment has no effect before adoption. More explicitly, for each period \(t\) being compared, the parallel-trends condition relative to the reference period is

E[Y_{it}(0)-Y_{i,T_0-1}(0)\mid D_i=1] = E[Y_{it}(0)-Y_{i,T_0-1}(0)\mid D_i=0].

The condition allows the groups to have different baseline outcome levels. It requires their average untreated changes from the reference period to be equal. No anticipation requires that outcomes before adoption are not affected by the future treatment. Under that condition, the treated group’s treatment effect at the reference period is zero. With these conditions, for \(k\ge 0\), the event-study coefficient equals the treatment effect on the treated at event time \(k\), for the population and weighting represented by the regression.

Write the observed treated-group change.
From the reference period to \(k\), it includes the change in untreated potential outcomes and the treatment effect at \(k\), minus any effect at the reference period.

Subtract the comparison-group change.
Comparison units remain untreated, so their observed change estimates the untreated change that would have occurred for the treated group if parallel trends holds.

Apply no anticipation.
The reference-period treatment effect is zero. The remaining contrast is the treated group’s average treatment effect at \(k\).

The algebra makes the identification step explicit. The treated group’s mean change from the reference period to \(k\) is its untreated-potential-outcome change plus the change in its treatment effect. The comparison group’s mean change is its untreated-potential-outcome change. Subtracting the latter from the former gives

\beta_k = \left\{E[Y_{it}(0)-Y_{i,T_0-1}(0)\mid D_i=1] - E[Y_{it}(0)-Y_{i,T_0-1}(0)\mid D_i=0]\right\} +\tau_k-\tau_{-1}.

Parallel trends sets the bracketed difference to zero. No anticipation gives \(\tau_{-1}=0\), so \(\beta_k=\tau_k\) for post-adoption periods. For a pre-adoption event time, no anticipation instead implies that there is no treatment effect yet. Its coefficient measures a difference in pre-period changes relative to the reference period; it is not an effect of treatment.

Worked Examples

Worked Example: Reading a Coefficient Table

Suppose a hypothetical regression uses event time \(k=-1\) as its reference period and reports \(\hat{\beta}_{-2}=-1.2\), \(\hat{\beta}_{0}=2.4\), and \(\hat{\beta}_{1}=4.1\). The outcome is weekly earnings measured in hundreds of dollars. Interpret each estimate.

At \(k=0\), the estimated treated-versus-control change from the last pre-adoption period is \(2.4\) hundred dollars, or \(240\) dollars per week. If parallel trends and no anticipation hold, this is an estimated treatment effect on the treated in the adoption period.

At \(k=1\), the corresponding estimated change is \(4.1\) hundred dollars, or \(410\) dollars per week. Under the same assumptions, it estimates the treatment effect one period after adoption. It is not an additional \(410\) dollars on top of the \(240\)-dollar estimate; each coefficient compares its period with the same omitted period.

At \(k=-2\), the estimated contrast is \(-1.2\) hundred dollars, or \(-120\) dollars per week, relative to \(k=-1\). Since this is before adoption, it is not a treatment effect under no anticipation. It reports how the treated group’s earlier outcome change differs from the comparison group’s change. The causal reading of the post-adoption coefficients still requires the identifying assumptions; the regression table alone does not establish them.

Worked Example: Calculating an Event-Study Coefficient

Consider hypothetical mean outcomes measured in units of an index. At the reference period \(k=-1\), the treated-group mean is \(50\) and the comparison-group mean is \(48\). At adoption, \(k=0\), their means are \(58\) and \(51\), respectively. Calculate the event-study coefficient for \(k=0\).

The treated group’s change is \(58-50=8\) index points. The comparison group’s change is \(51-48=3\) index points. Therefore,

\hat{\beta}_0=(58-50)-(51-48)=8-3=5.

The estimated coefficient is \(5\) index points. It says that the treated group’s increase from the reference period to adoption exceeds the comparison group’s increase by \(5\) index points. If the untreated potential outcomes satisfy parallel trends over these periods, there is no anticipation, and the other design conditions are credible, this contrast estimates the average treatment effect on the treated at adoption. Without those conditions, \(5\) remains a difference-in-differences association, not an established causal effect.

This calculation also shows why a difference in levels is not the coefficient: at \(k=0\), the difference in group means is \(58-51=7\), not \(5\). The coefficient additionally subtracts the pre-existing difference in means, \(50-48=2\).

Worked Example: Changing the Reference Period

Suppose a hypothetical event study uses \(k=-1\) as its reference period. Its estimated coefficient for \(k=-2\) is \(-1.5\) index points, and its coefficient for \(k=0\) is \(4.0\) index points. What would the \(k=0\) coefficient mean if \(k=-2\) were instead the reference period?

Changing the reference period expresses each event-time contrast relative to \(k=-2\), rather than relative to \(k=-1\). The new coefficient for \(k=0\) is the old contrast from \(k=-1\) to \(k=0\), minus the old contrast from \(k=-1\) to \(k=-2\):

\hat{\beta}^{\,(-2)}_0 = \hat{\beta}^{\,(-1)}_0-\hat{\beta}^{\,(-1)}_{-2} = 4.0-(-1.5)=5.5.

Thus the re-referenced coefficient is \(5.5\) index points, not \(4.0\). This is an algebraic change in the comparison period, not a change in the observed outcomes or the fitted regression. The fitted values are unchanged when the normalization is changed consistently. The new coefficient is a causal effect only if the assumptions justify interpreting the contrast against \(k=-2\); choosing a different reference does not create or repair identification.

Conditions and Scope of the Interpretation

The causal interpretation above relies on a particular design, not merely on using event-time indicators. In this common-adoption setting, check that the following conditions match the research question and data:

Common adoption date for treated units. The event-time index must correspond to the same calendar adoption date for all treated units in the analysis.
Untreated comparison group. Comparison units must not receive the treatment during the periods used to construct the contrasts, and should not be indirectly affected by treated units’ treatment.
Parallel trends in untreated outcomes. In the absence of treatment, average changes from the reference period would have been equal across groups. Different baseline levels are allowed.
No anticipation. Outcomes are not affected by future treatment before adoption, including in the chosen reference period.
Consistent outcome measurement and composition. The outcome must be defined comparably across groups and periods. In a panel, the same units’ outcomes should be followed as intended; in repeated cross-sections, changes in group composition must not distort the group means being compared.
A well-defined treatment and suitable estimation sample. Treatment status and event time must be measured correctly, and the included periods must provide the treated and comparison observations needed for the contrasts.

The regression’s functional form also matters. The displayed interpretation assumes the event-time interactions are specified for the periods of interest and that the fixed effects and estimation sample support the stated comparison. If event times are combined into bins, the resulting coefficient summarizes the bin’s specified contrast rather than a separate effect at every event time. If units receive unequal regression weights or the panel is unbalanced, state clearly which units and weighting scheme the estimates represent.

Common Mistakes

Calling every post-adoption coefficient an effect by definition. A regression coefficient is a relative change. It becomes a causal effect only under the design assumptions, especially parallel trends and no anticipation.
Reading the omitted coefficient as a zero treatment effect established by data. The omitted coefficient is set to zero for normalization. No anticipation is a separate assumption that says the treatment has not affected outcomes before adoption.
Comparing a coefficient with zero as if it were a level difference. Each coefficient is relative to the reference period and the change in the comparison group. The group-level outcome difference in a given period is generally not the coefficient.
Calling a pre-adoption coefficient a treatment effect. With no anticipation, treatment has not yet affected outcomes. The coefficient instead records a treated-versus-control difference in changes between two pre-periods.
Treating a new normalization as new evidence. Changing the reference period changes the numerical coefficients and their interpretation, but does not change the underlying fitted comparisons or make the causal assumptions more credible.

Key takeaway: An event-study coefficient is a treated-versus-control change relative to an omitted event-time period. For post-adoption periods, it identifies the corresponding average treatment effect on the treated only when the untreated trends, treatment timing, comparison group, and measurement conditions support that interpretation.

Check Your Understanding

Use the reference-period and causal-interpretation principles to answer these questions.

In an event study with \(k=-1\) omitted, what comparison does \(\beta_2\) represent?
A post-adoption coefficient is positive. Name two assumptions needed before interpreting it as a causal effect on the treated.
Why does omitting \(k=-1\) not, by itself, establish that the treatment effect in that period is zero?
If a coefficient for \(k=-3\) is nonzero and treatment has not yet begun, what does that coefficient describe under no anticipation?
Suppose the reference-period coefficient for \(k=-2\) is \(-2\) and the coefficient for \(k=0\) is \(3\). What is the \(k=0\) coefficient when \(k=-2\) becomes the reference period?