Difference‑in‑Differences (DiD) is a common way to estimate effects of policies or treatments by comparing how outcomes change over time for treated and untreated groups. A key assumption behind many DiD studies is conditional parallel trends (CPT): after accounting for some observed variables, the treated and untreated groups would have followed the same time path without the treatment. Michael C. Knaus and Henri Pfleiderer introduce a new graphical tool that helps researchers decide which variables they must condition on to make CPT plausible.

The authors build on Single World Intervention Graphs (SWIGs) and create a transformed version they call Δ‑SWIGs. At a high level, these graphs let researchers read off certain independence relationships from a picture. The technical rule they use is called d‑separation, which is a way to tell from the graph whether one set of variables provides enough information to block confounding paths between treatment and outcome. Using Δ‑SWIGs, the paper translates the CPT requirement into graphical conditions. It then studies how those conditions change when there are many time periods and when covariates change over time.

A central practical finding is about time‑varying covariates — variables that change after treatment starts. If such covariates affect the outcome, the paper shows that researchers often need to control for post‑treatment variables to identify effects. The authors use simulations to illustrate several patterns. For example, using only pre‑treatment covariates can give unbiased short‑term effects but biased longer‑run effects; controlling for covariates measured just before each outcome often matches the results of controlling for the whole sequence of covariates; and when the treatment affects future covariates (treatment–covariate feedback), many dynamic effect estimates become biased. The simulations reported use a very large sample (10 million) to make these patterns clear.