updated analaysis with offshore considerations

analysis/draft.md

@@ -0,0 +1,236 @@
+# Heterogeneous Enforcement of Transparency: Evidence from the Texas Railroad Commission
+
+## Introduction
+
+How does transparency alter regulatory enforcement in high-capacity but locally discretionary bureaucracies? We study the January 2019 Texas Railroad Commission (RRC) disclosure change that made well-level violation information publicly searchable. The policy constitutes a statewide transparency shock, but implementation and enforcement remain district-administered. This setting allows us to test both system-wide effects and district-level heterogeneity in policy response.
+
+Our core empirical finding is a two-part pattern. First, we find evidence of **gradual post-policy acceleration** in enforcement timing at the statewide level (significant post-policy trend improvement) rather than a sharp immediate level break in 2019. Second, district-level responses are strongly heterogeneous, and offshore-jurisdiction districts (02/03/04) exhibit systematically different post-policy dynamics once district-specific post effects are modeled.
+
+## Theory and Hypotheses
+
+Transparency may alter enforcement through reputational, political, and managerial channels. Public disclosure can increase the salience of noncompliance and create incentives for agencies to accelerate case movement through the regulatory pipeline. But local implementation discretion can mediate this effect, producing district-level divergence.
+
+We test:
+
+- **H1 (Regulatory Pipeline Acceleration)**
+  - **H1a:** Disclosure reduces time from violation discovery to enforcement action.
+  - **H1b:** Disclosure improves compliance verification (resolution on re-inspection).
+- **H2 (Bureaucratic Heterogeneity):** Post-policy effects vary across districts.
+- **H3 (Structural Moderators):** Capacity, baseline performance, EJ context, geology, border proximity, and rurality explain variation.
+- **H4 (Spatial Dynamics):** District treatment effects are spatially autocorrelated.
+- **H5 (Offshore Jurisdiction Moderator):** Districts 02/03/04 exhibit differential post-2019 response.
+
+## Data and Measures
+
+We construct a district-year panel (2015-2025, 13 RRC districts) from administrative inspection and violation records. Well-level integration uses `api_norm` as the normalized identifier across sources.
+
+Primary outcomes:
+
+- `log_days_to_enf`: log mean days from violation discovery to enforcement action.
+- `resolution_rate`: percent compliant on re-inspection.
+- `compliance_rate`: percent compliant at inspection.
+- `violations_per_inspection`.
+
+## Empirical Strategy
+
+We estimate policy effects in three layers.
+
+### Model 1: All-district policy-year shift (H1)
+
+\[
+Y_{dt}=\alpha_d + \beta_1 \text{YearNum}_t + \beta_2 \text{Post2019}_t + \beta_3 \text{PostTrend}_t + \varepsilon_{dt}
+\]
+
+Where \(\text{PostTrend}_t = \max(0, t-2018)\). This distinguishes an immediate post-2019 level shift (\(\beta_2\)) from post-policy slope change (\(\beta_3\)).
+
+### Model 2: District heterogeneity (H2)
+
+\[
+Y_{dt}=\alpha_d + \gamma_t + \sum_d \theta_d (\text{District}_d\times \text{Post2019}_t) + \varepsilon_{dt}
+\]
+
+This yields district-specific post-policy effects and a joint heterogeneity test.
+
+### Model 3: Offshore moderation (H5)
+
+\[
+Y_{dt}=\alpha_d + \gamma_t + \sum_d \theta_d (\text{District}_d\times \text{Post2019}_t) + \phi(\text{Post2019}_t\times \text{Offshore}_d) + \varepsilon_{dt}
+\]
+
+Where `Offshore_d = 1` for districts 02/03/04.
+
+All models use district-clustered standard errors.
+
+## Results
+
+### Descriptive pipeline trends
+
+Pre/post means indicate lower average enforcement delay post-policy (174.3 to 112.3 days), but reduced inspection frequency intensity (higher days between inspections).
+
+Figure 1 visualizes these system-level changes across the regulatory pipeline. The key descriptive pattern is that timeliness improves over the post-policy period even as inspection cadence shifts, motivating a design that separates immediate policy breaks from post-policy trend effects.
+
+![Regulatory Pipeline Trends](pipeline_trends_over_time.png)
+**Figure 1.** Regulatory pipeline trends, 2015-2025.
+
+### H1: Policy-year effects (all districts)
+
+**Model 1 (timing outcome):**
+
+- `post_2019` level shift: **0.1514**, p=0.3294.
+- `post_trend` slope shift: **-0.3603**, p=0.0010.
+
+Interpretation: no statistically significant immediate level break in 2019, but a significant post-policy acceleration trend in enforcement timing.
+
+**Table 1. Core policy-year and moderator estimates**
+
+| Model | Parameter | Coefficient | P-value | Interpretation |
+| :--- | :--- | ---: | ---: | :--- |
+| Model 1 (All districts, interrupted panel) | `post_2019` | 0.1514 | 0.3294 | No immediate level break |
+| Model 1 (All districts, interrupted panel) | `post_trend` | -0.3603 | 0.0010 | Significant post-policy acceleration trend |
+| Model 3 (District heterogeneity + offshore) | `post_2019:offshore_jurisdiction` | 0.3819 | <0.001 | Offshore districts relatively slower post-policy timing |
+
+Table 1 provides the baseline inferential results for the article’s identification strategy. The table shows that the main all-district effect appears in the post-policy slope term rather than a one-time post-2019 level break, and it also shows that offshore jurisdiction remains a statistically important differential once district heterogeneity is modeled.
+
+Event-study decomposition (relative to 2018) corroborates this dynamic pattern:
+
+- No significant pre-policy years (2015-2017).
+- Significant negative deviations in 2022, 2024, and 2025.
+
+**Table 2. Event-study coefficients (all districts, reference year = 2018)**
+
+| Year | Coefficient | P-value | Significant (p<0.05) |
+| :--- | ---: | ---: | :--- |
+| 2022 | -0.5853 | 0.0333 | Yes |
+| 2024 | -0.7829 | 0.0057 | Yes |
+| 2025 | -1.4800 | <0.001 | Yes |
+
+Table 2 highlights the years where post-policy deviations are most pronounced. Substantively, these estimates indicate that the policy response intensifies over time instead of materializing immediately in 2019.
+
+![Event Study](event_study_plot.png)
+**Figure 2.** All-district event-study decomposition and offshore differential annual effects.
+
+Figure 2 complements Table 2 by displaying the full time path (including pre-policy years), making the absence of pre-trend significance and the later post-policy acceleration visually transparent.
+
+### H2: District heterogeneity
+
+District-level post-policy responses are strongly heterogeneous and jointly significant. Estimated effects range from substantial acceleration (e.g., District 09) to substantial slowdown (e.g., Districts 03 and 04).
+
+From district effect summaries used in mapping:
+
+- Best improvement: District 09 (about -52.6%).
+- Largest deterioration: District 04 (about +138.5%).
+
+Figure 3 presents the estimated district-specific effects directly, while Figure 4 maps those effects geographically. Together they demonstrate that heterogeneity is not a minor perturbation around a common effect but a core empirical feature of the policy response.
+
+![District Effects](district_treatment_effects_psj.png)
+**Figure 3.** District-specific post-2019 treatment effects.
+
+To show where these effects are concentrated geographically, Figure 4 maps district-level percent changes in enforcement timing.
+
+![District Treatment Effects Map](district_treatment_effects_map_psj.png)
+**Figure 4.** Geographic distribution of district treatment effects (percent change in days to enforcement).
+
+The map indicates that large positive and negative effects coexist across regions, reinforcing the need to model district-level discretion explicitly rather than assuming uniform policy implementation.
+
+### H5: Offshore moderation
+
+In the conditional heterogeneity model (Model 3):
+
+- `post_2019:offshore_jurisdiction = 0.3819`, p<0.001.
+
+This indicates that, net of district-specific post effects, offshore-jurisdiction districts experience relatively slower post-policy enforcement timing.
+
+### H3: Structural moderators
+
+Main moderator block:
+
+- H3a Capacity: coef -0.0188, p=0.9415.
+- H3b Baseline performance: coef -0.0884, p=0.7144.
+- H3e Border proximity: coef -0.2768, p=0.3082.
+
+Deep-dive TWFE block:
+
+- H3c EJ context: coef 0.1818, p=0.4866.
+- H3f Rurality: coef 0.2213, p=0.4649.
+- H3e Border proximity: coef -0.3626, p=0.1669.
+- H3d Geology: mixed basin interactions; some terms significant (p<0.001).
+
+Overall, H3 receives limited support except partial geology effects.
+
+**Table 3. Structural moderator tests**
+
+| Hypothesis | Term | Coefficient | P-value | Result |
+| :--- | :--- | ---: | ---: | :--- |
+| H3a Capacity | `post_2019:high_capacity` | -0.0188 | 0.9415 | Not supported |
+| H3b Baseline performance | `post_2019:low_baseline_compliance` | -0.0884 | 0.7144 | Not supported |
+| H3c EJ context | `post_2019:high_eji` | 0.1818 | 0.4866 | Not supported |
+| H3e Border proximity | `post_2019:border_competition` | -0.3626 | 0.1669 | Not supported |
+| H3f Rurality | `post_2019:high_rural` | 0.2213 | 0.4649 | Not supported |
+| H3d Geology | `C(primary_basin):post_2019` | Mixed | Mixed | Partial support |
+
+Table 3 summarizes why structural accounts are only partially successful in this run: most moderators are imprecisely estimated, while geology shows selective basin-specific effects. Figure 5 and Figure 6 then provide visual context for these moderator patterns.
+
+![Moderators](heterogeneous_effects.png)
+**Figure 5.** Moderator interaction estimates.
+
+![Demographics and Geography](district_demographics_geography.png)
+**Figure 6.** Demographic/geographic correlates of district effects.
+
+### H4: Spatial dynamics
+
+Moran’s I on district effects:
+
+- \(I = -0.0493\), permutation p=0.8550.
+
+No evidence of statistically significant global spatial autocorrelation.
+
+Figure 7 visually corroborates the spatial test by showing no systematic clustering pattern consistent with strong spillovers.
+
+![Spatial Spillovers](spatial_spillovers.png)
+**Figure 7.** Spatial spillover diagnostics.
+
+## Robustness
+
+### Placebo policy years (all-district interrupted model)
+
+- 2017 placebo: coef 0.6565, p=0.0020.
+- 2021 placebo: coef -0.0245, p=0.9191.
+
+### Alternative outcomes (all-district interrupted model)
+
+- Resolution rate: post 4.3721 (p=0.2104), post-trend -2.9371 (p=0.1424).
+- Compliance rate: post -0.1311 (p=0.9316), post-trend -0.5562 (p=0.1870).
+- Violations/inspection: post -0.0082 (p=0.6690), post-trend 0.0106 (p=0.0600).
+
+### Sample restrictions
+
+- Full sample: post 0.1514 (p=0.3294), post-trend -0.3603 (p=0.0010).
+- Excluding extreme districts: post-trend remains negative/significant.
+- Excluding 2015-2016: post-trend remains negative (weaker significance).
+- Excluding 2020-2021: post-trend remains negative/significant.
+
+### Specification sensitivity
+
+- Linear interrupted model: post-trend -67.0420 days (p=0.0100).
+- Winsorized interrupted model: post-trend -0.3147 (p=0.0016).
+
+Across variants, the post-policy **slope** result is more stable than the immediate **level** effect.
+
+**Table 4. Robustness summary (interrupted panel framework)**
+
+| Check | `post_2019` (p) | `post_trend` (p) | Read |
+| :--- | :--- | :--- | :--- |
+| Full sample | 0.1514 (0.3294) | -0.3603 (0.0010) | Slope effect robust; level break weak |
+| Exclude extreme districts | 0.1917 (0.1930) | -0.2972 (0.0133) | Slope remains significant |
+| Exclude 2015-2016 | 0.1942 (0.1958) | -0.2313 (0.0950) | Slope negative, marginal |
+| Exclude 2020-2021 | 0.1516 (0.2959) | -0.3599 (0.0016) | Slope remains significant |
+| Linear interrupted | -41.9298 (0.3104) | -67.0420 (0.0100) | Same directional pattern |
+| Winsorized interrupted | 0.2137 (0.1021) | -0.3147 (0.0016) | Slope remains significant |
+
+Table 4 consolidates robustness evidence in one place: level-shift estimates are sensitive, but the negative post-policy slope remains comparatively stable across sample restrictions and alternative functional forms.
+
+## Discussion
+
+The transparency reform is associated with a gradual statewide acceleration in enforcement timing rather than a single immediate break at implementation. At the same time, district responses diverge sharply, confirming bureaucratic heterogeneity. Offshore jurisdiction explains a meaningful share of that heterogeneity once district-specific post effects are included, while most other structural moderators are weak or inconsistent in this run. Spatial diffusion across neighboring districts is not supported by global autocorrelation tests.
+
+These findings suggest that transparency reforms in decentralized regulatory systems should be evaluated as dynamic, district-conditioned processes, not monolithic statewide shocks.

analysis/draft_appendix.md

@@ -0,0 +1,178 @@
+# Appendix: Heterogeneous Enforcement of Transparency
+## Evidence from the Texas Railroad Commission
+
+## Appendix A. Data Construction and Variables
+
+### A1. Data integration
+
+The analysis combines inspection and violation administrative records (2015-2025) into a district-year panel. Well-level linkage is done via `api_norm`.
+
+### A2. Core variables
+
+| Variable | Definition |
+| :--- | :--- |
+| `log_days_to_enf` | Log of district-year mean days from violation discovery to enforcement action |
+| `resolution_rate` | Share of violations compliant on re-inspection |
+| `compliance_rate` | Share of inspections marked compliant |
+| `violations_per_inspection` | Total violations divided by inspections |
+| `post_2019` | Indicator for years >= 2019 |
+| `post_trend` | Piecewise linear trend after policy (`max(year-2018,0)`) |
+| `offshore_jurisdiction` | Indicator for districts 02/03/04 |
+| `high_capacity` | District above median pre-policy inspection volume |
+| `low_baseline_compliance` | District below median pre-policy compliance |
+| `high_eji` | District above median EJ score |
+| `high_rural` | District above median RUCA |
+| `border_competition` | Operationalized border-proximity indicator |
+| `primary_basin` | Dominant basin category |
+
+## Appendix B. Econometric Specifications
+
+### B1. Interrupted panel (all districts; H1)
+
+\[
+Y_{dt}=\alpha_d + \beta_1 \text{YearNum}_t + \beta_2 \text{Post2019}_t + \beta_3 \text{PostTrend}_t + \varepsilon_{dt}
+\]
+
+### B2. District heterogeneity (H2)
+
+\[
+Y_{dt}=\alpha_d + \gamma_t + \sum_d \theta_d (\text{District}_d\times \text{Post2019}_t) + \varepsilon_{dt}
+\]
+
+### B3. Offshore moderation (H5)
+
+\[
+Y_{dt}=\alpha_d + \gamma_t + \sum_d \theta_d (\text{District}_d\times \text{Post2019}_t) + \phi(\text{Post2019}_t\times \text{Offshore}_d) + \varepsilon_{dt}
+\]
+
+All models report district-clustered standard errors.
+
+## Appendix C. Main Run Outputs
+
+### C1. H1 (all-district timing outcome)
+
+| Parameter | Coefficient | P-value |
+| :--- | ---: | ---: |
+| `post_2019` | 0.1514 | 0.3294 |
+| `post_trend` | -0.3603 | 0.0010 |
+
+Interpretation: no significant immediate level shift; significant post-policy acceleration slope.
+Substantively, this table supports the main-text conclusion that the policy effect is best characterized as gradual acceleration through the enforcement pipeline rather than a single break at policy adoption.
+
+### C2. Event-study decomposition (all districts; ref=2018)
+
+| Year | Coefficient | P-value |
+| :--- | ---: | ---: |
+| 2015 | -0.4592 | 0.0658 |
+| 2016 | -0.3359 | 0.1615 |
+| 2017 | -0.0385 | 0.7502 |
+| 2019 | -0.1149 | 0.2843 |
+| 2020 | -0.1666 | 0.3878 |
+| 2021 | -0.4192 | 0.1072 |
+| 2022 | -0.5853 | 0.0333 |
+| 2023 | -0.4899 | 0.1160 |
+| 2024 | -0.7829 | 0.0057 |
+| 2025 | -1.4800 | <0.001 |
+
+Pre-policy years are jointly non-significant in this decomposition.
+The coefficient pattern reinforces parallel-pretrend credibility while showing that the post-policy effect strengthens in later years, consistent with delayed organizational adaptation.
+
+### C3. Offshore differential annual effects (ref=2018)
+
+| Year | Offshore differential coef | P-value |
+| :--- | ---: | ---: |
+| 2019 | 0.3479 | 0.1581 |
+| 2020 | 0.1796 | 0.7089 |
+| 2021 | 0.9121 | 0.1095 |
+| 2022 | 0.7532 | 0.0652 |
+| 2023 | 0.9166 | 0.0325 |
+| 2024 | 1.0693 | 0.0280 |
+| 2025 | 0.7233 | 0.2091 |
+
+These estimates indicate that offshore jurisdictions diverge from non-offshore districts in specific post years rather than uniformly across the entire post period. The strongest differentials appear in 2023-2024.
+
+### C4. H5 offshore moderator (conditional model)
+
+| Parameter | Coefficient | P-value |
+| :--- | ---: | ---: |
+| `post_2019:offshore_jurisdiction` | 0.3819 | <0.001 |
+
+See **Figure 4** in the main text (`district_treatment_effects_map_psj.png`) for the geographic distribution of district treatment effects.
+Read alongside C3, this pooled interaction should be interpreted as an average offshore differential in the post period after district heterogeneity is already modeled, not as a claim that offshore status is the dominant driver of all district variation.
+
+### C5. H3 moderator tests
+
+Main block:
+- H3a Capacity: -0.0188 (p=0.9415)
+- H3b Baseline performance: -0.0884 (p=0.7144)
+- H3e Border proximity: -0.2768 (p=0.3082)
+- H5 (same block estimate): 0.6317 (p=0.1055)
+
+Deep-dive block:
+- H3c EJ: 0.1818 (p=0.4866)
+- H3f Rurality: 0.2213 (p=0.4649)
+- H3e Border proximity: -0.3626 (p=0.1669)
+- H3d Geology: mixed basin interactions, with significant terms including:
+  - `C(primary_basin)[0]:post_2019 = 0.5322` (p<0.001)
+  - `C(primary_basin)[3]:post_2019 = -0.5707` (p<0.001)
+
+Taken together, these moderator results imply that broad structural covariates provide limited explanatory leverage in this run, while basin composition remains the clearest structural correlate of differential policy response.
+
+## Appendix D. Spatial Test (H4)
+
+Moran's I on district treatment effects:
+
+- Moran’s I = -0.0493
+- Permutation p-value = 0.8550
+
+Conclusion: no significant global spatial autocorrelation.
+The sign and magnitude of Moran’s I are both small, indicating no evidence that high- or low-response districts are systematically clustered in ways consistent with regional diffusion.
+
+## Appendix E. Robustness Tables
+
+### E1. Placebo policy years (all-district interrupted model)
+
+| Placebo year | Coefficient (`post`) | P-value |
+| :--- | ---: | ---: |
+| 2017 | 0.6565 | 0.0020 |
+| 2021 | -0.0245 | 0.9191 |
+
+The significant 2017 placebo estimate suggests that single-cut timing designs can produce spurious break effects, which is why the main analysis emphasizes trend-change evidence and event-study diagnostics instead of level shifts alone.
+
+### E2. Alternative outcomes (all-district interrupted model)
+
+| Outcome | `post` coef (p) | `post_trend` coef (p) |
+| :--- | :--- | :--- |
+| Resolution rate | 4.3721 (0.2104) | -2.9371 (0.1424) |
+| Compliance rate | -0.1311 (0.9316) | -0.5562 (0.1870) |
+| Violations per inspection | -0.0082 (0.6690) | 0.0106 (0.0600) |
+
+This table shows that timing acceleration does not mechanically translate into improvements across all compliance-oriented outcomes in the same period, highlighting outcome-specific channels of policy response.
+
+### E3. Sample restrictions (all-district interrupted model)
+
+| Restriction | `post_2019` coef (p) | `post_trend` coef (p) |
+| :--- | :--- | :--- |
+| Full sample | 0.1514 (0.3294) | -0.3603 (0.0010) |
+| Exclude extreme districts | 0.1917 (0.1930) | -0.2972 (0.0133) |
+| Exclude 2015-2016 | 0.1942 (0.1958) | -0.2313 (0.0950) |
+| Exclude 2020-2021 | 0.1516 (0.2959) | -0.3599 (0.0016) |
+
+Across restrictions, the post-trend estimate remains negative and generally significant, while the post level term stays weak. This stability is central to the article’s interpretation of gradual policy-induced acceleration.
+
+### E4. Specification sensitivity
+
+| Specification | `post` effect | `post_trend` effect |
+| :--- | :--- | :--- |
+| Linear interrupted | -41.9298 (p=0.3104) | -67.0420 (p=0.0100) |
+| Winsorized interrupted | 0.2137 (p=0.1021) | -0.3147 (p=0.0016) |
+| Year FE + district post terms | 13 interaction terms | N/A |
+
+Specification checks again point to the same empirical hierarchy: slope effects are more robust than level effects, and district-specific post terms remain necessary to represent the observed heterogeneity.
+
+## Appendix F. Interpretation Notes
+
+1. The strongest system-wide evidence in this run is a **post-policy slope change**, not a one-time 2019 level shift.
+2. District heterogeneity is substantial and statistically material.
+3. Offshore jurisdiction contributes meaningfully in conditional models, but placebo behavior indicates caution in purely timing-based causal claims.
+4. Spatial diffusion is not supported by global autocorrelation tests.

analysis/gemini_draft.md

@@ -1,122 +0,0 @@
-# Heterogeneous Enforcement of Transparency: Evidence from the Texas Railroad Commission
-
-## Methods
-
-### Data and Sample
-
-To evaluate the impact of the January 2019 disclosure policy (Rule 8 compliance transparency) on regulatory enforcement, we constructed a novel panel dataset integrating administrative records from the Texas Railroad Commission (RRC) with demographic and geographic controls. The observation period spans January 2015 to December 2025, providing 4 years of pre-policy and 7 years of post-policy data.
-
-**Administrative Data:** Our primary dataset is drawn from the Texas Railroad Commission (RRC) Resource Center (<https://www.rrc.texas.gov/resource-center/>). We utilized:
-
-1. **Inspection Records (N = 2,151,839):** Site-level inspection logs detailing the date, district office, operator, and compliance determination.
-2. **Violation Records (N = 242,899):** Detailed violation events, including the date of discovery, rule violated, severity classification, and dates of subsequent enforcement actions (e.g., Notice of Violation, Severance, Sealing).
-
-**Demographic and Geographic Controls:** To test for environmental justice implications and structural moderators, we merged well-level locations with:
-
-1. **American Community Survey (ACS 2021):** Well-level demographics including Environmental Justice Index (EJI) scores, poverty rates, and median household income at the census tract level. The **EJI score** is constructed as a composite social vulnerability index, calculated as the mean of percentile-ranked indicators including minority population share, poverty rate, unemployment rate, linguistic isolation, and educational attainment. Higher scores indicate greater cumulative social vulnerability.
-2. **Rural-Urban Commuting Area (RUCA):** 2020 codes to classify well locations as metropolitan, micropolitan, or rural.
-3. **Geologic Basins:** Spatial joins identifying the major shale play (e.g., Permian, Eagle Ford) associated with each well.
-
-Data were aggregated to the district-month and district-year levels to align with the administrative structure of the RRC, which operates through 13 geographically distinct district offices.
-
-### Hypotheses
-
-Drawing on theories of bureaucratic reputation and transparency-as-regulation, we test four primary hypotheses:
-
-* **H1 (Regulatory Pipeline Acceleration):** We conceptualize the enforcement process as a "regulatory pipeline" consisting of four distinct stages: (1) inspection, (2) violation discovery, (3) enforcement action, and (4) compliance verification. We hypothesize that the 2019 disclosure policy will accelerate the movement of violations through this pipeline, specifically by reducing the administrative delay between violation discovery and enforcement action ($H1a$) and increasing the rate of compliance verification ($H1b$).
-* **H2 (Bureaucratic Heterogeneity):** The impact of the policy will vary significantly across administrative districts, reflecting local discretion rather than uniform implementation.
-* **H3 (Structural Moderators):** Variation in policy response will be explained by district structural characteristics, specifically:
-  * *Capacity ($H3a$):* High-capacity districts will be more responsive.
-  * *Baseline Performance ($H3b$):* Low-compliance districts will improve most (catch-up effect).
-  * *Environmental Justice ($H3c$):* Districts with higher Environmental Justice Index (EJI) scores will experience different policy impacts.
-  * *Geology ($H3d$):* Enforcement patterns will vary by the dominant oil and gas basin.
-  * *Border Proximity ($H3e$):* Districts bordering other states or Mexico will exhibit different policy responsiveness due to inter-jurisdictional competition.
-  * *Rurality ($H3f$):* Rural districts (higher RUCA codes) will respond differently to the disclosure policy than metropolitan districts.
-* **H4 (Spatial Dynamics):** We expect positive spatial autocorrelation in district-level treatment effects, indicating that neighboring districts will exhibit similar responsiveness due to regional diffusion of best practices.
-
-### Econometric Strategy
-
-We employ a Difference-in-Differences (DiD) framework with two-way fixed effects.
-
-**Equation 1: Baseline Dynamic Effect (Event Study)**
-$$Y_{dt} = \alpha + \sum_{k \neq 2018} \beta_k \mathbb{1}(Year_t = k) + \gamma_d + \epsilon_{dt}$$
-Where $\gamma_d$ represents district fixed effects. This specification tests the parallel trends assumption and maps temporal evolution.
-
-**Equation 2: Heterogeneous Treatment Effects**
-$$Y_{dt} = \alpha + \lambda (District_d \times Post2019_t) + \delta_t + \gamma_d + \epsilon_{dt}$$
-This estimates a unique treatment effect $\lambda$ for each of the 13 district offices.
-
-**Equation 3: Moderator Analysis (Triple Difference)**
-$$Y_{dt} = \alpha + \beta_1 Post_t + \beta_2 (Post_t \times Moderator_d) + \delta_t + \gamma_d + \epsilon_{dt}$$
-This interacts the policy shock with structural moderators (Capacity, Compliance, Demographics) to explain heterogeneity.
-
-## Analysis and Results
-
-### Descriptive Trends
-
-Figure 1 illustrates the aggregate trends in the regulatory pipeline. Following the 2019 policy, we observe a structural break in enforcement behavior. While inspection frequency fluctuated, the average days to enforcement (top right panel) shows a marked post-2019 decline.
-
-![Regulatory Pipeline Trends](pipeline_trends_over_time.png)
-*Figure 1: Trends in inspection, compliance, and enforcement metrics (2015-2025).*
-
-### H1: Aggregate Policy Impact
-
-Table 1 presents the pooled Difference-in-Differences results testing **H1**. The policy is associated with a statistically significant reduction in enforcement delays, supporting **H1a**. Specifically, the log-linear specification indicates a **30.8% reduction** in time to enforcement action ($p < 0.05$). Robustness checks confirm this with a linear reduction of ~62 days. **H1b** is also supported, with a 3.7 percentage point increase in compliance rates.
-
-**Table 1: Baseline Difference-in-Differences Results**
-
-| Outcome | Coefficient | Std. Error | P-Value | Result |
-| :--- | :--- | :--- | :--- | :--- |
-| **Log(Days to Enforcement)** | **-0.369*** | 0.157 | 0.019 | **Supported (H1a)** |
-| Compliance Rate (%) | +3.740* | - | 0.002 | **Supported (H1b)** |
-| Violations per Insp. | -0.052* | - | 0.001 | Supported |
-
-*Note: Standard errors clustered at the district level. $* p < 0.05$.*
-
-The event study (Figure 2) validates the design. Coefficients for pre-policy years are null, while significant negative effects (faster enforcement) emerge consistently starting in 2022.
-
-![Event Study Plot](event_study_plot.png)
-*Figure 2: Event study coefficients relative to the 2018 baseline.*
-
-### H2: Bureaucratic Heterogeneity
-
-**H2 is strongly supported.** The average effect masks profound variation across the 13 districts. As shown in Figure 3, the policy impact ranges from a **65.9% improvement** in District 09 to a **71.9% decline** in District 04. Ten districts improved, while three exhibited backsliding.
-
-![District Treatment Effects](district_treatment_effects_psj.png)
-*Figure 3: Heterogeneous treatment effects by district (Percent change in days to enforcement).*
-
-### H3: Structural Moderators
-
-We tested **H3** using Triple-Difference models to explain this divergence. Table 2 presents the results for the four structural hypotheses. **We find strong support for H3c (Environmental Justice).** Districts with higher Environmental Justice Index (EJI) scores—indicating greater social vulnerability—saw significantly *slower* improvements in enforcement speed (Coefficient: +0.412, p<0.05). This suggests that the transparency benefits of the policy were unequally distributed, potentially exacerbating existing inequities.
-
-However, other structural factors failed to explain the heterogeneity. Capacity ($H3a$), baseline compliance ($H3b$), the underlying oil and gas basin ($H3d$), border proximity ($H3e$), and rurality ($H3f$) were not statistically significant predictors of policy response.
-
-**Table 2: Triple-Difference Moderator Analysis**
-
-| Moderator Hypothesis | Interaction Coef. | P-Value | Result |
-| :--- | :--- | :--- | :--- |
-| **H3a: High Capacity** | -0.285 | 0.363 | Not Supported |
-| **H3b: Low Baseline Compliance** | -0.134 | 0.628 | Not Supported |
-| **H3c: EnviroJustice Score (EJI)** | **+0.412** | **0.038** | **Supported (Inequity)** |
-| **H3d: Oil & Gas Basin** | -0.220 | 0.154 | Not Supported |
-| **H3e: Border Proximity** | -0.441 | 0.118 | Not Supported |
-| **H3f: Rurality (RUCA)** | +0.093 | 0.770 | Not Supported |
-
-Figure 4 visualizes these results. Figure 5 confirms a strong correlation between treatment effects and the district EJI scores.
-
-![Heterogeneous Effects](heterogeneous_effects.png)
-*Figure 4: Interaction effects for key district moderators.*
-
-![Demographics and Geography](district_demographics_geography.png)
-*Figure 5: Correlation of treatment effects with district demographic and geographic features (highlighting EJ Score).*
-
-### H4: Spatial Dynamics
-
-**H4 is not supported in the expected direction.** Instead of positive spillovers (clustering of high performance), spatial analysis reveals **significant negative spatial autocorrelation** ($I = -0.549$). Figure 6 maps this pattern, showing that high-performing districts are frequently adjacent to low-performing ones. This suggests distinct, localized administrative cultures rather than regional diffusion of best practices.
-
-![Spatial Map of Effects](district_treatment_effects_map_psj.png)
-*Figure 6: Spatial distribution of enforcement speed changes.*
-
-### Discussion
-
-These findings present a paradox for "transparency as regulation." While the policy succeeded in aggregate (**H1**), its implementation was heavily filtered through local discretion (**H2**). Although most structural variables failed to explain this variation, the significant finding regarding Environmental Justice (**H3c**) offers a critical caveat. The fact that high-vulnerability districts saw slower improvements suggests that transparency mechanisms may rely on community capacity to be effective, potentially widening the regulatory gap for disadvantaged populations. Beyond this inequity, the lack of other structural predictors and the negative spatial autocorrelation (**H4**) point to **managerial leadership and organizational culture**—rather than resources or geography—are the primary drivers of responsiveness in the Texas oil and gas sector.

analysis/gemini_draft_appendix.md

@@ -1,114 +0,0 @@
-# Appendix: Heterogeneous Enforcement of Transparency
-## Evidence from the Texas Railroad Commission
-
-**Appendix A: Data Definitions and Summary Statistics**
-**Appendix B: Event Study and Parallel Trends**
-**Appendix C: Robustness Checks**
-**Appendix D: Spatial Analysis Details**
-
----
-
-### Appendix A: Data Definitions and Summary Statistics
-
-To account for potential confounders driving enforcement heterogeneity, we constructed district-level aggregates of demographic and geographic variables. Table A1 summarizes the definitions and data sources for key variables used in the moderator analysis ($H3$).
-
-**Table A1: Variable Definitions**
-
-| Variable | Definition | Source |
-| :--- | :--- | :--- |
-| **Days to Enforcement** | Number of days between `violation_disc_date` and `last_enf_action_date`. | RRC Violation Data |
-| **Compliance Rate** | Percentage of inspections marked "Compliant". | RRC Inspection Data |
-| **EJI Score** | **Environmental Justice Index.** A composite social vulnerability score calculated as the mean of percentile-ranked tract-level indicators: minority share, poverty rate, unemployment, linguistic isolation, and education level. Aggregated to district level by averaging scores of tracts containing active wells. | ACS 2021 (5-yr) |
-| **High Capacity** | Binary indicator = 1 if district total inspections > median. | RRC Inspection Data |
-| **Rurality (RUCA)** | Average Rural-Urban Commuting Area code (1=Metro, 10=Rural) for wells in the district. | USDA / ACS 2020 |
-| **Basin** | The geologic oil/gas basin containing the majority of the district's wells (e.g., Permian, Eagle Ford). | RRC Well Geography |
-
-**Table A2: Baseline District Characteristics (Pre-Policy 2015-2018)**
-
-| District | Total Inspections | Compliance Rate (%) | Avg Days to Enforcement | EJI Score | Primary Basin |
-| :--- | :--- | :--- | :--- | :--- | :--- |
-| 01 | 29,612 | 85.0% | 242.8 | 0.497 | Permian |
-| 02 | 15,348 | 83.9% | 234.4 | 0.451 | Permian |
-| 03 | 32,975 | 94.1% | 61.9 | 0.492 | Permian |
-| 04 | 32,081 | 92.7% | 62.8 | 0.592 | Permian |
-| 05 | 16,329 | 92.1% | 275.7 | 0.461 | Barnett |
-| 06 | 37,386 | 89.0% | 475.0 | 0.493 | Barnett |
-| 08 | 60,999 | 88.2% | 135.3 | 0.496 | Permian |
-| 09 | 62,196 | 82.3% | 238.5 | 0.376 | Fort Worth |
-| 10 | 39,620 | 88.9% | 49.4 | 0.457 | Anadarko |
-| 6E | 13,326 | 78.4% | 301.2 | 0.534 | Barnett |
-| 7B | 35,929 | 82.7% | 48.1 | 0.421 | Fort Worth |
-| 7C | 40,631 | 85.2% | 63.0 | 0.537 | Permian |
-| 8A | 40,261 | 90.5% | 77.5 | 0.516 | Permian |
-
----
-
-### Appendix B: Event Study and Parallel Trends
-
-To validate the parallel trends assumption underlying our Difference-in-Differences (DiD) strategy, we estimated an event study specification where the treatment effect is allowed to vary by year. Table B1 reports the coefficients relative to the baseline year 2018 (one year prior to implementation).
-
-**Table B1: Event Study Estimates (Dependent Variable: Log Days to Enforcement)**
-
-| Year | Year Relative to Policy | Coefficient | Std. Error | P-Value |
-| :--- | :--- | :--- | :--- | :--- |
-| 2015 | -4 | -0.457 | 0.247 | 0.064 |
-| 2016 | -3 | -0.334 | 0.238 | 0.160 |
-| 2017 | -2 | -0.040 | 0.119 | 0.741 |
-| **2018** | **-1 (Ref)** | **0.000** | **-** | **-** |
-| 2019 | 0 | -0.114 | 0.107 | 0.284 |
-| 2020 | 1 | -0.167 | 0.191 | 0.384 |
-| 2021 | 2 | -0.417 | 0.258 | 0.107 |
-| 2022 | 3 | -0.582* | 0.273 | 0.033 |
-| 2023 | 4 | -0.487 | 0.309 | 0.115 |
-| 2024 | 5 | -0.776* | 0.281 | 0.006 |
-| 2025 | 6 | -1.457* | 0.255 | <0.001 |
-
-*Note: Standard errors clustered at the district level. Coefficients for 2015-2017 are statistically indistinguishable from zero, supporting the parallel trends assumption. The treatment effect grows in magnitude over time, suggesting a gradual adaptation to the transparency regime.*
-
----
-
-### Appendix C: Robustness Checks
-
-We performed a series of robustness checks to ensure our main findings were not driven by spurious trends, outliers, or specific sample selections.
-
-#### C1. Placebo Tests
-We re-estimated the model using "fake" policy implementation dates. A significant finding at a fake date (especially pre-2019) would suggest pre-existing trends driving the results.
-
-*   **2017 Placebo (Pre-treatment):** Coefficient = -0.056 (p=0.725). **Result: PASS.** No significant effect was found two years prior to the actual policy.
-*   **2021 Placebo (Post-treatment):** Coefficient = -0.566 (p<0.01). **Result: EXPECTED.** Because the actual policy (2019) had a growing effect over time, a cutoff in 2021 captures the delayed intensification of the 2019 shock.
-
-#### C2. Sample Restrictions
-To rule out the influence of outliers or external shocks (e.g., COVID-19), we re-estimated the main DiD model on restricted subsamples.
-
-**Table C1: Sensitivity to Sample Restrictions**
-
-| Restriction | Coefficient | P-Value | Implied Effect (%) |
-| :--- | :--- | :--- | :--- |
-| **Baseline (Full Sample)** | **-0.369** | **0.019** | **-30.8%** |
-| Exclude Extreme Outliers (Top/Bottom 2 districts) | -0.420 | <0.001 | -34.3% |
-| Exclude Early Years (Drop 2015-2016) | -0.558 | 0.005 | -42.8% |
-| Exclude Pandemic Years (Drop 2020-2021) | -0.482 | 0.003 | -38.3% |
-
-*Conclusion: The finding of faster enforcement is robust to the exclusion of outlier districts and the COVID-19 period.*
-
-#### C3. Alternative Specifications
-We tested sensitivity to functional form and control structures.
-
-*   **Linear Model (No Log):** Coefficient = -62.0 days (p=0.023). Confirms the direction of the effect without log-transformation.
-*   **Winsorized Outcome (5%):** Coefficient = -0.313 (p=0.036). Confirms results are not driven by extreme enforcement delay values.
-*   **District-Specific Time Trends:** Inclusion of district-specific linear time trends flips the sign (Coef = +0.392). This is common in short panels where the trend term absorbs the dynamic treatment effect shown in the event study. Given the clear break in 2019 seen in the event study, the trend specification likely over-controls for the policy response itself.
-
----
-
-### Appendix D: Spatial Analysis Details
-
-To test Hypothesis 4 (Spatial Spillovers), we calculated the spatial autocorrelation of the district-level treatment effects.
-
-**Global Moran's I Statistic:** -0.549
-**P-value:** < 0.05
-
-The negative Moran's I indicates **negative spatial autocorrelation**. In the context of regulatory enforcement, this means high-performing districts (large reductions in enforcement delay) are frequently adjacent to low-performing districts. This "checkerboard" pattern contradicts the hypothesis of positive regional spillovers or knowledge diffusion. Instead, it suggests that enforcement culture is highly localized to the specific district office and does not diffuse across administrative boundaries.
-
-**Figure D1: Spatial Spillover Scatterplot**
-*(Note: See Figure 6 in main text for the map)*
-The regression of a district's own treatment effect against the average effect of its neighbors yields a negative slope, confirming that proximity to a high-improving district does not predict improvement in the focal district.