技能 数据科学 经济地理学空间因果识别指南

经济地理学空间因果识别指南

v20260724
jegeo-identification
本指南为撰写经济地理学(JEG)领域学术论文提供全面的方法论指导。它专门针对空间数据分析中的核心挑战,包括空间自相关、空间溢出效应和结构性弹性识别。内容涵盖了空间双重差分法(DID)、网络均衡模型(NEG)和定性案例研究,旨在帮助作者确保因果推断的科学严谨性。
获取技能
334 次下载
概览

Identification Strategy (jegeo-identification)

When to trigger

  • A spatial regression rests on OLS + region fixed effects, or TWFE on staggered place-based policy
  • A quantitative-spatial / NEG model is estimated but it is unclear what in the spatial data identifies the key elasticities
  • Treatment in one region plausibly spills over to "control" regions (SUTVA across space is violated)
  • A qualitative/comparative-case paper makes a causal-sounding claim with no explicit logic of inference
  • You are unsure the strategy reads as credible to BOTH an economist and a geographer

The JEG identification bar

Because JEG bridges geographical economics and human geography, "identification" means different things by branch — but in all of them the spatial structure of the data is part of the identification problem, not a nuisance. Two threats are nearly universal at JEG and referees expect them confronted head-on: spatial autocorrelation in errors (inference) and spatial spillovers / general-equilibrium leakage across units (SUTVA). Pick the branch and make the data-to-claim mapping explicit.

Branch A: Spatial causal design (place-based policy, regional treatment)

  • Spatial DID / event study: with staggered adoption move beyond TWFE (Callaway–Sant'Anna, Sun–Abraham, de Chaisemartin–D'Haultfœuille); show clean event-study leads; report a Goodman-Bacon decomposition.
  • Spillovers / SUTVA across space: the control region is often the treated region's neighbor. Use donut/ring specifications, model spatial spillovers explicitly, or argue why leakage is bounded — do not assume independence across adjacent units.
  • Spatial RDD / border designs: geographic discontinuities (administrative borders) are powerful but demand a continuity argument across the border and attention to what else changes at it.
  • IV with a spatial instrument: Bartik/shift-share and geography-based instruments are common; defend exogeneity of the shares (Goldsmith-Pinkham et al.) or of the shocks, not just first-stage strength.
  • Inference: cluster at the spatial-treatment level AND address spatial correlation across clusters with Conley spatial-HAC standard errors; report how the cutoff distance was chosen.

Branch B: Quantitative-spatial / NEG model identification

  • Name what identifies each structural elasticity (trade elasticity, agglomeration elasticity, migration elasticity) — tie it to specific spatial variation or moments, not "the estimator converged."
  • Calibration vs. estimation: if elasticities are borrowed, say from where and show the counterfactual is not driven by an indefensible borrowed value; report sensitivity.
  • General-equilibrium counterfactuals: the headline welfare/relocation number depends on the model's spatial linkages — show which parameters and which spatial structure move it.

Branch C: Case-based / qualitative geographic inference

  • Make the logic of inference explicit: comparative cases, process tracing, or theory-building from a critical case — and state what would have falsified the claim.
  • Justify case selection on substantive spatial grounds; address generalizability rather than claiming it.

Shift-share / Bartik instruments in a spatial setting

Shift-share instruments are pervasive in economic geography (regional exposure to national shocks via local industry mix), and JEG referees scrutinize them closely. Two defenses, two literatures:

  • Exogenous shares (Goldsmith-Pinkham–Sorkin–Swift): identification rests on the pre-period industry shares being as-good-as-random; defend the shares' exogeneity and report the Rotemberg weights that show which industries drive the estimate.
  • Exogenous shocks (Borusyak–Hull–Jaravel): identification rests on many quasi-random national shocks; defend the shocks and the equivalent shock-level regression.

State which justification you rely on — "we use a Bartik instrument" without naming the identifying assumption is exactly the move a JEG referee flags.

Execution bridge (StatsPAI / Stata MCP)

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

  • detect_designrecommend → fit with as_handle=trueaudit_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

  • Branch chosen; the spatial-data-to-claim mapping stated in one sentence
  • Spatial autocorrelation addressed in inference (Conley SEs / appropriate clustering; cutoff justified)
  • Cross-unit spillovers / SUTVA across space confronted, not assumed away
  • Staggered designs use a modern estimator; pre-trends/leads shown
  • Structural: each key elasticity tied to identifying spatial variation; counterfactual sensitivity shown
  • Qualitative: explicit inference logic + falsification condition + case-selection justification
  • The claim never exceeds what the spatial design supports

Anti-patterns

  • Default heteroskedastic SEs (or clustering on one dimension) when errors are spatially correlated
  • Treating neighboring regions as clean controls while the treatment spills across the border
  • TWFE on staggered place-based policy with no heterogeneity-bias discussion
  • "The estimator converged" offered as structural identification of agglomeration/trade elasticities
  • A qualitative paper making a causal claim with no stated logic of inference or falsifier
  • Reporting significance with asterisks instead of standard errors and confidence intervals

Worked vignette (illustrative)

A special economic zone is rolled out across regions and the paper estimates its effect on firm entry with TWFE and region-clustered SEs. Two JEG referees object: the economist says the zones were placed where growth was already accelerating (selection) and neighboring regions absorbed displaced firms (spillover inflates the gap); the geographer says "region" is the wrong scale because clusters cross administrative lines. The fix routes through all three: a Callaway–Sant'Anna estimator with clean leads (selection on trends), a ring specification isolating displacement (spillover), Conley SEs at a justified distance (spatial correlation), and a re-aggregation to commuting zones (scale). Only then is the entry effect — say a 6% rise, illustrative — credible to both readers.

Referee pushback mapped to the identification fix

  • "Your control regions are the treated region's neighbors — spillover inflates the effect." → Add ring/donut specs or a spatial-lag model; report the bounded effect net of displacement.
  • "Standard errors ignore that adjacent units co-move." → Conley spatial-HAC SEs over a range of cutoffs; show residual Moran's I.
  • "The agglomeration elasticity is calibrated, not identified." → Name the spatial variation that pins it; show the counterfactual is not driven by a borrowed value.
  • "This is a region case study calling itself causal." → State the inference logic and the falsifier explicitly, or downgrade the causal language.
  • "The result is an artifact of the spatial unit." → Re-estimate at another scale (the MAUP test) — partly a robustness move, but raised at identification.

Why spatial inference is non-negotiable at JEG

Economic-geography data violate the independence assumption almost by construction: nearby places share shocks, labor markets, and institutions. A JEG referee from the economics side treats overstated inference as a fatal flaw, and one from the geography side treats "space as iid error" as conceptually naive. Confronting spatial autocorrelation and spillovers is therefore not a robustness afterthought here — it is part of whether the design identifies anything at all. Decide the spatial error structure and the spillover structure before you read the point estimate, so the inference is not reverse-engineered to keep significance.

Output format

【Branch】spatial-causal / quantitative-spatial-model / qualitative-case
【Spatial-data-to-claim mapping】one sentence
【Spatial autocorrelation】inference fix (Conley / clustering; cutoff)
【Spillovers / SUTVA across space】how confronted
【Identification evidence】leads+Bacon / elasticity-to-variation / inference logic
【What it does NOT identify】[...]
【Next skill】jegeo-theory-model
信息
Category 数据科学
Name jegeo-identification
版本 v20260724
大小 9.26KB
更新时间 2026-07-28
语言