Repeated-Period DiD With a Common Intervention Date

Why Use More Than Two Periods?

A two-period difference-in-differences design compares one pre-intervention outcome with one post-intervention outcome for treated and control groups. When outcomes are observed in several periods on each side of a common intervention date, the same comparison can use information from all selected periods. The key is to specify how the periods are summarized: the repeated-period contrast below gives equal weight to each pre-period and equal weight to each post-period.

This setup does not require treated and control groups to have the same outcome level before intervention. It does require a credible comparison of how their outcomes would have changed without intervention. The repeated observations make the outcome pattern richer, but do not by themselves make the causal assumption true.

Definition and Estimator

Let $G_i=T$ indicate membership in the treated group and $G_i=C$ indicate membership in the control group. Suppose both groups are observed in $K$ pre-intervention periods and $L$ post-intervention periods. The intervention occurs between the last pre-period and the first post-period, and the date is common to both groups. For example, the pre-periods might be $t=-2,-1$, and the post-periods $t=1,2,3$.

Write $\mu_{g,t}$ for the mean outcome in group $g$ during period $t$, and $\widehat{\mu}_{g,t}$ for its sample estimate. Define the average pre-period and post-period means for each group by taking the arithmetic mean of that group's period means. Thus every selected period receives equal weight, regardless of how many observations were recorded in it.

Definition: The repeated-period DiD estimator with a common intervention date and equal period weights is $$ \widehat{\tau}_{\mathrm{DiD}} = \left( \frac{1}{L}\sum_{t\in\mathcal{P}}\widehat{\mu}_{T,t} - \frac{1}{K}\sum_{t\in\mathcal{R}}\widehat{\mu}_{T,t} \right) - \left( \frac{1}{L}\sum_{t\in\mathcal{P}}\widehat{\mu}_{C,t} - \frac{1}{K}\sum_{t\in\mathcal{R}}\widehat{\mu}_{C,t} \right), $$ where $\mathcal{R}$ is the set of selected pre-intervention periods and $\mathcal{P}$ is the set of selected post-intervention periods. It is the treated group's average post-minus-pre change minus the control group's average post-minus-pre change.

The estimator is a sample contrast. To interpret it as an effect, define $Y_{it}(0)$ as unit $i$'s outcome in period $t$ without the intervention, and $Y_{it}(1)$ as its outcome when exposed to the intervention. In the treated group, the intervention is absent before the common date and present in the post-periods; the control group remains untreated. The target under the assumptions below is the equally weighted average treatment effect for the treated group across the selected post-periods.

Worked Example: Calculating the Repeated-Period Contrast

Suppose a hypothetical program is introduced between periods 0 and 1. The outcome is hospital admissions per 1,000 residents. The table gives the group mean in each period.

Group	Period -2	Period -1	Period 1	Period 2	Period 3
Treated	42	44	50	53	55
Control	38	39	40	41	42

The treated group's average pre-period mean is $(42+44)/2=43$ admissions per 1,000. Its average post-period mean is $(50+53+55)/3=52.667$, so its average change is $52.667-43=9.667$ admissions per 1,000.

For the control group, the average pre-period mean is $(38+39)/2=38.5$, and the average post-period mean is $(40+41+42)/3=41$. Its average change is $41-38.5=2.5$ admissions per 1,000. Therefore, $$ \widehat{\tau}_{\mathrm{DiD}} = (52.667-43)-(41-38.5) = 9.667-2.5 = 7.167. $$ The estimate is 7.167 additional admissions per 1,000 residents, as a treated-minus-counterfactual change, averaged equally across the three post-periods. This causal interpretation requires the assumptions discussed below; the table alone cannot establish them.

Why the Contrast Can Identify an Effect

The identifying idea is that the control group's change helps estimate how the treated group's outcomes would have changed without intervention. Define $\mu^0_{T,t}$ as the treated group's mean untreated potential outcome in period $t$, and $\mu^0_{C,t}$ as the corresponding mean for the control group. Define the period-specific treated effect by $\tau_t=E[Y_{it}(1)-Y_{it}(0)\mid G_i=T]$.

The assumptions needed for the basic interpretation are:

Common intervention timing and treatment status: the treated group begins exposure at the same date, and the control group is not exposed during the selected periods.
No anticipation: before the intervention date, treated outcomes equal their untreated potential outcomes. The intervention does not affect outcomes in advance.
Parallel trends in untreated outcomes: the difference between treated and control mean untreated outcomes does not change across the selected periods. In particular, the average treated-control gap over the post-periods equals its average over the pre-periods.
Consistency and no spillovers: observed outcomes correspond to the treatment status actually received, and the intervention does not change control outcomes through spillovers from the treated group.
Comparable group composition: the group-period means refer to stable, comparable target populations, or the design otherwise handles changes in who is represented. This matters especially when different units are sampled in different periods.

The parallel-trends condition allows treated and control groups to start at different levels. It restricts changes in their untreated outcomes, not equality of their observed outcome means. Under these assumptions, a short algebraic decomposition shows what remains in the contrast.

Write observed treated post-period means as untreated outcomes plus effects.
For each post-period $t$, the treated mean is $\mu^0_{T,t}+\tau_t$. Before intervention, no anticipation makes the observed treated mean equal to $\mu^0_{T,t}$.

Separate the effect from the untreated-outcome contrast.
The population DiD equals the average post-period treated effect plus the change in the treated-control gap in untreated outcomes, where each pre- and post-period set is averaged equally.

Apply parallel trends.
Parallel trends makes the average post-period untreated gap equal the average pre-period untreated gap. Their difference is zero, leaving only the equally weighted average of the post-period effects.

In symbols, the decomposition and its implication are

\begin{aligned} \tau_{\mathrm{DiD}} &= \frac{1}{L}\sum_{t\in\mathcal{P}}\tau_t + \left[ \frac{1}{L}\sum_{t\in\mathcal{P}}(\mu^0_{T,t}-\mu^0_{C,t}) - \frac{1}{K}\sum_{t\in\mathcal{R}}(\mu^0_{T,t}-\mu^0_{C,t}) \right],\\ \text{parallel trends} &\implies \frac{1}{L}\sum_{t\in\mathcal{P}}(\mu^0_{T,t}-\mu^0_{C,t}) = \frac{1}{K}\sum_{t\in\mathcal{R}}(\mu^0_{T,t}-\mu^0_{C,t}),\\ \therefore\quad \tau_{\mathrm{DiD}} &= \frac{1}{L}\sum_{t\in\mathcal{P}}\tau_t. \end{aligned}

The derivation explains why a single repeated-period estimate need not be the effect in every post-period. If the effect changes over time, this equal-period-weighted contrast estimates the average of those period-specific effects, not their sum and not necessarily the effect in the final period.

Worked Example: Effects That Change Across Post-Periods

Consider a hypothetical outcome measured in points. Suppose the control group's means are 20 and 22 in the two pre-periods, then 24 and 26 in the two post-periods. The treated group's observed means are 30 and 32 before intervention, then 39 and 44 after it.

The treated group's pre-period average is $(30+32)/2=31$, and its post-period average is $(39+44)/2=41.5$. Its average change is $41.5-31=10.5$ points. The control group's pre-period average is $(20+22)/2=21$, and its post-period average is $(24+26)/2=25$, for a change of $25-21=4$ points. Thus the repeated-period estimate is $10.5-4=6.5$ points.

For interpretation, suppose the untreated treated-group means would have continued as 34 and 36 in the post-periods. These are each 10 points above the control means in the corresponding period, as they were in the pre-periods. The observed treated means, 39 and 44, are therefore 5 and 8 points above those untreated counterfactual means. Their equal-weighted average is $(5+8)/2=6.5$ points, matching the DiD contrast. This illustrates the target under parallel trends; in an actual study, the counterfactual means are not observed.

Period Weighting Is Part of the Estimand

Averaging period means gives each period equal influence. This is a deliberate choice, not a consequence of having repeated data. If sample sizes vary substantially across periods, pooling all observations before calculating a pre- or post-period mean generally gives more weight to periods with more observations. That can produce a different contrast and target. The two calculations coincide only under conditions such as equal period sample sizes or identical period means within the relevant pre- and post-period sets.

Worked Example: Why Unequal Period Counts Matter

Suppose four hypothetical period means are recorded for each group. The treated group has pre-period means 10 and 12 and post-period means 17 and 20. The control group has pre-period means 8 and 9 and post-period means 10 and 12. The equal-period calculation gives a treated change of $(17+20)/2-(10+12)/2=18.5-11=7.5$, and a control change of $(10+12)/2-(8+9)/2=11-8.5=2.5$. The repeated-period DiD is $7.5-2.5=5$ outcome units.

Now suppose each group's sample sizes are 100 in the first pre-period, 10 in the second pre-period, 10 in the first post-period, and 100 in the second post-period. Pooling observations gives a treated pre mean of $(100(10)+10(12))/110=10.182$ and post mean of $(10(17)+100(20))/110=19.727$, a change of 9.545. The pooled control pre mean is $(100(8)+10(9))/110=8.091$, and post mean is $(10(10)+100(12))/110=11.818$, a change of 3.727. The pooled contrast is $9.545-3.727=5.818$ outcome units, not 5.

Neither arithmetic is automatically incorrect: they use different period weights. The equal-period estimator targets a contrast in which the selected periods count equally. A pooled estimator emphasizes periods with more observations. State the weighting rule so the estimand is clear.

What the Estimate Does and Does Not Establish

Repeated-period DiD is useful when outcomes are measured more than once before and after a shared intervention date. It summarizes the treated and control trajectories without requiring their baseline levels to match. The pre-period observations also show whether the groups had different observed levels or visible changes, but they cannot prove that their untreated outcomes would have followed parallel trends after intervention.

The causal interpretation depends on the design and assumptions, not on the subtraction alone. In particular, a control group exposed to a related policy, anticipation by the treated group, or changing group composition can undermine the counterfactual comparison. A numerical estimate should therefore be described as causal only with a defensible account of why the control group's change represents the treated group's untreated change.

Common Mistakes

Comparing only post-period levels. A higher treated post-period mean may reflect a pre-existing level difference. The correction is to compare changes in the two groups.
Assuming equal baseline levels are necessary. They are not. Parallel trends concerns untreated changes; the groups may begin at different outcome levels.
Calling the estimate the effect in every post-period. With changing effects, equal weighting estimates their average across the selected post-periods.
Ignoring the period weights. Equal averaging and pooling observations can produce different estimates when counts vary. Report which contrast was used.
Treating observed pre-period patterns as proof of parallel trends. They can inform the plausibility of the assumption, but untreated post-period outcomes for the treated group remain unobserved.

Key takeaway: With a common intervention date, repeated-period DiD subtracts the control group's average pre-to-post change from the treated group's average pre-to-post change. Under no anticipation, parallel trends in untreated outcomes, and a well-defined weighting and target population, equal period weights identify the equally weighted average treated effect across selected post-periods.

Check Your Understanding

Answer each question by stating the relevant comparison or assumption.

There are three pre-periods and two post-periods. What does equal period weighting mean for the contribution of each period?
Why can treated and control groups have different average outcome levels in every pre-period while still satisfying parallel trends?
If the treatment effect is 4 units in the first post-period and 10 units in the second, what post-period effect does an equally weighted contrast target?
Why might pooling all observations produce a different contrast from averaging period means when sample sizes differ by period?
Which assumption rules out an intervention effect on treated outcomes before the common intervention date?