← Every exhibit
Why must causal inference name the cohort-time estimand before running regressions?
| Field | Question | Why it matters |
|---|---|---|
| Cohort | Which units first receive treatment in this group? | Treatment timing is part of the object. |
| Calendar time | In which period is the outcome observed? | A cohort can have different effects over time. |
| Event time | How far before or after treatment is the period? | Dynamic effects are not the same as one post-period average. |
| Comparison | Which untreated or not-yet-treated units provide the contrast? | Already-treated units may not be valid controls for a later cohort. |
| Outcome | What is measured, in which unit, and at what horizon? | A coefficient cannot repair an unclear outcome. |
| Aggregation | Which cohort-time effects are combined, and with what weights? | The overall number depends on the aggregation target. |
| Inference | Where was treatment assigned, and where should uncertainty be clustered? | Precision is part of the design, not an afterthought. |
Swipe or scroll horizontally if the table is wider than your screen.
Reference & Evidence
Source: Table from this essay. Sources and interpretation are given in the article.
Related exhibits
-
The staggered DiD estimand release gate
From the essay Staggered difference-in-differences needs a cohort-time estimand
-
The synthetic interference exposure ledger
From the essay What is interference? When one unit changes another unit's outcome
-
The measurement-error indicator audit
From the essay What is measurement error? The gap between a construct and its indicator