A GII score is defined by its specification set, not a single number. Before any headline, the dossier reports the primary specification, the plausible range, the contribution breakdown, and any rank or category instability. The figures below run the @gii/sensitivity harness live on the fictional example.
Plausible range across specifications#
| Specification | GII-M | GII-D | Headline |
|---|---|---|---|
| Primary — equal domains, equal families | 11.65 | -5.7 | withheld |
| Equal indicator weights | 11.89 | -28.09 | withheld |
| Equal family weights | 11.89 | -28.09 | withheld |
| Exposure-weighted | 14.92 | -50.96 | withheld |
| Stakeholder-weighted | 11.84 | -7.29 | withheld |
| Empirical redundancy (ρ=0.5) | 11.65 | -5.7 | withheld |
Leave-one-out influence#
Change in the primary GII-M when each indicator is removed.
| Indicator removed | GII-M without it | Δ |
|---|---|---|
| compliance-hours | 9.12 | +2.53 |
| market-response | 14.18 | -2.53 |
| info-access | 13.35 | -1.7 |
| admin-capacity | 10.43 | +1.22 |
Method assignment#
Each policy pattern maps to preferred candidate designs and required cautions. The assignment guides confidence but does not mechanically determine it; a nationwide rule with no clean comparison carries an explicit confidence ceiling.
| Policy pattern | Preferred designs | Cautions |
|---|---|---|
| Staggered adoption across jurisdictions | Group-time difference-in-differences, event study | heterogeneous effects, treatment timing, pretrends, spillovers |
| One treated jurisdiction or national intervention | Synthetic control, comparative time series | donor contamination, fit, concurrent shocks, inference |
| Eligibility or funding threshold | Regression discontinuity | manipulation, bandwidth, local estimand, compliance |
| Clearly timed population intervention | Interrupted time series | serial correlation, seasonality, concurrent events, anticipation |
| Short-window market information shock | Financial event study | leakage, event confounds, expected-policy component, welfare interpretation |
| Tax, transfer, or benefit rule | Microsimulation plus observed validation | behavioral response, take-up, incidence, baseline assumptions |
| Monetary-policy decision | High-frequency identification, local projections, structural time series | endogeneity, information effects, distributed lags |
| Judicial change with geographic or temporal variation | Difference-in-differences, event study, legal epidemiology | legal reach, enforcement, later treatment, spillovers |
| Nationwide rule without a clean comparison · ceiling | Structural/contribution model, multiple controls, negative controls, triangulation | strong assumptions, confidence ceiling may apply |
Run this on any record from the sandbox via POST /api/v1/sensitivity or gii sensitivity. See the methodology for the full uncertainty and sensitivity requirements.