Spatial Identification (jue-identification)

When to trigger

• A causal claim rests on OLS + region controls, or TWFE on staggered place-based policy
• A shift-share/Bartik instrument's exogeneity (shares vs shocks) is asserted, not argued
• A boundary/border discontinuity lacks a continuity-of-confounders defense
• The estimate could be driven by spatial sorting/selection across locations rather than the treatment
• Control areas may be contaminated by spillovers/displacement (SUTVA failure)
• Inference ignores spatial autocorrelation (Conley/HAC) and overstates precision

The JUE identification bar

JUE referees are sophisticated about the failure modes that are specific to space. A clean national design is not enough; you must defend it against the three spatial confounds that recur in every urban paper: (1) sorting/selection — people, firms, and developers choose locations, so cross-location comparisons mix treatment with composition; (2) spillovers/SUTVA — treating one place moves activity to or from neighbors, contaminating controls and biasing reduced forms; (3) spatial autocorrelation — nearby units are correlated, so naive SEs are too small. State the estimand, name the identifying variation, show the diagnostic that could have failed, and address all three confounds explicitly.

Design paths

Path A: Boundary / spatial discontinuity (school zones, jurisdiction borders, corridors)

• Continuity defense: show pre-determined covariates are smooth across the boundary; the running variable is geographic distance.
• Local-linear with data-driven bandwidth; bias-corrected robust CIs; donut to drop units at the exact line; bandwidth sensitivity.
• Defend that the boundary is not also a discontinuity in something else (other jurisdiction services, zoning, natural features).
• The estimand is the local effect at the border — resist extrapolating to the whole city.

Path B: Shift-share / Bartik (local labor demand, immigration, trade exposure)

• State whether identification rests on exogenous shares (Goldsmith-Pinkham–Sorkin–Swift; report Rotemberg weights) or exogenous shocks (Borusyak–Hull–Jaravel).
• Show the implied just-identified instruments and which industries/shocks drive the estimate.
• Address that initial shares reflect prior sorting; defend pre-period balance and exclusion.

Path C: Historical / geographic IV (terrain, soil, historical infrastructure, lines on maps)

• Argue the instrument affects the outcome only through the spatial channel of interest — the exclusion restriction is where these papers live or die.
• Falsification on placebo outcomes and never-treated channels; control for the geography the instrument correlates with.

Path D: Place-based policy DiD / event study (zones, infrastructure, rezoning)

• With staggered rollout, move beyond TWFE (Callaway–Sant'Anna, Sun–Abraham, de Chaisemartin–D'Haultfœuille); clean event-study leads; Goodman-Bacon decomposition.
• Spillover-robust controls: ring/donut specifications around treated areas; estimate displacement to neighbors rather than assuming SUTVA.
• Rambachan–Roth honest-DID sensitivity to parallel-trend violations.

Path E: Mobility / neighborhood experiment (MTO-style, voucher lotteries)

• Lottery/randomization as the source of variation; ITT and LATE/TOT distinguished; non-compliance handled.
• Selection into take-up addressed; neighborhood exposure measured, not assumed.

Cross-cutting spatial inference

• Spatial autocorrelation: Conley spatial-HAC SEs (with a defended distance cutoff) or cluster at the spatial-market level; report how SEs change versus naive.
• Sorting/selection: show composition is balanced or model the sorting; never present a cross-location comparison as if locations were randomly assigned.
• SUTVA/spillovers: define the treated and control geography so that spillovers do not contaminate controls; estimate the spillover itself where possible.

Execution bridge (StatsPAI / Stata MCP)

Estimate and audit the design, don't only describe it. Full map: execution-with-mcp. JUE is urban/spatial economics — sorting and spatial dependence; identification + Conley/spatial-robust inference.

• detect_design → recommend → fit with as_handle=true → audit_result.
• Observational causal claims: staggered DiD (callaway_santanna / sun_abraham + bacon_decomposition + honest_did_from_result); IV (effective_f_test + anderson_rubin_ci); RDD (rdrobust + mccrary_test).
• Experiments: randomization-based inference + romano_wolf for many-outcome control.
• Sensitivity: oster_delta / sensemakr for observational claims.

Report the magnitude in interpretable units; route the full battery to the appendix. A run end-to-end (synthetic data, real returns) is in the JF execution walkthrough.

Checklist

• Estimand named (local-at-boundary / ATT / LATE) and matched to the design
• The spatial variation is stated in one sentence; the diagnostic that could have failed is shown
• Sorting/selection addressed (balance, composition, or explicit sorting model)
• Spillovers/SUTVA addressed (rings/donuts; displacement estimated, not assumed away)
• Spatial autocorrelation in inference (Conley/spatial cluster), reported vs naive SEs
• Shift-share: shares-vs-shocks identification stated; Rotemberg weights or BHJ reported
• Modern estimator where TWFE/2SLS would bias; the claim never exceeds the local estimand

Anti-patterns

• TWFE on staggered place-based policy with no heterogeneity-bias discussion
• A boundary RD with no covariate-smoothness test or with the boundary confounding other services
• "Plausibly exogenous" shares/geography asserted with no falsification or Rotemberg/BHJ diagnostic
• Treating control regions as clean when treatment plausibly displaced activity into them
• Naive (non-spatial) standard errors on geographically clustered data
• Reading a local boundary or LATE estimate as a city-wide or national effect

Referee pushback mapped to the identification fix

• "This is sorting, not the treatment." → Show pre-period composition is balanced across treated/control, or model the location choice; add a placebo on pre-trends.
• "Your controls are contaminated by displacement." → Add a spillover ring; estimate the displacement to neighbors rather than assuming SUTVA; show controls outside the ring give the same answer.
• "Your standard errors ignore spatial correlation." → Report Conley spatial-HAC SEs at a defended distance cutoff and contrast them with the naive SEs.
• "Staggered TWFE is biased here." → Re-estimate with Callaway–Sant'Anna or Sun–Abraham; show flat event-study leads and a Goodman-Bacon decomposition.
• "Your shift-share shares are not exogenous." → Report Rotemberg weights (which industries drive it) or move to a Borusyak–Hull–Jaravel shock-exogeneity argument.

Worked vignette (illustrative)

A new light-rail line is evaluated with a DiD comparing corridor tracts to the rest of the metro. A referee flags two spatial confounds: richer households sort into the corridor, and demand displaced from control tracts contaminates them. The JUE fix: a boundary-distance design with ring controls (0–800m treated, 800–1600m as a spillover ring, >1600m control), covariate-smoothness across the ring boundary, Conley SEs with a 5km cutoff, and a sorting check on pre-period demographics. The capitalization estimate settles at 4.5% (Conley s.e. 1.3) and the spillover ring shows measurable displacement — reported, not hidden.

Output format

【Design】boundary-RD / shift-share / historical-IV / place-based-DiD / mobility-experiment
【Spatial variation → estimand】one sentence
【Estimand】local-at-boundary / ATT / LATE
【Sorting/selection】how addressed
【Spillovers/SUTVA】rings/donut; displacement estimated?
【Spatial inference】Conley/spatial cluster + cutoff; vs naive SEs
【What it does NOT identify】[...]
【Next skill】jue-theory-model (if a model is needed) or jue-robustness