Group-Time ATT Estimation With Never-Treated Controls

Why Estimate Effects by Cohort and Time?

When units adopt treatment in different periods, the effect at a particular calendar time need not be the same for every adoption cohort. A group-time estimate keeps those distinctions visible: it compares one cohort’s outcome change with the change among units that are never treated. This avoids treating all treated units as if they shared one treatment date or one effect.

The comparison is a difference-in-differences (DiD). It is causal only under assumptions about the outcome trend the treated cohort would have experienced without treatment. The never-treated group supplies an estimate of that counterfactual trend; it does not make the assumption unnecessary.

Potential Outcomes and the Group-Time ATT

Let $G_i$ denote unit $i$’s first treatment period, with $G_i=\infty$ for a unit that is never treated. Treatment is absorbing: a unit with $G_i=g$ is untreated before $g$ and treated from $g$ onward. Write $Y_{it}(g)$ for the potential outcome at time $t$ if unit $i$ first adopts in period $g$, and $Y_{it}(\infty)$ for its potential outcome if it remains untreated throughout. The observed outcome is the potential outcome corresponding to the unit’s actual adoption path.

Definition: For a cohort first treated in period $g$, its group-time average treatment effect is $$ ATT(g,t)=E\left[Y_{it}(g)-Y_{it}(\infty)\mid G_i=g\right],\qquad t\ge g. $$ It is the average effect, at calendar time $t$, of adopting in period $g$ rather than remaining untreated, among units in cohort $g$.

The natural baseline for cohort $g$ is period $g-1$, immediately before adoption. Under no anticipation, adoption in $g$ has not affected outcomes at that baseline. Using a common baseline for that cohort, the DiD compares its outcome change from $g-1$ to $t$ with the same-period change among never-treated units.

Identification Conditions and the DiD Formula

The basic formulation below uses a panel in which outcomes are observed for the same units at the baseline and at $t$. In repeated cross-sections, the corresponding quantities are changes in population means, estimated from samples representative of the same groups at each period. In either setting, the causal interpretation requires the following conditions.

Consistency and no interference: observed outcomes follow each unit’s actual treatment path, and one unit’s treatment does not change another unit’s outcome.
No anticipation: for cohort $g$, outcomes before treatment begins are unaffected by the future adoption. In particular, $Y_{i,g-1}(g)=Y_{i,g-1}(\infty)$ for units with $G_i=g$.
Parallel trends for the untreated potential outcome: for each target $t\ge g$, the mean change that cohort $g$ would have had without treatment equals the mean change among never-treated units: $$ E\left[Y_{it}(\infty)-Y_{i,g-1}(\infty)\mid G_i=g\right] = E\left[Y_{it}(\infty)-Y_{i,g-1}(\infty)\mid G_i=\infty\right]. $$
Overlap and observed support: the cohort and never-treated group both have positive probability, and outcomes are observed for both groups at $g-1$ and $t$. Thus the never-treated comparison exists for the target cohort and period.

Theorem: Under consistency, no interference, no anticipation, parallel trends as stated above, and overlap, $$ ATT(g,t) = E\left[Y_{it}-Y_{i,g-1}\mid G_i=g\right] - E\left[Y_{it}-Y_{i,g-1}\mid G_i=\infty\right]. $$

The first expectation is the treated cohort’s observed mean change. The second is the never-treated group’s observed mean change. Subtracting removes the change that, under parallel trends, represents the cohort’s counterfactual untreated trend. The remaining difference identifies the cohort’s average treatment effect at time $t$.

From Identification to a Sample Estimate

With panel data, let $n_g$ be the number of observed units in cohort $g$, and let $n_\infty$ be the number of observed never-treated units. The sample analogue is

\widehat{ATT}(g,t) = \frac{1}{n_g}\sum_{i:G_i=g}(Y_{it}-Y_{i,g-1}) - \frac{1}{n_\infty}\sum_{i:G_i=\infty}(Y_{it}-Y_{i,g-1}).

This formula estimates the population contrast; it does not guarantee that a particular finite-sample estimate equals the population ATT. The arithmetic units are the outcome’s units, such as percentage points, dollars, or days. In applied work, uncertainty should also be quantified using an inferential procedure appropriate to the sampling and treatment-assignment structure.

Worked Example: A Cohort’s Effect at One Calendar Time

Suppose a hypothetical panel study observes a cohort first treated in period 3 and a group never treated. At period 2, the cohort’s mean outcome is 50 points and the never-treated group’s mean is 48 points. At period 4, their respective means are 64 and 55 points. Assume outcomes are observed for the same units in both periods.

The cohort’s change is $64-50=14$ points. The never-treated group’s change is $55-48=7$ points. The sample DiD is $$ \widehat{ATT}(3,4)=(64-50)-(55-48)=14-7=7\text{ points}. $$

Under the stated identification conditions, this estimates the average effect at period 4 for units first treated in period 3, relative to those units remaining untreated. The estimate is not simply the 9-point difference between the groups at period 4: the baseline gap was 2 points, and the DiD compares changes rather than endpoint levels. The calculation alone cannot establish that parallel trends or the other assumptions hold.

Why the Same Cohort Can Have Different Effects Across Time

For a fixed cohort $g$, estimate $ATT(g,t)$ separately at each post-adoption period $t$. Each estimate compares the cohort’s change from $g-1$ to that target period with the never-treated group’s change over the same interval. The comparison group and baseline remain the same, but the outcome period changes. This provides cohort-specific effects at calendar times; it does not require the effect to be constant over time.

Worked Example: Two Post-Adoption Periods for One Cohort

Consider a hypothetical cohort first treated in period 3. Its mean outcomes in periods 2, 3, and 4 are 40, 48, and 55 points. The corresponding means for the never-treated group are 42, 46, and 49 points. Period 2 is the cohort’s $g-1$ baseline.

For period 3, the cohort changes by $48-40=8$ points, while the never-treated group changes by $46-42=4$ points. Therefore, $$ \widehat{ATT}(3,3)=(48-40)-(46-42)=8-4=4\text{ points}. $$

For period 4, the changes from the same baseline are $55-40=15$ points and $49-42=7$ points. Therefore, $$ \widehat{ATT}(3,4)=(55-40)-(49-42)=15-7=8\text{ points}. $$

The two estimates differ: 4 points at period 3 and 8 points at period 4. Under the assumptions, this pattern represents a larger average effect at the later calendar time for this cohort. It is not obtained by changing the baseline or comparing period 4 only with period 3.

Aggregating Cohort-Specific Effects

A single cohort-time ATT answers a specific question. If a summary is needed for all units treated by calendar time $t$, the relevant aggregate weights each eligible cohort by its share of those treated units. Let the set of cohorts treated by $t$ be $\mathcal{G}_t=\{g:g\le t\}$. When those cohort effects are identified and the treated population is represented,

ATT^{\text{treated}}_t = \sum_{g\in\mathcal{G}_t}P(G_i=g\mid G_i\le t)\,ATT(g,t).

The weights are nonnegative and sum to one. This aggregate is the average effect at time $t$ among units already treated by $t$, with each cohort represented according to its share of that treated population. It is not an unweighted average of cohort effects unless the cohorts happen to have equal sizes. Nor does it combine effects from different calendar times.

Worked Example: Combining Two Cohorts at Period 4

Suppose a hypothetical study estimates $ATT(2,4)=6$ points for a cohort of 30 units and $ATT(4,4)=2$ points for a cohort of 70 units. Assume these are the only cohorts treated by period 4 and both estimates use valid never-treated comparisons.

The shares among units treated by period 4 are $30/(30+70)=0.30$ and $70/(30+70)=0.70$. The treated-population aggregate is $$ ATT^{\text{treated}}_4 = 0.30(6)+0.70(2) = 1.8+1.4 = 3.2\text{ points}. $$

The equal-cohort average would instead be $(6+2)/2=4$ points. That number answers a different question: it gives each cohort equal weight, regardless of its size. The 3.2-point estimate is the average effect for the 100 treated units under the stated cohort sizes and assumptions.

Aggregation requires attention to which cohorts are included and how they are weighted. If a cohort lacks outcome support at a target time, it cannot simply be assigned a group-time estimate from this design; omitting it changes the population represented by the aggregate. Reporting the cohort-specific estimates alongside any summary makes that target clearer.

Common Mistakes and Why This Design Matters

Using endpoint differences as the effect. A treated-control difference at $t$ includes any baseline gap. The DiD subtracts the groups’ changes from the cohort-specific baseline.
Assuming never-treated controls guarantee parallel trends. Never-treated units provide an untreated comparison path, but the identification claim still requires the treated cohort’s counterfactual untreated trend to match theirs on average.
Changing the baseline without changing the interpretation. The formula here uses $g-1$, immediately before adoption, and relies on no anticipation. A different baseline defines a different comparison and needs its own justification.
Giving cohorts equal weight by default. Equal weighting estimates an average across cohorts, not necessarily the average effect among all units treated by $t$. State the target population and choose weights accordingly.
Calling the estimate causal without its conditions. A calculated DiD is a contrast in observed changes. Its causal interpretation follows only if the identification assumptions are credible for the cohort, comparison group, and periods used.

Key takeaway: Estimate each $ATT(g,t)$ by comparing cohort $g$’s change from $g-1$ to $t$ with the never-treated group’s change over that same interval. Parallel trends and overlap make this contrast informative about the causal effect; explicit aggregation weights determine which treated population a summary describes.

Check Your Understanding

Use the definitions and conditions above to answer these questions.

In words, what population does $ATT(g,t)$ describe, and what two potential outcomes does it compare?
A cohort’s mean outcome rises by 9 points while the never-treated group’s mean rises by 5 points over the same interval. What is the DiD estimate, and what additional assumptions are needed for a causal interpretation?
Why is the never-treated group’s observed outcome change relevant to a cohort’s counterfactual untreated trend?
At a given time, two cohorts have effects of 5 and 1 points and sizes of 20 and 80 units. What is the treated-population weighted aggregate?
How does an equal-cohort average differ in interpretation from an aggregate weighted by cohort shares among treated units?