Why Parallel Trends Matters
A difference-in-differences comparison subtracts the change in an untreated comparison group from the change in a treated group. The subtraction is useful only if the comparison group’s change represents what would have happened to the treated group without treatment. Parallel trends is the assumption that gives this interpretation a precise form: the groups may start at different outcome levels, but their average outcomes in the absence of treatment would have changed by the same amount.
The key phrase is “in the absence of treatment.” After treatment begins, the treated group’s untreated outcome is a counterfactual: it cannot be observed for the same units at the same time as their treated outcomes. Parallel trends is therefore an identifying assumption, not a fact established simply by calculating a difference-in-differences statistic.
Potential Outcomes and the Target Effect
Consider a two-period design with a pre-treatment period \(t=0\) and a post-treatment period \(t=1\). Let \(D_i=1\) indicate that unit \(i\) is in the group assigned to treatment, and let \(D_i=0\) indicate membership in the comparison group. Write \(Y_{it}(0)\) for unit \(i\)'s outcome at time \(t\) under no treatment, and \(Y_{it}(1)\) for its outcome under treatment. The treatment is introduced between periods 0 and 1.
Assume there is no anticipation: treatment does not affect outcomes in the pre-treatment period, so \(Y_{i0}(1)=Y_{i0}(0)\). Treatment is absent in both periods for comparison units, and treated units receive it only in period 1. Under consistency, the observed outcome is the potential outcome corresponding to the treatment status actually received. Thus, for treated units, \(Y_{i0}=Y_{i0}(0)\) and \(Y_{i1}=Y_{i1}(1)\); for comparison units, \(Y_{i0}=Y_{i0}(0)\) and \(Y_{i1}=Y_{i1}(0)\).
The ATT is not generally the difference between the treated and comparison groups’ observed post-treatment means. Those groups can differ for reasons other than treatment, including differences already present before treatment. Difference-in-differences uses changes over time to remove a stable level difference, but it needs an assumption about the groups’ counterfactual changes.
Parallel Trends as an Identification Assumption
This statement concerns average changes in the potential outcome under no treatment. It does not say that treated and comparison units have equal outcomes in either period. Nor does it require every treated unit’s untreated change to match every comparison unit’s untreated change. The equality is between conditional group averages.
To use the assumption for identification, also require that treatment status and outcomes are well-defined, there are no spillovers that make one unit’s outcome depend on another unit’s treatment, and the groups represent comparable populations over the two periods. These conditions support interpreting observed group means as means for the intended units and treatment regimes. In a panel, the same units are followed; in designs using repeated samples, the group means must instead represent consistently defined populations across time.
Why the Difference-in-Differences Contrast Identifies the ATT
Let \(E[Y_{it}\mid D_i=d]\) denote the mean observed outcome in group \(d\) at time \(t\). The population difference-in-differences contrast is the treated group’s observed mean change minus the comparison group’s observed mean change:
For treated units, the observed post-treatment outcome is \(Y_{i1}(1)\), while their pre-treatment outcome is \(Y_{i0}(0)\). For comparison units, both observed outcomes are untreated potential outcomes. Substituting these quantities gives
Add and subtract the treated group’s mean untreated change. The first resulting term is the ATT, and the difference between untreated changes is zero under parallel trends:
The first equality follows by expanding the two terms: the treated group’s \(Y_{i1}(0)\) and \(Y_{i0}(0)\) cancel, leaving its observed treated-versus-untreated post-treatment difference and the difference between untreated changes. The second equality uses parallel trends. This is the identification result: under the stated conditions, the observed difference-in-differences contrast equals the ATT.
The result identifies a population estimand. In data, a sample difference-in-differences estimate replaces each population mean with a sample mean. Its numerical value is not guaranteed to equal the population ATT exactly; sampling variation remains. The identification assumption concerns the relationship between population counterfactual trends, not whether a particular finite sample produces perfectly equal changes.
Worked Example: Identifying an ATT From Group Means
Suppose a hypothetical program is introduced for one group between two periods. The mean outcome is measured in outcome-points, with the same target populations represented in each period. The observed means are:
| Group | Pre-treatment mean | Post-treatment mean |
|---|---|---|
| Treated | 50 | 62 |
| Comparison | 45 | 53 |
The treated group’s observed change is \(62-50=12\) outcome-points. The comparison group’s observed change is \(53-45=8\) outcome-points. The difference-in-differences contrast is therefore
If consistency, no anticipation, no spillovers, and parallel trends are appropriate for this design, the contrast identifies an ATT of 4 outcome-points. One way to see the counterfactual calculation is that parallel trends assigns the treated group the comparison group’s untreated change of 8: its predicted mean without treatment would be \(50+8=58\). The treated group’s observed post-treatment mean is 62, so the difference from that counterfactual mean is \(62-58=4\) outcome-points. The observed means alone cannot prove that the treated group’s untreated change would have been 8; that equality is the maintained identifying assumption.
Different Levels Are Compatible With Parallel Trends
Parallel trends concerns changes, not starting points. A treated group can have a higher or lower mean than the comparison group before treatment and still satisfy the assumption. A fixed level gap cancels when each group’s pre-treatment mean is subtracted from its post-treatment mean. What matters for identification is whether the groups’ average untreated changes would have been equal.
Worked Example: Unequal Baseline Means
In a hypothetical study, the treated group has a pre-treatment mean of 80 outcome-points and a post-treatment mean of 89. The comparison group’s means are 60 before and 65 after. Assume the treatment is absent before the second period, there is no anticipation or spillover, and the groups represent consistent populations. The baseline gap is \(80-60=20\) outcome-points. The observed changes are
The difference-in-differences contrast is \(9-5=4\) outcome-points. Under parallel trends in untreated potential outcomes, the treated group would have gained 5 points without treatment, so its counterfactual post-treatment mean is \(80+5=85\). Its observed mean of 89 is 4 points higher. The baseline gap of 20 points does not prevent identification: it is not the condition that parallel trends requires. The conclusion is conditional on the assumption about untreated changes, not on equality of observed levels.
Group-Average Trends Need Not Match Unit by Unit
The equality in parallel trends is an equality of conditional means. Individual untreated changes may vary substantially, and particular treated units need not follow the same trajectory as particular comparison units. The relevant requirement is that the average untreated change for the treated group equals the average untreated change for the comparison group.
Worked Example: Unequal Individual Changes, Equal Group Trends
The following fully hypothetical table displays all potential outcomes for four units. \(Y_{i0}(0)\) and \(Y_{i0}(1)\) are equal because there is no anticipation. Comparison units remain untreated, while treated units receive treatment in period 1. The period-1 treatment potential outcome for each comparison unit is included for completeness, though it is not observed.
| Unit | Group | \(Y_{i0}(0)\) | \(Y_{i0}(1)\) | \(Y_{i1}(0)\) | \(Y_{i1}(1)\) |
|---|---|---|---|---|---|
| A | Treated | 10 | 10 | 15 | 19 |
| B | Treated | 20 | 20 | 23 | 29 |
| C | Comparison | 8 | 8 | 10 | 13 |
| D | Comparison | 18 | 18 | 24 | 25 |
For treated units A and B, the untreated changes are \(15-10=5\) and \(23-20=3\), averaging \((5+3)/2=4\) outcome-points. For comparison units C and D, the untreated changes are \(10-8=2\) and \(24-18=6\), also averaging \((2+6)/2=4\) outcome-points. Thus the group-average untreated trends are equal, even though no pairing of individual changes is identical.
The treated group’s observed pre-treatment mean is \((10+20)/2=15\), and its observed post-treatment mean is \((19+29)/2=24\), giving a change of 9. The comparison group’s observed means are \((8+18)/2=13\) and \((10+24)/2=17\), giving a change of 4. Therefore,
The individual treatment effects among A and B are \(19-15=4\) and \(29-23=6\) outcome-points. Their average is \((4+6)/2=5\), matching the difference-in-differences result. This example illustrates why “parallel” describes group-average untreated trends, not identical paths for every unit.
What the Assumption Does Not Establish
Parallel trends is not directly testable in the two-period design. For treated units in period 1, the data reveal \(Y_{i1}(1)\), not \(Y_{i1}(0)\). The treated group’s post-treatment untreated mean—the quantity needed to check its counterfactual change—is unobserved. Observed pre-treatment outcomes may inform whether a design seems plausible, but they cannot establish that the unobserved post-treatment counterfactual trend satisfies the assumption.
The assumption also does not assert that treatment was randomly assigned, that group membership is unrelated to outcome levels, or that treatment effects are identical across units. It says specifically that, absent treatment, the groups would have had equal average changes over the relevant interval. Other design features and substantive knowledge are needed to judge whether that claim is credible.
Common Mistake
- Wrong reasoning: “The treated and comparison groups have different pre-treatment means, so parallel trends fails.” Correction: Different levels are allowed. Parallel trends equates their mean changes under no treatment, not their levels.
- Wrong reasoning: “Some treated units’ untreated outcomes would rise more than some comparison units’ outcomes, so the assumption fails.” Correction: The assumption compares group-average untreated changes; unit-level changes need not match.
- Wrong reasoning: “The comparison group’s observed change proves what would have happened to the treated group.” Correction: That interpretation follows only if parallel trends and the other identifying conditions are credible. The treated group’s post-treatment untreated outcome is counterfactual and unobserved.
Check Your Understanding
Answer each question using the potential-outcome interpretation of the two-period design.
- Write parallel trends as an equality involving conditional expectations of untreated potential-outcome changes.
- Why can treated and comparison groups have unequal pre-treatment means while satisfying parallel trends?
- A treated group’s observed mean changes by 11 points, and the comparison group’s mean changes by 7 points. What ATT is identified if the required conditions hold?
- Does parallel trends require each treated unit’s untreated change to equal each comparison unit’s untreated change? Explain.
- Which outcome needed to evaluate the treated group’s untreated post-period change is not observed, and why does that matter?