Event-Study Panel Plot¶
Metric / dashboard — instantiates Time Series Cross-Section Analysis
Shows trajectories around an event, treatment, or adoption time.
An event-study panel plot re-centers time on the moment each unit is treated: calendar dates are thrown away and replaced by event time — periods since or until each unit's own event, with the event itself at zero. Because different units are treated on different calendar dates, stacking them all on this relative axis turns a scattered, staggered rollout into a single aligned picture. The plot then shows the outcome path across event time, and that path does two jobs at once: the stretch before zero is a visual test of the design's core assumption (it should be flat, since nothing has happened yet), and the stretch after zero shows the effect's dynamic shape as it emerges, grows, or fades. Its defining move is alignment on the event, not the fitting of any single summary number — it is a picture whose flat left half earns the right to believe its rising right half.
Example¶
A city opens bike-share docking stations across sixty neighborhoods over three years, on staggered dates that depend on permitting and construction — no two neighborhoods start on the same month. To see whether stations lift nearby retail activity, an analyst re-indexes each neighborhood's monthly foot-traffic to months relative to its own station opening, from −12 to +18, then averages across neighborhoods at each relative month and plots it. The left half of the curve, months −12 through −1, hovers near zero with no drift — reassuring, because it means neighborhoods getting stations were not already on the rise for other reasons. At month zero the line lifts, climbing to roughly +8% by month +12 before leveling. The value is in the shape: a flat pre-period licenses the causal reading, and the gradual post-period rise shows the effect is cumulative rather than instant — something a single before/after number for one neighborhood could never reveal.
How it works¶
- Re-index to event time. For each unit, relabel periods by distance to its treatment date; pool units on that shared relative axis.
- Estimate leads and lags against a reference. Fit an effect for each event-time offset relative to an omitted period (usually −1), so every point is read against the eve of treatment.
- Read the pre-period as a test. Flat, near-zero leads support the design; sloping leads signal a pre-trend that undermines it.
- Guard the staggered comparisons. With units treated at different times, avoid using already-treated units as controls for later ones.
Tuning parameters¶
- Reference period — which pre-event offset is the baseline the rest are measured against. A poorly chosen reference can visually manufacture or hide a pre-trend.
- Event-window width — how many leads and lags to show. Wider captures slow dynamics but thins the sample at the extremes where few units are observed.
- Endpoint binning — whether distant leads/lags are pooled into "≤−k" and "≥+k" bins to stabilize sparse tails.
- Estimator for staggered timing — a naive two-way specification or a robustness-aware one; the choice governs whether the plotted path is trustworthy under heterogeneous timing.
When it helps, and when it misleads¶
Its strength is that it makes an assumption visible: instead of asserting that treated and untreated paths were parallel, it shows the pre-period and lets the eye judge, while also exposing whether an effect is a jump, a ramp, or a fade. The failure modes are twofold. First, the pre-trend test is often underpowered — a flat pre-period with wide confidence bands proves little, and a genuine slow drift can hide inside the noise. Second, when adoption is staggered and the plot is built from a naive two-way fixed-effects specification, already-treated units silently serve as controls for later-treated ones[1], producing the "forbidden comparisons" that the Goodman-Bacon decomposition exposes as a source of sign-flipping bias. The classic misuse is reading an eye-catching post-event bump without noticing a sloping pre-period beneath it. The guarding discipline is to plot confidence bands, use a staggered-robust estimator, and treat a noisy-but-flat pre-period as weak reassurance rather than proof.
How it implements the components¶
visual_trajectory_display— the plotted leads-and-lags path is the mechanism's actual artifact; the trajectory is the output.staggered_adoption_structure— it exploits units adopting at different calendar times, aligning them on a common event axis so the rollout becomes analyzable.within_unit_change_contrast— each unit's path is read against its own pre-event reference period, so the display is built from within-unit change.
It renders no raw calendar-time grid for monitoring: it does not implement unit_time_observation_matrix — that is unit_time_dashboard, its nearest twin, which stays in calendar time and never re-centers on a treatment. The separator is the event alignment above: this plot has a treatment axis, the dashboard has none.
Related¶
- Instantiates: Time Series Cross-Section Analysis — it supplies the archetype's event-aligned visual test of effect timing.
- Consumes: fixed_effects_panel_model — the event-time coefficients it plots are typically estimated from a within-unit regression with event-time dummies.
- Sibling mechanisms: fixed_effects_panel_model · difference_in_differences_design · unit_time_dashboard · lagged_panel_regression · balanced_panel_completeness_check · peer_trajectory_benchmarking
Editorial Notes¶
Form Classification¶
Form family: Analysis, Modeling & Optimization
Rationale: The mechanism re-indexes units to event time, estimates lead and lag effects against a reference period, and uses pre-period coefficients to test the design's identifying assumptions.
Nearest alternative: Interface, Display & Cue — The panel plot makes trajectories legible, but the operative contribution is the event-study estimation and pretrend inference rather than the display surface alone.
Review outcome: Adjudicated after independent review; high confidence.
Origin Attribution¶
Primary origin: Economics & Finance
Origin pattern: Convergent development
Present-day reach: Multi-domain
Rationale: Econometrics established relative-time event-study coefficient plots for panel treatments and policy adoption, including pre-event trend diagnosis and dynamic effects.
Related originating lineages:
- Political Science — Policy evaluation generalized the display to staggered institutional adoption.
- Statistics & Experimental Design — Longitudinal causal inference supplies identification assumptions and uncertainty intervals.
Review resolution: AEA and NBER sources treat event-study plots as econometric panel designs and show their relationship to pretrend testing and staggered adoption, supporting economics over a generic statistics origin.
Attribution caveat: Finance also uses event windows, but the named panel plot follows the causal-panel lineage.
Review outcome: Researched adjudication after independent review; high confidence.
Sources consulted:
- American Economic Review: Pre-Event Trends in the Panel Event-Study Design
- NBER: Difference-in-Differences with Variation in Treatment Timing
References¶
[1] Goodman-Bacon, Andrew. "Difference-in-Differences with Variation in Treatment Timing". Journal of Econometrics 225(2), 254–277 (2021). Decomposes staggered-adoption TWFE estimates into two-group comparisons and shows how already-treated controls can bias estimates, including reversing their sign when effects evolve over time. registry ↩