jue-robustness

Spatial Robustness Suite (jue-robustness)

When to trigger

• The main spatial estimate is in hand and must be shown not to be an artifact of one specification
• A referee asks "is this robust to the spatial scale / the buffer / the geography you chose?"
• The result depends on a bandwidth, a ring radius, a fixed-effects geography, or a clustering choice
• You need to rule out that spatial sorting or MAUP (modifiable areal unit problem) drives the result
• The estimate could move under a spatial-spillover or boundary-definition change

The JUE robustness bar

JUE referees probe whether the spatial estimate is stable across the spatial choices the researcher made — the scale of the units, the boundaries, the buffer/ring radii, the fixed-effects geography — and whether inference accounts for spatial dependence. Robustness here is not a wall of regressions; it is a targeted set of checks, each tied to a spatial threat, reported so the reader sees the point estimate barely moves.

Robustness craft

1. Lock the primary spatial specification first — the scale, FE geography, and bandwidth you prefer — then perturb around it. Do not present five co-equal spatial specs.
2. One spatial threat → one check. Each robustness exhibit reads "here is the worry (MAUP / spillover / sorting / spatial SEs), here is the evidence it is not the story."
3. Show stability of the point estimate, not just that significance survives — across scales, radii, and FE geographies the coefficient should barely move.
4. Stress-test inference for spatial dependence. Report Conley SEs at several distance cutoffs; wrong (non-spatial) SEs are the most common JUE robustness failure.
5. Report honestly where it weakens. A check that shifts the estimate is information — bound the implication rather than hiding the specification.

Execution bridge (StatsPAI / Stata MCP)

Run the battery, don't just enumerate it. Full map: execution-with-mcp. JUE is urban/spatial economics — sorting and spatial dependence; identification + Conley/spatial-robust inference.

• Many outcomes / specifications: romano_wolf (step-down FWER) or benjamini_hochberg.
• OVB sensitivity: oster_delta / sensemakr.
• Inference: wild_cluster_bootstrap (few clusters), twoway_cluster / conley.
• Re-fit off one handle: audit_result(result_id) lists missing checks + the exact suggest_function for each.
• Exhibits: etable / did_summary_to_latex from the handle — no retyped numbers.

Decisive checks in the body, exhaustive battery in the appendix. JF execution walkthrough.

Checklist

• Primary spatial spec declared (scale, FE geography, bandwidth) before perturbations
• MAUP addressed: estimate stable across at least two spatial scales
• Boundary/ring choices varied; result insensitive to radius and donut width
• Sorting/selection check (pre-period balance / placebo pre-trends)
• Spillover ring estimated; controls outside the spillover zone give the same answer
• Conley/spatial-cluster SEs reported across distance cutoffs, vs naive
• Finer geographic FE / local trends show the point estimate barely moves
• A spatial placebo (fake boundary / pre-period / unaffected outcome) is shown to be null
• For a QSM, counterfactual sensitivity to the least-identified elasticity is reported
• Any check that moves the estimate is reported and bounded honestly

Anti-patterns

• A 20-column robustness table with no map from check to spatial threat (kitchen-sink robustness)
• Reporting one spatial scale only, leaving MAUP unaddressed
• Naive standard errors on geographically clustered data, then claiming robustness
• Hand-picking the ring radius or bandwidth that maximizes significance
• Reporting that significance survives while the point estimate wanders across scales
• Hiding the FE geography or boundary definition that breaks the result

Referee pushback mapped to the robustness fix

• "This is an artifact of your spatial scale." → Re-estimate at tract / block-group / commuting-zone; show the coefficient is scale-stable (MAUP not driving it).
• "Your boundary/ring radius is arbitrary." → Vary radii and donut widths; show insensitivity across the grid of cuts.
• "Did you cluster for spatial dependence?" → Conley SEs at several distance cutoffs (and a spatial-cluster alternative), contrasted with naive SEs.
• "This is specification search." → Declare the primary spatial spec; show a specification curve over scale × FE × bandwidth in which the point estimate barely moves.

Robustness for a structural / QSM paper

When the JUE paper is a quantitative spatial model, robustness shifts from specification perturbations to parameter sensitivity and counterfactual stability. Report how the headline counterfactual moves as the least-identified elasticities (migration, commuting, agglomeration) are varied across plausible ranges from the literature; show that the qualitative conclusion and the order of magnitude survive. A counterfactual that is fragile to one elasticity is a finding to bound and disclose, not to bury — the same honesty norm as reduced-form stability.

Worked vignette (illustrative)

A density-wage elasticity is 0.045 (Conley s.e. 0.012). The spatial robustness suite: (i) re-estimated at tract, block-group, and commuting-zone scale, the elasticity stays in [0.041, 0.049] — MAUP is not driving it; (ii) Conley SEs at 50/100/200 km cutoffs keep the CI away from zero; (iii) adding commuting-zone fixed effects moves it to 0.043; (iv) a placebo on pre-period wage growth is flat, arguing against sorting on trends; (v) the spillover specification shows neighboring-area contamination is small. The point estimate barely moves — the JUE target.

Spatial placebo and falsification

Beyond perturbing the main spec, the most persuasive JUE robustness evidence is a placebo that should show nothing and does. Useful spatial placebos: assign the treatment to a pre-period and show no effect (rules out pre-trends/sorting on trends); apply the design to an outcome that the mechanism should not move (rules out a generic local shock); shift the boundary or corridor to a fake location and show the discontinuity vanishes. A clean placebo is often worth more to a referee than another robustness column, because it tests the design rather than re-running it — and a placebo that unexpectedly does fire is critical information to report, not suppress.

Output format

【Primary spatial spec】scale / FE geography / bandwidth — estimate: ___ (Conley s.e. ___)
【MAUP】scales tested: ___ ; range: [___, ___]
【Boundary/ring】radii/donut varied? result stable? [Y/N]
【Sorting check】pre-period balance / placebo pre-trend: ___
【Spillover】ring estimated; controls-outside-zone consistent? [Y/N]
【Spatial inference】Conley cutoffs: ___ ; vs naive: ___
【Estimate stability】range across checks: [___, ___]; checks that move it: ___
【Next skill】jue-tables-figures