Already-Treated Controls and Dynamic Contamination

Why Already-Treated Controls Need Special Attention

In a staggered-adoption design, different cohorts begin treatment in different periods. A standard two-period difference-in-differences comparison uses one group that changes from untreated to treated and another group that remains untreated. But some comparisons in a multi-period analysis use a group that is already treated as the control. That group’s outcome change can reflect not only the change that would have occurred without treatment, but also a change in its treatment effect.

The resulting problem is dynamic contamination: the comparison subtracts the evolution of the control cohort’s treatment effects from the effect for the cohort adopting treatment. It is not enough to assume that effects are heterogeneous across cohorts. The key question is whether the effects for the already-treated control change over the comparison interval.

Definition: Let $G_i$ denote unit $i$’s first treatment period, with $G_i=\infty$ for a never-treated unit. Write $Y_{it}(g)$ for the outcome unit $i$ would have in period $t$ if treatment began in period $g$, and $Y_{it}(\infty)$ for its outcome if it remained untreated. For cohort $g$ in a period $t\ge g$, define its cohort-time average treatment effect as $$ \tau_{g,t}=E\!\left[Y_{it}(g)-Y_{it}(\infty)\mid G_i=g\right]. $$ This is the average effect for cohort $g$ in calendar period $t$, not necessarily a constant effect over time.

Assume treatment is absorbing: after adoption, a unit remains treated. Assume no anticipation, so a unit’s outcome before its adoption equals its never-treated outcome. The derivation below also requires parallel trends in untreated outcomes for the cohorts being compared over the chosen interval, and no spillovers between units. These conditions allow the untreated outcome changes to cancel; they do not require treatment effects to be constant.

The Already-Treated-Control Contrast

Consider two periods $t_0<t_1$. Cohort $h$ is untreated at $t_0$ and adopts by $t_1$, while cohort $g$ adopted by $t_0$ and is treated in both periods. The observed two-period contrast is

D_{h,g}(t_0,t_1) = \bigl(E[Y_{it_1}\mid G_i=h]-E[Y_{it_0}\mid G_i=h]\bigr) - \bigl(E[Y_{it_1}\mid G_i=g]-E[Y_{it_0}\mid G_i=g]\bigr).

For cohort $h$, no anticipation means its observed outcome at $t_0$ is its untreated outcome. At $t_1$, its observed outcome includes $\tau_{h,t_1}$. Cohort $g$’s observed outcomes include its treatment effects in both periods. Under parallel trends in the untreated outcomes, the cohorts’ untreated changes are equal and cancel in the contrast. Therefore,

D_{h,g}(t_0,t_1) = \tau_{h,t_1} - \left(\tau_{g,t_1}-\tau_{g,t_0}\right).

The first term is the effect for the adopting cohort at $t_1$. The second term is the change in the already-treated cohort’s effect between $t_0$ and $t_1$. If the control cohort’s effect grows, the contrast is smaller than $\tau_{h,t_1}$; if its effect shrinks, the contrast is larger. Only when its treatment effect does not change over this interval does this particular source of contamination vanish.

Key result: With an already-treated cohort as the control, a valid parallel-trends comparison identifies the adopting cohort’s effect minus the control cohort’s change in effects. It generally does not identify the adopting cohort’s effect alone.

Worked Comparisons

Worked Example: A Growing Effect Attenuates the Contrast

Suppose an early cohort adopted in period 2 and a later cohort adopts in period 3. Compare their outcome changes from period 2 to period 3, using the early cohort as the control. The following hypothetical cohort means have untreated outcomes that rise by 2 units for both cohorts. Treatment effects are shown only when treatment has begun.

Cohort	Period 2 untreated outcome	Period 3 untreated outcome	Period 2 effect	Period 3 effect
Early, $G=2$	10	12	3	7
Later, $G=3$	8	10	Untreated	5

The early cohort’s observed mean is $10+3=13$ in period 2 and $12+7=19$ in period 3, a change of $19-13=6$. The later cohort is untreated in period 2, so its observed mean is 8; in period 3 its observed mean is $10+5=15$, a change of $15-8=7$. Thus the observed difference-in-differences contrast is $$ D_{3,2}(2,3)=(15-8)-(19-13)=7-6=1. $$

The algebra gives the same result: $$ \tau_{3,3}-(\tau_{2,3}-\tau_{2,2}) =5-(7-3)=1. $$

The later cohort’s effect in period 3 is 5 outcome units, but the comparison is only 1 unit because the early cohort’s effect grew by 4 units over the interval. The assumed parallel untreated changes hold here: each cohort’s untreated outcome rises by 2. The discrepancy is therefore due to the already-treated control’s changing effect, not a difference in untreated trends.

Worked Example: A Declining Control Effect Inflates the Contrast

Now suppose the early cohort’s effect declines between periods 2 and 3, while the later cohort’s period-3 effect is 4 units. Let both cohorts’ untreated outcomes rise by 1 unit over the interval. For the early cohort, take untreated outcomes of 20 and 21, with effects of 8 and 5. For the later cohort, take untreated outcomes of 15 and 16, with no treatment in period 2 and an effect of 4 in period 3.

The early cohort’s observed outcomes are $20+8=28$ and $21+5=26$, so its change is $26-28=-2$. The later cohort’s observed outcomes are 15 and $16+4=20$, so its change is $20-15=5$. The comparison is $$ D_{3,2}(2,3)=(20-15)-(26-28)=5-(-2)=7. $$

The effect-based calculation is $$ \tau_{3,3}-(\tau_{2,3}-\tau_{2,2}) =4-(5-8)=7. $$

The contrast exceeds the later cohort’s effect because the early cohort’s effect fell by 3 units. Its observed outcome change is reduced by that decline, making the difference between the cohorts’ changes larger. Both cohorts have the same 1-unit untreated trend, and the later cohort has no anticipation before adoption. A large contrast in this example is not evidence that the later cohort’s effect is 7 units.

Worked Example: A Collection of Pairwise Comparisons

Consider three hypothetical cohorts, A, B, and C, adopting in periods 2, 3, and 4, respectively. Suppose untreated outcomes for each cohort rise by 1 unit per period. Their effects are: cohort A has effects 2, 6, and 10 in periods 2, 3, and 4; cohort B has effects 3 and 5 in periods 3 and 4; cohort C has an effect of 4 in period 4. Each comparison below uses a cohort already treated at the start of its interval as control.

Compare B with A from period 2 to period 3. Cohort B’s period-3 effect is 3, while A’s effect rises from 2 to 6. Hence $$ D_{B,A}=3-(6-2)=-1. $$

Compare C with A from period 3 to period 4. Cohort C’s period-4 effect is 4, while A’s effect rises from 6 to 10. Thus $$ D_{C,A}=4-(10-6)=0. $$

Finally, compare C with B from period 3 to period 4. Cohort B’s effect rises from 3 to 5, so $$ D_{C,B}=4-(5-3)=2. $$

The three contrasts are $-1$, 0, and 2, even though the adopting cohorts’ effects used as the first terms are 3, 4, and 4. The different contrasts arise because each comparison subtracts a different control cohort’s effect change. The untreated trends are parallel by construction, and each adopting cohort is untreated at its comparison’s baseline period. This example shows why a collection of staggered comparisons cannot automatically be read as a collection of clean cohort-time effects.

How the Contamination Enters a Staggered-Adoption Estimate

With several adoption cohorts and periods, a two-way fixed-effects (TWFE) regression can combine comparisons across different cohort pairs and time intervals. Some comparisons use groups that are untreated throughout the interval; others use groups that are already treated throughout it. The latter comparisons have the form derived above. Consequently, the regression’s aggregate coefficient can include terms involving changes in treatment effects for cohorts serving as controls.

The exact contribution of each comparison depends on the design and the regression’s weighting structure. It is not generally correct to describe the coefficient as a simple average of the $\tau_{g,t}$. The earlier discussion of TWFE decompositions and negative weights concerns how the regression can combine comparisons and effects; the additional point here is what an already-treated-control comparison contains before aggregation: a target effect minus a change in another cohort’s effects.

This is a distinct issue from whether a cohort’s treatment effect is larger or smaller than another cohort’s effect. Suppose a cohort has a constant effect of 5 in both periods. It can differ from the adopting cohort’s effect, yet its change is zero, so it does not create this particular dynamic-contamination term. Conversely, even if all cohorts have similar effects at one date, a control cohort whose effect changes substantially across the interval can contribute substantial contamination.

The result also clarifies why assumptions and interpretation must be kept separate. Parallel trends in untreated outcomes justify cancellation of the untreated trends. They do not make an already-treated cohort untreated, and they do not remove its evolving treatment effects. Thus a comparison can satisfy parallel trends and still fail to identify the adopting cohort’s effect by itself.

Common Pitfall and Practical Reading

Wrong reasoning: “The two cohorts would have followed parallel trends without treatment, so the difference-in-differences contrast is the later cohort’s treatment effect.” This skips the fact that the control cohort is treated in both periods.

Correction: Write out the control cohort’s treatment effects at both dates. Under the stated assumptions, the contrast is $\tau_{h,t_1}-(\tau_{g,t_1}-\tau_{g,t_0})$. To interpret it as the adopting cohort’s effect, one needs a reason that the control cohort’s effect change is zero, or another design or estimator that avoids using that treated cohort as the control.

When reading a staggered-adoption estimate, identify which cohorts are untreated at each baseline and follow-up date. For any comparison in which a control cohort is already treated at both dates, ask how its treatment effect could evolve over that interval. A single aggregate coefficient can obscure that question by combining multiple comparisons and effect changes.

Classify adoption timing.
For each comparison interval, identify which cohort is untreated at baseline, which adopts by follow-up, and whether the control cohort was already treated at both dates.

Check the untreated-outcome comparison.
State the parallel-trends condition for the cohorts over that interval, along with no anticipation and no spillovers where needed.

Write the effect decomposition.
For an already-treated control, distinguish the adopting cohort’s effect from the control cohort’s change in effects.

Check Your Understanding

Use the cohort-time effects and comparison formula to answer each question.

Cohort A is already treated at both dates, and its effect rises from 2 to 5. Cohort B adopts by the second date and has an effect of 8 then. What does the already-treated-control contrast equal?
If the already-treated cohort’s effect is constant over the comparison interval, which contamination term vanishes? Does that fact imply effects are identical across cohorts?
Can parallel trends in untreated outcomes alone ensure that an already-treated-control contrast identifies the adopting cohort’s effect? Explain.
An already-treated cohort’s effect falls by 4 units, and the adopting cohort’s effect is 6 units. What is the contrast, and why is it larger than the adopting cohort’s effect?
In a staggered design with several cohorts, why might a TWFE coefficient differ from a simple average of cohort-time effects even when untreated outcomes satisfy parallel trends?