Anticipation and Timing Violations in DiD

Why Timing Matters in Difference-in-Differences

A difference-in-differences (DiD) design compares outcome changes for a treated group with outcome changes for a comparison group. Its usual interpretation depends on the comparison period being genuinely untreated for the treated group. That condition can fail even when the recorded intervention has not yet begun: people or organizations may respond to an announcement, prepare for implementation, or change behavior because they expect the intervention.

This creates a timing problem. A date in an administrative record may mark implementation, while the relevant response began earlier. If that earlier response falls in the period used as the DiD baseline, the baseline is already affected. The resulting contrast need not represent the effect of the intervention after its recorded start.

The definition depends on what counts as the treatment. If the intervention is a policy package that includes its announcement, a response to the announcement may be part of the treatment effect. If the target is specifically the effect of implementation beyond the announcement, the announcement is a prior exposure that must be addressed. A date is not automatically the correct treatment onset simply because it is the date recorded in a policy database.

How Anticipation Enters the DiD Contrast

Let \(Y_{it}(0)\) denote unit \(i\)’s outcome at time \(t\) without the intervention, and let \(Y_{it}(1)\) denote its outcome with the intervention. Let \(T\) identify the treated group and \(C\) the comparison group. For the usual two-period design, the observed DiD contrast is the treated group’s post-minus-baseline change minus the comparison group’s post-minus-baseline change.

Suppose the untreated outcome trends are parallel: absent the intervention, the treated and comparison groups would have the same average change between \(t=0\) and \(t=1\). Suppose also that the intervention affects the treated group in the post-period, with average effect \(\tau_1\). If the treated group has already experienced an average anticipatory response \(a_0\) in the baseline period, its observed baseline mean includes that response.

Here, \(\mu_{g,t}(0)\) is group \(g\)’s mean outcome at time \(t\) in the absence of the intervention. Parallel trends gives \(\mu_{T,1}(0)-\mu_{T,0}(0)=\mu_{C,1}(0)-\mu_{C,0}(0)\). Substituting that equality cancels the untreated trend terms and leaves:

This is a derivation of the timing problem under the stated conditions: a baseline response is subtracted as part of the treated group’s change. If anticipation raises the baseline outcome, so \(a_0>0\), the contrast is smaller than the post-period effect \(\tau_1\). If anticipation lowers the baseline outcome, the contrast is larger. The direction of the distortion therefore depends on how the response changes the outcome; it is not always attenuation.

This expression is not a general correction formula. It isolates the baseline-anticipation component under parallel trends for untreated outcomes, no anticipation in the comparison group, and a well-defined post-period treatment effect. If untreated trends also differ, or the comparison group is affected, other discrepancies enter the contrast as well.

Assessing the Intervention Date

Timing should be assessed using both substantive knowledge and outcome data. Substantive knowledge helps identify when people first could have learned about, prepared for, or responded to an intervention. Pre-period outcomes can then reveal whether the treated group’s trajectory begins to diverge during a period that was classified as untreated. Neither source is decisive on its own: an outcome pattern may have other explanations, and a plausible opportunity to anticipate does not establish that outcomes actually changed.

A useful first step is to lay out the sequence of relevant events, not just the recorded implementation date. Depending on the setting, relevant dates may include announcement, eligibility determination, application opening, funding decisions, preparation, and service delivery. Ask which events could affect the outcome, which groups knew about them, and when that knowledge could plausibly change behavior. The answer should match the outcome’s mechanisms: a hiring decision may respond to an announced subsidy before the first subsidized payroll, while a measured service receipt may not.

Using Pre-Period Outcomes Without Overinterpreting Them

Pre-period outcomes help evaluate whether a chosen baseline looks plausible. In particular, if treated and comparison outcomes start to diverge after an announcement but before implementation, that pattern is consistent with anticipation. It is not proof: a coincident change in population composition, measurement, or another event could also produce divergence. Conversely, similar pre-period trajectories do not prove no anticipation; effects may be small, delayed, or obscured by noisy outcomes.

The key question is not merely whether a pre-period comparison is statistically significant. It is whether the timing and size of observed changes, alongside substantive information, make the proposed untreated baseline credible for the target effect. Earlier pre-periods can be informative about the untreated trajectory, while a period affected by anticipation should not be treated as an unaffected baseline simply because it precedes the implementation date.

When and How to Shift the Date

If substantive knowledge indicates that treatment begins when an announcement becomes actionable, one defensible choice may be to define treatment onset at that earlier point. This can make the timing consistent with the causal question about the entire policy package. The choice should be made on substantive grounds and reported clearly; it should not be selected only because one date produces a preferred estimate or a more favorable pre-period pattern.

Shifting the date also changes which periods count as pre-treatment and post-treatment, and may therefore change the estimand. A contrast with an earlier onset can capture effects of information and preparation as well as effects of implementation. If the target is the implementation effect alone, an earlier announcement response is not solved by simply renaming it treatment. Depending on the setting, the analyst may need a different comparison, a longer unaffected pre-period, or a design that separately models the announcement and implementation phases. If no credible unaffected baseline exists, the standard DiD claim may not be identified by the available contrast.

Common Mistakes

• Assuming that “before implementation” means untreated. That is incorrect when an announcement or expectation can already change outcomes. Consider when exposure to information or preparation began.
• Calling any pre-period divergence proof of anticipation. Divergence is a warning, not a diagnosis by itself. Check other events, measurement, and composition before attributing it to anticipation.
• Claiming that a nonsignificant pre-period comparison establishes no anticipation. Lack of a detectable difference does not establish that the baseline was unaffected. Interpret the outcomes in context and consider whether the data could reveal a meaningful response.
• Moving the date without stating what effect is now estimated. An earlier onset may include announcement responses. State whether the target is the effect of the full policy process or an effect beyond the announcement.
• Assuming that a shifted date repairs every problem. A different date cannot by itself establish parallel untreated trends or rule out other group-specific changes.