Anticipation and Treatment-Date Definitions

Why the Recorded Treatment Date May Not Be the Beginning

A policy can change behavior before its official implementation. Firms may hire in response to an announced subsidy; patients may change care when they learn a program will soon begin; local governments may prepare for a regulation before its effective date. In each case, a recorded adoption date is not necessarily the first date on which the policy affects outcomes.

This distinction matters for difference-in-differences (DID) and event studies because a pre-adoption observation is not automatically untreated. If treated units anticipate a policy during the period used as the baseline, the comparison can subtract away part of the policy’s effect or make an event-study lead look like a pre-existing trend. The key task is to define treatment timing to match the causal question and the information or exposure that can affect outcomes.

Definition: Let $G_i$ be unit $i$’s recorded adoption date, and let $Y_{it}(g)$ denote its potential outcome at time $t$ if adoption occurs at date $g$. Let $Y_{it}(\infty)$ denote its potential outcome if it never adopts. There is no anticipation for horizon $L_i$ if $$ Y_{it}(g)=Y_{it}(\infty)\quad\text{for every }t<g-L_i. $$ Thus, the outcome may respond during the anticipation window $g-L_i,\ldots,g-1$, as well as at or after adoption. The horizon $L_i$ is measured in the study’s time periods.

The definition allows anticipation effects to vary across units and dates; it does not require them to be constant or to begin at the same time for everyone. If the anticipation horizon is zero, all periods strictly before adoption are unaffected by the future treatment under this definition. If the horizon is positive, the earliest period guaranteed to be untreated is before $G_i-L_i$. The boundary matters: with horizon $L_i=2$, the period $G_i-2$ is potentially affected, not a clean untreated period.

A study should distinguish the date an intervention is announced, the date it becomes legally or administratively effective, and the date units can first respond to it. These dates may coincide, but that should be justified rather than assumed. The appropriate date depends on the estimand. An effect of the entire policy process may include responses to announcement; an effect of implementation conditional on prior information is a different question.

Untreated Periods and the DID Comparison

Consider treated units with adoption date $G_i$ and comparison units that remain untreated and unexposed during the comparison. For a simple two-period comparison, write $t_0$ for the baseline and $t_1$ for the later outcome. To use $t_0$ as an untreated baseline for treated units, it must satisfy $t_0<G_i-L_i$. Comparison units must also have no treatment or anticipation effects in either period. In addition, the relevant untreated outcome trends must be parallel: absent treatment, the treated and comparison groups would have had the same mean change from $t_0$ to $t_1$.

\widehat{\tau}_{DID} = (\bar{Y}_{T,t_1}-\bar{Y}_{T,t_0}) - (\bar{Y}_{C,t_1}-\bar{Y}_{C,t_0})

Here $\bar{Y}_{T,t}$ and $\bar{Y}_{C,t}$ are the group mean outcomes for treated and comparison units at time $t$. When $t_1$ is post-adoption, the DID contrast can identify the average treatment effect for the treated at $t_1$ under the stated conditions. If $t_0$ falls inside the anticipation window, it is instead an outcome under partial exposure; the contrast generally no longer identifies the effect relative to never-treated outcomes at $t_1$.

Key implication: If treatment may affect outcomes beginning at $G_i-L_i$, a clean pre-treatment observation must occur strictly before that date. The label “pre-adoption” alone does not establish that an observation is untreated.

Worked Example: An Announcement Before Implementation

Suppose a hypothetical city policy is officially implemented in period 3, but its announcement in period 2 leads businesses to make changes. The mean outcome (in units of a monthly index) is 100, 104, and 110 for treated areas in periods 1, 2, and 3. In comparison areas, the corresponding means are 80, 82, and 84. Assume period 1 is unaffected, the comparison areas are untreated and unexposed, and the untreated trends would be parallel.

The period-2 change in the treated areas is $104-100=4$, while the comparison-area change is $82-80=2$. The difference in changes is $4-2=2$ index points. Under the assumptions, this is evidence of a differential change before official implementation; it is not evidence that the period-2 observation is a clean placebo. The announcement may have affected outcomes.

Using period 1 as the baseline and period 3 as the post period gives $$ (110-100)-(84-80)=10-4=6 $$ index points. This comparison includes the total effect present by period 3, including any response that began after the announcement, provided period 1 was genuinely untreated and the other DID assumptions hold.

If period 2 were incorrectly used as the “pre-treatment” baseline, the calculation would be $$ (110-104)-(84-82)=6-2=4 $$ index points. This contrast is relative to an already affected baseline. It measures the additional differential change from period 2 to period 3, not the total effect by period 3 relative to an unaffected baseline. The numbers therefore answer different questions.

Event Time When Outcomes Can Respond Early

Event time is usually defined relative to recorded adoption: $k=t-G_i$. Under this convention, $k=0$ is the adoption period, $k<0$ are leads, and $k>0$ are lags. If anticipation is possible, some leads may be affected by the future treatment. A lead is a calendar position relative to adoption; it is not automatically an untreated observation or a valid placebo.

For an anticipation horizon $L_i$, periods with $k<-L_i$ are guaranteed to precede the anticipation window under the definition above. Periods with $-L_i\leq k<0$ are potentially affected. A researcher may show these periods as anticipation effects, omit them from a baseline comparison, or redefine the event around the onset of exposure. The choice should be explicit: redefining the event as the first possible response changes what “event time zero” means and can make the official adoption date less visible.

Worked Example: Selecting a Clean Baseline for a Monthly Event Study

A hypothetical regulation takes effect in July, so $G_i=\text{July}$. Firms learn the final regulation in May, and the analysis treats May as the earliest plausible response month. In event time relative to July, May is $k=-2$, June is $k=-1$, and April is $k=-3$.

Because May and June may be affected, a baseline period for an event study should be earlier than May. If April is the omitted category, its event time is $k=-3$. The coefficients for May and June compare those potentially affected months with April, subject to the model’s other assumptions. They can describe differential changes during the anticipation window; they should not be presented as clean tests of parallel trends.

Suppose the planned event-time window is $k=-4$ through $k=3$. It includes April and earlier as pre-anticipation periods, May and June as possible anticipation periods, July as the implementation month, and the three following months. The window should label these roles clearly. If the analysis instead used June as its omitted baseline, both the baseline and several apparent lead comparisons would already be potentially affected.

This example also shows why the event-time window is not only a plotting choice. Its early endpoint determines which pre-anticipation observations are available, and the omitted category determines what the displayed coefficients compare against.

Choosing Treatment Dates and Event-Time Windows

The treatment date should be defined using the institutional process and the causal question, not selected because one date produces a preferred estimate. A defensible timing decision usually specifies what event makes a unit eligible or exposed, when that event occurs in the data, and whether information or behavior can change earlier. For example, an evaluation of implementation may use the effective date while separately accounting for a public announcement; an evaluation of the policy’s total effect may treat the announcement as part of the intervention’s onset.

Event-time windows should make enough periods available to assess timing while avoiding unsupported claims from very distant or sparse observations. Long windows may have few observations at their endpoints or combine periods with different institutional conditions. Short windows may conceal earlier responses or leave too few clean pre-anticipation periods. Report the chosen window and the reference period, and state which leads are potentially affected by anticipation.

When the earliest plausible response date is uncertain, do not imply that the data identify a precise onset merely because one date is recorded. Use substantive information to specify a plausible range of anticipation horizons. Then report how conclusions change under defensible alternatives—for example, excluding one, two, or more periods immediately before adoption from the untreated baseline. This is a timing sensitivity analysis, not proof that any chosen horizon is correct.

Worked Example: Sensitivity to the Assumed Anticipation Horizon

Consider a hypothetical quarterly program adopted in quarter 4. Treated-group mean outcomes are 50, 53, 56, and 61 in quarters 1 through 4. Comparison-group means are 40, 42, 44, and 46. Suppose quarter 4 is post-adoption, comparison units remain unexposed, and untreated trends would be parallel. A policy announcement may have affected treated units before adoption, so compare two plausible anticipation assumptions.

Assume a two-quarter horizon. Quarters 2 and 3 may be affected, leaving quarter 1 as the available clean baseline. The DID estimate is $$ (61-50)-(46-40)=11-6=5 $$ outcome units. Under the assumptions, this estimates the total effect present in quarter 4 relative to the untreated outcome path.

Assume a one-quarter horizon. Quarter 3 may be affected, while quarter 2 is outside the anticipation window and can be used as the baseline. The estimate becomes $$ (61-53)-(46-42)=8-4=4 $$ outcome units. This contrast is relative to quarter 2, not quarter 1. If quarter 2 already contained an anticipatory response, the one-quarter assumption is not valid and the estimate omits that earlier response.

The estimates differ by 1 outcome unit because the baseline period changes. The arithmetic does not determine which horizon is credible. The analyst must justify the horizon using what units knew and when, and should explain that the estimate of 4 is not the total effect relative to an unaffected baseline if anticipation had already begun by quarter 2.

Common Mistakes and Practical Checks

A frequent mistake is to call every observation before adoption “pre-treatment,” use the final pre-adoption period as the baseline, and interpret every event-study lead as a placebo. The correction is to ask when outcomes could first respond, define the anticipation window, and classify periods relative to that window. A nonzero lead may reflect anticipation, differential trends, or other model problems; its value alone does not identify which explanation is correct.

Before estimating or interpreting an event study, make the timing choices visible. State the recorded treatment date and its institutional meaning, the earliest plausible response date or anticipation horizon, the clean pre-treatment periods, the event-time range, and the omitted reference period. If adoption dates are measured only by year or quarter, acknowledge that the data may not distinguish announcement, response, and implementation within that interval.

These definitions affect the estimand as well as the regression specification. An estimate comparing post-adoption outcomes with an unaffected baseline can capture effects accumulated from anticipation through implementation. An estimate using an anticipation-affected baseline instead asks how outcomes changed after that baseline. Neither contrast is automatically wrong, but they must not be described as the same treatment effect.

Key takeaway: Define treatment timing by the earliest exposure relevant to the causal question. Use only periods before plausible anticipation as untreated baselines, and label event-study leads according to whether they are clean pre-periods or potentially affected anticipation periods.

Check Your Understanding

Use the treatment date, anticipation horizon, and comparison assumptions explicitly in your answers.

A policy is adopted in period 8, and outcomes may respond beginning two periods earlier. Which periods are potentially affected, and what condition must a baseline period satisfy to be guaranteed untreated?
Why can an event-study lead be nonzero even if there was no differential trend before units learned about the policy?
A researcher uses the last pre-adoption period as the baseline despite a known announcement one period earlier. What does the resulting post-period contrast fail to measure relative to a clean baseline?
Give one reason the adoption date and the date outcomes can first respond might differ, and explain which date would be relevant to an evaluation of implementation versus the policy’s total effect.
If the anticipation horizon is uncertain, describe one principled way to assess whether the estimate depends on the timing assumption.