Treatment Timing and Dynamic Effects in Synthetic Control

Why Treatment Timing Matters

A policy or intervention may have effects that change over time. A new transit service might affect employment gradually; a regulation may cause an immediate adjustment followed by a slower response. A single post-treatment average can conceal this pattern. Synthetic control can instead provide a sequence of estimated gaps, one for each period after treatment begins.

That sequence is meaningful only if the treatment date matches the intervention being studied. An announcement can change behavior before formal implementation, and a policy may be introduced in stages. Moreover, the post-treatment gaps should not be used to choose the specification that makes the estimated effects look most favorable. The specification and summaries should be grounded in pre-treatment information and a clearly stated estimand.

Dynamic Effects and Synthetic-Control Gaps

Let unit 1 be the treated unit, observed in periods $t=1,\ldots,T$. Let $T_0$ be the first period in which it is treated under the chosen definition of treatment. Write $Y_{1t}$ for its observed outcome. Let $Y_{1t}(0)$ denote the outcome it would have had in period $t$ under no treatment throughout the relevant period, and let $Y_{1t}(1)$ denote its outcome under the treatment path being evaluated. For periods from treatment onset onward, the unit-specific effect is $\tau_{1t}=Y_{1t}(1)-Y_{1t}(0)$.

A synthetic control uses comparison units and weights chosen using pre-treatment information to estimate the treated unit’s no-treatment outcome. Write that estimate as $\widehat{Y}_{1t}(0)$. The observed gap in period $t$ is the treated unit’s outcome minus this estimated counterfactual. Event time $k$ counts periods relative to treatment onset: $k=t-T_0$. Thus $k=0$ is the first treated period, and positive values are subsequent periods.

Definition: The estimated dynamic synthetic-control effect at event time $k\geq 0$ is the treated unit’s observed outcome minus its estimated no-treatment outcome in the corresponding calendar period: $$ \widehat{\tau}_{1,T_0+k} = Y_{1,T_0+k} - \widehat{Y}_{1,T_0+k}(0). $$ The sequence of these gaps describes how the estimated effect changes over post-treatment time. It is an estimate of a causal effect only to the extent that the synthetic control credibly represents the treated unit’s no-treatment path.

For this interpretation, the treatment and its timing must be defined consistently, and the treated unit’s observed outcome must correspond to the treatment actually received. Periods used to fit the synthetic control should not already be affected by the intervention or its anticipation. The pre-treatment fit and the rationale for the selected comparison units and weights should be assessed using pre-treatment information. As in synthetic-control analysis generally, a close pre-treatment fit is useful evidence, but it does not establish that the estimated post-treatment counterfactual is correct.

The comparison units must also provide a meaningful untreated reference over the periods being evaluated. If the treatment definition or timing changes, then the estimand changes too: an effect measured from an announcement date includes responses to the announcement, whereas an effect defined from an implementation date uses a different event-time origin. Those are not interchangeable interpretations of the same sequence.

Worked Example: Calculating Period-by-Period Gaps

Suppose a hypothetical region introduces a workforce program in period 5. The outcome is jobs per 100 eligible residents. A synthetic control uses fixed weights of 0.6 on comparison region A and 0.4 on comparison region B; these weights were selected using pre-treatment data. The following values are hypothetical.

Period	Treated region	Region A	Region B
1	9	8	10
2	11	10	12
3	12	11	13
4	14	13	15
5	18	16	14
6	20	17	15
7	22	18	16

In each period, the estimated synthetic outcome is $0.6$ times region A’s outcome plus $0.4$ times region B’s outcome. For period 5, for example: $$ \widehat{Y}_{1,5}(0) = 0.6(16)+0.4(14) = 9.6+5.6 = 15.2, \qquad \widehat{\tau}_{1,5} = 18-15.2 = 2.8. $$ The same calculation gives synthetic outcomes of $16.2$ and $17.2$ in periods 6 and 7. The gaps are therefore $20-16.2=3.8$ and $22-17.2=4.8$, respectively.

The pre-treatment gaps are $0.2$ in each of periods 1 through 4. Their root mean squared value is $$ \sqrt{\frac{0.2^2+0.2^2+0.2^2+0.2^2}{4}} = \sqrt{0.04} = 0.2 $$ jobs per 100 eligible residents. This is a descriptive measure of pre-treatment fit, not proof of the post-treatment counterfactual. If the program’s timing and assumptions are credible, the post-treatment estimates indicate gaps of 2.8, 3.8, and 4.8 jobs per 100 eligible residents in periods 5, 6, and 7. They do not, by themselves, show why the gaps increased.

Anticipation and the Choice of Treatment Date

The treatment date is not always the date an administrative record labels “implementation.” People or organizations may respond to an announcement, eligibility notice, or credible expectation of a future intervention. This is anticipation: outcomes can change because treatment is expected, before the formal start date. If a period used to fit the synthetic control is already affected, it is not an untreated pre-treatment period for the intended counterfactual.

Define the intervention according to the causal question. If the question concerns the effect of a policy process beginning with its announcement, use that announcement as the event-time origin. If the question concerns implementation, state that target precisely, but do not treat announcement-affected periods as untreated data. A gap before the formal start can be an important sign of anticipation or of a failing counterfactual; it should not simply be hidden by redefining the graph’s horizontal axis.

When the onset is gradual, a single date may summarize a more complicated treatment path. The event-time estimates then describe effects relative to a chosen onset convention, not necessarily the effect of a uniform treatment dose at every event time. Report the operational definition—such as first eligibility, first service availability, or full implementation—and explain why it matches the question.

Worked Example: An Announcement Before Implementation

Suppose a hypothetical benefit is announced in week 6 and becomes available in week 8. A researcher wants to estimate the effect of the policy process, including responses to the announcement. The synthetic control was fitted using weeks 1–5 only. Assume the estimated gaps, in dollars of weekly spending per household, are $1.5$ in week 6, $2.0$ in week 7, $3.0$ in week 8, and $3.5$ in week 9.

For this question, the appropriate event-time origin is week 6, when the policy process first became known. Week 6 has $k=0$, week 7 has $k=1$, and weeks 8 and 9 have $k=2$ and $k=3$. The sequence therefore describes estimated gaps of $1.50, $2.00, $3.00, and $3.50 per household per week, beginning with the announcement.

If the researcher instead labels week 8 as $k=0$, weeks 6 and 7 become apparent “pre-treatment” periods despite the announcement. Using those weeks to fit the synthetic control would risk fitting to outcomes already affected by the policy. If the target were specifically the effect of service availability after week 8, the researcher could report post-availability gaps, but should retain weeks 1–5 as the untreated fitting period and explain that the estimates may include earlier anticipation. These data alone do not separate an announcement effect from an implementation effect without additional assumptions or design information.

Summarizing a Dynamic Trajectory

A plot of the estimated gaps is often the clearest way to display the full trajectory. A summary can help answer a focused question, but it should specify the horizon and the scale. For example, an average over the first $H$ post-treatment periods is the mean of those period-specific estimates. It summarizes the average estimated effect over that window; it does not claim that the effect was constant within it.

The cumulative sum of period-specific effects answers a different question when outcomes are measured as a flow, such as dollars spent per month. It can represent a total over months, whereas the average remains in the original per-period units. For outcomes that are stocks or rates, the substantive interpretation of a sum may not be useful. In every case, distinguish the arithmetic summary from the underlying event-time pattern.

Choose the primary horizon or summary based on the research question, preferably before inspecting post-treatment estimates. If several horizons are informative, label them as such and explain their role. Averages over windows of different lengths are not directly comparable without considering which periods each includes.

Worked Example: Average and Cumulative Effects Over a Fixed Horizon

Consider a hypothetical program with estimated monthly gaps in employment, measured in percentage points, of $1, 2, 4, 3,$ and $5$ over its first five months. Suppose the five-month average is the prespecified primary summary. Its calculation is $$ \overline{\widehat{\tau}}_{1:5} = \frac{1+2+4+3+5}{5} = \frac{15}{5} = 3 $$ percentage points.

The result means that the average estimated gap across these five months is 3 percentage points. It does not mean that the estimated effect was 3 percentage points in every month: the sequence ranges from 1 to 5 percentage points. If the outcome is interpreted as a monthly flow and summing across months is substantively appropriate, the sum is $1+2+4+3+5=15$ percentage-point-months. That cumulative unit is not the same as a 15-percentage-point effect in any single month.

The average for the first two months would be $(1+2)/2=1.5$ percentage points, while the average for months 3–5 would be $(4+3+5)/3=4$ percentage points. Those summaries suggest different phases, but selecting whichever window gives the largest estimate after seeing the sequence would turn the window into a data-driven choice. The five-month primary summary remains interpretable because its horizon was specified in advance.

Why Post-Treatment Fit Must Not Choose the Specification

The synthetic control is intended to estimate what the treated unit would have experienced without treatment. The treatment may therefore cause the treated unit to diverge from that counterfactual. A specification that minimizes post-treatment gaps, or that is chosen because its post-treatment path looks attractive, uses outcomes that may contain the effect being estimated. This can pull the estimated effect toward a desired pattern and make uncertainty calculations misleading.

A sound workflow makes timing, donor eligibility, predictors, and tuning choices using institutional knowledge and pre-treatment information. Then it checks pre-treatment fit and reports the post-treatment trajectory under that specification. If several defensible timing conventions or specifications are relevant, show them as planned sensitivity analyses with their substantive rationale; do not quietly designate the most favorable result as primary. The earlier material on placebo inference provides tools for evaluating evidence, but a series of event-time estimates also raises a multiplicity issue: isolated period-by-period claims should not be treated as though only one outcome was examined. A prespecified summary statistic can make the inferential target clearer.

Justification: The gap’s interpretation follows by adding and subtracting the treated unit’s true no-treatment outcome. In any post-treatment period, $$ \widehat{\tau}_{1t} = Y_{1t}-\widehat{Y}_{1t}(0) = \bigl(Y_{1t}(1)-Y_{1t}(0)\bigr) + \bigl(Y_{1t}(0)-\widehat{Y}_{1t}(0)\bigr). $$ The first term is the causal effect $\tau_{1t}$; the second is counterfactual estimation error. Averaging over a prespecified set $\mathcal{H}$ of $H$ periods gives $$ \frac{1}{H}\sum_{t\in\mathcal{H}}\widehat{\tau}_{1t} = \frac{1}{H}\sum_{t\in\mathcal{H}}\tau_{1t} + \frac{1}{H}\sum_{t\in\mathcal{H}} \bigl(Y_{1t}(0)-\widehat{Y}_{1t}(0)\bigr). $$ Thus the average gap estimates the average causal effect over that horizon when the average counterfactual estimation error is sufficiently small. Excellent pre-treatment fit can support the choice of a counterfactual, but does not guarantee that this error remains small after treatment.

Common Pitfalls

Using an official date without checking for anticipation. A formal launch date does not ensure earlier periods were unaffected. Check announcements, eligibility changes, and other plausible sources of early response.
Calling every gap a causal effect. A gap is an observed outcome minus an estimated counterfactual. Its causal interpretation depends on the counterfactual being credible.
Treating a positive average as a constant effect. An average summarizes a chosen window; it does not erase variation across periods.
Choosing weights, dates, or horizons to improve the post-treatment story. Such choices can make the estimate depend on the outcomes it is meant to evaluate. Use pre-treatment information and a stated rationale.
Interpreting a pre-treatment gap as a treatment effect. Before the chosen onset, a gap is a diagnostic unless the period is itself affected by anticipation and the estimand explicitly includes that response.

Key takeaway: Define the intervention and its onset, estimate a post-treatment gap for each event time using a counterfactual constructed from pre-treatment information, and summarize only a clearly specified horizon. A dynamic pattern is informative when its timing, units, assumptions, and uncertainty are reported together.

Check Your Understanding

Use these questions to check your interpretation of treatment timing and dynamic gaps.

A policy is announced two periods before implementation. What risk arises if both announcement periods are used to fit the synthetic control?
In one period, the treated outcome is 24 units and the synthetic estimate is 20 units. What is the estimated gap, and what additional condition is needed to interpret it causally?
What does an average effect over the first four post-treatment periods summarize, and what does it not imply about each period’s effect?
Why can choosing synthetic-control weights to minimize post-treatment gaps compromise the analysis?
For a phased intervention, name one possible onset convention and explain why stating it matters for event-time interpretation.