Difference in differences and parallel trends

Let G=1 denote the group treated in period one, with period zero as baseline, while G=0 remains untreated in both periods. Write Yₜ(0) for untreated potential outcomes and Y₁(1) for the treated post outcome. No anticipation requires that baseline outcomes in the eventually treated group equal their untreated potential outcomes. Assume consistency, no interference, positive probabilities for both groups, and integrability. The target is the post-period effect on the treated group, not automatically an effect for everyone.

Parallel trends equates the two groups’ mean untreated changes between the periods. Add the observed comparison-group change to the treated group’s observed baseline mean to obtain its counterfactual untreated post mean. Subtracting that prediction from its observed post mean yields the DID contrast. Baseline levels may differ because the identifying restriction concerns changes. A common additive time shock with a fixed group offset satisfies this condition, while a group-specific untreated shock can violate it.

Several untreated pre-periods permit useful checks for earlier trend differences and dynamic behavior. A failure to reject a pretrend difference is not proof that the relevant post-period counterfactual trends would have been parallel. Tests can lack power, treatment may be anticipated, and a new shock can coincide with adoption. Conversely, identification for a specified post comparison need not assert identical group trends at every historical date. The exact temporal assumption should match the design and scientific context.

A useful sensitivity parameter is the difference between the treated and comparison groups’ untreated mean changes. The observed DID equals the treated effect plus that difference, so specifying a possible deviation immediately shows how the causal conclusion moves. This is not an estimated correction unless the deviation is identified by additional information. It makes the identifying content transparent: DID cancels a common trend, not an arbitrary differential trend. A regression coefficient with group and time indicators does not remove the assumption.

With repeated cross-sections, group means may change because different kinds of units are sampled at each date. Comparable group composition or an explicit covariate-standardization design is needed to interpret those changes as the desired population trajectories. A stable panel avoids that particular replacement issue but can still suffer attrition, interference, or changing measurement. Uncertainty also needs to reflect serial dependence and the level at which treatment varies; treating every repeated observation as independent can exaggerate precision.

With staggered adoption and heterogeneous effects, a simple two-way fixed-effects coefficient can combine comparisons involving already-treated groups and weight different group-time effects in unintuitive ways. One should define cohort-time targets and select genuinely appropriate never-treated or not-yet-treated comparisons under explicit anticipation and trend assumptions. The two-period formula does not automatically validate every multigroup regression using a similar name. The durable principle is to construct a credible untreated trajectory for a named treated target before aggregating effects across groups or dates.